Segmentation of Brain MRI Tumors by MedSAM with Prompts Generated by Object Detection

This research paper explores the use of deep learning technology in the segmentation of multimodal brain tumor medical images, particularly emphasizing the application of the customized and fine-tuned MedSAM (Segment Anything in Medical Images) model. MedSAM, an adaptation of the original SAM (Segment Anything Model), is specifically tailored for the nuances of medical imaging. In the model's training and fine-tuning phase, it retains the robust image feature extraction capabilities of the pre-trained model's image and prompt encoders. At the same time, the mask decoder is specifically modified to boost the precision of tumor segmentation. The results of the study indicate that the refined MedSAM model surpasses both the standard SAM model and the initial version of MedSAM in Dice coefficient scores, confirming the method's efficacy in enhancing brain tumor image segmentation accuracy. This research demonstrates that a general model, when appropriately modified, can be effectively utilized for specialized medical image segmentation tasks, not only elevating the accuracy of brain tumor segmentation but also offering innovative approaches and insights for future clinical diagnostic applications. Furthermore, the study sheds light on the model's versatility and its potential applicability to diverse dataset types, providing a foundation for subsequent research avenues.