DeepDoseSens: dose prediction for glioblastoma, and how sensitive it is to contour changes

This repository accompanies two papers:

• ISBI 2023 — How sensitive are deep learning based radiotherapy dose prediction models to variability in Organs At Risk segmentation? Amith Kamath, Robert Poel, Jonas Willmann, Nicolaus Andratschke, Mauricio Reyes
• Cancers 2023 — Deep-Learning-Based Dose Predictor for Glioblastoma — Assessing the Sensitivity and Robustness for Dose Awareness in Contouring Robert Poel, Amith Kamath, Jonas Willmann, Nicolaus Andratschke, Ekin Ermiş, Daniel M. Aebersold, Peter Manser, Mauricio Reyes

Data, model weights and predicted dose volumes are archived outside this repository and shared on request — see Configuring where the data lives.

See a short video description of this work here:

🔗 Project Website

---

Overview

A cascaded 3D U-Net predicts the full 3D dose distribution for glioblastoma VMAT treatment from a planning CT and the target and organ-at-risk contours, in seconds instead of the hours a planning system needs. The question both papers ask is whether such a model is sensitive enough to be useful for contour quality assurance: if an organ contour is edited, does the predicted dose move the way the re-optimised clinical plan does? If it does, a contour can be judged by its dosimetric consequence rather than by geometry alone.

• ISBI 2023 establishes the model's accuracy and its sensitivity to ten plausible left optic nerve contours: predicted and planned dose differences correlate at 0.926, while the Dice coefficient correlates at only −0.471. Dice says a contour is different; only the dose says whether that matters.
• Cancers 2023 stress-tests the same model on out-of-distribution target shapes — concave targets and targets split into several lesions — and shows that adding six such cases to the training set recovers the conformity it loses.

---

Demonstration videos

docs/videos/ holds one clip per experiment. See README-videos.md for the layout and the appearance choices.

Video	Content
dose_sweep_DLDP_086.mp4	Planned versus predicted dose, sweeping caudal to cranial. Largest target of the 20 test cases (278 cc) and the most organs at risk in play (7 above 20 Gy) — also the worst dose score
dose_sweep_DLDP_083.mp4	The same, for the next-largest target (235 cc)
dose_sweep_DLDP_097.mp4	The same, for the demanding case whose fine streak structure the model reproduces best (216 cc, 6 OARs above 20 Gy)
optic_nerve_sensitivity.mp4	Ten plausible left optic nerve contours: the re-optimised plan beside the prediction, with the mean dose each delivers to the contour
robustness_concave.mp4	A concave target: planned dose, initial model, model retrained with concave cases
robustness_multiple.mp4	A multi-lesion target, same three panels

Each frame puts the planning CT in grey under a dose heat map, outlines the target volume and the organs at risk, and carries a structure legend and a dose colour bar. Cases are chosen by measured properties rather than by eye — target size, how many organs at risk the plan has to work around, and how well the prediction reproduces the plan's fine streak structure — and the selector prints the full ranking of all 20 test cases before picking. The retrained-model clips additionally skip any case the model was retrained on, which the scores give away.

Rebuild them with:

scripts/fetch_artifacts.sh
scripts/make_videos.sh

---

Model

A two-level cascaded 3D U-Net (Liu et al., Med Phys 2021), the architecture that won the OpenKBP challenge. The second U-Net takes the first one's output concatenated with its input.

• Input: 15 channels at 128³ — the normalised planning CT, the target volume and 13 organ-at-risk masks.
• Output: a continuous dose distribution, scaled to 0–70 Gy.
• Training: 60 cases, 15 validation, 80 000 iterations, 0.5·L1(ref, coarse) + L1(ref, refined).
• Inference: four-flip test-time augmentation. About 45 s per case on Apple silicon (MPS), 15 s on an A5000 GPU.

Prescription was 60 Gy in 30 fractions, normalised so 100% of the dose covers 50% of the target volume.

---

Results, as reproduced from the archived artifacts

scripts/reproduce.sh recomputes every published number from the NIfTI volumes and prints it beside the paper's value. Headlines:

Quantity	Paper	Reproduced
Dose score, 20 test cases (Cancers)	0.94 (0.36) Gy	0.94 (0.36) Gy
DVH score, 20 test cases (Cancers)	1.95 Gy	1.96 Gy
Per-OAR dose and DVH scores (ISBI Table 1, 26 values)	—	all match to ≤ 0.002 Gy
Optic nerve sensitivity (ISBI Table 2, 27 values)	—	all match to ≤ 0.001
Correlation, predicted vs planned dose shift	0.926	0.926
Correlation, planned dose shift vs Dice	−0.471	−0.471
Cancers Table 3 (48 values)	—	47 match to ≤ 0.01 Gy

Inference from the archived weights regenerates the archived predictions to 3 × 10⁻⁴ Gy on a 0–70 Gy scale (python -m deepdosesens.analyze.verify_inference), so the tables above are checked against the model, not only against saved predictions.

