Advancing Brain MRI Segmentation using Segment Anything Model

Advancements in medical imaging, particularly within the realm of brain MRI segmentation, have traversed a transformative journey from manual to automated methodologies. This evolution signifies a pivotal shift in the detection and diagnosis of conditions such as brain cancer. The emergence of the U-Net architecture marked a significant milestone, offering enhanced efficiency and accuracy in segmentation tasks. However, the landscape of medical imaging continues to evolve rapidly, and the introduction of the Segment Anything Model (SAM) represents a paradigm shift in this domain. This paper seeks to examine the effectiveness of SAM in the context of brain MRI imaging. The objective is to utilize SAM’s advanced segmentation capabilities to create accurate and detailed masks for brain MRI scans. This concept significantly enhances the accuracy and efficiency of medical diagnoses and clinical application in medical imaging. Unlike traditional approaches relying solely on predetermined algorithms, SAM harnesses the power of prompt-based learning, enabling more dynamic and adaptable segmentation models. Also, it provides increased efficiency and flexibility, reducing manual tuning, minimum effort and adapting quickly to new tasks. Continuous improvement and regular updates from Meta AI ensure that SAM remains at the forefront of technology, offering increasingly advanced capabilities for segmentation tasks. This study investigates SAM’s efficacy, particularly in the context of brain MRI imaging, by comparing SAM with and without bounding box techniques against DeepLabV3, U-Net, and DenseNet121. The overarching goal is to revolutionize medical diagnoses and treatment planning through precise segmentation, providing insights into the strengths and limitations of different segmentation methodologies.