State of the art deep learning methods for melanoma classification

Melanoma is a skin cancer, which spreads at the faster rate with higher motality if left untreated. Latest advancements in deep learning (DL) are changing the way of melanoma diagnosis employing traditional clinical and dermoscopic investigations performed by medical experts. Emphasising on the convolutional neural networks (CNNs), and its utilization in the field of melanoma diagnosis, this study examines state-of-the-art approaches used till date. By using the large annotated image samples, CNNs can capture subtle patterns in shape, texture and color of lesions, resulting in comparable accuracy as that of dermatologists. Also, the performance of diagnosis could be further improved by models like pre-trained transfer learning. But, issues of restricted demographic validation, interpretability of models, and possible biases against people having darker tones still exists. Although, there are still gaps in the use of more recent designs such as transformers, few recent work include merging of segmentation and classification after pre-processing to create end-to-end pipelines. Broad deployment requires addressing clinical integration, data diversity, and resilience. Building on these developments, DL can improvise early diagnosis of melanoma, lower mortality and reduce patients across a range of demographics.