3D Attention UNet++ Brain Tumor Segmentation

🏆 2nd Place, Hack for Research 2025
Advanced 3D Attention UNet framework for multimodal MRI brain tumor segmentation and survival prediction, designed for precise volumetric analysis of gliomas. The model leverages attention mechanisms to selectively focus on tumor-relevant regions, achieving high Dice similarity and robust boundary delineation on the BraTS dataset.

Key Features:

• 🧠 3D Attention UNet Architecture: Integrates attention gates into UNet3D to suppress irrelevant regions and enhance tumor-focused feature learning.
• 🧬 Multi-Modal MRI Fusion: Supports FLAIR, T1, T1CE, and T2 modalities for comprehensive tumor representation.
• 🎯 High Segmentation Accuracy: Achieved Dice scores > 0.97 across tumor sub-regions (WT, TC, ET).
• 📊 Survival Prediction Module: Includes optional regression/classification head for patient survival estimation.
• ⚙️ Deep Supervision & Residual Learning: Improves gradient flow and stabilizes training for 3D volumes.

Architecture Overview:

• Encoder–Decoder UNet3D Backbone: Captures hierarchical spatial features from volumetric MRI inputs.
• Attention Gates: Dynamically weight skip-connection features based on tumor relevance.
• Multi-Scale Feature Aggregation: Preserves fine-grained tumor boundaries across resolutions.
• Joint Learning Setup: Enables simultaneous tumor segmentation and survival prediction.

Technical Highlights:

• Built with PyTorch using a modular and extensible training pipeline
• Advanced preprocessing: normalization, patch extraction, rotation, flipping, and intensity augmentation
• Evaluated using Dice Score, Hausdorff Distance, Sensitivity, and Specificity
• Adapted from PyTorch-3DUNet and optimized for BraTS MRI datasets