Xing Yao, Han Liu, Dewei Hu, Daiwei Lu, Ange Lou, Hao Li, Ruining Deng, Gabriel Arenas, Baris Oguz, Nadav Schwartz, Brett C. Byram, and Ipek Oguz. “FNPC-SAM: Uncertainty-guided False Negative/Positive Control for SAM on Noisy Medical Images.” Proceedings of SPIE Medical Imaging 2024: Image Processing, vol. 12926, 1292602, 2024, San Diego, California, United States.
Researchers have improved the Segment Anything Model (SAM) to better handle challenging medical images, like low-contrast, noisy ultrasound images. Originally, SAM needed clear manual hints to accurately segment images. The new method enhances SAM’s accuracy and reliability by using multiple bounding boxes and correcting mistakes using a strategy that identifies uncertain areas. This enhancement doesn’t require additional training of the model. Additionally, they introduced a technique that allows creating 3D segmentations from a single 2D image slice. These improvements make SAM more effective and efficient for analyzing difficult medical images. The source code is available online for further use.
