Melanoma Lesion Segmentation and Classification Using SegNet

Abstract

Melanoma is one of the worst forms of skin cancers that should be detected early for proper treatment. Usually the dermatologists inspect lesion region via optical inspection but this method is time-consuming and error-prone. Furthermore, over the past few years, due to advent of Machine Learning (ML) based systems, the researchers have developed automatic skin cancer diagnosis techniques. However, they rely heavily on manual image segmentation and handcrafted feature extraction techniques. Moreover, the performance of these systems is also degraded due to hair, blood vessels, poor contrast and hazy tumor boundaries. In this paper, we propose a deep learning-(DL) based melanoma lesion segmentation framework using SegNet. The proposed technique is trained and evaluated on dermoscopic images taken from ISIC-2016. We also evaluated the performance of our proposed methodology on a cross data scenario using ISIC-2017 database. The performance of the proposed framework is evaluated using various evaluation metrics such as accuracy, precision, Intersection over Union (IoU) and recall. The proposed framework succeeded in achieving 89% accuracy and is robust to presence of artefacts such as blood vessels or hair. The experimental results demonstrate the robustness of the suggested melanoma lesion segmentation and classification method. Hence, the system can be deployed in clinical settings to automatically detect melanoma lesions from dermoscopic images.