Skin lesion segmentation in dermoscopic images using CNN architecture

Semantic image segmentation is the process of labeling each image pixel with its associated class. An encoder-decoder-based approach, like U-Net and its variants, is a popular strategy for solving medical image segmentation tasks. We propose an encoder-decoder-based approach with novel convolutions called spatial group convolutions (SGC) for skin lesion segmentation. The SGC employ multiple convolution kernels such as 1×1, 3×3, 5×5, 7×7, where each of these k×k kernel operation is split as the combination of k×1 convolution followed by 1×k convolution. The large kernel implemented using two 1-D kernels increases the receptive field of the model for better feature extraction along with the reduced learnable parameters. Furthermore, the use of multi size kernel helps in discriminative feature learning. SGC based proposed model shows improved segmentation accuracy while requiring decreased model parameters. The performance of the proposed method is analyzed on the publicly available PH2 dataset. The mean Jaccard index for the proposed method is 88.5%, the mean Dice score is 93.48%, and the mean accuracy value is 95.81%, with 0.4 million parameters for the proposed method on the PH2 test dataset.