DRAW — Deep Radiotherapy Autosegmentation Workflow

End-to-end auto-segmentation pipeline for radiotherapy (RT) planning. DRAW reads DICOM CT studies, predicts organs-at-risk and clinical target volumes for multiple cancer sites, and writes back DICOM RT-Struct files for use in clinical treatment planning systems. Built on top of nnU-Net v2 with custom logic for overlapping clinical labels and parallel multi-model inference.

---

Status

DRAW is deployed across three hospitals including Tata Medical Centre, Kolkata, where the pipeline supports clinical RT planning workflows on 300+ CT scans per day.

The prostate-cancer configuration (TSPrime) is the published flagship and is the reference for evaluating new cancer-site configurations.

Paper

End-to-End Prostate Cancer Segmentation for RT Planning Sandip Dutta⋆, Surajit Kundu, Santam Chakraborty, Indranil Mallick, Sougata Maity, Aranya Sarkar, Soumyajit Das, Sanjoy Chatterjee, Rimpa Basu Achari, Moses Arunsingh, Tapesh Bhattacharyya, Jayanta Mukhopadhyay, Nishant Chakravorty. Computer Vision and Image Processing (CVIP), 2024. Springer CCIS, Vol. 2477, pp. 192–204. IIT Kharagpur (Dept. of CSE) × Tata Medical Centre, Kolkata (Dept. of Radiation Oncology). DOI: 10.1007/978-3-031-93709-5_14 | [Google Scholar] ⋆ Correspondence: [email protected]

Headline results (prostate, TSPrime configuration)

Eight RT structures segmented end-to-end on a 78-patient prospective trial dataset (split 48 / 12 / 18). Best configuration: PlainConvUNet + connected-component post-processing.

Structure	Dice	Clinical acceptability (acceptable / minor edits)
Bladder	0.972	63% / 32%
Anorectum	0.873	95% / 5%
Bag_Bowel	0.910	37% / 53%
Femur_Head_L	0.895	53% / 47%
Femur_Head_R	0.883	78% / 22%
Penilebulb	0.677	47% / 47%
CTVp	0.850	42% / 32%
CTVn	0.852	47% / 37%

The 0.677 Dice on penile bulb matches the published inter-observer variability among expert human delineators (0.67–0.70). It reflects the structural ambiguity of the anatomy, not a model failure.

End-to-end inference time: 30 min → 5 min per case on training hardware (Xeon E5-1660 v4, GTX 1080 Ti 11GB). On a hospital workstation (Intel i7-12700K, NVIDIA T1000), the same pipeline runs in 15 minutes — still a 50% saving over manual delineation, with no clinician time required.

Supported cancer sites

DRAW uses a YAML-driven model registry: each cancer site is configured via a file in config_yaml/, and adding a new site is a config change rather than a code change.

Site	Config	Notes
Prostate	ts_prime.yml	Published reference (CVIP'24)
Gynaecological	ts_gyne.yml	Full-bladder protocol
Breast	ts_breast_new.yml
Head & Neck	ts_headneck.yml
CNS	ts_cns.yml
Lung	draw_lung.yml
Cranio-spinal (CSI)	draw_csi.yml

Architecture

DRAW is a Click-based CLI with five top-level commands (main.py):

Command	Module	What it does
preprocess	draw/preprocess/	Convert DICOM to NIfTI, normalise labels, prepare nnU-Net dataset
train-single-gpu	draw/train/	Train an nnU-Net model on a prepared dataset
predict	draw/predict/	Run inference on a study and write DICOM RT-Struct
start-pipeline	draw/pipeline/	Continuous mode: watch a directory and process incoming studies
zip-model	draw/impex/	Export a trained model bundle for deployment

