Focal Loss-Based Skin Lesion Classification and Segmentation Using Deep Learning: An Augmented Approach

Melanoma is very life-threatening and is also likely to be underdiagnosed, hence posing a great challenge to public health. Classification and skin lesion segmentation are very critical tasks that have to be accomplished accurately to make early diagnosis possible along with suitable treatments. The authors of this paper introduce here a framework of deep learning using state-of-the-art models trained on ISIC 2018 and ISIC 2019 datasets. For classification, we employ ResNet152 along with data augmentation, focal loss, and adaptive learning rate, achieving 92% classification accuracy on varied lesion types. For segmentation, we employ a variant of U-Net consisting of an encoder-decoder along with skip connections, employing the Adam optimizer along with binary cross-entropy loss. Our framework is able to achieve 94% of the Dice metric along with 93% of IoU metric, establishing its usefulness in automated skin lesion interpretation.