|
|
|
Submit Paper / Call for Papers
Journal receives papers in continuous flow and we will consider articles
from a wide range of Information Technology disciplines encompassing the most
basic research to the most innovative technologies. Please submit your papers
electronically to our submission system at http://jatit.org/submit_paper.php in
an MSWord, Pdf or compatible format so that they may be evaluated for
publication in the upcoming issue. This journal uses a blinded review process;
please remember to include all your personal identifiable information in the
manuscript before submitting it for review, we will edit the necessary
information at our side. Submissions to JATIT should be full research / review
papers (properly indicated below main title).
|
|
|
Journal of
Theoretical and Applied Information Technology
May 2026 | Vol. 104
No.10 |
|
Title: |
ENHANCING LUNG DISEASE CLASSIFICATION USING CONVNEXT: AN IMPROVEMENT OVER
CNN-ELM-BASED MODELS |
|
Author: |
Y. SUREKHA, NALLA AKHILA, DR. K. KOTESWARA RAO, RAKESH KANCHARLA, P. THRINATH,
S.V. SYAM PRASAD REDDY, S. SARATH CHANDRA, P. MAHESH |
|
Abstract: |
The process of using medical images to classify lung diseases serves two main
purposes which include establishing early diagnosis methods and supporting
clinical decision-making procedures. Recent studies have demonstrated the
effectiveness of hybrid deep learning models that combine Convolutional Neural
Networks (CNNs) with Extreme Learning Machines (ELMs) to achieve efficient and
accurate classification. The base paper for this research project development
which establishes a strong baseline for lung disease classification uses a
CNN–ELM-based model as its first implementation. The reimplemented CNN-based
model achieves an accuracy of approximately 97.8% under optimized experimental
conditions. The ConvNeXt architecture-based model introduces an advanced model
which delivers better classification results. ConvNeXt uses current
convolutional design methods together with its transfer learning system to
identify important and unique lung image features. The proposed ConvNeXt model
achieves approximately 98.5% accuracy which exceeds the performance of the
CNN-ELM baseline according to experimental results. The results of the
comparative analysis show that advanced deep convolutional architectures provide
better lung disease classification results when compared to traditional
CNN-based hybrid methods |
|
Keywords: |
Lung Disease Classification, CNN-ELM, ConvNeXt, Transfer Learning, Medical Image
Analysis. |
|
DOI: |
https://doi.org/10.5281/zenodo.20497051 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
ARTIFICIAL INTELLIGENCE–DRIVEN CONTROL STRATEGIES FOR ADAPTIVE CYBER-PHYSICAL
SYSTEMS |
|
Author: |
DR JEEVAN JALA, SAI SRINIVAS VELLELA, SRAVANTHI JAVVADI, THALAKOLA SYAMSUNDARA
RAO, KOYA HARITHA, RAJU THOMMANDRU, DR SHOBANA GORINTLA |
|
Abstract: |
Cyber-Physical Systems (CPS) have emerged as an indispensable part of
intelligent infrastructures today, which allows close interaction between
computation and communication and physical processes. Nonetheless, rising
complexity in the system, dynamic operating conditions and uncertainty are some
of the challenges associated with some traditional control strategies which in
most cases depend on fixed parameter and precision system models. To overcome
such limitations, the current paper suggests an Artificial Intelligence-powered
adaptive control model of Cyber-Physical Systems that will help to improve their
adaptability, resilience, and real-time decision-making in the context of
uncertain and nonlinear conditions. The approach suggested incorporates
learning-based control especially reinforcement learning into a model-driven CPS
platform to allow the control actions to be continuously adjusted in response to
feedback of the system. The experimental setting is created using simulation and
is used to assess the effectiveness of the AI-driven controller and compare it
with the conventional PID and model-oriented control measures. The key metrics
used to evaluate the performance are tracking accuracy, convergence speed,
stability margin, energy consumption, disturbance recovery capability, and
scalability as system complexity increases. Experimental data prove that the
AI-based controller is always superior to the traditional ones with lower
tracking error, convergent faster, stronger in stochastic disturbances, lower
control energy, and high scalability. The results prove the fact that the
introduction of artificial intelligence into CPS control loops can greatly
increase the resilience and the efficiency of the systems. All in all, this
paper reveals AI-based control as a scalable and viable answer to
next-generation adaptive Cyber-Physical Systems in fields like intelligent
manufacturing, smart infrastructure, and automation of industry. |
|
Keywords: |
Artificial Intelligence, Cyber-Physical Systems, Adaptive Control, Reinforcement
Learning, Intelligent Systems. |
|
DOI: |
https://doi.org/10.5281/zenodo.20497089 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
EXPLORING METHODS AND TRENDS IN EVACUATION SAFETY RISK ASSESSMENT IN HIGH-RISE
BUILDINGS: A SYSTEMATIC LITERATURE REVIEW AND BIBLIOMETRIC MAPPING |
|
Author: |
NUR MAISARAH NOR AZHARLUDIN, KHYRINA AIRIN FARIZA ABU SAMAH, AMIR HAIKAL ABDUL
HALIM, RASEEDA HAMZAH, LALA SEPTEM RIZA, SITI SALWA SALLEH |
|
Abstract: |
Evacuation safety risk assessment (ESRA) in high-rise buildings is a critical
area of study due to the complex structural, functional, and occupancy
characteristics that influence evacuation performance during emergencies. Over
the past decade, researchers have introduced a range of simulation techniques,
computational models, and analytical frameworks to examine evacuation
efficiency, human behavior, and system performance. However, limited work has
consolidated these approaches in a systematic manner. In addition, existing
studies remain fragmented across simulation, behavioral, and technological
domains, with limited integration of emerging artificial intelligence
approaches, creating a gap in achieving a holistic understanding of ESRA. This
study addresses this gap by providing a critical and integrated synthesis,
contributing new knowledge through the identification of structural research
limitations and uncovering relationships between key themes using a combined
systematic literature review (SLR) and bibliometric mapping approach. This study
conducted a SLR on ESRA in high-rise buildings. Using Preferred Reporting Items
for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, 25 articles
published between 2020 and 2025 were identified from Scopus, Web of Science, and
ScienceDirect. Bibliometric mapping using VOSviewer revealed four dominant
clusters, including fire safety, high-rise building risk factors, evacuation
strategies, and emerging deep learning applications. The findings show that
simulation-based and engineering-focused studies remain central, while
artificial intelligence approaches are growing but remain underutilized. The
systematic review also identified four major themes: building type, technology
and smart systems, evacuation strategies, and methods used. Overall, this review
provides a structured overview of current research trends, identifies gaps in
existing approaches, and offers insights to guide future studies and practical
risk management efforts in high-rise buildings. |
|
Keywords: |
Evacuation Safety Risk Assessment, High-Rise Buildings, PRISMA, Risk Management,
Systematic Literature Review |
|
DOI: |
https://doi.org/10.5281/zenodo.20497112 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
A BLOCKCHAIN ASSISTED ADAPTIVE BOOSTING GRAPH LSTM FRAMEWORK WITH FIREFLY
OPTIMIZATION FOR ROBUST MALICIOUS ACTIVITY PREDICTION IN CLOUD ENVIRONMENTS |
|
Author: |
J NIVITHA, R ANANDAN |
|
Abstract: |
Cloud infrastructures with their unmatched scalability and flexibility are
becoming the target of advanced malicious operations, which carry serious
security threats to confidential data and critical services. Conventional types
of detection techniques usually fight with a dynamic, voluminous, and
complicated nature of the cloud-based threats that cause a high false alarm
percentage or failure to identify. In this paper, a new communication system
called Blockchain-Assisted Graph-LSTM Framework with Attention and Firefly
Optimization (BAG-LSTMAFO) is suggested to ensure a robust malicious
activity-detecting system in cloud environments. The framework uses blockchain
technology to provide tamper-proof recording of activities within the system and
exchange threat intelligence to increase the integrity and auditability of data.
