Assessing MedSAM's Generalization to White Matter Hyperintensities in Magnetic Resonance Imaging

The field of medical imaging segmentation develops in two different directions, that of specialized models attempting to be the best at a specific task, such as automated segmentation of white matter hyperintensities (WMH), and that of artificial general intelligence (AGI). Early AGI attempts have the ambition to generalize to new tasks and domains effortlessly, in a similar way to how humans learn new tasks with few examples and prompts. MedSAM claims to be a model able to perform most medical segmentation tasks without the need for additional training, given an image and prompt. In this paper, we examine MedSAM's WMH segmentation performance in five different datasets when provided with a good initial prompt. The method's generalization is observed under variations of magnetic resonance sequences, acquisition parameters, orientation, prompt variations and other factors. Among other findings, this study found that MedSAM performs better for WMH with the axial view and FLAIR sequence, obtaining Dice values of 0,57 in MICCAI-2016, 0,54 in MICCAI-2017 and 0,69 in MICCAI-2018. Specialized methods from those challenges had performances of 0,67 in MICCAI-2016 and 0,80 in MICCAI-2017. This suggests that MedSAM is still inferior to current automated methods in this task even when provided with good initial prompts.