Vestibular Schwannoma Segmentation

An ML pipeline for automatic segmentation of vestibular schwannomas (VS) from 3D MRI scans using size-aware data augmentation and transformer-enhanced nnUNet architectures.

🎯 Highlights

• 2.5D UNet with Transformer bottleneck + gated residuals
• Conditional Diffusion (Med-DDPM) for generating rare tumor samples
• Cross-modality fusion decoder (T1 + T2)
• Dice score: 0.926 on T1, 0.906 on T2

🧪 Data

• Public VS-SEG dataset (242 patient scans)
• T1-weighted and T2-weighted MRIs
• Tumor contours in JSON, converted to NIfTI using 3DSlicer

🔧 Setup

pip install niftynet tensorflow

🧠 Architecture

• nnUNet backbone
• Optional Transformer block between encoder/decoder
• Fusion Decoder uses pretrained Swin Transformer

🖼 Sample Outputs

• Synthetic tumors in rare size bins
• Segmentation visualizations coming soon

📝 Citation

Work based on: Poe, Wu, Conlin, Longhitano et al. (2024), BU NLP