Tumor Generation in Longitudinal Mammograms via Transformer GAN with Adaptive Blending

PyTorch implementation of a projection-aware longitudinal tumor synthesis framework for mammograms. The model uses paired prior and current mammograms, view/side embeddings, temporal cross-attention, a variational latent unit, anatomically constrained blending, and a Swin-based discriminator.

Repository Structure

projection_aware_longitudinal_tumor_synthesis/
├── train.py
├── test.py
├── infer.py
├── configs/
├── data/
├── models/
├── losses/
├── engine/
├── metrics/
├── utils/
├── scripts/
└── assets/

Expected Dataset Layout

dataset/
├── train/
│   ├── metadata.csv
│   ├── prior/
│   ├── current/
│   ├── breast_masks/
│   └── tumor_masks/
├── val/
│   ├── metadata.csv
│   ├── prior/
│   ├── current/
│   ├── breast_masks/
│   └── tumor_masks/
└── test/
    ├── metadata.csv
    ├── prior/
    ├── current/
    ├── breast_masks/
    └── tumor_masks/

metadata.csv

case_id,prior_path,current_path,breast_mask_path,tumor_mask_path,view,side,label
0001,train/prior/0001.png,train/current/0001.png,train/breast_masks/0001.png,train/tumor_masks/0001.png,CC,Left,1
0002,train/prior/0002.png,train/current/0002.png,train/breast_masks/0002.png,,MLO,Right,0

• label=1 indicates cancer case.
• tumor_mask_path can be empty for normal cases.
• view should be one of CC, MLO.
• side should be one of Left, Right.

Installation

conda create -n proj_tumor python=3.10 -y
conda activate proj_tumor
pip install -r requirements.txt

Training

python train.py \
  --data_root /path/to/dataset \
  --train_csv train/metadata.csv \
  --val_csv val/metadata.csv \
  --output_dir runs/exp1

Testing

python test.py \
  --data_root /path/to/dataset \
  --test_csv test/metadata.csv \
  --checkpoint /path/to/checkpoints/best_generator.pt

Inference

python infer.py \
  --data_root /path/to/dataset \
  --csv_path test/metadata.csv \
  --checkpoint /path/to/checkpoints/best_generator.pt \
  --save_dir outputs/inference