CDR-Net

Change Detection and Reconstruction Network for Post-operative Glioma Follow-up

CDR-Net predicts spatially-resolved tumor change maps — where edema (ED) and active tumor (AT) will increase or decrease — directly from a single time-point MRI, without requiring a follow-up scan at inference.

---

How it works

Input:  Time1 MRI  (T1, T1ce, T2, FLAIR)  →  4-channel 2D slice
             │
     Shared Residual Encoder
            ├──────────────────┐
       ED Decoder          AT Decoder
            │                  │
    [ED_INC, ED_DEC]   [AT_INC, AT_DEC]   ←  predicted change maps

Evaluation:  T2_pred = (T1_mask ∪ INC_pred) \ DEC_pred
             Dice(T2_pred, T2_GT)           ←  Recon Dice

Ground truth is derived from paired expert segmentation masks:

INC = time2_mask & ~time1_mask   (newly appeared region)
DEC = time1_mask & ~time2_mask   (disappeared region)

---

Repository Structure

project_root/
├── dataset.py         ← NIfTI slice dataset, patch crop, augmentation
├── utils.py           ← Loss functions, evaluation metrics
└── CDR-Net/
    ├── model.py       ← Dual-decoder U-Net (ED / AT)
    ├── train.py       ← Standard training
    ├── train_ema.py   ← EMA training (recommended)
    ├── run.py         ← CLI for train.py
    ├── run_ema.py     ← CLI for train_ema.py
    ├── test.py        ← Evaluation (standard)
    └── test_ema.py    ← Evaluation (EMA)

---

Setup

Dataset — UCSF Post-operative Glioma Dataset:

https://imagingdatasets.ucsf.edu/dataset/2

Each patient folder should contain:

{pid}_time1_t1.nii.gz      {pid}_time1_t1ce.nii.gz
{pid}_time1_t2.nii.gz      {pid}_time1_flair.nii.gz
{pid}_time1_seg.nii.gz     {pid}_time2_seg.nii.gz

Dependencies:

pip install torch nibabel numpy tqdm

---

Usage

# Train with EMA (recommended)
python run_ema.py --root /path/to/dataset

# Train standard
python run.py --root /path/to/dataset

# Evaluate
python test_ema.py   # or test.py for standard checkpoint

---

Key Design Choices

	Detail
Split	Patient-level 8 / 1 / 1 (no slice-level leakage)
Sampling	AT-change slices ×10, ED-change slices ×3
Class imbalance	Per-channel BCE pos_weight (cap 100)
Loss — ED	BCE + Dice + Tversky
Loss — AT	BCE + Focal + Dice + Tversky + overlap penalty
Best model	Val Recon Dice (not Delta Dice)
EMA decay	0.995

---