Generalist Vision Foundation Models for Medical Imaging: A Case Study of Segment Anything Model on Zero-Shot Medical Segmentation

https://arxiv.org/pdf/2304.12637.

Abstract: We examine the recent Segment Anything Model (SAM) on medical images, and report both quantitative and qualitative zero-shot segmentation results on nine medical image segmentation benchmarks, covering various imaging modalities, such as optical coherence tomography (OCT), magnetic resonance imaging (MRI), and computed tomography (CT), as well as different applications including dermatology, ophthalmology, and radiology. Our experiments reveal that while SAM demonstrates stunning segmentation performance on images from the general domain, for those out-of-distribution images, e.g., medical images, its zero-shot segmentation performance is still limited. Furthermore, SAM demonstrated varying zero-shot segmentation performance across different unseen medical domains. For example, it had a 0.8704 mean Dice score on segmenting under-bruch's membrane layer of retinal OCT, whereas the segmentation accuracy drops to 0.0688 when segmenting retinal pigment epithelium. For certain structured targets, e.g., blood vessels, the zero-shot segmentation of SAM completely failed, whereas a simple fine-tuning of it with small amount of data could lead to remarkable improvements of the segmentation quality. Our study indicates the versatility of generalist vision foundation models on solving specific tasks in medical imaging, and their great potential to achieve desired performance through fine-turning and eventually tackle the challenges of accessing large diverse medical datasets and the complexity of medical domains.

The involved code and datasets will be available very soon