Brain Tumour Segmentation with Medical Foundation Models

A comparison of a task-specific 3D U-Net and a promptable medical foundation model (LiteMedSAM, the laptop version of MedSAM) for segmenting gliomas and their sub-regions in four-sequence brain MRI - trained, evaluated and deployed entirely on a student laptop CPU. On 96 held-out BraTS studies the CPU-trained U-Net reaches Dice 0.870 / 0.727 / 0.676 (whole tumour / core / enhancing tumour), and LiteMedSAM prompted automatically by the U-Net reaches 0.819 / 0.694 / 0.555 after fine-tuning its decoder. A secure web application lets a clinician upload a study, run any of the models, compare overlays slice by slice, review and report.

Category: Medical Image Analysis / Deep Learning / Computer Vision / Healthcare AI

CSE Project 13 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.