Skin cancer detection, localization, and classification are ongoing research areas, and segmentation is a critical step in the localization of skin cancer. Binary segmentation is specifically used in skin cancer localization. The article explores binary segmentation with the HAM10000 dataset of 10,015 skin lesion images and their corresponding masks. A convolution neural network, specifically an encoder-decoder architecture with Resnet34 and pre-trained weights on ImageNet, is used for the task. The output is a 2D mask image of size 128×128, and the dice loss is used as the loss function, with the Jaccard Index/Intersection over Union as the tracking metric. The images are resized to 128×128 to balance the trade-off between image quality and computational time. The distribution of lesion sizes is found to be skewed towards the left side, indicating the most frequent fraction of image covered by skin lesion is 10-20%.

source update: How I built Supervised Skin Lesion Segmentation on HAM10000 Dataset – Towards AI