1. SAM2-PATH: A better segment anything model for semantic segmentation in digital pathology
- Author
-
Zhang, Mingya, Wang, Liang, Gu, Limei, Li, Zhao, Wang, Yaohui, Ling, Tingshen, and Tao, Xianping
- Subjects
Electrical Engineering and Systems Science - Image and Video Processing ,Computer Science - Computer Vision and Pattern Recognition - Abstract
The semantic segmentation task in pathology plays an indispensable role in assisting physicians in determining the condition of tissue lesions. Foundation models, such as the SAM (Segment Anything Model) and SAM2, exhibit exceptional performance in instance segmentation within everyday natural scenes. SAM-PATH has also achieved impressive results in semantic segmentation within the field of pathology. However, in computational pathology, the models mentioned above still have the following limitations. The pre-trained encoder models suffer from a scarcity of pathology image data; SAM and SAM2 are not suitable for semantic segmentation. In this paper, we have designed a trainable Kolmogorov-Arnold Networks(KAN) classification module within the SAM2 workflow, and we have introduced the largest pretrained vision encoder for histopathology (UNI) to date. Our proposed framework, SAM2-PATH, augments SAM2's capability to perform semantic segmentation in digital pathology autonomously, eliminating the need for human provided input prompts. The experimental results demonstrate that, after fine-tuning the KAN classification module and decoder, Our dataset has achieved competitive results on publicly available pathology data. The code has been open-sourced and can be found at the following address: https://github.com/simzhangbest/SAM2PATH., Comment: 6 pages , 3 figures
- Published
- 2024