Evaluation of the Segment Anything Model (SAM) for Brain Tumor Segmentation

Mohammad Peivandi, Jason Zhang, Michael Lu, Chengyin Li, Dongxiao Zhu, Zhifeng Kou

Abstract: Brain tumor segmentation is a complex task where deep learning models, though useful, fall short compared to human expert segmentation. We explore the application of the Segment Anything Model (SAM), originally trained on diverse natural images, to this domain. Our paper presents an enhancement of SAM’s mask decoder via fine tuning with the Decathlon brain tumor dataset. We also employ data augmentation techniques like rotations and elastic deformations. Performance is evaluated using the Dice Similarity Coefficient and the Hausdorff Distance 95th Percentile. Compared to the original SAM and nnUNetv2, our fine tuned SAM shows a considerable improvement, particularly in the challenging cases. While nnUNetv2 maintains overall higher accuracy, our SAM based model gives more consistent results, suggesting a strong potential for future advancements in brain tumor segmentation. The source code is available from GitHub.