Journal of Electronics, Electromedical Engineering, and Medical Informatics

HAREN: A Hybrid Attention Residual Ensemble Network for PCOS classification and Prediction

Pragati Patil — 2026-05-07

Polycystic Ovary Syndrome (PCOS) is one of the most prevalent endocrine disorders affecting women of reproductive age and is a leading cause of infertility. Ultrasound imaging is widely used for PCOS diagnosis; however, visual assessment of ovarian morphology is highly subjective, time-consuming, and dependent on clinical expertise. Quality differences in ultrasound images, very near to similar visual patterns among PCOS and NOT PCOS images, and noise in the images increase the threat of improper diagnosis. These problems suggest a need for an accurate, automatic, and computer-assisted PCOS diagnostic system. This research aims to create a deep learning-assisted automatic PCOS diagnostic system which can detect and classify the Polycystic Ovary Syndrome from the gray-scale ultrasound ovarian images. In addition to high classification accuracy, the proposed framework incorporates an explicit explainability pipeline that highlights diagnostically relevant ovarian regions, such as follicular distributions and stromal patterns, thereby supporting clinically interpretable decision making. The proposed HAREN framework addresses the limitations of single backbone models, and attention augmented variants, such as vanilla ResNet50 and ResNet50 with hybrid attention by leveraging ensemble learning and residual feature fusion. HAREN combines three architecturally diverse and complementary pretrained CNN backbones (ResNet50, DenseNet121, and EfficientNetB0) to enhance feature diversity. In addition, a novel hybrid attention mechanism combining channel, spatial, and cross-scale attention is introduced to emphasize diagnostically relevant ovarian regions. A residual fusion strategy is employed to preserve discriminative features and stabilize training, and an explicit explainability pipeline is incorporated to support Grad CAM-based visual interpretation. This network first converts the ultrasound grayscale ovarian images to RGB , followed by the extraction of important features applying backbones, which are augmented with attention mechanisms. The network, trained with categorical crossentropy loss, was evaluated using comprehensive performance metrics on 11,784 ultrasound images (6,784 PCOS and 5,000 NOT PCOS). HAREN achieved 99.33% accuracy, 98.96% precision, 98.97% recall, 98.96% F1 score, and an AUC of 99.93%, outperforming conventional models. Overall, it delivers an accurate, reliable, and interpretable solution for automated PCOS detection, demonstrating strong potential for clinical decision support systems

Predicting the Severity of Thyroid Nodules with YOLOv8 and CA+LSR Architecture

Kalpana Devi — 2026-05-16

The rise in thyroid cancer has significantly increased the burden on radiologists to diagnose thyroid nodules using sonography accurately. To address this challenge, a highly precise and efficient automatic computer-aided diagnosis system is needed. A retrospective analysis was conducted on a dataset consisting of 200 ultrasound images from 161 patients (84 benign and 77 malignant) at Wenzhou Central Hospital. This study presents an enhanced version of the You Only Look Once version 8 (YOLOv8) neural network, specifically designed to improve the accuracy of thyroid nodule diagnosis. YOLO has been objective in handling the required elements from the given input images or frames, and the article discusses the extensive benefits of the same. The proposed network incorporates a Coordinate Attention (CA) module and a Label Smoothing Regularization (LSR) module, which facilitate the extraction of positional information and enhance overall performance. The improved neural network demonstrates high accuracy in identifying lesion areas and classifying nodule types, achieving a mean average precision (mAP) of 90% with an average inference time of 8 milliseconds on the test dataset. The ablation experiment revealed that incorporating the CA and LSR modules adds 1.2 milliseconds of computational time per image while providing a significant 4.1% improvement in mean average precision (mAP). Compared with state-of-the-art networks, the enhanced YOLOv5 network performed exceptionally well in diagnosing benign and malignant thyroid nodules, even with a limited dataset. Furthermore, its high accuracy and efficiency suggest potential applicability to other sonographic diagnostic tasks, aiding radiologists in improving diagnostic accuracy and patient outcomes.