Graph Long Short-Term Memory (G-LSTM) networks are a type of network that models
the intricate spatiotemporal interrelations of cloud system interactions, which
are modelled as dynamic graphs. To enhance the interpretability and accuracy of
detection an attention mechanism is incorporated to enable the model to
concentrate on the most salient features and time steps that may be indicative
of malicious behaviour. Moreover, Firefly Optimization is applied to optimize
the hyperparameters of G-LSTM model automatically which guarantees optimal
performance and generalization. The synergistic solution will aim to attain a
high detection rate, lower false positives, and offer a flexible and adaptive
defence system to changing cyber threats in the complex cloud infrastructures. |
|
Keywords: |
G-LSTM, BAG-LSTMAFO, Blockchain Technology, Grey Wolf Optimizer (GWO),
Graph Neural Networks (GNNs), BGLA-FO, IDS |
|
DOI: |
https://doi.org/10.5281/zenodo.20497126 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
A HYBRID LEGAL-FORENSIC FRAMEWORK FOR ENSURING DIGITAL EVIDENCE INTEGRITY IN
CRIMINAL PROCEEDINGS UNDER INTERNATIONAL STANDARDS |
|
Author: |
OLEKSIY ODERIY, VIKTOR VASYLYNCHUK, YEVHEN SHAPOVALENKO, OLEKSANDR SAVKA, TYMUR
LOSKUTOV |
|
Abstract: |
Ensuring the integrity and admissibility of digital evidence has become a
critical challenge in contemporary criminal proceedings due to the rapid
expansion of digital technologies and the increasing complexity of cyber-enabled
crimes. This study proposes a hybrid legal–forensic framework that integrates
cryptographic verification, automated chain-of-custody management, and
procedural safeguards to ensure continuous integrity validation across the
entire lifecycle of digital evidence. The research adopts an interdisciplinary
methodology combining doctrinal legal analysis with computational modeling. The
proposed framework incorporates SHA-256-based hashing, structured audit logging,
and modular evidence processing to align technical verification mechanisms with
international legal standards and due process requirements. The framework is
evaluated using a dataset of 120 digital evidence objects under controlled
conditions. The results demonstrate significant improvements over traditional
and hash-based approaches, achieving an Integrity Preservation Rate (IPR) of
99.2%, Chain-of-Custody Completeness (CCC) of 97.8%, and Verification Accuracy
(VA) of 98.9%. These findings confirm that continuous cryptographic verification
combined with automated procedural controls substantially enhances the
reliability, traceability, and admissibility of digital evidence. Although the
proposed approach introduces a moderate increase in processing time, the
improvement in evidentiary integrity and procedural compliance outweighs this
limitation. The scientific contribution of the study lies in the development of
a unified operational model that bridges the gap between digital forensics and
procedural law by establishing a direct link between technical verification
processes and legal admissibility criteria. The proposed framework provides a
scalable and legally compliant solution for digital evidence management, with
practical implications for law enforcement agencies, forensic experts, and
judicial authorities. Future research should focus on large-scale empirical
validation, optimization of computational performance, and the integration of
advanced technologies such as blockchain and artificial intelligence to further
enhance digital evidence processing in complex investigative environments. |
|
Keywords: |
Digital Evidence Integrity, Procedural Safeguards, Chain of Custody, Digital
Forensics |
|
DOI: |
https://doi.org/10.5281/zenodo.20497139 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
INTEGRATING GRAPH CONVOLUTIONAL NETWORKS FOR ENHANCED TRUST-AWARE CLUSTER HEAD
SELECTION IN WIRELESS SENSOR NETWORKS |
|
Author: |
GAJJALA SAVITHRI, N. RAGHAVENDRA SAI |
|
Abstract: |
Wireless Sensor Networks (WSNs) play an important role in many applications,
necessitating strong trust mechanisms to ensure reliable communication and data
integrity. This paper presents GLENET, a novel model that incorporates Graph
Convolutional Networks (GCN) into the Low-Energy Adaptive Clustering Hierarchy
(LEACH) protocol to improve cluster head selection in WSNs. GLENET's trust model
employs adaptive penalty coefficients, which allow for dynamic adjustments in
response to abnormal behavior, thereby fostering network trust. This work aims
to improve the effectiveness of WSNs by addressing the limitations of
traditional clustering protocols. GLENET's uniqueness lies in its comprehensive
trust assessment, which combines direct, indirect, and energy trust metrics. The
model responds dynamically to changing network conditions, aligning penalties
with contextual abnormalities to discourage malicious behavior. The results show
that GLENET achieves a throughput of 18520 kbps while outperforming comparable
methods in computation time (48.34 seconds) and residual energy conservation
(13.18 Joules). The model's adaptability and prioritization of nodes with
sustainable energy levels help to ensure long-term stability, making it a
promising approach for future WSN applications. |
|
Keywords: |
Wireless Sensor Networks, Trust Model, Glenet, Graph Convolutional Networks,
Leach Protocol. |
|
DOI: |
https://doi.org/10.5281/zenodo.20497157 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
METHODOLOGICAL PROVISION OF TECHNICAL AND FORENSIC DOCUMENTATION OF DIGITAL AND
PHYSICAL EVIDENCE IN PRE-TRIAL CRIMINAL PROCEEDINGS |
|
Author: |
INHA KALANCHA, VASYL SMIKH, NADIIA MORHUN, SERHII BARHAN, EDUARD USHKANENKO |
|
Abstract: |
Ensuring proper documentation of digital and physical evidence today determines
not only the quality of the pre-trial investigation, but also the real
possibility of making a fair court decision, which necessitates the need for
procedures capable of guaranteeing the authenticity, integrity and traceability
of evidentiary information in accordance with international ISO and NIST
standards. The aim of the study is the substantiation and testing of the
author’s three-level conceptual model of methodological provision of documenting
evidence, which integrates the technical, organizational, and legal levels of
evidentiary information management. The model is built on the basis of the
Forensic Documentation Integrity Index (FDII), which includes technical
reliability, procedural admissibility, integrity of the storage chain, and
methodological standardization. The methodology is based on a combination of
comparative legal, functional analytical, and expert modelling approaches using
Delphi survey, analytical hierarchy process (AHP), and correlation analysis (n =
27 experts). The results confirmed that the implementation of the model provides
increased consistency between technical, procedural, and methodological
parameters. The highest FDII value was recorded in the United Kingdom (0.94) due
to the full digitalization of the chain of custody in accordance with the
Forensic Science Regulator’s Code and the Digital Forensics Science Strategy. In
Germany (0.90), the stability of the system is ensured by the codification of
procedures in the Strafprozessordnung (StPO) and BKA standards. In Ukraine
(0.81), the highest increase (ΔFDII = +0.15) was recorded after the model was
tested. Correlation analysis (r = 0.79–0.95) demonstrated the effectiveness of
the integration structure. The academic novelty is the creation and confirmation
of the effectiveness of a three-level model of documenting evidence. The
practical value is the formation of a toolkit for harmonizing the Ukrainian
system with international standards ISO/IEC 27037, 27041 and NIST IR 8387. |
|
Keywords: |
Digital Evidence, Forensic Documentation, Methodological Model, Chain of
Custody, Criminal Justice, Innovation, Legal Governance |
|
DOI: |
https://doi.org/10.5281/zenodo.20497207 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
A MULTI-SCALE CROSS-ATTENTION VISION TRANSFORMER FRAMEWORK WITH CLASS-IMBALANCE
OPTIMIZATION FOR AUTOMATED DIABETIC FOOT ULCER DIAGNOSIS |
|
Author: |
GOWRI MANOHARI V, MERCY PAUL SELVAN |
|
Abstract: |
Diabetic foot ulcers (DFUs) are a serious complication of diabetes that, if left
untreated, can result in the loss of a lower limb. In contrast to traditional
clinical approaches for classifying DFUs, automated methods based on deep
learning architectures have shown promising results. However, current deep
learning techniques frequently fail to capture both global contextual
information and fine-grained local characteristics, which reduces generalization
in real-world scenarios. This study suggests CrossViT versions with multi-scale
feature learning and a weighted cross-entropy loss function to successfully
handle class imbalance in order to address these issues. This paper presents a
novel approach that improves feature extraction and representation learning on
DFU images by utilizing CrossViT across multiple scales. The multi-scale
CrossViT architecture enables the model to learn both local lesion features and
global image-level representations, which is crucial for detecting ulcer patches
of varying sizes, textures, and colors. To further address class imbalance in
DFU datasets, a weighted cross-entropy loss (WCEL) is employed during training
to emphasize underrepresented classes, such as early-stage or small ulcers. This
enhances the model’s sensitivity and overall diagnostic accuracy. Experimental
results demonstrate that the proposed CViT-WCEL approach achieves an accuracy of
98.19% and outperforms conventional CNN and single-scale transformer models in
terms of F1-score, sensitivity, specificity, and area under the receiver
operating characteristic curve (AUC-ROC). These findings highlight the potential
of the proposed method as a scalable and efficient solution for reliable and
early DFU diagnosis in real-world clinical environments. |
|
Keywords: |
Diabetic Foot Ulcer, CrossViT, Multi-scale Vision Transformer, Weighted
Cross-Entropy, Deep Learning, Medical Image Analysis, Automated Diagnosis,
Explainable AI |
|
DOI: |
https://doi.org/10.5281/zenodo.20497227 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
ENHANCING SECURITY IN ELECTRONIC MEDICAL RECORDS USING GENETIC ALGORITHM-DRIVEN
BLOCK CHAIN ENCRYPTION |
|
Author: |
DR. MANAL AL KHAMMASH, SUBUHI KASHIF ANSARI, DR. RAWIA ELARABI, ANNE ANOOP,
YASIR AHMED, DR. NOHA MOSTFA, VAIBHAV SHARMA |
|
Abstract: |
Electronic Medical Records (EMRs) are now central to modern healthcare, enabling
seamless data exchange and coordinated patient management. Yet their sensitivity
and growing interconnectivity make them increasingly attractive targets for
cyberattacks. Conventional protection mechanisms static encryption paired with
centralized access control suffer from two major limitations: fixed
cryptographic parameters cannot adapt to heterogeneous data types, device
capabilities, or evolving threat conditions, and centralized management
introduces single points of failure with limited auditability.
