Zombies-learning-progression-PGGAN

Overview

An end-to-end pipeline for augmenting longitudinal brain MRI data with a Two-Generator Progressive GAN (PGGAN). The GAN generates anatomically consistent baseline and follow-up MRI pairs, and this synthetic data is used to augment limited real datasets for two downstream tasks: 3D tumor segmentation and RANO classification.

Research paper (Overleaf): read-only link

What it does

The pipeline runs in five stages:

1. Train the GAN on preprocessed real MRI volumes (CACHED128).
2. Generate synthetic data: baseline–follow-up pairs written to generated1070 as .npz files (each containing baseline and follow-up T1ce).
3. Train a 3D segmentation model (CoTrSeg / nnU-Net / DynUNet) on real MRI data.
4. Run inference on the synthetic images with the trained segmentation model (visualisation128aug).
5. Train a RANO classifier with XGBoost on real plus synthetic images, their segmentation masks and radiomics features, using SMOTE to handle class imbalance.

How it's built

• Dual generators: G_B generates a baseline MRI from a latent vector z; G_F generates the follow-up MRI from z plus baseline features, conditioned at all scales.
• Dual discriminators: D_B for baseline and D_F for follow-up, trained with WGAN-GP for stability.
• Loss: adversarial + λ₁·L1 + λ₂·SSIM, chosen to enforce anatomical consistency and temporal realism.
• Volumes: all processing is on 128 × 128 × 128 3D volumes.
• Data format: each real sample is a 10-channel .npz of shape (10, 128, 128, 128): baseline T1/T1ce/T2/FLAIR (channels 0–3), follow-up T1/T1ce/T2/FLAIR (4–7), and baseline/follow-up masks (8–9) with labels 0 = background, 1 = edema, 2 = tumor. The T1ce channels are the ones used by the GAN and segmentation.
• Code layout: pggan/ (config, dataset, networks, train, generate), train_cotrseg.py / model_cotrseg.py for segmentation, and train_classifier_radiomics.py for the RANO classifier.

Results

Quantitative results are not given in the README; they are deferred to the linked paper.