Quantum-Inspired Feature Engineering and Explainable AI for Robust Heart Disease Classification

Rashmi Mothkur — 2026-05-18

Early and accurate prediction of cardiovascular disease is essential to improve patient outcomes and reduce healthcare costs. This research presents a hybrid classical–quantum machine learning framework for heart disease prediction using the Cleveland dataset. The proposed pipeline integrates advanced feature engineering, bio-inspired optimization, and quantum-inspired learning to improve classification performance and interpretability. The system applies multiple feature selection techniques followed by a hybrid feature fusion strategy. Orthogonal Component Analysis is then used for dimensionality transformation, while quantum-inspired feature mapping simulates quantum state coding. A feature selection mechanism based on a Genetic Algorithm optimizes the subset of features. Classical and quantum machine learning models are evaluated, including Random Forest, Gradient Boosting, K-Nearest Neighbors, Logistic Regression, Quantum Support Vector Classifier, Variational Quantum Classifier, Quantum KNN, and Quantum Neural Networks. Model performance is evaluated using accuracy metrics. To ensure transparency and trustworthiness, explainable AI techniques such as SHAP, LIME and DiCE are integrated to provide local and global interpretability of predictions. Experimental results demonstrate that the proposed hybrid framework improves predictive performance by achieving 90% accuracy compared to traditional machine learning approaches, while maintaining model explainability. The model achieved an overall accuracy of 90%, indicating strong predictive capability in cardiovascular disease risk classification. A detailed analysis of class-wise performance shows that for Class 0, the model obtained a precision of 0.85, a recall of 0.97, and an F1-score of 0.90, demonstrating excellent ability to correctly identify negative cases with minimal false negatives. For Class 1, the model achieved a precision of 0.96, a recall of 0.84, and an F1-score of 0.90, indicating high confidence in positive predictions, though with slightly lower recall compared to Class 0. This study highlights the potential of combining classical feature engineering, evolutionary optimization and quantum-inspired learning for next-generation medical decision support systems. The integration of quantum-inspired techniques also provides a promising direction for improving computational efficiency and model robustness in healthcare analytics. The findings suggest that hybrid classical–quantum learning approaches can support clinicians in making faster and more reliable diagnostic decisions.

Classification of Lenke Scoliosis using GLCM Feature Extraction and Support Vector Machine

Anna Nur Chamim — 2026-05-19

Lenke scoliosis is a spinal deformity that is classified into six types by the Lenke classification system. Traditionally, clinicians undertake classification based on manual visual examination of X-ray images, which is time-consuming, requires high skill and is subject to errors caused by human fatigue. To overcome these constraints, the current work presents an automated and reliable classification system to boost the efficiency and accuracy of diagnosis. The method is based on the application of the Grey Level Co-occurrence Matrix (GLCM) for the feature extraction and of a Support Vector Machine (SVM) classifier. The main contribution is the optimisation of SVM kernel functions (Quadratic, Cubic and Coarse Gaussian) using advanced pre-processing methods to achieve very good accuracy while preserving compute efficiency suitable for clinical applications. The approach combines picture pre-processing (grey scale conversion, resize, contrast improvement by adaptive histogram equalisation, segmentation, augmentation) and GLCM-based feature extraction and classification using multiple SVM kernels. The model's performance is evaluated based on accuracy, precision, recall, F1 Score, and execution time. The testing results demonstrate that the Quadratic SVM has the best classification accuracy of 92.26% with a processing time of 20.44 seconds, which outperforms the Cubic SVM (90.97%, 19.30 seconds) and the Coarse Gaussian SVM (60.64%, 21.70 seconds). The results show that the quadratic SVM has the optimum compromise between accuracy and processing efficiency. In conclusion, the proposed GLCM-SVM approach has tremendous potential to support doctors in the automatic categorisation of Lenke scoliosis, improving the accuracy and speed of diagnosis without requiring large computational resources. In future work, we will aim to expand the dataset and include additional features to further improve the model's resilience and generalisability.

A Cross-Scale Spatial–Channel Attention Inception Network for Efficient Medical Image Segmentation

