Authors:
Bibin Bose Sivananthan, B. Dhanush, Amarilys González García, Bushra Rehman, Kotova Anna Sergeevna, Cheng Ziming

Addresses:
School of Nursing, Ingram School of Nursing, McGill University, Montreal, Quebec, Canada. Department of Artificial Intelligence and Machine Learning, SRM Institute of Science and Technology, Ramapuram, Chennai, Tamil Nadu, India. Department of Research and Development, Placental Histotherapy Center, Playa, Havana, Cuba. Institute of Pathology and Diagnostic Medicine, Khyber Medical University, Peshawar, Khyber, Pakistan. Department of Pathological Physiology, Kyrgyz-Russian Slavic University named after the first President of the Russian Federation Boris Nikolaevich Yeltsin, Bishkek, Kyrgyzstan, Russia. Department of Computing and Information Systems, Singapore Management University, Singapore.

Abstract:

Automated interpretation of dermoscopic skin images is difficult because visually similar lesions can belong to different diagnostic groups, and the HAM10000 benchmark is markedly imbalanced. DermaLens AI addresses this problem through a browser-accessible platform that combines a MobileNetV2 classifier with contextual clinical information and reporting services. The model starts from ImageNet weights and is adapted to the seven HAM10000 categories: melanocytic nevi, melanoma, benign keratosis, basal cell carcinoma, actinic keratosis, vascular lesion, and dermatofibroma. Its training workflow uses augmentation, lesion-level deduplication, stratified partitioning, weighted sampling, focal loss, AdamW, cosine-annealing warm restarts, gradient clipping, and early stopping. A FastAPI/PyTorch inference service accepts JPEG, PNG, and WEBP files, standardizes them to 224 x 224 pixels, ranks the five leading class probabilities, and produces a structured report with risk information, ABCD-related descriptors, recommendations, and differential conditions. The platform also contains a broader disease knowledge base, optional symptom and anatomical-location inputs, a browser assistant, clinic search with maps, and printable reports. The accessible project snapshot does not contain a verified final training-history artifact or final test metrics; consequently, this manuscript reports the reproducible procedure and system-level behavior without fabricating accuracy Figures 1 and 2. The work demonstrates an integrated architecture for educational and research-oriented skin-lesion analysis, not a clinically validated diagnostic service.

Keywords: Skin Lesion Analysis; MobileNetV2 and HAM10000; Dermoscopy and Deep Learning; Medical Image Classification; Fast-API and Focal Loss; Clinical Decision Support; Explainable AI.

Received on: 02/08/2025, Revised on: 15/10/2025, Accepted on: 08/11/2025, Published on: 03/06/2026

DOI: 10.69888/FTSBE.2026.000791

FMDB Transactions on Sustainable Biomedical Engineering, 2026 Vol. 1 No. 2, Pages: 96-112

• Views : 25
• Downloads : 7

Download PDF