The two papers report the overall dose score over slightly different scoring regions; the value reproduced here, 0.94 Gy, is the one the journal version reports.

---

Getting started

Requirements

• Python 3.9+
• PyTorch (CUDA, MPS or CPU)
• SimpleITK, NumPy, SciPy, pandas, matplotlib
• ffmpeg, for the videos

git clone https://github.com/amithjkamath/deepdosesens.git
cd deepdosesens
uv venv .venv
source .venv/bin/activate
uv pip install -r pyproject.toml

Configuring where the data lives

No paths are hardcoded. Copy .env.example to .env and point it at your copy of the artifacts (or set the same variables in the environment):

DEEPDOSESENS_ISBI_ARCHIVE=/path/to/2022-11-ISBI/artifacts
DEEPDOSESENS_CANCERS_ARCHIVE=/path/to/2023-08-Cancers/artifacts
#DEEPDOSESENS_DATA=/mnt/big-disk/deepdosesens/data
#DEEPDOSESENS_CHECKPOINTS=/mnt/big-disk/deepdosesens/checkpoints

Everything falls back to directories inside the repository, so the defaults work once scripts/fetch_artifacts.sh has unpacked the archives. Check what is in effect with:

python -m deepdosesens.config

Only code and the demonstration videos are committed here. Planning CTs, contours, reference plans, predicted dose volumes and model weights live in the artifact archives and are shared on request; each archive carries its own manifest describing the layout. Fetch from it with:

WHAT=isbi scripts/fetch_artifacts.sh   # ISBI data, predictions and weights
scripts/fetch_artifacts.sh             # adds the Cancers cases and models

Predicting dose

The model is built and its weights read once; predicting a test set is one load and N forward passes:

from deepdosesens.config import checkpoint_path, data_path
from deepdosesens.inference import DosePredictor

predictor = DosePredictor(checkpoint_path("dose-predictor", "weights.pt"))
for case in ["DLDP_081", "DLDP_082"]:
    dose = predictor.predict(data_path("glioblastoma", case))  # Gy, (Z, H, W)

or from the command line:

python -m deepdosesens.inference --run my-run          # the 20 test cases
python -m deepdosesens.analyze.verify_inference        # check against the archive

Training

python train_C3D.py --batch_size 2 --max_iter 80000    # the cascaded 3D U-Net
python train_UNet.py                                   # single U-Net baseline

---

Repository layout

Path	Contents
deepdosesens/config.py	every path the project uses, from env vars or .env
deepdosesens/inference.py	DosePredictor — load once, predict many
deepdosesens/model/	the cascaded 3D U-Net, loss, and checkpoint loading/slimming
deepdosesens/data/	reading, preprocessing and augmentation
deepdosesens/analyze/	score definitions, the two reproduction scripts and the inference check
deepdosesens/visualization/	the video builders and their shared drawing code
scripts/	fetch_artifacts.sh, reproduce.sh, make_videos.sh
examples/	notebooks from the original analysis

If this is useful in your research, please consider citing:

@article{poel2023deep,
  title={Deep-Learning-Based Dose Predictor for Glioblastoma--Assessing the Sensitivity and Robustness for Dose Awareness in Contouring},
  author={Poel, Robert and Kamath, Amith J and Willmann, Jonas and Andratschke, Nicolaus and Ermi{\c{s}}, Ekin and Aebersold, Daniel M and Manser, Peter and Reyes, Mauricio},
  journal={Cancers},
  volume={15},
  number={17},
  pages={4226},
  year={2023}
}

@inproceedings{kamath2023doseprediction,
  title={How sensitive are deep learning based radiotherapy dose prediction models to variability in Organs At Risk segmentation?},
  author={Kamath, Amith and Poel, Robert and Willmann, Jonas and Andratschke, Nicolaus and Reyes, Mauricio},
  booktitle={2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI)},
  pages={1--4},
  year={2023},
  organization={IEEE}
}

Credits

Major props to the code and organization in https://github.com/LSL000UD/RTDosePrediction, which is what this model is based on (looks like this repo is not maintained/available anymore!)