Krishnakumar B — 2026-06-01

Medical image segmentation plays a crucial role in modern computerized diagnosis, as accurate delineation of anatomical structures directly impacts clinical decision-making and treatment planning. However, segmenting anatomically complex regions at a fine-grained level remains challenging, especially when computational efficiency is a key requirement. To address these challenges, the authors propose a novel, lightweight medical image segmentation framework, CSA-IncepLiteNet, designed to achieve high segmentation accuracy without imposing a significant computational burden. The CSA-IncepLiteNet architecture integrates two key innovations: cross-scale feature extraction and unified spatial channel attention learning. Central to this framework is the newly introduced Cross-Scale InceptionLite module, which efficiently captures multi-scale contextual information. This module is built using depth-wise separable convolutions and point-wise convolutions, enabling effective feature extraction while significantly reducing the number of trainable parameters. By learning features across multiple spatial scales, the network can better represent anatomically complex structures present in medical images. In addition, the authors propose a Cross-Scale Spatial Channel Attention (CSA) module that jointly models spatial saliency and channel-wise interdependencies within a unified attention-learning paradigm. This dual attention mechanism allows the network to focus on the most informative regions and feature channels simultaneously, leading to improved segmentation precision. The performance of CSA-IncepLiteNet was evaluated on the BUSI breast ultrasound dataset and multiple CT image modality-based datasets. Experimental results demonstrate that the proposed framework consistently outperforms existing state-of-the-art methods across all evaluated datasets. Notably, CSA-IncepLiteNet achieves an accuracy of 92.1% and a Dice coefficient of 82.94% on the BUSI dataset, while utilizing over 26 million fewer parameters than a conventional U-Net. These results highlight the model’s effectiveness, robustness, and suitability for resource-constrained medical imaging applications.

Hybrid CNN-Transformer Architecture for Robust Liver Tumor Segmentation in 2D CT Slices

Huda Dham Bader — 2026-06-01

Liver tumor segmentation from CT scans is a task affected by class imbalance, low contrast, and small lesion size. Manual segmentation is time-consuming and also suffers from inter-observer variability. We propose a 2D CNN-Transformer model with 20.3M parameters in an encoder–decoder structure with four transformer layers (8 heads, 2048 feedforward dimension). The model processes 2D axial slices due to GPU memory limits. The loss function combines Cross-Entropy, Dice, and Focal losses with α = 0.25 and γ = 2.0. Preprocessing includes CLAHE (clip limit = 2.0, 8×8 tiles) and gamma correction (γ = 1.2). From the LiTS dataset (131 volumes), 11 volumes with 1,688 slices were selected based on tumor presence, annotation quality, and artifact removal. A patient-level split of 80% for training, 10% for validation, and 10% for testing was used to prevent data leakage. The model achieved liver Dice = 0.916 ± 0.122 and tumor Dice = 0.810 ± 0.304. The 95% confidence intervals using bootstrapping (1,000 resamples) were [0.897–0.934] for liver and [0.765–0.856] for tumor. Best validation results at Epoch 98 were liver Dice = 0.938, tumor Dice = 0.823, and accuracy = 0.992. Pixel accuracy was 99.20% and was not used as the main metric due to class imbalance, where background pixels exceed 90%. An ablation study showed that CLAHE and gamma correction improved tumor Dice by 8.6% and liver Dice by 3.3% compared to a baseline without preprocessing. The model shows performance for liver tumor segmentation on a LiTS subset. External validation on the full dataset and multi-center data is required before clinical use

Comparative Analysis of Attention Mechanisms in Pix2Pix for Multimodal MRI Fusion

Ali-Abdelatif Betouil — 2026-06-12

Medical image fusion (MIF) is a key technique in medical imaging, which combines complementary information from different imaging modalities, thereby improving the accuracy of diagnosis, particularly for lesion detection and treatment planning. Deep learning has significantly advanced this area, with the development of generative models and transformers leading to improvements in fidelity and accuracy, although the study of the influence of attention mechanisms on these models remains limited to a single type or a single architectural placement. This paper offers an analytical examination of the architectures of Pix2Pix with three attention mechanisms (spatial attention, channel attention (Squeeze-and-Excitation), and self-attention), where they are tested in three different placement strategies (encoder-only, decoder-only, and encoder-decoder), using the BraTS2020 dataset, with training supervised by a pseudo-ground-truth derived from arithmetic averaging. We fused six MRI modality pairs (FLAIR-T1, FLAIR-T1ce, FLAIR-T2, T1-T1ce, T1-T2, T1ce-T2), evaluating them using different metrics, including SSIM, PSNR, NMI, Entropy, and Q^AB/F. Results show that, in all cases, attention integration can significantly improve the quality of fusion over baseline methods, including cGAN and standard Pix2Pix. Spatial attention with encoder-decoder placement shows the best results, with SSIM values up to 0.91 and PSNR superior to 25 dB for the heterogeneous modality pair FLAIR-T1. Similarly, channel and self-attention demonstrate their effectiveness, especially with encoder-decoder placements. Based on these findings, attention-based fusion systems can be practically designed in a way that enhances MMIF, and the importance of designing attention in accordance with the nature of the modality is emphasized for optimal fusion performance. Our study demonstrates its effectiveness and may serve as a foundation for future research.

