Towards a Portable Deep Learning-based Application for Melanoma Cancer Classification

Melanoma is an aggressive skin cancer that can rapidly spread to other parts of the body if not diagnosed and treated promptly. Current diagnostic methods include visual evaluation, biopsy, and histopathological analysis, but can be subjective and require significant time and resources. This work proposes the development of a melanoma classification protocol based on small and large MobileNetV3 architectures combined with two fine-tunning schemes. The best performance was achieved by the large MobileNetv3 architecture with the fine-tuning 2 schema. Training evaluation on 2003 images reported a successful mean of the area under the receiver characteristic operating curve score of 0.906. Additionally, the test on 223 images provided a competitive score of 0.917. Both results were obtained using a stratified ten-fold cross-validation mechanism. The best model was implemented on two mobile emulators to analyze its feasibility in terms of power consumption, resulting in a mean of 0.45 mAh per image, indicating high-quality performance. Furthermore, the model was implemented in a web app, and the average response time of 115.44 ms with an average of 15kb transferred over the network per image demonstrated efficient utilization of computational resources. These findings demonstrate the possibility of developing and deploying successful deep CNN models with transfer learning into limited-resource devices, serving as a valuable secondary diagnostic tool for the early self-diagnosis of melanoma in patients.