To overcome
these constraints, this study introduces a Genetic Algorithm (GA) driven
blockchain encryption framework for EMR security. The key innovation lies in
dynamically optimizing encryption parameters such as key length, cipher mode,
and rotation frequency using GA, while employing a permissioned blockchain to
deliver decentralized, tamper proof access control and audit trails. This dual
architecture enhances both security resilience and operational efficiency.
Encrypted EMRs are stored off chain, while smart contracts manage access rights,
cryptographic profile identifiers, and hybrid AES–ECC key distribution. The GA
encodes encryption configurations as chromosomes and evolves them using a multi
objective fitness function balancing confidentiality, latency, and storage
overhead. Experiments using de identified EMR datasets on a Hyperledger Fabric
testbed demonstrate that the proposed framework outperforms fixed parameter
AES–ECC baselines, improving encryption–decryption performance, strengthening
cryptographic robustness, and reducing blockchain transaction latency. These
results validate the practicality of an adaptive, blockchain enabled security
model for real world healthcare environments. |
|
Keywords: |
Blockchain Encryption, Electronic Medical Records, Genetic Algorithm
Optimization, Healthcare Data Security, Hybrid AES–ECC Cryptosystem |
|
DOI: |
https://doi.org/10.5281/zenodo.20497255 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
CYBER RESILIENCE OF THE DIGITAL STATE: THE ROLE OF AI TECHNOLOGIES IN PROTECTING
ELECTRONIC PUBLIC SERVICES |
|
Author: |
YELYZAVETA TYMOSHENKO, MARYNA DZEVELIUK, ANDRII DZEVELIUK, NATALIIA
CHERNYSHCHUK, IRYNA SKICHKO |
|
Abstract: |
The paper analyzes the cyber resilience of public e‐services and public digital
platforms in an environment of escalating cyber aggression and assaults.
Traditional cyber security methods based on technical protection cannot
guarantee that public services are continuous and reliable when facing great
uncertainty. This paper views artificial intelligence assets as a tool to
increase the state's ability at the intra-state level in cyber defense. The
analysis is a mixture of indexing, mathematical model and econometric
verification. The study applies standardized global indicators in cybersecurity,
digital governance and readiness to adopt AI in public sectors. A multiplicative
model is used to test the effect of AI integration on the degree of cyber
resilience. Management effects are then tested using regression analysis. The
introduction of artificial intelligence on defense systems of electronic public
services possesses management nature. The findings illustrate cyber resilience's
force multiplier effect through AI which decreases response times, increases the
consistency of management decisions and lowers the magnitude of service
disruptions. Econometric validation establishes the statistical significance of
the model and its dependence on management maturity level of digital
organizations. The cyber robustness of electronic public services is a key
outcome that stems from management decisions supported with AI analysis powers.
The approach opens up a wider scientific debate on e-governance. The method
provides a basis for the evidence in the formation of evidence-based public
policies on national cybersecurity. |
|
Keywords: |
Digital State, Electronic Public Services, Artificial Intelligence, Public
Sector Cybersecurity, Algorithmic Decision-Making |
|
DOI: |
https://doi.org/10.5281/zenodo.20497278 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
HOW ONLINE COMMENTS ON MARRIAGE TOPICS ON WEIBO INFLUENCE MARRIAGE INTENTIONS OF
UNMARRIED WOMEN IN CHINA |
|
Author: |
YE CHEN, AINI AZEQA MAR OF, SEYEDALI AHRARI, ZEINAB ZAREMOHZZABIEH,HASLINDA
ABDULLAH, HANINA HALIMATUSAADIAH HAMSAN |
|
Abstract: |
The study aimed to use a combination of Information Adoption Model and Theory of
Reasoned Action to explain the underlying mechanisms by which online comments on
marriage topics in Weibo influence unmarried women's marital intentions.
Argument quality and source credibility were identified as content features of
online comments on marriage topics in Weibo. Perceived information usefulness,
and marital attitudes were identified as the internal processes by which content
features of online comments influence marital intention. The study collected
data from 422 un-married women through a questionnaire on the Weibo platform.
PLS-SEM with SmartPLS 4.0 was utilized to test the structural model. The results
showed that the content characteristics (argument quality and source
credibility) of online comments on marriage topics had a positive effect on
perceived information usefulness and a negative effect on marriage intention. In
addition, the perceived information usefulness of online comments had a negative
effect on marital attitudes. Positive marital attitudes had a positive effect on
marital intentions. Perceived information usefulness and marital attitudes acted
as serial mediators between the content characteristics of online comments on
marital topics and marital intentions. This study extends the literature on
online commenting research by examining the effects of content features of
online comments on marriage topics on marital intentions, fills a theoretical
gap in the scope of related re-search, and provides valuable theoretical and
practical implications. |
|
Keywords: |
Argument quality; Source credibility; Perceived Information Usefulness;
Attitude; Intention |
|
DOI: |
https://doi.org/10.5281/zenodo.20497292 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
ARTIFICIAL INTELLIGENCE AND THE PRODUCTIVITY PARADOX: MODELING STRUCTURAL
TRANSFORMATION AND LABOR MARKET POLARIZATION ACROSS ECONOMIC GROUPS |
|
Author: |
MYKHAYLO VIDYAKIN, TETIANA MELNYK, OKSANA NAZARCHUK, NATALIIA POCHERNINA,
OLEKSANDR AKIMOV |
|
Abstract: |
The active introduction of artificial intelligence into the work process causes
significant changes in the labor market, accounting and management control
systems, as well as in economic growth models. With the advent of new
information technologies, it became obvious that national economies need to
increase labor productivity through the process of digitalization. This requires
scientific substantiation of energy-efficient technologies that affect the use
of labor resources in terms of their effectiveness. The purpose of this research
was to study the impact of artificial intelligence technologies on labor
productivity and changes in the employment structure, especially in the context
of the digital economy. The methodology was based on a systematic,
logical-structural, comparative and econometric analysis of panel data for the
period from 2018 to 2024. The analysis revealed that the introduction of
automation, investments in human capital development and spending on scientific
research and development have a statistically significant positive effect on
labor productivity growth. An increase in the level of automation by 1% leads to
a 0.44% increase in productivity, while an increase in spending on education
provides an increase of 0.31%. In 2024, the average gross domestic product per
capita in developed countries will be $ 133.2 thousand. In countries with
economies in transition, this figure is $57.300, while in developing countries
it is only $34.000. The digital integration index in developed countries is
0.85, while in developing countries it is only 0.52. The level of accounting
automation in economically developed countries is 74%, while in less developed
regions this figure is only 29%. The practical value of the results obtained
lies in the formation of policies that contribute to the growth of labor
productivity, the development of human resources and the improvement of
management structures in the conditions of digital transformation. |
|
Keywords: |
Automation, Accounting, Control, Personnel, Digitalization, Management,
Efficiency. |
|
DOI: |
https://doi.org/10.5281/zenodo.20497307 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
REGISTRATION AWARE SEMI SUPERVISED MULTIMODAL LEARNING FOR PROSTATE CANCER
DETECTION AND GRADING |
|
Author: |
SAMANA JAFRI, GAJANAN BIRAJDAR |
|
Abstract: |
Accurate detection and grading of prostate cancer are critical for clinical
decision making, particularly in assessing tumor aggressiveness using gleason
grading. Magnetic resonance imaging (MRI) provides noninvasive anatomical
information, while histopathology offers definitive cellular level diagnosis
however, existing studies suffer from three key limitations (i) reliance on
unimodal data that fails to capture complementary anatomical and cellular
information, (ii) lack of explicit spatial correspondence modeling between MRI
and histopathology, and (iii) dependence on large scale annotated datasets,
which are difficult to obtain in clinical settings. This creates a critical need
for a unified multimodal framework that can leverage limited annotations while
preserving anatomical consistency across modalities. This study proposes Multi
modal Pathology Gleason Network as MultiPathGleasoNet, a novel registration
aware semi supervised multimodal learning framework for prostate cancer
detection, tumor segmentation, and three class gleason grading using spatially
aligned T2 weighted MRI and histopathology images. The novelty of the proposed
approach lies in integrating modality specific Vision Transformer encoders,
graph based spatial modeling of histopathology via a Graph Attention Network
(GATv2), and a registration aware cross modal fusion transformer that explicitly
captures anatomical correspondence between modalities. To address limited
annotations, a student teacher learning strategy with adaptive pseudo labeling
is employed to effectively utilize unlabeled data. The framework is evaluated on
654 registered MRI histopathology slice pairs from 152 patients and demonstrates
strong performance across tasks, achieving an Area Under Curve (AUC) of 0.986
for cancer detection, a Dice similarity coefficient of 0.903 for tumor
segmentation, and a Cohen’s Kappa of 0.91 for Gleason grading. Additionally, the
model achieves an average inference time of approximately 80 ms per sample,
indicating computational efficiency. These results suggest that combining
registration aware multimodal fusion with graph based spatial reasoning and semi
supervised learning enhances prostate cancer diagnosis while reducing annotation
dependency, highlighting its potential for clinical and computational pathology
applications. This study is motivated by the need to bridge the gap between
