A Dual-Branch Fusion Network with Asymmetric Depthwise Convolutions for Whole-Body Tumor Segmentation

Wenze Fan

Keywords: Dual branch，Tumour segmentation，Asymmetric Depthwise Convolutions

TL;DR: Developed a dual branch encoder to solve the problem of whole-body tumor segmentation

Abstract:

Although deep learning models have made significant progress in medical segmentation, they have encountered challenges in tumor detection and segmentation tasks, particularly in complex whole-body scanning scenarios.To address this issue, this paper proposes an improved nn UNet model. Specifically, an efficient dual branch encoder has been developed, with one branch using a special "Inception style depthwise separable convolution" inspired by the Google Inception network, and the other branch using ordinary residual connections. This model effectively improves segmentation efficiency.Experiments on the MICCAI FLARE 2025 dataset demonstrated significant improvements in both segmentation accuracy and efficiency. Our method achieved an average organ Dice Similarity Coefficient (DSC) of 13.4\% and a Normalized Surface Dice
(NSD) of 5.98\% on the public validation set.The average inference time is 11 seconds.

Submission Number: 8