Automated Detection and Grading of Tuberculosis Bacilli in Ziehl Neelsen-Stained Sputum Using YOLO with IUATLD-Based Classification

Syevana Dita Musvika — 2026-06-30

Tuberculosis (TB) remains one of the most pressing global health challenges, particularly in low- and middle-income countries, where diagnostic capacity is often limited. Accurate and efficient detection of Mycobacterium tuberculosis bacilli in sputum smear samples stained with Ziehl-Neelsen remains the cornerstone of TB diagnosis. However, conventional microscopic examination is inherently labor-intensive, subject to interobserver variability and prone to human error, leading to inconsistent diagnostic outcomes. Addressing these limitations, this study proposes the development of an automated bacilli detection and quantification system utilizing the YOLO (You Only Look Once) object detection framework, specifically the YOLOv8 architecture, to improve diagnostic accuracy, consistency, and efficiency in TB identification. The research methodology encompasses image acquisition of Ziehl Neelsen-stained sputum samples from the Microbiology Laboratory of Universitas Airlangga Hospital (RSUA) and publicly available repositories, followed by meticulous annotation using Roboflow. The annotated dataset was employed to train the YOLOv8 model, and performance was evaluated through key metrics, including accuracy, precision, and error rate. The developed model achieved an overall accuracy of 73.33%, with class-wise accuracies of 100% for BTA 1+, 80% for BTA 2+, and 40% for BTA 3+ categories, conforming to IUATLD classification standards. The suboptimal performance observed in the BTA 3+ category was attributed to discrepancies in Field of View (FOV) alignment between the microscope’s ocular lens and the attached digital camera, affecting image consistency. Despite this limitation, the results demonstrate the potential of YOLO-based automated detection systems to reduce dependence on manual analysis, enhance diagnostic objectivity, and accelerate TB screening workflows. Future work should prioritize hardware calibration, particularly FOV synchronization, and dataset diversification to further refine model performance and clinical applicability. The proposed approach represents a significant step towards scalable, rapid, and reliable TB diagnosis, with implications for broader adoption in resource-constrained healthcare environments.

Wavelength Configuration and Signal Duration for Low-Complexity PPG-Based Anemia Detection: A Preliminary Validation Study

Mulia Rahmah — 2026-07-01

Anemia remains a major global health problem, while standard diagnosis still depends on invasive hemoglobin testing, which may be less practical for repeated and resource-limited screening. Photoplethysmography (PPG) offers a potential non-invasive alternative, but the contribution of different wavelength configurations to anemia classification remains unclear. This preliminary subject-based validation study evaluated the effect of PPG wavelength configuration and recording duration on low-complexity anemia classification. A public dataset containing green, red, and infrared PPG recordings from 52 subjects was used, consisting of 42 normal and 10 anemia subjects. Eight morphological and temporal features were extracted from each wavelength. Seven signal configurations, namely Green, Red, IR, Green+Red, Green+IR, Red+IR, and all channels, were evaluated across 30, 45, 60, and 90 s recording durations. Support Vector Machine, Logistic Regression, Random Forest, and Extra Trees classifiers were trained using class-weighted learning and assessed with 5-fold subject-based cross-validation to reduce subject-level data leakage. The Red+IR configuration with a class-weighted SVM at 90 s achieved the best pooled performance, with a macro F1-score of 0.754, F1-Anemia of 0.588, anemia recall of 0.500, anemia precision of 0.714, accuracy of 0.769, and an error rate of 0.231. Fold-wise analysis showed substantial variability, with a macro F1-score of 0.617 ± 0.251, sensitivity of 0.467 ± 0.506, specificity of 0.846 ± 0.144, ROC-AUC of 0.864 ± 0.150, and PR-AUC of 0.694 ± 0.344. These findings suggest that adding more PPG wavelengths does not necessarily improve classification performance. However, the model still missed 5 of 10 anemia cases, and the limited anemia recall, small minority class, and demographic imbalance indicate that the results should be interpreted as preliminary and require validation on larger, more balanced datasets.

