Skin Lesion Segmentation using Residual U-NET

Skin lesion segmentation is an essential diagnosis procedure for various skin related diseases. Along with these minor medical conditions, malignant melanoma which is a type of skin cancer also has similar skin lesions as initial symptoms. Hence automatic segmentation of skin lesions is vital for diagnosis of various skin diseases saving time as well as medical resources. Advancements in the field of deep learning could be useful for mitigating this medical image segmentation constraint. Convolutional neural networks (CNN) have been a state of art for feature extraction from biomedical images. This paper proposes a ResNetl01 backboned UNET architecture for accurate segmentation of lesions from the healthy portion of the skin. The efficiency and the performance of the model was drastically improved by extensive data augmentation techniques. With ResNet architecture providing efficient feature extraction and retention, the model was trained successfully achieving training and test jaccard scores of 96.28% and 84.20% outperforming the existing results.