Skin Lesion Image Segmentation Algorithm Based on MC-UNet

Abstract

Aiming at the situation of dermatoscopic images with fuzzy lesion boundaries, variable morphology and high similarity to background, this paper proposes a skin lesion segmentation algorithm that achieves higher segmentation accuracy by combining existing convolutional neural network methods. The algorithm begins by using a Multiscale Residual Block (MRB) with different-sized convolutional kernels to enlarge the receptive field and extract multi-scale features of dermatoscopic images. Secondly, the skip connections are enhanced with a Bidirectional Information Fusion Module (BFM) to refine features by bidirectionally fusing semantic information from high-level feature maps and spatial information from low-level feature maps. Finally, the network’s segmentation accuracy is improved through the use of a new loss function called MixLoss, which combines BceLoss and DiceLoss. Specifically, it achieves a Dice coefficient of 92.37% and an accuracy of 95.32% with a sensitivity of 93.41% on the ISIC2016 dataset. On the ISIC2017 dataset, it achieves a Dice coefficient of 89.43%, an accuracy of 94.81%, and a sensitivity of 90.41%. The experimental results show that the proposed algorithm outperforms other mainstream algorithms and exhibits superior performance in skin lesion segmentation.