radiological and histopathological analysis while reducing annotation dependency
in clinical workflows. |
|
Keywords: |
Graph Attention Network, Multimodal Deep Learning, Prostate cancer detection,
Semi Supervised Learning, Vision Transformer |
|
DOI: |
https://doi.org/10.5281/zenodo.20497327 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
AMAT-IDS: A HYBRID FRAMEWORK COMBINING GA-SUS FEATURE SELECTION AND DYNAMIC TWIN
AUTO-ENCODERS FOR INTRUSION DETECTION |
|
Author: |
RADHARANI AKULA, GS NAVEEN KUMAR |
|
Abstract: |
Network intrusion detection has become increasingly vital as cyber threats grow
in sophistication and scale. Current approaches often achieve strong aggregate
metrics yet fail to detect rare but critical attacks, while relying on
computationally expensive architectures unsuitable for resource-constrained
deployments. This paper introduces AMAT-IDS, a multi-stage framework that
addresses these limitations through two complementary innovations: Enhanced
Genetic Algorithm with Stochastic Universal Sampling (GA-SUS) for
multi-objective feature optimization, and Dynamic Twin Auto-Encoders (DTAE) for
class-specific representation learning. Evaluated on NSL-KDD using a stratified
70/15/15 train-validation-test protocol, GA-SUS reduced the feature space from
41 to 11 attributes (73% reduction) while maintaining 96.49% test accuracy. DTAE
further enhanced minority-class detection, elevating U2R precision from 0.500 to
0.778 and R2L precision from 0.563 to 0.987, achieving 96.02% overall accuracy
with a compact 9-dimensional representation. Five-fold cross-validation
confirmed model stability (96.07% ± 0.35%). Statistical tests validated
significant improvements in minority-class metrics (bootstrap CI: U2R precision
[0.741, 0.815], R2L recall [0.923, 0.967] at 95% confidence). The framework
processes samples at 1.2 ms latency on standard CPU hardware, enabling real-time
deployment. By reconciling efficiency, interpretability, and balanced detection
across imbalanced classes, AMAT-IDS provides a practical solution for edge and
IoT environments where computational resources are limited yet security
requirements remain stringent. |
|
Keywords: |
Intrusion Detection System, GA-SUS, Dynamic Twin Auto-Encoder, Feature
Selection, Class Imbalance, Cybersecurity. |
|
DOI: |
https://doi.org/10.5281/zenodo.20497340 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
ADAPTIVE SURVIVABLE ROUTING IN LARGE-SCALE WDM OPTICAL NETWORKS USING
TRAFFIC-REGIME-CONDITIONED MACHINE LEARNING |
|
Author: |
G PRAVEEN BABU, K V RAMANA |
|
Abstract: |
Large-scale Wavelength Division Multiplexed (WDM) optical networks are becoming
more and more essential to be deployed in heterogeneous and fast-changing
traffic regimes induced by cloud interconnection services, elastic data center
synchronization and backbone applications with strict latency requirements. The
routing and protection in these networks are primarily reliance on static
heuristics, shortest-path formulations, fixed-alternate policies and etc. which
are mostly designed to highly predictable traffic behaviours and stable
operation conditions. But under dynamic traffic regimes, such approaches have
low adaptability because of, among other factors, the strong coupling of
wavelength contentions, congestion propagation and survivability. This
restriction often leads to higher blocking probability, less efficient use of
the spectrum and unpredictable restore situation for large scale optical
backbone networks. In this paper, to tackle this problem, Traffic Regime
Conditioned Survivable Routing Policy Learning (TRC-SRPL) framework for adaptive
routing and protection in WDM optical networks is proposed. The framework model
problem of routing and survivability decisions as a regime-aware supervised
learning problem, where a protection-available wavelength is selected based on
the learnt policy that conditions on network state variables like the occupancy
status of wavelengths, congestion indicators, path feasibility, protection
availability, and characteristics of the traffic regime. Routing-protection
decisions generated by Oracle under the two scenarios (static and dynamic) are
used to train a deterministic low-latency policy suitable to be deployed within
realistic optical control planes. For the proposed framework, large-scale
simulation analyses on NSFNET and USNET topologies are carried out under mixed
traffic regime with around 48,000 connection requests for both static and
dynamic operating scenarios are performed. The results show that the proposed
TRC-SRPL policy always performs better than the traditional SRPL policy, SPP and
FPR. The proposed framework achieves significantly better blocking probability
reduction, mean (spectrum) wavelength utilization, and congestion exposure
reduction - up to 26.5%, ~12.3% and ~16%, respectively, under dynamic traffic
conditions, while demonstrating stable spectrum fragmentation behavior while
keeping the inference latency non-prohibitive in operation. In addition, the
learnt policy demonstrates consistent routing behaviour and restoration
performance, no matter how large the traffic volume is, without the need of
regime-specific re-training. The results obtained show that explicit
traffic-regime awareness is indeed a wholesome, non-subsidiary, factor in
survivable optical routing. The above adaptive survivable control requirements
thus suggests an immediate direction for the proposed TRC-SRPL framework for
survivable control in next generation large-scale WDM optical backbone networks. |
|
Keywords: |
Traffic-Regime-Aware Routing, WDM Optical Networks, Machine Learning Routing
Policy, Survivable Optical Networking, Dynamic Traffic Modeling |
|
DOI: |
https://doi.org/10.5281/zenodo.20497373 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
A MEDOID-BASED SEMI-SUPERVISED CLUSTERING APPROACH FOR WORD SENSE DISAMBIGUATION
IN TELUGU |
|
Author: |
DURGAPRASAD PALANATI , SUNITHA K.V.N , PADMAJA RANI B |
|
Abstract: |
Word Sense Disambiguation (WSD) is critical component in all the Natural
Language Processing related tasks. Due to limited availability of sense
annotated corpus and morphologically richness of Telugu language its’ very
challenging to develop WSD systems. This paper proposes a semi-supervised
clustering technique for Telugu WSD that minimizes the utilization of large
annotated corpora. The proposed method uses IndicBERT-based sentence embeddings
to find contextual semantics. A novel seed selection approach based on
sum-of-squared error (SSE) is proposed to make sure the initialization of sense
clusters, followed by a formula-based medoid selection mechanism and an adaptive
similarity thresholding scheme for sense propagation. The approach is evaluated
on a manually annotated corpus and also compared with baseline methods Most
Frequent Sense, Most Common Sense and simple K-medoid. The performance metrics
show that the proposed system surpasses the baseline approaches. This shows the
effectiveness of proposed approach for Telugu WSD and applicability to
Information retrieval tasks. |
|
Keywords: |
Contextual Sentence Embeddings, Lexical Ambiguity Resolution, Low-Resource
Language Processing, Medoid-Based Clustering, Semi-Supervised Learning, Word
Sense Disambiguation. |
|
DOI: |
https://doi.org/10.5281/zenodo.20497394 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
SENTIMENTTRENDAI: AN AI-DRIVEN FRAMEWORK FOR SENTIMENT ANALYSIS AND SOCIAL MEDIA
TREND DETECTION |
|
Author: |
GUJJETI NAGARAJU, DR. BODDUPALLY JANAIAH, SHIVANI YADAO, VASAVI OLETI, DR. Y.
SOWMYA REDDY |
|
Abstract: |
The rapid growth of multilingual social media platforms has created a critical
need for intelligent systems capable of accurately analyzing sentiment and
detecting emerging trends across diverse linguistic and cultural contexts.
However, existing sentiment analysis models often struggle with multilingual,
code-mixed, and temporally evolving data, while also failing to capture the
influence of key users and the spread of misinformation, limiting their
effectiveness in real-world applications. To address these challenges, this
paper proposes SentimentTrendAI, a unified deep learning framework that
integrates sentiment classification, temporal trend forecasting, influencer
identification, and misinformation detection within a scalable and explainable
architecture. The framework employs a hybrid CNN-BiLSTM model with attention for
robust sentiment analysis and incorporates temporal modeling and graph-based
techniques to capture sentiment evolution and social influence dynamics. The
proposed system is evaluated on multilingual datasets comprising over 120,000
social media posts. Experimental results demonstrate strong performance,
achieving an accuracy of 92.3%, a macro F1-score of 91.7%, and an MCC of 0.89
for sentiment classification, along with high effectiveness in trend prediction
and misinformation detection (AUC = 0.94). Ablation studies further validate the
contribution of key architectural components. The proposed framework provides a
practical and scalable solution for real-time monitoring of public sentiment,
enabling governments, media organizations, and public health agencies to
identify emerging trends, assess societal reactions, and mitigate the spread of
misinformation across multilingual environments. |
|
Keywords: |
Sentiment Analysis, Trend Forecasting, Multilingual Social Media, Deep Learning,
Misinformation Detection |
|
DOI: |
https://doi.org/10.5281/zenodo.20497401 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
SPEECH EMOTION RECOGNITION USING DEEP LEARNING TECHNIQUES |
|
Author: |
ZABER AL HASSAN AYON, MUHAMMAD ALIFF AHMAD ZAINUDIN, NUR HAFIEZA ISMAIL, NUR
SHAZWANI KAMARUDIN, MUHAMMAD ARY MURTI |
|
Abstract: |
Accurate recognition and interpretation of emotions from speech signals
represent a critical frontier in human-computer interaction and affective
computing. This research presents a novel Deep Learning (DL) framework for
speech emotion recognition that achieves state-of-the-art performance by
integrating advanced acoustic feature extraction with sophisticated sequential
modeling. The proposed approach leverages Mel-frequency Cepstral Coefficients
(MFCC) for robust feature extraction and employs Long Short-Term Memory (LSTM)
networks to effectively capture the temporal dynamics inherent in emotional
speech patterns. Through rigorous experimentation on the Toronto Emotional
Speech Set (TESS), our model demonstrates exceptional accuracy of 98.21%,
significantly outperforming traditional Machine Learning (ML) approaches.