Three-Arm Robotic Diagnostic Coordination Using Artificial Neural Network-Based Decision Support

Hariprasath Manoharan — 2026-07-01

The growing demand for smart healthcare systems and increasing burden on healthcare professionals have necessitated the need for autonomous diagnostic technologies that can facilitate real-time clinical decision-making. Current robotic diagnostic systems are often limited to discrete tasks, including sensing, monitoring, and diagnostic support. This results in limited coordination, transparency, and decision-making capabilities. The aim of the proposed method is to design a three-arm diagnostic robot with Artificial Neural Network (ANN) intelligence to improve healthcare support. The proposed framework includes dedicated robotic arms for sensing, visualization, and diagnostic tool manipulation, along with a coordinated communication architecture. A decision-support module based on an ANN gathers diagnostic information from the different subsystems of a robot and offers intelligent diagnostic evaluations. A seven-axis coordination approach is implemented to improve the synchronous performance of robotic components and to reduce the operational liabilities during diagnostic operations. The proposed framework was evaluated with four scenarios, and the performance was assessed in terms of transparency, coordination efficiency, association error, diagnostic accuracy, sensing latency, and communication delay. The experimental results showed that the proposed system achieved a diagnosis accuracy of 93% versus 71% for the baseline method. Moreover, the framework achieved 93% of transparency rate, 85% of coordination efficiency, 12% of reduction of association error, a 40 ms sensing latency, and a 15 ms communication delay. Statistical analysis reported consistent performance with deviation values of 1.2%, 1.7%, and 1.3% for arm coordination, visualization, and diagnostic tool management, respectively. The results confirm that the combination of ANN-based decision support and synchronized multi-arm robotic work can significantly improve the diagnostic efficiency and the operational reliability. The proposed architecture provides a strong foundation for future intelligent healthcare systems and enables the development of autonomous robotic diagnostics for advanced medical applications

Dynamic Fine-Tuning Strategy of Deep Learning Models for Lung Disease Classification on Chest X-ray Images

Phuoc-Hai Huynh — 2026-07-01

Lung diseases remain a leading cause of life-threatening illnesses worldwide, particularly in developing countries with limited healthcare resources. In recent years, deep convolutional neural networks (CNNs) have demonstrated strong potential in the automatic interpretation of chest X-ray (CXR) images. However, existing approaches often rely on rigid two-stage fine-tuning or fixed-step progressive unfreezing strategies, which may fail to effectively adapt pretrained representations or destabilize optimization, especially when applied to imbalanced real-world datasets. This study proposes a validation-driven dynamic fine-tuning strategy for transfer learning that adaptively unfreezes network layers based on convergence signals observed on the validation set rather than predefined training epochs. By coupling the timing and depth of adaptation to generalization behavior, the proposed method enables controlled knowledge transfer while mitigating catastrophic forgetting and improving training stability. Experiments were conducted on a large-scale, real-world clinical dataset comprising 15,416 CXR images collected at An Giang Regional General Hospital, Vietnam. The proposed strategy was systematically evaluated across multiple CNN backbones, including Xception, DenseNet121, EfficientNetV2S, InceptionV3, MobileNet, ResNet50, and VGG16. Performance was assessed using overall accuracy and macro-F1 score to address class imbalance. Results demonstrate consistent improvements across all architectures, with a mean accuracy gain of 3.18% compared to conventional fine-tuning (p = 0.02). MobileNet achieved the best performance with 85.1% accuracy and 66.8% macro-F1, while maintaining a compact model size of 73.05 MB. These findings indicate that validation-driven dynamic fine-tuning provides a stable, statistically significant, and deployment-feasible transfer learning mechanism suitable for real-world clinical environments.

Skin Cancer Classification by Applying Different Models of Artificial Intelligence

Wajid Dawood Alwan — 2026-07-06

Accurate multiclass classification of dermoscopic skin lesions remains challenging because of high inter-class visual similarity, substantial intra-class variability, and frequent acquisition artifacts (black borders, hair occlusions, noise). We propose a unified, reproducible framework that systematically coordinates four stages: (i) artifact-aware preprocessing (field-of-view circular cropping, hair removal, CLAHE, bilateral filtering); (ii) lesion-focused segmentation via GrabCut-refined fusion and a U-Net with EfficientNet-B3 encoder; (iii) compact deep-feature extraction (EfficientNet-B7) refined by principal component analysis and Neural Spline Flow density calibration; and (iv) robust machine-learning classification. The HAM10000 dataset (n = 10,015, seven diagnostic classes) was partitioned once by stratified random sampling into training (70 %, n = 7010), validation (15 %, n = 1502), and test (15 %, n = 1503) subsets under a strictly sequential anti-leakage protocol with patient-level isolation; the test set was sequestered until terminal evaluation. External generalization was assessed on an independent ISIC 2019 subset (n = 350, 50 per class) without retraining. On the held-out HAM10000 test set, XGBoost achieved the highest accuracy of 99.47 % with an F1-score of 98.99 %, followed by LightGBM (98.20 %) and MLP (97.67 %). Ablation analysis confirmed incremental gains of +2.55 % (preprocessing), +1.75 % (segmentation), and +1.32 % (Neural Spline Flow refinement). On the external ISIC 2019 data, MLP attained the best cross-domain accuracy of 95.43 %, demonstrating that the feature backbone generalizes beyond the training distribution. The demonstrated synergy of artifact suppression, lesion-centered segmentation, and density-calibrated feature learning yields highly discriminative and generalizable representations, providing a robust foundation for reliable computer-aided dermatologic screening

