Segmenting melanoma Lesion using Single Shot Detector (SSD) and Level Set Segmentation Technique

Melanoma is a lethal type of skin cancer that orginates fron melanocytes cells of skin and it is responsible of several deaths annually due to exposure of ultraviolet radiations. Early diagnosis and proper treatment of melanoma significantly improves the patient's survival rate. In the computer aided diagnosis, the automatic segmentation is first step in early and accurate diagnosis of the Melanoma lesion area. However, the presence of natural or clinical artifacts hinders the precise lesion segmentation. The goal of our work is to establish a novel pipeline that automatically pre-process, localize and then segment the melanoma lesion precisely and improve its segmentation accuracy. In our proposed method, dermoscopic images are segmented in three steps: 1. Preprocessing using morphological operations to remove hair. 2. Localization of melanoma lesion by utilizing a deep convolutional neural network named as Single Shot Detection (SSD) network, 3. Segmentation using level set algorithm. The proposed approach was evaluated on ISBI 2016 challenge dataset (Skin Lesion Analysis Towards Melanoma Detection Challenge Dataset). On ISIC 2016, our method achieved an average of Jc, Di and Ac as 0.82, 0.901 and 0.90 respectively. The results of the segmentation are also compared with the state-of-the-art methods to justify the effectiveness of the proposed approach.