Engineering decisions worth surfacing

1. Multi-model split for overlapping RT labels. DICOM RT-Struct supports overlapping segmentations; NIfTI does not. CTVn (nodal target) routinely overlaps with Bag_Bowel. Rather than dropping overlap voxels by priority — which contaminates either bowel or CTVn — DRAW trains separate nnU-Net models per non-overlapping label group and recombines the outputs into a multi-label DICOM RT-Struct downstream. The TSPrime configuration uses three sub-models (OARs / CTVp / CTVn).
2. Parallel multi-model inference. A single PlainConvUNet does not saturate the GPU on a 512³-class CT volume. DRAW co-schedules the label-group models on the same GPU at inference, sharing input loading and post-processing. The constraint accepted for clinical correctness (label-group split) becomes the lever that fills the GPU. ~3× wall-clock speedup vs serial execution.
3. TTA disabled by default. nnU-Net's default test-time augmentation runs each volume eight times with negligible Dice impact on the trial dataset. Disabling TTA gives an additional ~8× speedup.
4. Largest-3D-component post-processing for paired structures. Femur_Head_L / Femur_Head_R get occasionally cross-classified across the midline. Retaining only the largest connected component per label eliminates the strays. Femur_Head_L Dice: 0.874 → 0.895. Neutral or positive on every other structure.
5. Continuous prediction pipeline. start-pipeline runs as a long-lived watcher that picks up new DICOM studies, runs inference, writes RT-Struct back, and records workflow state in the configured database. This is what powers the 300+ scans/day clinical deployment.
6. DICOM ↔ NIfTI conversion glue via dcmrtstruct2nii (forward) and rt_utils (reverse).

The custom training logic lives in Dutta-SD/nnunet_draw, which DRAW pins as a dependency in requirements.txt.

Installation

git clone https://github.com/CHAVI-India/draw.git
cd draw
pip install -r requirements.txt
cp template.env.draw.yml env.draw.yml   # then fill in the values

Required environment variables (in env.draw.yml): DB_URL, DB_NAME, TABLE_NAME, WATCH_DIR, MODEL_DEF_ROOT.

Documentation

Comprehensive documentation lives under documentation/:

• architecture.md — system design, module structure, key engineering decisions
• installation.md — prerequisites, setup, hardware requirements
• cli_reference.md — full CLI reference for all 5 commands
• configuration.md — environment file, YAML model configs, tunable constants
• data_flow.md — format conversions (DICOM ↔ NIfTI ↔ RT-Struct), axis handling
• training_guide.md — end-to-end training workflow, cloud training, resumption
• pipeline.md — continuous prediction pipeline, state machine, protocol routing
• database.md — schema, queue operations, monitoring queries
• deployment.md — single-machine and multi-hospital deployment, troubleshooting

Citation

If you use DRAW or its prostate configuration in academic work, please cite the paper:

@inproceedings{dutta2024drawrt,
  title     = {End-to-End Prostate Cancer Segmentation for RT Planning},
  author    = {Dutta, Sandip and Kundu, Surajit and Chakraborty, Santam and Mallick, Indranil and Maity, Sougata and Sarkar, Aranya and Das, Soumyajit and Chatterjee, Sanjoy and Achari, Rimpa Basu and Arunsingh, Moses and Bhattacharyya, Tapesh and Mukhopadhyay, Jayanta and Chakravorty, Nishant},
  booktitle = {Computer Vision and Image Processing (CVIP)},
  series    = {Communications in Computer and Information Science},
  volume    = {2477},
  pages     = {192--204},
  year      = {2024},
  publisher = {Springer},
  doi       = {10.1007/978-3-031-93709-5_14}
}

Please also cite the underlying framework:

@article{isensee2021nnunet,
  title   = {nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation},
  author  = {Isensee, Fabian and Jaeger, Paul F and Kohl, Simon A A and Petersen, Jens and Maier-Hein, Klaus H},
  journal = {Nature Methods},
  volume  = {18},
  number  = {2},
  pages   = {203--211},
  year    = {2021}
}

License

Apache License 2.0.

Acknowledgements

This work was carried out under grant IIT/SRIC/CS/RFE/2023-2024/016 (19-04-2023). Patient data was used with prior consent and de-identified per Kundu et al., J. Digital Imaging 2022. Built on nnU-Net v2 by the Applied Computer Vision Lab at Helmholtz Imaging and DKFZ.