Deep Learning CT Reconstruction

This repository contains a CT reconstruction workflow built around two stages:

1. Classical reconstruction from projection images
2. Deep-learning preparation and training on degraded-vs-reference reconstructions

Current pipeline

• data_loader.py: load sample 1 projections and scan settings
• geometry.py: parse CT geometry from settings.cto
• reconstruct_fbp.py: early slice-wise baseline
• reconstruct_fdk_astra.py: main cone-beam FDK reconstruction using ASTRA
• simulate_degradation.py: create sparse-view, limited-angle, and noisy projection datasets
• build_training_pairs.py: build degraded/input and reference/target reconstruction pairs
• train_unet.py: train a 2D U-Net on axial slice pairs
• export_for_colab.py: export compact training data for Colab

Recommended workflow

1. Reconstruct the full reference volume with reconstruct_fdk_astra.py
2. Generate degraded projection datasets with simulate_degradation.py
3. Build aligned input-target reconstruction pairs with build_training_pairs.py
4. Train the first model with train_unet.py

Notes

• Raw datasets, generated outputs, checkpoints, previews, and exported training archives are excluded from Git.
• The main 3D reconstruction path uses ASTRA with CUDA.