Design and Mechanical Evaluation of a Polymer Keel SACH Foot Using Finite Element Analysis and Experimental Validation

Agus Setyo Nugroho — 2026-07-10

Conventional Solid Ankle Cushion Heel (SACH) foot commonly uses wooden keels, which may exhibit variability in mechanical properties and limited long-term durability. This study aimed to develop and evaluate a polymer-based keel SACH foot as a low-cost alternative with improved mechanical performance. An integrated methodology combining finite element analysis (FEA) and experimental validation was employed. Three polymer keel SACH foot configurations were designed and analyzed under loading conditions that represent heel strike, mid-stance, and terminal stance, in accordance with ISO 10328. The optimal design was subsequently fabricated and tested under static loading conditions to validate the numerical model. The models utilized ABS for the keel, HDPE for the footplate, and vulcanized rubber for the foot body. Mechanical performance was assessed through total deformation, von Mises stress, strain, and safety factor analysis. Among the three configurations, Model A demonstrated the best mechanical performance, with the lowest average deformation (20.12 mm), the lowest stress concentration (12.67 MPa), and the highest safety factor (1.59). The selected design was subsequently fabricated and validated experimentally under static loading conditions up to 1176.78 N. Experimental validation showed strong agreement with FEA predictions, with deviations below 5% across all gait phases, confirming the accuracy of the FEA model. Comparative testing against a conventional wooden-keel SACH foot revealed significantly lower deformation values for the polymer-based keel SACH foot (p < 0.05), indicating improved structural stiffness and more efficient load distribution during loading process. These findings suggest that replacing conventional wooden keels with polymer-based structures can enhance mechanical consistency and structural reliability, while maintaining manufacturability and cost-effectiveness. The proposed design offers a promising approach to developing an affordable and durable prosthetic foot, particularly for use in low- and middle-income countries

Deep Learning Based Pain Recognition via Facial Expression Using Feature Fusion of Visual and Thermal Images

Raihan Islamadina — 2026-07-14

Objective pain assessment in non-verbal patients remains a significant clinical challenge. While automated facial expression analysis offers a promising solution, existing methods often rely on simple cross-modality concatenation or shallow attention mechanisms, which typically fail to fully capture complex, interdependent cross-modal dynamics. To address this limitation, this study proposes a deep learning-based multimodal feature fusion framework that combines visual and thermal images to significantly enhance pain detection accuracy. The proposed framework integrates pretrained convolutional neural network backbones, specifically VGGFace, ResNet50, and DEYOLO, with two novel attention modules: Dual Semantic Enhancing Channel Weight Assignment (DECA), and Dual Spatial Enhancing Pixel Weight Assignment (DEPA) to adaptively optimize joint feature representations. For comparison, a hybrid baseline model that processes visual and thermal images separately was also developed, allowing a comparative analysis between fusion-based and non-fusion-based approaches. The model performance was systematically evaluated on the MIntPain dataset, which comprises 20 healthy subjects experiencing five distinct levels of pain intensity. To ensure data independence and prevent identity leakage across the training and testing phases, a strict subject-wise data split protocol was implemented with 15 subjects allocated for training and 5 subjects for testing. Experimental results demonstrate that the proposed multimodal fusion framework achieves superior performance, attaining a peak F1 score of 0.938 using the VGGFace backbone. Furthermore, external validation on the UNBC McMaster Shoulder Pain dataset yields a classification accuracy of 0.872, confirming the strong generalization capability and stability of the framework across unseen subjects. These findings highlight the effectiveness of visual-thermal synergy and the efficacy of the proposed DECA and DEPA modules, showcasing high potential for robust, non-invasive clinical pain-monitoring and assessment applications

