Manually outlining a tumor across hundreds of MRI slices is slow and inconsistent. This system does it automatically – a deep-learning model that segments brain tumors from MRI scans to support faster, more reliable diagnosis.
A 3D U-Net trained on multimodal MRI (T1, T2, and FLAIR) reaches 0.85 Dice, supported by a full preprocessing pipeline and 3D visualization tools that render predictions volume by volume. Every experiment is tracked end to end in MLflow for full reproducibility.
Predictions are served through a FastAPI inference service, and targeted model and pipeline optimization cut inference time by 40% – bringing research-grade segmentation closer to practical clinical use.