Skin lesion segmentation using an improved framework of encoder-decoder based convolutional neural network

AbstractAutomatic lesion segmentation is a key phase of skin lesion analysis that significantly increases the performance of subsequent classification steps. Segmentation is a highly complex task due to the varying nature of lesions, such as unique shapes, different colors, and structures. In this study, a two‐step system is proposed comprising preprocessing algorithm and lesion segmentation network. The hairlines removal algorithm is designed using morphological operators to eliminate noise artifacts. The resulting output images are fed to the convolutional neural network (CNN) to perform lesion segmentation. The proposed CNN network is a new framework designed from scratch based on encoder‐decoder architecture. The layers are stacked in a unique sequence to perform downsampling and upsampling, generating a high‐resolution segmentation map. Additionally, a Tversky loss function is implemented to reduce the error rate between predicted and target output. In this study, we focus on the challenges related to the extraction of accurate lesion regions and the presence of hairlines in lesion images that can occlude important information resulting in poor segmentation. The proposed model is evaluated on four publicly available datasets, namely ISIC 2016–2018. The mean intersection over union (IoU) obtained for ISIC 2016–2018 and is 87.1%, 77.8%, 85.2%, and 88.8%, respectively. The proposed model demonstrated higher performance as compared to other state‐of‐the‐art methods. The hair removal step aids in the improvement of the overall performance.