IMedSeg: Towards efficient interactive medical segmentation

Zidi Shi

doi:10.1016/j.neucom.2025.129419

2025-04-01

PRE

DOI：10.1016/j.neucom.2025.129419

原文链接

原文求助

摘要

En 中文

Click-based Interactive Medical Segmentation (IMS) aims to segment target objects under limited user interactions. Although existing deep learning-based methods achieve promising performance, they still exhibit deficiencies: (1) they fail to fully and explicitly utilize click information and (2) when used to refine preexisting masks, they cannot avoid destroying correctly segmented regions. Against these issues, we propose IMedSeg to formulate the IMS task into a progressive loop process, where each click by the user results in a precise and efficient modification. In particular, we design a CNN/Transformer-based model to perform the initial segmentation on the target image. To enhance segmentation accuracy, we especially propose a memory- efficient interactive medical segmentation backbone called ICS-Unet by incorporating CSWin Transformer with Unet. Then, a Similarity Mask Correction Module (SMCM) is proposed to refine segmentation only on the local region with user clicks to avoid corrupting the correct mask. Moreover, a Compressed Similarity Aggregator (CSA) is constructed for computing the correlations between the deep features of clicked and unclicked points, making full use of click information for pixel-wise classification of the local region. Note that our proposed method is a generic framework that can be readily combined with existing interactive segmentation backbones. Extensive experiments show that our method attains superior performance over state-of-the-arts in terms of accuracy, memory and inference time. The results on ISIC-2017 and OAI-ZIB datasets reveal average Dice coefficients of 96.18% and 87.44%, and IoUs of 82.97% and 78.19%, respectively.

Keyword:

Interactive medical segmentation

Progressive loop process

Transformer

Refinement

UNet