Claim Missing Document
Check
Articles

Found 3 Documents
Search

Multi-modal transformer and convolutional attention architectures for melanoma detection in dermoscopic images Guidoum Amina; Maamar Bougherara; Amara Rafik
IAES International Journal of Robotics and Automation (IJRA) Vol 15, No 1: March 2026
Publisher : Institute of Advanced Engineering and Science

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.11591/ijra.v15i1.pp136-148

Abstract

The deadliest type of skin cancer, melanoma, requires early and accurate detection for a successful course of treatment. Traditional diagnostic techniques, which rely on visual inspection and dermoscopy, are frequently arbitrary and prone to human error. Automated melanoma detection exemplifies the integration of multimedia, a truly interdisciplinary field that melds visual data processing, human-computer interaction, and digital technologies. This study presents a multi-modal architecture: a multi-modal transformer network (MMTN) and a convolutional attention mechanism multi-modal (CAMM) that combines clinical data and dermoscopy images to enhance melanoma detection. The models achieve higher performance compared to other approaches by utilizing the strengths of architecture based on transformers, an encoder for image processing, dense layers for clinical data also Spatial Attention for the second architecture proposed. We evaluate the models on the entire set of ISIC 2019 data, showing significant improvements in accuracy and AUC. The models achieve high accuracy and AUC using CPU in both architectures. Our findings highlight the potential of a multi-modal learning architecture to enhance clinical decision-making and diagnostic accuracy in dermatology. To our knowledge, this is the first implementation combining MobileNet, transformer encoder attention, and clinical data fusion for the ISIC 2019 dataset, providing a significant advancement in the automated categorization of skin malignancies.
EdgeRetina: Hybrid multimedia architecture for diabetic retinopathy screening on low-cost mobiles Guidoum Amina; Achour Soltana; Maamar Bougherara; Amara Rafik; Mhamed Tayeb
IAES International Journal of Robotics and Automation (IJRA) Vol 15, No 1: March 2026
Publisher : Institute of Advanced Engineering and Science

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.11591/ijra.v15i1.pp234-246

Abstract

Diabetic retinopathy (DR) is a major cause of preventable blindness, particularly in areas with limited medical resources where access to ophthalmologists is critical. Existing automated solutions struggle to balance clinical performance, cost-effectiveness, and robustness in the face of fundus image variability—including lighting differences, artifacts, and uneven capture quality. To address this challenge, we propose EdgeRetina, an integrated solution for diabetic retinopathy screening on low-cost mobiles. Our approach combines lightweight preprocessing (128×128 resizing, intensity normalization, and targeted augmentations simulating real-world conditions) with a hybrid SqueezeNet-MobileViT architecture (1.4 million parameters), optimized by dynamic threshold calibration (median: 0.3), maximizing clinical utility. Clinically calibrated INT8 quantization reduces the model to 8.27 MB (-92%) without altering diagnostic performance (sensitivity of 90.7% for referable diabetic retinopathies), while preserving compatibility with floating point 32 (FP32)-based gradient-weighted class activation mapping (Grad-CAM) visualizations. Evaluated on the APTOS 2019 dataset, this solution achieves an AUC of 0.96 with a latency (inference time) of 15.43 ms, reducing CPU consumption by 43% compared to FP32. The dynamic threshold/INT8 coupling decreases false positives by 71.4%. This pipeline thus enables accurate, accessible, and early screening of diabetic retinopathy on low-cost mobile devices, combining operational efficiency and diagnostic reliability in constrained environments, which is crucial to prevent avoidable blindness.
CPU-optimized regularized convolutional neural network architecture for multimedia-based breast cancer metastasis detection Guidoum Amina; Bouhenni Abdelhake
International Journal of Informatics and Communication Technology (IJ-ICT) Vol 15, No 4: December 2026
Publisher : Institute of Advanced Engineering and Science

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.11591/ijict.v15i4.pp1881-1892

Abstract

The histopathological diagnosis of breast cancer metastases, a subjective and lengthy process, benefits from advances in artificial intelligence for its automation. This study proposes an automatic classification method based on a convolutional neural network (CNN) architecture with progressive regularization (Batch normalization and adaptive dropout), trained on the PatchCamelyon (PCam) dataset. Our approach integrates a data augmentation strategy specific to medical images and advanced optimization for CPU environments, allowing us to balance performance, robustness, and computational accessibility. The architecture comprises only 94,000 parameters, 223 times fewer than leading models. Multi-criteria evaluation shows competitive results: area under the curve (AUC) of 92.92%, F1-score of 82.39%, accuracy of 88.74%, and specificity of 90.61%. A bootstrap robustness analysis provides 95% confidence intervals (AUC: [89.56% 92.91%]), and an ablation study quantifies the reduction in overfitting through progressive dropout (33.6%). This research validates that a well regulated and CPU-optimized CNN architecture can achieve clinically relevant performance while remaining accessible on standard environments, thus opening up concrete prospects for the development of practical AI tools in digital pathology.