Intelligent Fusion of Multi-Modal Medical Imaging: A Comprehensive Review of Methods, Challenges, and Clinical Integration

Majda Maatallah — 2026-06-12

Multimodal Medical Imaging Fusion (MMIF) is defined as the incorporation of information from multiple imaging modalities in a way that is mutually supplementary, thereby addressing limitations associated with using a single imaging modality to evaluate a patient and increasing diagnostic accuracy. Further, this review provides a dedicated synthesis of deep learning architectures in MMIF, examining CNN-based hybrids, attention-enhanced transformers, GAN-driven unsupervised fusion, and emerging diffusion models. The state of the art in MMIF can be classified into three levels of fusion: (1) pixel level, fusion of raw pixel intensity values to preserve spatial detail; (2) feature level, features are derived from textures, edges, and region-of-interest (ROI) descriptors; (3) decision level, fusing independent outputs of each source using ensemble or rule-based methods to produce a single, integrated output from all sources, potentially improving interpretability of the integrated output. The use of AI algorithms improves fusion outcomes by yielding higher-quality results. However, clinicians' confidence in deep-learning-based models is limited due to their inability to generalise across multiple scanners, protocols, and medical systems. This analysis demonstrates that clinical AI systems must be developed with interpretability as a core attribute, to provide an explanation of how each modality is contributing to the final decision, and to establish a fusion policy that preserves the ability to make accurate diagnostic determinations based on fused images. In addition to developing more sophisticated algorithms, future developments in MMIF will require collaborative partnerships between developers and clinicians to develop fused images into reliable diagnostic tools to be used in precision medicine.

Cloud-Edge Collaborative Computing Framework for Stroke Disease Classification Using Machine Learning

I Made Suartana — 2026-07-04

Stroke is the second leading cause of death and the third leading cause of disability worldwide. Artificial intelligence-based early detection in distributed environments faces three main obstacles: high latency in centralized cloud approaches, risks to patient data privacy during data transmission, and class imbalance in stroke datasets. This study proposes a three-layer collaborative computing framework, Cloud-Edge Collaborative Computing (CECC), which intelligently distributes the computational workload between edge nodes and the cloud for IoMT-based stroke risk classification. The primary novelty of this study lies in the hierarchical computing collaboration that enables real-time preprocessing at the edge layer, centralized model training at the cloud layer, and a local differential privacy mechanism (LDP, ε=0.5) that preserves patient data confidentiality during transmission, all entirely evaluated within a single unified multi-criterion benchmarking protocol. Gradient Boosting achieved the best performance in the hold-out evaluation with an accuracy of 95.01% and an AUC-ROC of 0.994. The CECC framework reduced inference latency by 44.9% (286.2ms to 157.8ms), bandwidth by 73.9% (3,240 to 847 Kbps), and memory by 84.4% (312.4 to 48.7 MB) with an accuracy degradation of only 0.30% compared to cloud only. This study is a simulation-based framework evaluation using a public retrospective dataset prospective clinical validation in a real IoMT environment remains necessary before actual clinical implementation because the dataset used is retrospective, small, highly imbalanced, and was not collected from a real IoMT system

Semantic-Filtered SMOTE-PSO for Breast Cancer Trial Eligibility Classification

Taslim — 2026-07-05

This study addresses breast cancer clinical trial eligibility classification from free-text criteria under severe class imbalance, a condition that biases learning toward the majority class and complicates screening decisions when false positives and false negatives carry different operational costs. The study evaluates whether semantic plausibility control and optimization improve classification performance and screening-oriented error trade-offs under imbalanced conditions. The main contribution of this study is the proposed BEACoN framework, which integrates semantic-filtered augmentation and PSO-guided optimization within a unified screening-oriented eligibility classification setting. Four BioBERT-BiLSTM variants were evaluated using fixed train-validation-test partitions across three random seeds: a baseline model (M1), SMOTE augmentation (M2), SMOTE with cosine filtering (M2.5), and the proposed BEACoN framework (M3). Performance was evaluated using Precision, Recall, F1, AUROC, and AUPRC with pooled multi-seed statistical analysis to improve robustness and reduce single-seed bias. The evaluated augmentation-based configurations achieved pooled F1 scores up to 0.9381 ± 0.0005, AUROC up to 0.9976 ± 0.0001, and AUPRC up to 0.9808 ± 0.0004, indicating improved screening-oriented classification performance relative to the baseline. However, SMOTE with cosine filtering behaved broadly similarly to standard SMOTE under the evaluated embedding setting, indicating that the selected cosine threshold functioned largely as a permissive constraint, although modest seed-dependent prediction differences were still observed. Although BEACoN did not demonstrate statistically significant superiority over SMOTE in aggregate performance, it provided a more balanced false-positive and false-negative trade-off under comparable classification performance. Overall, the findings suggest that plausibility-controlled augmentation may provide practical value for screening-oriented eligibility classification under severe class imbalance

