Deep Learning-Based Skin Lesion Classification Using Augmented Clinical Datasets from Bangladeshi Hospitals

Skin cancer represents one of the most prevalent forms of oncological malignancies globally; however, early diagnostic intervention dramatically optimizes prognostic outcomes. Traditional diagnostic modalities depend heavily on the availability of expert dermatologists, inducing systematic bottlenecks in resource-constrained regions. This paper presents a mobileoptimized deep learning framework engineered for automated skin lesion classification in low-resource settings. Our architecture leverages a custom-augmented hybrid dataset integrating international benchmarks with local clinical repositories to mitigate domain representation biases. The proposed MobileNetV2 network achieves an empirical training accuracy of 99.01% and a testing accuracy of 82.7%. Quantized via TensorFlow Lite (TFLite), the deployment model maintains a compact structural footprint of 14.3 MB and a mean inference latency of 83 ms on consumer edge hardware. The structural framework demonstrated robust operational efficacy during field validation across 12 rural healthcare centers in Bangladesh, establishing a scalable paradigm for equitable, edge-computed dermatological diagnostics.