.png){kind=icon}
.png){kind=icon}
Return
Robust lesion segmentation based on segment anything model
.png){kind=icon}
PRE
AI
DOI：10.1016/j.bspc.2026.110156.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
As a foundation segmentation model, the Segment Anything Model (SAM) has demonstrated remarkable zero-shot capability in natural image domain, and has garnered significant attention in the field of medical images. While preliminary efforts have showcased SAM’s great potential for medical image segmentation, the existing methods often rely on high-quality prompts, such as precise bounding boxes for lesion targets, or the segmentation performance will be decreased due to the attention shift caused by the low-quality prompts. In this paper, we present Robust LeSAM, a novel model tailored for lesion segmentation to ensure consistent accuracy across prompts of varying qualities. Firstly, we address the attention shift issue through a Feature Resampling Module (FRM), which predicts feature offsets and resamples the feature positions to effectively redirect model’s attention to target regions. Secondly, we introduce a Feature Weighting Module (FWM) that assigns different weights to the resampled features and the original features, thereby enhancing the feature resampling and avoiding unnecessary feature deformations when precise prompts are provided. Additionally, for enhanced lesion edge delineation, we introduce a Multi-Feature Fusion Module (MFFM) that integrates bottleneck features from various levels of the image encoder alongside output from the original mask decoder. We have conducted comprehensive experiments on five publicly available lesion segmentation datasets that has been proved sensitive to prompt variations in previous study. Our experimental results demonstrate that Robust LeSAM significantly enhances the performance compared to previous state-of-the-art methods when handling low-quality prompts.
Keywords:
LeSAM
Feature Resampling Module
Feature Weighting Module
Multi-Feature Fusion Module
lesion segmentation
Journal
IF:
4.9
Papers:
9.7K
Citations:
2.4W