3D Attention U-Net for Brain Tumor Segmentation

Year

2025

Tech & Technique

PyTorch, 3D Attention U-Net (Islam et al., 2019), BraTS 2019

Description

🏆 2nd Place, Hack for Research 2025
Brain tumor segmentation on multimodal MRI from the BraTS 2019 dataset with a 3D Attention U-Net in PyTorch. I trained and modified public code from Islam et al. (2019) (3D_Attention_UNet); the architecture is theirs.

Key Facts:

🏆 2nd place, Hack for Research 2025 (team lead).
🧠 Model: 3D U-Net with attention gates, from Islam et al. (2019).
🧬 Data: BraTS 2019 multimodal MRI.
⚙️ Framework: PyTorch.

My Role

Team Lead

✅ Led the team that placed 2nd at Hack for Research 2025.
🧠 Trained and modified the public 3D Attention U-Net code from Islam et al. (2019) on BraTS 2019.

Figure from Islam et al. (2019)

BraTS 2019 examples and Dice comparison (3D Attention UNet vs 3D UNet) from the Islam et al. (2019) paper — the authors' results, not mine.

Sample Predictions

T1ce / FLAIR / T2 slices with ground-truth and predicted tumor masks.

Reference Architecture (third-party)

Third-party figure of a 3D U-Net with CoT attention layers — shown for reference; not this project's model.

3D U-Net Concept (Çiçek et al., 2016)

Figure from the original 3D U-Net paper by Çiçek et al. (2016) — credit to the authors.