TransVert is a deep learning model for reconstructing 3D vertebra volumes from sagittal and coronal 2D medical slices. It uses a dual-encoder architecture followed by a 3D volume expander and is trained with a custom Dice Loss function. This repository includes training, validation, evaluation, and interactive visualization tools.
- Dual-view 2D input (sagittal + coronal) → 3D volume output
- Dice Loss-based optimization
- Training & validation loss tracking with
wandb - Inline evaluation loop with Dice Score logging
- Visual mid-slice predictions (wandb + matplotlib)
- Interactive 3D slice viewer with
ipywidgets
pip install torch torchvision matplotlib tqdm wandb ipywidgets scikit-image