SegNet Architecture for Dermoscopic Image Segmentation

The lesion segmentation is a fundamental step in developing a Computer-Aided Diagnosis system for skin cancer detection. However, accurate segmentation is challenging due to the variability of lesion characteristics (e.g., shapes, sizes, colors, etc.) and artifacts in a dermoscopic image (e.g., hair, color charts, etc.). Therefore, most of the existing methods heavily depend on pre-processing procedures that often cause an information loss and, consequently, an inaccurate segmentation. This paper proposes a deep learning-based method using a SegNet architecture for pixel-wise segmentation. First, we trained the SegNet on the PH2 dataset. However, this dataset is small, and the images are almost clean of artifacts. Unlike the PH2 dataset, the ISIC 2018 dataset contains a more significant number of images with different types of artifacts. Therefore, we trained the proposed model on the ISIC2018 dataset in a second experiment. As a result, we achieved an IOU of 95.72% for the PH2 dataset and an IOU of 95.06% for the ISIC2018 dataset. A combination of quantitative and qualitative experimental evaluation has been conducted on two publicly available datasets. The achieved results of the empirical analysis demonstrate that the proposed method is highly accurate.