This repository is the 2023 final project for the NTU Medical Image Processing course. We aim to experiment with different state-of-the-art (SOTA) models, exploring their strengths and weaknesses. We also compare their performances in terms of WT, TC, ET, and visualize the segmentation results for a comprehensive evaluation.
Brain Tumor Segmentation(BraTS2020) challenge is the semantic segmentation of brain tumors. This is achieved by providing a 3D MRI dataset with voxel-wise ground truth labels annotated by medical professionals. It contains training dataset comprises 369 subjects, each with four 3D MRI modalities:
- native (T1)
- post-contrast T1-weighted (T1Gd)
- T2-weighted (T2)
- T2 Fluid-attenuated Inversion Recovery (T2-FLAIR)
We chose four different types of popular models: 3D-CNN-based (3D-UNet, SegResNet), Transformer-based (Swin-UNETR), and Diffusion-Based (Diff-UNet)
- 3D-UNet (2016 Jun)
- SegResNet (2018 Nov)
- Swin-UNETR (2022 Jan)
- Diff-UNet (2023 Mar)