Comparative analysis reveals the LSTM-based model has superior ability to
differentiate between acoustically similar emotions, a persistent challenge in
speech emotion recognition systems. The architecture's computational efficiency
and robustness to acoustic variability make it particularly well-suited for
real-time applications across diverse domains including healthcare monitoring,
customer experience management, and intelligent human-machine interfaces. This
research advances the field of affective computing by establishing a
comprehensive framework that combines acoustic feature engineering with deep
sequential learning, offering a reliable solution for emotion recognition from
speech signals and laying the groundwork for more intuitive and emotionally
intelligent computing systems. |
|
Keywords: |
Deep learning, Speech signal, Emotion recognition, LSTM, MFCC. |
|
DOI: |
https://doi.org/10.5281/zenodo.20497408 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
SOLARSOILINGNET: A DUAL-HEAD VISION TRANSFORMER FRAMEWORK FOR AUTOMATED SOLAR
PANEL SOILING CLASSIFICATION AND RAIN-AWARE CLEANING ACTION SELECTION |
|
Author: |
JAYALAKSHMI MURUGAN, RAJERMANI THINAKARAN, KALIAPPAN M, MAHARAJAN K, MARIAPPAN E |
|
Abstract: |
Soiling of photovoltaic (PV) panel surfaces reduces power output by 7% to over
50%, depending on geographic location and prevailing climate. Accurate
identification of soiling type — dry dust, sticky mud, bird droppings, water
spots, or mixed contamination — is indispensable for optimal cleaning scheduling
and cost-efficient operations. This paper presents SolarSoilingNet, a dual-head
deep learning framework that couples a ViT-B/16 Vision Transformer backbone with
a rain-aware auxiliary action head to simultaneously classify soiling type and
recommend a cleaning intervention. On a curated six-class benchmark of 2,050 RGB
images, SolarSoilingNet achieves 98.2% test accuracy and a weighted F1-score of
97.9%, surpassing all evaluated CNN and transformer baselines by at least seven
percentage points. A real-time weather context module — encoding next-day
precipitation, temperature, wind speed, and relative humidity — conditions the
decision layer so that unnecessary cleaning operations are deferred when natural
precipitation is forecast. Extensive ablation experiments, exploratory data
analysis, and comparisons with state-of-the-art methods published between 2021
and 2026 confirm the model's superiority in accuracy, inference speed, and
operational practicality. The framework is compatible with IoT-enabled
monitoring platforms and drone-based inspection systems. |
|
Keywords: |
Solar Panel Soiling Classification, Vision Transformer, Dual-Head Architecture,
Rain-Aware Decision Support, Photovoltaic Maintenance, Transfer Learning, Deep
Learning |
|
DOI: |
https://doi.org/10.5281/zenodo.20497418 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
HYBRID QUANTUM DEEP LEARNING FRAMEWORK FOR MULTI-CLASS KIDNEY DISEASE
CLASSIFICATION AND DEMOGRAPHIC ANALYSIS FROM CT IMAGES |
|
Author: |
CHALUMURU SURESH, RAVI KANTH MOTUPALLI, SIVAMANI SELVARAJU, A N V K SWARUPA,
DHARMA TEJA LANKA, ASHRAYA YELISETTY, SWATHI ANNAM |
|
Abstract: |
Kidney diseases like cysts, stones and tumors are renal conditions. Accurately
identifying them from CT scans is crucial for treatment.Deep learning models
have made progress in image analysis. However traditional methods still struggle
to distinguish between kidney conditions at once.This paper suggests a model
that combines a classical CNN for image feature extraction with quantum circuits
to improve classification.The model was tested on a dataset of around 12,000 CT
images across four categories: Normal Cyst, Stone and Tumor.It achieved 98% test
accuracy and a ROC-AUC of 0.985.This model consistently outperformed models in
controlled experiments.The results also show that the model produces probability
outputs.Additionally the paper looks at how disease types vary across age groups
and genders.This information can be useful for screening programs.The model can
help doctors diagnose Kidney diseases, such as Kidney cysts, Kidney stones and
Kidney tumors accurately.The new model is an improvement over existing methods,
for detecting Kidney diseases. |
|
Keywords: |
Quantum Deep Learning (QDL), Quantum Neural Networks, Computed Tomography,
Multi-Class Medical Image Classification, Demographic Analysis. |
|
DOI: |
https://doi.org/10.5281/zenodo.20497438 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
OPTIMIZING FEEDBACK IN DIGITAL FITNESS ENVIRONMENTS: CAN THE INTEGRATION OF
REAL-TIME POSE ESTIMATION AND RETRIEVAL-AUGMENTED GENERATION ARCHITECTURES
IMPROVE USER ACCURACY? |
|
Author: |
LEONARDO CARLOS VELIZ ARCE, JAIDER PARAGUAY JUNCO, BRAD JHOMERS ROSALES TAPIA,
ROSARIO DELIA OSORIO CONTRERAS |
|
Abstract: |
Physical inactivity and the lack of technical supervision in home-based training
lead to increased injury risks and reduced exercise effectiveness. This study
addresses this problem by developing FitPoseTracker, an intelligent platform
that integrates computer vision and conversational artificial intelligence to
optimize physical training. The proposed system, FitPoseTracker, performs
real-time posture analysis and automatic repetition counting using human pose
estimation and vectorial geometry, while a specialized conversational assistant
based on a Retrieval-Augmented Generation (RAG) architecture provides
contextualized technical guidance in natural language. The posture analysis
module leverages MediaPipe to detect anatomical keypoints and compute joint
angles, enabling accurate movement state classification. System performance was
evaluated using a two-factor ANOVA to analyze the effects of exercise type and
user experience level on repetition-counting accuracy. Results indicate that
exercise type significantly influences accuracy (F(2,141)=166.98, p<2e−16),
whereas user experience level shows no significant effect, demonstrating
consistent performance across different skill levels. Squats achieved the
highest accuracy, followed by sit-ups and push-ups. Additionally, the
conversational assistant was evaluated through automated testing on 50
representative queries, achieving high accuracy (0.92), recall (0.78), and
F1-score (0.84), with strong contextual coherence. Overall, the results confirm
that integrating real-time pose analysis with Large Language Models (LLMs)
effectively bridges the gap between movement quantification and specialized
technical guidance. |
|
Keywords: |
Artificial Intelligence; Human Pose Estimation; Computer Vision; Fitness
Training; Retrieval-Augmented Generation; Conversational Assistant |
|
DOI: |
https://doi.org/10.5281/zenodo.20497466 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
A HYBRID DIFFUSION DRIVEN SPATIO TEMPORAL DEEP NEURAL FRAMEWORK FOR
PROBABILISTIC HIGH RESOLUTION GLOBAL WEATHER FORECASTING |
|
Author: |
KOTESWARARAO BADITHALA, RAJA KRISHNAMOORTHI |
|
Abstract: |
True prediction of weather throughout the world, due to the nonlinear and
chaotic and multi-scale characteristics of the atmosphere systems is an uphill
task. The solutions of the NWP models though have a physical basis, do cost to
calculate and cannot solve the finescaled dynamics, limiting their effectiveness
in high social and spatial scale conditions. To overcome these shortcomings, the
current paper shall present a Hybrid Diffusion-Driven Spatio-Temporal Deep
Neural Framework that can be used to achieve high-resolution and
probabilistically valid predictions over the planet. The architecture involves
the use of convolutional neural networks (3D-CNNs) in combination to learn to
extract spatial features at multiple scales in a hierarchical manner,
spatio-temporal attention transformers to learn to model long-range temporal
dynamics, and probabilistic refinement module using diffusion to facilitate
uncertainty-aware prediction through the assistance of a generative denoising
process. It is based on the most recent developments in diffusion-based
time-series modeling, interpretable process system identification and
transformer-modulated probabilistic learning, and applies them during a large
atmospheric prediction on the first instance. As it has been shown in multiple
experiments, ERA5 and ECMWF reanalysis information (20102022) are significantly
stronger in hybrid model use compared to the state-of-the-art baselines such as
GraphCast, FourCastNet, SwinRDM and Chronos. As shown in the quantitative
assessment, the RMSE and CRPS were up to 29 and 30 percent higher, respectively,
in comparing the accuracy and the statistical reliability of the forecast of the
24 hours, 36 hours and the 60 hours forecast lead times. Moreover, the
calibration diagnostic, ablation analysis demonstrates that the decoder with
diffusion produces well-structured uncertainty estimates, which should be the
reason why it is more reliable to utilize in practice. The overall results
indicate that the suggested hybridization can be effective in capturing the
multiscale nature of the atmosphere dynamics, and at the same time effective in
computation and interpretation. The article presents scalable
uncertainty-constrained deep learning models to the next-generation global
weather forecast and gives substantial ground to more general applications to
the climate analytics, environmental monitoring and model the earth system using
data. |
|
Keywords: |
Forecasting the weather on a global level, Spatio-Temporal Deep Learning,
Diffusion Models, 3D Convolutional Neural Networks (3D-CNNs), Attention
Transformers, Probabilistic Forecasting, ERA5 Reanalysis Data, Hybrid Neural
Architecture, Multiscale Atmospheric Modeling |
|
DOI: |
https://doi.org/10.5281/zenodo.20497470 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
ANALYZING THE IMPACT OF HYBRID DEEP LEARNING FOR NFT CLASSIFICATION IN IOT
ENABLED METAVERSE ENVIRONMENTS |
|
Author: |
DR. SHAHANAWAJ AHAMAD, DR. VEERA TALUKDAR, DR. A.PANKAJAM, DR. MAMATHA G,
CHANDRA SHEKHAR NAGENDRA, DR. VIPIN JAIN, DR. S. DHAMODARAN, DR. S. PADMAPRIYA |
|
Abstract: |
Research is becoming harder to classify NFTs in metaverse settings that use IoT,
and this research tries to help with that. In these kinds of situations,
traditional deep learning models usually have trouble with scalability, feature
extraction, and computing efficiency. It utilized a dataset to look at how
changes in epochs affected the validity, efficiency, and accuracy of four
distinct neural network models: CNN, DenseNet, Inception, and a Hybrid model.
