Skin Lesion Segmentation and Classification: A Deep Learning and Markovian Approach

Abstract

Skin cancer is a very dangerous type of cancer which can spread in other organs if not detected early. Skin Melanoma is a type of cancer where melanocytes mutate rapidly. It is estimated that 99% of people can survive this type of cancer if it is detected at an early stage. A feedforward neural network with fully connected dense layers and Gray Level Co-occurrence Matrix as features is used which gives an AUC score of 0.63 on the private leader board of SIIM ISIC 2020 challenge. A pre-trained EfficientNETB7 is used to achieve an AUC score of 0.73after the images from the SIIM ISIC challenge was augmented with ISIC 2019 challenge images. Segmentation of the skin cancer images from PH2 dataset is done using Markov Random Field, An Encoder-Decoder network is proposed to segment the images in the PH2 dataset. A precision of 0.91 and a dice coefficient of 0.73 are achieved on the validation dataset using the Encoder-Decoder network. Segmentation can be followed by feature extraction of the lesions and then classification can be performed using machine learning.