Accuracy Enhancement of a Hybrid CNN–VGG16 Architecture through Dropout Regularization Strategy for Breast Cancer Histopathology Classification

Fawaidul Badri Fawaid — 2026-07-08

Breast cancer remains the leading cause of cancer-related mortality among women globally, necessitating accurate diagnosis through histopathological image analysis. However, manual examination of these images is time-consuming and susceptible to inter-observer variability, highlighting the critical need for reliable automated computer-aided diagnostic (CAD) systems. This study was conducted to systematically evaluate and optimize convolutional neural network (CNN) architectures for automated classification of breast cancer histopathology images, with a focus on mitigating overfitting and enhancing diagnostic accuracy through hybrid deep learning methodologies. The principal innovation is the development of a CNN-VGG16 hybrid architecture that strategically integrates pre-trained feature extraction with a customized CNN framework, hypothesized to substantially improve classification accuracy and model generalization. Three model configurations were developed and comparatively analyzed: (1) baseline CNN, (2) CNN with dropout regularization, and (3) hybrid CNN-VGG16 model. Input images underwent preprocessing, including resizing to 150×150 pixels, normalization, and data augmentation. All models were trained with identical hyperparameters: an Adam optimizer with a learning rate of 0.001, a batch size of 32, and 10 epochs. Dropout regularization with a fixed rate of 0.5 was applied to fully-connected layers to mitigate overfitting. Model evaluation was conducted utilizing standard performance metrics. The proposed CNN-VGG16 hybrid model achieved superior performance: accuracy of 85.19%, precision of 87.16%, recall of 92.75%, and F1-score of 88.37%. These metrics represent significant improvements of 4.2% relative to baseline CNN and 3.4% compared to the dropout-regularized variant, indicating substantially enhanced diagnostic capability and reduced false-negative rates. Strategic integration of pre-trained feature extraction with customizable CNN architectures significantly improves generalization and classification performance in histopathological image analysis. Future investigations should incorporate larger heterogeneous datasets, attention mechanisms, and explainable artificial intelligence (XAI) to enhance clinical interpretability and strengthen practitioner confidence in digital pathology systems

Efficient VGA-Net Modification Using ConvNeXt-Tiny and GATv2 for Retinal Vessel Segmentation

Billie Zandra Widiyanto — 2026-07-09

Retinal blood vessel segmentation plays a crucial role in the early detection of ocular diseases such as diabetic retinopathy, glaucoma, and macular degeneration. Existing hybrid architectures, such as VGA-Net, suffer from high computational complexity due to the VGG-16 backbone and limited attention expressiveness due to its static GAT module, yet no prior work has examined replacing both components within a patch-based graph architecture in which backbone feature quality directly conditions graph attention effectiveness. This study aims to improve the computational efficiency and topological modeling of VGA-Net by replacing VGG-16 with ConvNeXt-Tiny and substituting GAT with GATv2. The primary contribution is a 55% parameter reduction through the ConvNeXt-Tiny backbone substitution and improved vessel topology modeling through GATv2's dynamic attention mechanism, which produces fully dynamic attention coefficients per query node. Experiments were conducted on the DRIVE and STARE datasets using a consistent preprocessing pipeline, one-factor-at-a-time hyperparameter tuning, and a unified evaluation protocol across all compared methods. The proposed model achieves the lowest parameter count (5.3M) and GFLOPs (3.2443), with a competitive inference time of 61.00 ms per image, among all compared methods, while achieving competitive performance in sensitivity and topological continuity. On the DRIVE dataset, the model achieved the highest sensitivity of 0.8718 and the highest clDice of 0.8446. On the STARE dataset, the model achieved the highest sensitivity of 0.9383 and the highest clDice of 0.9055. These results demonstrate that the proposed model achieves a favorable efficiency-performance trade-off, leading to sensitivity and topological continuity at the lowest computational cost among all compared methods, at the expense of lower specificity, accuracy, Dice, and MCC relative to certain compared methods.