The Hybrid model always has better validation accuracy than its competitors.
With an accuracy of 95.1 percent, it beat the previous record at epoch 30. The
Hybrid approach works better than other methods. The Hybrid model does well on a
number of classification tasks, as seen by accuracy measures, which back up this
tendency. Efficiency research looked at testing and training times,
computational complexity, and resource use, and it found that the Hybrid model
was the best way to classify NFTs. The Hybrid model only needed 2.5 hours of
training, whereas Inception and DenseNet needed 3.2 and 2.8 hours, respectively.
The Hybrid model cut testing time by a lot, only taking 18 seconds for each
batch. This is a big difference over CNN and DenseNet. The Hybrid approach makes
the most of processing time by maximizing performance and efficiency. It does
require a little more computing power since its average GPU use is 78% compared
to CNN's 70%. Our research shows that a Hybrid model is a good option for
classifying NFTs using deep learning since it combines several architectures to
improve accuracy and efficiency. The main outcome of this study is a hybrid
architecture that efficiently integrates complementary strategies to address the
shortcomings of individual deep learning models. |
|
Keywords: |
Metaverse, IoT, NFT, Deep learning, Machine learning, Image classification |
|
DOI: |
https://doi.org/10.5281/zenodo.20497492 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
PERFORMANCE ANALYSIS OF INFINITE FAILURE NHPP SOFTWARE RELIABILITY MODEL USING
LOG-BASED LIFETIME DISTRIBUTIONS |
|
Author: |
DAE YU KIM |
|
Abstract: |
This study analyzes the performance of an infinite-failure NHPP software
reliability model using log-based lifetime distributions (Gompertz,
Log-Logistic, and Pareto), in which the failure rate varies nonlinearly over
time. Software failure-time data are used for model construction, and the
parameters are estimated using maximum likelihood estimation (MLE). Through
various evaluation methods, including model efficiency assessed using MSE and
R², predictive accuracy evaluated by the m(t) function, failure occurrence
intensity examined using the λ(t) function, and model reliability analyzed
through the R ̂(τ) function, the Log-Logistic model was confirmed to be the most
effective among the proposed models. These findings indicate that the proposed
approach can be effectively applied to reliability assessment across diverse
software environments. Consequently, this study clarifies the previously
underexplored reliability characteristics of log-based lifetime distributions
and is expected to provide developers with a fundamental analytical technique
for assessing software failure rates during the early stages of development. |
|
Keywords: |
Gompertz, Infinite Failure, Log-based, Log-Logistic, NHPP, Pareto. |
|
DOI: |
https://doi.org/10.5281/zenodo.20497501 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
DESIGN OF AN INTEGRATED EXPLAINABLE PREEMPTIVE AND ADAPTIVE FINANCIAL FRAUD
DETECTION MODEL FOR REAL-TIME TRANSACTION SYSTEMS |
|
Author: |
MUDIMELA MADHUSUDHAN, PRAMODA PATRO |
|
Abstract: |
The rise in the number of digital payment systems, online banking and financial
transaction systems has greatly augmented fraudulent activities. The traditional
fraud detection models are mainly based on fixed machine learning models which
are not able to keep up with changing fraud trends and they are not easily
interpretable which restricts their applicability in real time financial
settings. In this paper, a combination of explainable and adaptive fraud
detection system will be proposed to detect fraudulent transactions beforehand
and to ensure transparency in the decisions. The architecture suggested includes
Drift-Aware Contrastive Embedding (DACE) which is proposed to represent features
adaptively, Intrinsic Explainable Neural Tree (IxENTree) which is suggested to
classify using explanations, Meta-Adaptation by Few-Shot Fraud Transfer (MAFT)
which is proposed to detect new fraud schemes quickly, Adversarial and
Counterfactual Explainability Stress Testing (ACEST) that should ensure the
robustness of an explanation, and Continuous Feedback and Audit Loop (CFAL) that
An evaluation of the system is done by financial transaction datasets and
compared with conventional machine learning and deep learning models. The
experimental findings indicate that the given framework is characterized by a
better accuracy of detection, higher recall, and lower false-positive rates, and
the Area Under the ROC Curve (AUC) reaches about 0.97. Adaptive learning and
explainable artificial intelligence allow the suggested system to achieve
trustworthy and understandable fraud detection in live financial transactions
circumstances. Existing fraud detection systems struggle to adapt to concept
drift and often lack interpretability required for regulatory compliance.
Experimental results demonstrate that the proposed framework improves fraud
detection robustness, achieving superior recall and reduced false-positive rates
while maintaining transparent decision-making suitable for real-time financial
systems. |
|
Keywords: |
Financial Fraud Detection, Explainable Artificial Intelligence, Concept Drift
Detection, Adaptive Learning, Counterfactual Explanations. |
|
DOI: |
https://doi.org/10.5281/zenodo.20497533 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
A SCALABLE PRIVACY-PRESERVING DATA MINING FRAMEWORK FOR MULTI-SITE CLOUD
ENVIRONMENTS |
|
Author: |
ANKITA SINGH, KANIKA GARG |
|
Abstract: |
The increasing adoption of cloud-based data analytics has enabled organizations
to perform large-scale data mining using distributed resources; however, this
raises several concerns about data privacy, security, and legal/compliance
issues with regard to cloud-based data mining. For many real-world applications,
the owners of sensitive data will be multiple independent locations that do not
want or are not willing to send their raw data to a central cloud due to privacy
issues. Existing data mining techniques in the cloud typically rely on either
direct data outsourcing or limited trust, making them impractical for use in
privacy-sensitive environments where data is scattered across multiple
independent sites. Thus, this article presents a scalable framework for
privacy-preserving data mining within a multi-site cloud environment that allows
for collaboration between sites via data mining while keeping confidential the
sensitive data during the entire data mining process. Specifically, each owner
of data at a site will preprocess its data locally (i.e., at the site) and apply
a form of homomorphic encryption to its data prior to sending it to the cloud.
