NR-Align

Official code release for our MICCAI 2026 accepted paper:

NR-Align: Non-Rigid Alignment for Non-simultaneous Two-View 3D Coronary Reconstruction in Complex Cardiac Interventions

This repository provides the open-source implementation of the data generation and training pipeline used in NR-Align. The toolkit builds physiology-aware synthetic non-simultaneous two-view training cases, converts them into backprojection volumes, and trains a compact 3D non-rigid alignment network for improving two-view coronary reconstruction under asynchronous acquisition.

What Is Included

• generate_cases.py: create asynchronous non-rigid two-view vessel masks.
• build_bp.py: project masks and build SIRT backprojection volumes.
• train_nr_align.py: train the NR-Align correction network.
• nr_align_models.py: model definitions.
• deform.py: deformation and mask-processing utilities.
• geometry.py: metadata parsing and TIGRE geometry helpers.

Install

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

TIGRE is required only for build_bp.py:

python -c "import tigre; print('TIGRE OK')"

Input Data

Prepare one binary 3D vessel mask and one geometry file per case:

data/gt128/{ID}.npy       # shape=(128,128,128), uint8 {0,1}, XYZ order
data/meta/{ID}.npz        # projection geometry metadata

Generated training cases are written as:

data/cases/ID001/amp6/k00/
|-- view1_mask.npy
|-- view2_mask.npy
|-- meta.json
|-- bp012_nr.npy
`-- bp012_gt.npy

bp012_nr.npy is the non-simultaneous two-view input. bp012_gt.npy is the aligned target.

Quick Start

Generate non-rigid two-view cases:

python scripts/generate_cases.py \
  --ids 1-100 \
  --gt_root data/gt128 \
  --meta_npz_dir data/meta \
  --out_root data/cases \
  --amp_list_mm 6,10,14,18 \
  --ks 0-2 \
  --variant_map sym_small,sym_large,same_side_small \
  --device cuda:0 \
  --overwrite 1

Build backprojection volumes:

CUDA_VISIBLE_DEVICES=0 python scripts/build_bp.py \
  --data_root data/cases \
  --gt_dir data/gt128 \
  --meta_npz_dir data/meta \
  --ids 1-100 \
  --amps 6,10,14,18 \
  --ks 0-2 \
  --nDetector 512 \
  --sirt_iters 1 \
  --save_proj 1 \
  --overwrite 1

Train NR-Align:

python scripts/train_nr_align.py \
  --data_root data/cases \
  --work_dir runs/nr_align_dual_ds_amp6_18 \
  --ids 1-100 \
  --amps 6,10,14,18 \
  --ks 0-2 \
  --variant dual_ds \
  --epochs 120 \
  --batch 1 \
  --grad_accum 4 \
  --lr 2e-4 \
  --wd 1e-4 \
  --base 24 \
  --amp 1 \
  --seed 0

Training writes:

runs/nr_align_dual_ds_amp6_18/
|-- nr_align_best.pt
|-- nr_align_ckpt_ep*.pt
|-- nr_align_run_meta.json
`-- nr_align_train_history.json

Model Variants

• single: union output only.
• dual: union and overlap outputs.
• dual_ds: union and overlap outputs with 64^3 / 32^3 deep supervision.

dual_ds is the recommended default.

More Details

• Data format
• Parameters
• Recommended public parameters

Choose an open-source license before publishing.