基于SAM的零样本多模态舌体分割方法
Zero-Shot Multimodal Tongue Image Segmentation Based on SAM

舌诊通过观察舌体特征评估健康状态，而舌体分割作为智能舌诊的关键步骤，需要准确分离舌体与背景，为后续特征提取和健康分析奠定基础。然而，舌体分割目前面临着两大挑战：一是数据的稀缺性，二是现有的分割大模型(如SAM模型)对人工提示的依赖性。为了解决以上问题，本文提出了一种零样本多模态的分割方法。该方法结合SAM模型和多模态提示技术，通过两阶段框架实现：1) 初步分割和相似度聚类，利用SAM模型生成初步分割结果，并通过相似度聚类解码器筛选潜在有效分割；2) 精细化分割，利用多模态大语言模型分析舌体特征，生成精确点提示，再次输入到SAM模型中以实现高精度分割。该方法在无需特定任务训练或标注数据的情况下，实现了SAM模型在舌诊领域的智能分割应用。实验结果显示，相比于原始的SAM模型，该方法在三个舌诊数据集上的mIoU指标分别提升了27.3%，18.2%，29.7%。
Tongue diagnosis assesses health status by observing tongue characteristics, and tongue segmentation, as a key step in intelligent tongue diagnosis, requires accurately separating the tongue body from the background to lay a foundation for subsequent feature extraction and health analysis. However, tongue segmentation currently faces two main challenges: data scarcity and the dependency of existing large segmentation models (such as the segment anything model) on manual prompts. To address these issues, this paper proposes a zero-shot multimodal segmentation method. This method combines the SAM model with multimodal prompt techniques and implemented in a two-stage framework: 1) initial segmentation and similarity clustering, where the SAM model generates initial segmentation results, followed by a similarity clustering decoder to filter out potentially effective segmentations; 2) refined segmentation, where a multimodal large language model analyzes tongue characteristics to generate precise point prompts, which are re-entered into the SAM model to achieve high-precision segmentation. This method enables intelligent segmentation with the SAM model in tongue diagnosis without the need for task-specific training or annotated data. Experimental results show that, compared to the original SAM model, this method improves the mIoU metric on three tongue diagnosis datasets by 27.3%, 18.2%, and 29.7%, respectively.