Further, secure aggregation will be employed to aggregate (i.e., combine) the
contributions of the encrypted data from many different site owners together
(i.e., without decryption). Finally, all of the data mining operations will be
performed directly on the encrypted data in the cloud using an
honest-but-curious model for threat. Plaintext cannot be transmitted to the
cloud, as only an authorized analyst can decrypt it. The innovative nature of
this new way of working is in the way that it has been designed at a framework
level to consider privacy preservation, multiple-user workloads across multiple
sites, scalability, and a myriad of other requirements, rather than simply
implementing one algorithm or application at a time. Additionally, scalability
in terms of the number of users participating and size of data being processed
both are taken into account within the proposed framework making it feasible to
support real-world cloud deployments. Testing against the UCI Adult Census
Income Dataset demonstrates that the proposed framework produces a similar
classification performance to other solutions that do not have any encryption
but provides additional computational burden due to encryption. Moreover,
further analysis confirms the ability to scale and the ability to validate
privacy has been addressed using this new approach. Therefore, the proposed
framework provides an efficient, scalable solution for conducting
privacy-sensitive data mining using distributed cloud environments. The primary
research contribution of this work is the design and implementation of a
unified, framework-level solution that simultaneously addresses three critical
challenges — privacy preservation, multi-site collaboration, and scalability —
which have not been collectively addressed in prior literature. Unlike existing
algorithm-specific approaches, this framework introduces a generalised
architecture that integrates homomorphic encryption and secure aggregation into
a cohesive pipeline applicable to diverse cloud-based data mining scenarios. |
|
Keywords: |
Privacy-Preserving Data Mining; Homomorphic Encryption; Secure Aggregation;
Multi-Site Data Analytics; Cloud Computing; Data Privacy |
|
DOI: |
https://doi.org/10.5281/zenodo.20497544 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
EARLY PREDICTION AND ANALYSIS OF FINANCIAL CRISIS USING LIGHT-WEIGHTED RESIDUALS
OF MULTI-GATED RECURRENT UNIT NETWORKS (MULTI-GRU) |
|
Author: |
RAMSHEEL M, SHANMUGAPRIYA M |
|
Abstract: |
This study investigates the vital necessity for early detection and analysis of
financial crises, which are necessary for reducing their negative consequences
on world economic stability. Traditional forecasting methods can lack timely and
accurate predictions due to simplicity or single-layered neural networks. To
enhance the forecast of financial crises, especially with regard to stock market
movements, we provide a special solution combining Multi-Gated Recurrent Unit
Networks (MULTI-GRU) with lightweight residuals in reaction to these
limitations. Our approach includes model creation and integration following a
thorough data collection and preparation strategy. Applying this approach to
past financial data produces a forecast accuracy of 99.04%, far exceeding
present models in terms of memory efficiency, accuracy, and general predictive
capacity. Furthermore, supported by comparative performance evaluation metrics,
we provide a comprehensive analysis of performance indicators, including
standard deviation and performance comparison metrics. This work informs
financial market participants and regulatory authorities with practical ideas
for proactive risk management strategies and helps to clarify financial crisis
dynamics. The enlarged model significantly improves financial crisis
forecasting, hence promoting more resilience in worldwide financial
institutions. The findings of this study are expected to be beneficial for
financial institutions, investors, regulatory agencies, policymakers, and
researchers working in financial risk analysis and economic forecasting. The
proposed MULTI-GRU framework can assist financial analysts and market
participants in identifying early warning signals of financial instability and
improving proactive risk management strategies. Regulatory authorities and
policymakers may utilize the model for monitoring market behaviour and reducing
the impact of potential financial crises. In addition, researchers in artificial
intelligence and financial analytics can use the proposed framework as a
reference for developing advanced deep learning models for financial forecasting
applications |
|
Keywords: |
Financial Crisis Prediction, Multi-Gated Recurrent Unit Networks (MULTI-GRU),
Stock Market Analysis, Early Warning Systems, Neural Network Residual Analysis. |
|
DOI: |
https://doi.org/10.5281/zenodo.20497551 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
AN ENHANCED VGG16 BASED DEEP LEARNING FRAMEWORK FOR DIABETIC RETINOPATHY
SEVERITY GRADING |
|
Author: |
ASHOK KUMAR KAVURU, RAJESH KUMAR PATJOSHI, RAKHEE PANIGRAHI |
|
Abstract: |
Diabetic retinopathy (DR) is a common and severe effect of diabetes that
contributes significantly to vision impairment worldwide. Retinal fundus imaging
is widely used for the detection and monitoring of this condition; however,
accurate interpretation of these images requires substantial clinical expertise.
Hence, achieving consistent and reliable assessment is challenging, particularly
given the limited availability of trained specialists. To address this issue,
this work presents a revised VGG16 model for automatically classifying DR types.
The proposed approach was assessed using standard evaluation metrics and
achieved 95.43% accuracy on the preprocessed APTOS2019 fundus image dataset. The
results suggest that the proposed method is well-suited for early screening and
severity grading of diabetic retinopathy, enabling the timely diagnosis and
lowering the risk of vision loss. |
|
Keywords: |
Diabetic Retinopathy, VGG16, Fundus Images, APTOS2019 |
|
DOI: |
https://doi.org/10.5281/zenodo.20500622 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
ADAPTIVE MULTI-MODAL TRANSFORMER NETWORKS FOR TIME-SERIES FORECASTING WITH
UNCERTAINTY QUANTIFICATION |
|
Author: |
K SRI VIJAYA, DR SAURABH SHARMA, DR D.BHAVANA, DR.L.KANYA KUMARI, DR. HARI
JYOTHULA, DR SUBBA RAO POLAMURI, MEDARAMETLA ANUSHA RANI, SARATH CHANDRA B |
|
Abstract: |
This paper introduces Adaptive Multi-Modal Transformer Networks (AMMTN), a novel
deep learning architecture that addresses the challenges of time-series
forecasting with heterogeneous data sources. We propose a transformer-based
framework that dynamically adapts to varying input modalities while maintaining
robustness to missing data and temporal irregularities. Our approach
incorporates uncertainty quantification through a novel composite loss function
combining predictive accuracy with calibrated confidence intervals.
Experiments conducted on five benchmark datasets (MIMIC-III, M4, ETT, Weather,
and Financial Markets) demonstrate that AMMTN consistently outperforms existing
state-of-the-art methods. Quantitatively, our model achieves a 17.3% reduction
in mean absolute error, 19.8% improvement in root mean squared error, and 22.6%
enhancement in uncertainty calibration. The performance gains are particularly
pronounced for long-horizon forecasts and datasets with significant missing
values, where AMMTN exhibits 28.4% better prediction accuracy compared to the
best baseline models.
Our theoretical analysis establishes convergence
guarantees for the adaptive attention mechanism, providing mathematical insights
into why the model excels at integrating information across multiple time scales
and modalities. Ablation studies confirm that each component contributes
meaningfully to overall performance improvement, with the adaptive cross-modal
attention mechanism providing the largest marginal benefit.
The scientific
contributions of this work are threefold: (1) a novel adaptive cross-modal
attention mechanism with formal convergence guarantees; (2) a composite loss
function combining negative log-likelihood with Continuous Ranked Probability
Score (CRPS) for jointly optimizing accuracy and calibration; and (3) empirical
validation across five heterogeneous benchmark datasets demonstrating consistent
superiority over seven state-of-the-art baselines. |
|
Keywords: |
Multi-Modal Transformers, Uncertainty Quantification, Time-Series Forecasting,
Adaptive Attention, Multivariate Prediction, Heterogeneous Data Integration,
Calibrated Confidence Intervals |
|
DOI: |
https://doi.org/10.5281/zenodo.20497590 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
DACN-NET: ALZHEIMER’S DISEASE DETECTION USING DATA AUGMENTED CONVOLUTIONAL
NEURAL NETWORKS |
|
Author: |
NAGA DURGA SAILE K, BHARADWAJ V.Y., V. S. R. K. RAJU , P. LAKSHMI SRUTHI, Dr. K.
SYAMALA DEVI, Dr. VIJAYA LAKSHMI V. |
|
Abstract: |
Alzheimers disease develops progressively and leads to irreversible cognitive
decline as well as memory impairment. Therefore, the importance of early and
accurate diagnosis cannot be overstated. Hence, the need for clinical
intervention and disease management is well established. In this paper, we
evaluate the efficacy of various deep learning-based techniques to classify
patients with Alzheimer’s using structural magnetic resonance imaging (MRI)
data. Our study includes the assessment of the performance of multiple CNN
(Convolutional Neural Network) architectures, including two of the most popular
architectures in the transfer learning approach (i.e., ResNet50 and GoogLeNet).
In addition, we developed a new deep learning architecture, DACN-Net,
specifically for classifying patients with AD using MRI images, which were
preprocessed to grayscale and resized to 128 × 128 ×1 pixels. When comparing the
performance of the different models, we found that DACN-Net achieved the best
performance, with an accuracy of 93.6%, when classifying patients with AD from
the OASIS MRI dataset that included four clinically relevant categories of AD.
