Faster Region Based Convolutional Neural Network for Skin Lesion Segmentation

Abstract

The diagnostic interpretation of dermoscopic images is a complex task as it is very difficult to identify the skin lesions from the normal. Thus the accurate detection of potential abnormalities is required for patient monitoring and effective treatment. In this work, a Two-Tier Segmentation (TTS) system is designed, which combines the unsupervised and supervised techniques for skin lesion segmentation. It comprises preprocessing by the median filter, TTS by Colour K-Means Clustering (CKMC) for initial segmentation and Faster Region based Convolutional Neural Network (FR-CNN) for refined segmentation. The CKMC approach is evaluated using the different number of clusters (k = 3, 5, 7, and 9). An inception network with batch normalization is employed to segment melanoma regions effectively. Different loss functions such as Mean Absolute Error (MAE), Cross Entropy Loss (CEL), and Dice Loss (DL) are utilized for performance evaluation of the TTS system. The anchor box technique is employed to detect the melanoma region effectively. The TTS system is evaluated using 200 dermoscopic images from the PH 2 database. The segmentation accuracies are analyzed in terms of Pixel Accuracy (PA) and Jaccard Index (JI). Results show that the TTS system has 90.19% PA with 0.8048 JI for skin lesion segmentation using DL in FR-CNN with seven clusters in CKMC than CEL and MAE.