Deep Learning / Medical Vision
BrainSeg AI.
Brain tumor segmentation with a PyTorch Attention U-Net on BraTS 2020.
Problem
Locate tumor regions in medical images, beyond image-level classification.
Context
Deep learning project applied to medical segmentation. Experimental results without clinical validation.
Objective
Evaluate WT, TC and ET region segmentation using an Attention U-Net.
Dataset / Data
BraTS 2020. Training / validation / test splits and preprocessing protocol remain to be documented.
Architecture
Medical images → Attention U-Net (PyTorch) → segmentation masks → Dice evaluation.
Methodology
Train an Attention U-Net and evaluate the three tumor regions separately. Best reference epoch: 22.
My contribution
Project reported in the candidate’s background. Individual scope and team responsibilities remain to be confirmed.
Challenges
Segmenting regions of varying size and interpreting region-specific performance.
Limitations
Exact protocol and data separation need verification. These results do not establish clinical performance or real-world validation.
What I learned
The project exercises segmentation, attention-based architectures and Dice evaluation.
Project results
Average Dice
Project evaluation, mean across WT / TC / ET. Reference epoch: 22.
WT Dice
Whole Tumor, project evaluation.
TC Dice
Tumor Core, project evaluation.
ET Dice
Enhancing Tumor, project evaluation.
Patients in the dataset
Reported BraTS 2020 dataset size, not the test subset size.