Our results demonstrate that DACN-Net can accurately classify early-stage cases
of AD and suggest that the use of deep learning-based approaches could be a
viable tool for the early detection of AD. |
|
Keywords: |
Alzheimers disease, MRI, CNN, DACN-Net, ResNet50, Deep Learning. |
|
DOI: |
https://doi.org/10.5281/zenodo.20497603 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
WORKLOAD-AWARE DEPLOYMENT DECISION-MAKING IN MICROSERVICE ARCHITECTURES: AN
EMPIRICAL FRAMEWORK WITH PERFORMANCE INDEX |
|
Author: |
NIDHI VANIYAWALA, KAMLENDU KUMAR PANDEY |
|
Abstract: |
Microservice architectures support flexible deployment across heterogeneous
environments; however, selecting an appropriate deployment strategy for varying
workload characteristics remains a significant challenge. This study presents an
empirical workload-aware evaluation of monolithic, microserver, and orchestrated
deployment environments using a normalized Performance Index (PI) to measure
deployment efficiency. Two workload categories were experimentally evaluated:
read-intensive single-service operations and multi-service transactional
workflows. The proposed PI provides a unified quantitative metric for comparing
heterogeneous deployment environments based on performance behaviour under
varying workloads. Experimental results demonstrate a strong workload-dependent
variation in deployment efficiency. For read-intensive workloads, monolithic
deployment achieved the highest efficiency (PI = 0.85), outperforming
microserver (PI = 0.34) and orchestrated deployments (PI = 0.13). Conversely,
for multi-service transactional workloads, orchestrated deployment achieved the
best performance (PI = 0.70), while monolithic deployment exhibited minimal
efficiency (PI = 0.06) and microserver deployment showed moderate performance
(PI = 0.31). The findings confirm that no single deployment paradigm is
universally optimal for all workload types. The proposed PI-based evaluation
framework supports workload-aware deployment selection and enables data-driven
optimization of microservice systems within Intelligent DevOps environments. |
|
Keywords: |
Microservice Architecture, Empirical Software Engineering, Containerization,
Intelligent DevOps, Performance Benchmarking, Observability |
|
DOI: |
https://doi.org/10.5281/zenodo.20497613 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
ADAPTIVE AND DYNAMIC LEARNING STYLE CLASSIFICATION USING HYBRID LSTM-GNN
FRAMEWORK AND SELF SUPERVISED LEARNING IN E LEARNING |
|
Author: |
P.KIRUTHIKA, DR.V.VASANTHI |
|
Abstract: |
Learning style identification is a crucial part of enhancing personalization in
e-learning platforms. Due to dynamic and unstable user behavior, learning style
detection needs more appropriate labeled datasets. The contribution of the work
is developing a novel learning style classification model, “Dynamic Learning
Style Classification (DLSC)”, with a new learning style named as “Cumulative
learning style model (CLSM)”, which is based on implicit and explicit feedback,
access log, and contents from a custom-built e-learning platform. The key
process behind this model is collecting both implicit and explicit data from a
dynamic e-learning platform and applying it to the DLSC to effectively find the
learning style. The proposed learning style consists of 17 styles which are
developed from the analysis of existing learning styles. A different set of
attributes in the implicit, explicit category, Artificial Intelligence
(AI)-driven, and behavioral tracking attributes are gathered and used to find
the learner learning style. Learner’s navigation patterns, time spending
patterns, User Interface (UI) interactions, and more implicit attributes, such
as learning goals, preferred content format and other preference based explicit
attributes are gathered to make the learning style identification process much
stronger. For effortless training and effective feature extractions, Graph Auto
Encoder (GAE) with self-supervised learning (SSL) is proposed. Finally, the
integration of Long short-term memory (LSTM) and Graph neural network (GNN) made
to the better classification result. The experiment has been carried out and
collected many raw data and then it is processed through a collective machine
and deep learning algorithms. The results show the proposed model achieves a
higher amount of precision and accuracy of 97.8 % in dynamic learning style
detection. |
|
Keywords: |
Learning Style Classification, Adaptive E-Learning, Graph Autoencoder (GAE),
Self-Supervised Learning (SSL), Long Short-Term Memory (LSTM), Graph Neural
Network (GNN). |
|
DOI: |
https://doi.org/10.5281/zenodo.20497626 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
NEUROGRAPH-FUSE: A PHYSICS-INFORMED GRAPH NEURAL FRAMEWORK FOR MULTIMODAL BRAIN
TUMOR CLASSIFICATION |
|
Author: |
NAGARAJU ARUMALLA, VEERRAJU GAMPALA |
|
Abstract: |
Accurate classification of brain tumors from multi-parametric magnetic resonance
imaging (MRI) is of crucial importance to clinical diagnosis and treatment
planning, but is difficult to address due to tumor heterogeneity, cross-site
variability, and prediction uncertainty. To tackle these challenges, we present
NEUROGRAPH-FUSE, a graph-based multimodal framework for physics-inspired
intensity regularization reasoning in neurograph inference that combines
convolutional neural networks for local feature extraction, transformers for
global context modeling, and GNNs for relational learning over tumor subregions.
The model contains a physics-intensity consistency module that uses MRI
acquisition priors to enforce biophysically feasible representations, a gated
multi-sequence fusion module that adaptively fuses heterogeneous MRI modalities,
and a tumor-centered contrastive learning scheme to maximize inter-class
separability. In addition, an evidential-graph lookahead adaptive optimizer
enhances convergence stability and a dual-branch evidential classifier adds
Dirichlet-based belief scores for uncertainty-aware predictions. Extensive
experiments on benchmark MRI datasets show that NEUROGRAPH-FUSE consistently
beats the state-of-the-art baselines such as Res-BRNet, BrainMRNet, and ViT
Fusion+Ensemble with improved accuracy, calibration, and cross-site
generalization. These results indicate that NEUROGRAPH-FUSE is a potential
candidate for practical and interpretable computer-aid diagnosis in real-world
neuro-oncology. |
|
Keywords: |
Convolutional Neural Network (CNN), Transformers, Graph Neural Network (GNN),
Multimodal Fusion, Evidential Deep Learning, Uncertainty Estimation, Contrastive
Learning |
|
DOI: |
https://doi.org/10.5281/zenodo.20497644 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
AN EXPLAINABLE HYBRID CNN–VISION TRANSFORMER FRAMEWORK FOR OFFLINE PULMONARY
DISEASE SCREENING ON EMBEDDED EDGE DEVICES |
|
Author: |
M. LAKSHMANAN, G.S. ANANDHA MALA, MANIKANDAN MOOVENDRAN, GOWDHAM C, JOSHUVA
AROCKIA DHANRAJ, SRIRAMKUMAR R, R. TAMILAMUTHAN |
|
Abstract: |
Early detection of pulmonary diseases remains challenging in rural and
resource-constrained healthcare environments due to limited diagnostic
infrastructure, shortage of specialists, and unreliable network connectivity.
Although Artificial Intelligence (AI)-based pulmonary screening systems have
shown promising diagnostic capabilities, most existing approaches rely on
cloud-dependent architectures and computationally intensive models that are
unsuitable for offline and low-power deployment. This paper proposes an
explainable lightweight hybrid Convolutional Neural Network–Vision Transformer
(CNN–ViT) framework for real-time pulmonary disease screening on embedded edge
devices. The novelty of the proposed framework lies in integrating CNN-based
local feature extraction, transformer-based global contextual learning, and
Grad-CAM explainability within a unified offline edge-AI inference system. The
framework processes chest X-ray images locally, enabling privacy-preserving and
low-latency diagnosis without cloud dependency. Publicly available datasets
including NIH ChestX-ray8, CheXpert, and COVIDx were used to evaluate the system
for multi-class pulmonary disease classification involving normal, pneumonia,
COPD-related, and infectious lung disease patterns. Experimental results
achieved 91.2% accuracy, 89.7% sensitivity, and 93.1% specificity, with an
average inference latency below two seconds and power consumption under 8 W. The
proposed framework demonstrates an interpretable, computationally efficient, and
scalable solution for offline pulmonary disease screening in underserved
healthcare settings. |
|
Keywords: |
Edge AI, Hybrid CNN–Vision Transformer, Explainable AI Track, Chest X-Rays
Analysis, Off-Line Medical Diagnostics, Embedded Healthcare Systems,
Resource-Constrained Settings. |
|
DOI: |
https://doi.org/10.5281/zenodo.20497654 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
Title: |
DEEP ENSEMBLE-BASED WEED DETECTION IN SPINACH CROPS WITH RESNET-101 FEATURE
EXTRACTION AND XGBOOST CLASSIFICATION |
|
Author: |
K. SWARUPA RANI, Dr.VANKUDOTHU MALSORU, POORNAKUMAR D, Dr. SARANGAM KODATI, K V
SATYANARAYANA, DR. T. PREM CHANDER, PADMAJA INTURI, Dr. D.GANESH, Dr. SIVA KUMAR
PATHURI |
|
Abstract: |
One of the provisions in precision agriculture is precise identification of the
weed which enables herbicides used to be applied under control and is a
provision under sustainable farming methods. This paper presents a deep ensemble
framework to identify weed in spinach production fields using ResNet-101 in the
extraction of deep features and XGBoost in the identification with a high level
of accuracy. The ResNet-101 architecture can be used to extract more difficult
spatial and textural features of the field images and the XGBoost classifier can
be well-generalized to find weeds and spinach plants. It is a hybrid
architecture that combines the robustness of the representation grade of deep
learning on one hand with gradient boosting in terms of efficiency and
robustness; it is an optimal tradeoff between the richness of features and the
speed of classification. The experimental study implemented based on an academic
dataset of spinach fields helps us to demonstrate that the developed ResNet-101
+ XGBoost model demonstrates a relatively stable high degree of performance when
compared to the classical CNN-based classifiers. The ensemble methodology still
provides a high level of performance even where such problematic factors as
varying light conditions, thick foliage coverage and the presence of background
noise are involved, which makes this methodology more applicable to real-life
applications. The proposed method achieved high accuracy of 96%that was far much
better than the baseline classifier and Support Vector Machines (88%) and
Decision Tree (90%) classifier. All in all, this efficient and robust system
offers a strong and dependable automated system of weed recognition in spinach
crops, which makes its contribution to the use of precision agriculture. |
|
Keywords: |
Weed detection, Deep Learning, ResNet-101, Support Vector Machine, Decision
Tree. |
|
DOI: |
https://doi.org/10.5281/zenodo.20497679 |
|
Source: |
Journal of Theoretical and Applied Information Technology
31st May 2026 -- Vol. 104. No. 10-- 2026 |
|
Full
Text |
|
|
|
