Surface-Conditioned Implicit Anatomy Reconstruction

Code for Surface-Conditioned Implicit Reconstruction of Internal Anatomy for Automated Patient Positioning.

Overview

This repository implements a surface-conditioned implicit neural representation for reconstructing dense internal anatomical labels from an external body-surface point cloud. The model combines a hierarchical Point Transformer body encoder with a modulated SIREN decoder. At inference time, volumetric segmentations are generated by querying 3D coordinates and predicting anatomical class logits.

Repository Structure

• main.py: entry point for preprocessing, training, volume generation, and evaluation.
• data/: dataset loading and batching utilities.
• models/: body encoder, modulated SIREN decoder, and segmentation network wrapper.
• steps/segmentation/processing/: label mapping, query sampling, and volume generation.
• steps/segmentation/evaluation/: segmentation metrics and HTML report generation.
• steps/training/: training loop utilities.
• lib/pointops/: custom CUDA point operations used by the Point Transformer layers.
• common/ and util/: shared helpers for scoring, coordinate handling, I/O, and losses.

Installation

The environment can be created with the provided setup script:

bash env_setup.sh ia

Data

The dataset is not included due to privacy restrictions. The code expects preprocessed body-surface point clouds, volumetric segmentation masks, label metadata, and train/test subject lists.

Input paths are configured via command-line arguments. If your local file names differ from the defaults, use the optional path overrides listed below.

Usage

Preprocess labels and sample training queries:

python main.py --preprocess \
  --data_root /path/to/data \
  --save_path /path/to/experiments \
  --gpu=0

Train the model:

python main.py --train \
  --data_root /path/to/data \
  --save_path /path/to/experiments \
  --gpu=0

Generate segmentation volumes:

python main.py --generate \
  --data_root /path/to/data \
  --save_path /path/to/experiments \
  --gpu=0

Evaluate generated segmentations:

python main.py --evaluate \
  --data_root /path/to/data \
  --save_path /path/to/experiments

Use --gpu=auto to select the least-used GPU according to available memory.

Optional path overrides:

--training_list_file_path /path/to/training_patient_list.txt
--testing_list_file_path /path/to/test_patient_list.txt
--raw_label_json_file_path /path/to/organ_label_list.json
--label_json_file_path /path/to/label_organs.json
--body_data_path /path/to/data_1k_python37.npz
--mask_dir /path/to/masks_volumetric_preprocessed_v2
--metadata_csv /path/to/masks_volumetric_metadata.csv
--gt_dir /path/to/mapped_masks

Outputs

By default, experiment outputs are written below save_path:

<save_path>/implicit-anatomy/
  1_preprocess/
  2_train/
  3_generate/
  4_evaluate/

The generated volumes are stored as NIfTI files, and evaluation produces a CSV table together with an HTML segmentation report.

Acknowledgements

This repository builds upon the original Point Transformer implementation by Zhao et al. (2021): https://github.com/POSTECH-CVLab/point-transformer

Citation

If you use this code, please cite:

@inproceedings{krnjacasurface,
  title={Surface-Conditioned Implicit Reconstruction of Internal Anatomy for Automated Patient Positioning},
  author={Krnjaca, Denis and Heinrich, Mattias P},
  booktitle={Off-Grid: 1st Workshop on Continuous Representations and Grid-Free Methods in Medical Imaging},
  year={2026}
}

The official Springer LNCS citation will be added once the proceedings are published.