Abstract

Info

Title:: Multi-scale tissue segmentation in melanoma whole slide images using an improved U-net

Author(s):: LIU Jianwei1; LIU Liu1; LIU Mengmeng2; LI Yuxi1; AN Xiaodong1; 1. School of Mechanical Engineering, Zhengzhou University of Aeronautics, Zhengzhou 450046, China 2. Department of Pathology, the Third Peoples Hospital of Zhengzhou, Zhengzhou 450044, China

Keywords:: melanoma image segmentation deep learning whole slide image multi-scale feature

Abstract:: To meet the segmentation requirements for 3 major tissue types in acral melanoma whole slide images, namely melanoma, epidermal tissue, and interstitial tissue, the study constructs the first multi-scale tissue segmentation dataset for acral melanoma whole slide image. Furthermore, a dilated DynamicConv with RepVGG-SE U-net (DDCRS-Unet) model based on an improved U-net architecture is developed to address the corresponding segmentation challenges. This model adopts a dual dynamic dilated convolution module to achieve hierarchical pathological feature extraction through graded dilation rates and channel attention, embeds a multi-branch feature enhancement module into the skip connections, and leverages a channel attention mechanism to recalibrate key pathological features. Experimental results demonstrate that the model improves the average precision for segmenting the 3 tissue structures in acral metastic lesions to 0.840 3, the mean intersection over union to 0.525 8, and the average Dice coefficient to 0.621 1, representing improvements of 3.26%, 5.69%, and 4.70% over U-net, respectively.