Integrating a novel SRCRN network for segmentation with representative batch-mode experiments for detecting melanoma

Abstract Melanoma is a type of skin cancer that tends to spread to other parts of the body and can be fatal if not detected at an early stage. This paper proposes an automated and non-invasive methodology to assist clinicians to detect melanoma. A two-stage framework was suggested in the study. In the first stage, the Resnet 50-based novel SRCRN Network was designed, which extracts high-dimensional distinctive features for skin lesion segmentation, and uses the advantage of stride regulation effectively. In the framework of SRCRN, pixel maps of different sizes were obtained by upsampling and downsampling methods between block layers, and the performance of segmentation was improved by selecting the most appropriate pixel map. In the second stage, the Resnet-50 network was used again for melanoma detection. The classification network was operated through a proposed class balancing strategy and batch mode experiments to improve performance in the training process. The segmentation performance of the proposed SRCRN network was found to be the best among similar studies with 95% Acc. In the second stage of melanoma detection experiments, the best classification performance was achieved with 93.9% Acc and 97.3% Auc in batch size-32 trials. The proposed melanoma prediction model showed a 6.7% improvement in Acc value and a 9.9% improvement in Auc value compared to the study which took the top place in the ISBI 2017 competition.