人工智能
计算机科学
计算机视觉
分割
图像分割
图像(数学)
模式识别(心理学)
尺度空间分割
特征(语言学)
图像处理
代表(政治)
作者
Jiaxu Jiang,Sen Lei,Hui Li
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2026-04-08
卷期号:: 1-4
被引量:1
标识
DOI:10.1109/isbi61048.2026.11515535
摘要
Referring Medical Image Segmentation aims to segment target regions in medical images according to referring expressions. Current SAM-based approaches typically generate prompts via vision-language models by regressing positive and negative point coordinates or directly producing prompt embeddings. However, these methods often suffer from limited point diversity and insufficient semantic guidance during embedding generation. To overcome these limitations, we propose a novel SAM-based framework that leverages the strong representational capability of DINOv3 to guide prompt generation. Representative features of the target region, together with a similarity loss, are used to supervise the generation of prompt embeddings, providing semantically consistent and discriminative guidance for SAM. Extensive experiments demonstrate that our DINOv3-guided module improves the Dice score of a baseline SAM-based method by 3.33% and achieves state-of-the-art performance on the QaTa-COV19 dataset. The code will be available at https://github.com/aticejiang/ReferDinoSAM.
科研通智能强力驱动
Strongly Powered by AbleSci AI