This project focuses on developing a deep learning framework for predicting three-dimensional radiotherapy dose distributions for prostate cancer treatment planning.
The pipeline uses a 3D U-Net implemented using the MONAI framework and is trained on processed clinical radiotherapy datasets containing:
- CT scans,
- Planning Target Volume (PTV) masks,
- Bladder Signed Distance Maps (SDMs),
- Anorectum Signed Distance Maps (SDMs),
- and clinically delivered radiation dose distributions.
The long-term objective of the project is to explore physics-guided deep learning strategies for radiotherapy treatment planning.
Modern radiotherapy planning is a highly complex optimization problem involving:
- accurate tumor dose coverage,
- sparing of healthy organs-at-risk (OARs),
- spatial dose falloff,
- and clinically meaningful dose distributions.
Traditional voxel-wise regression losses such as Mean Squared Error (MSE) do not explicitly encode these physical and clinical constraints.
This project therefore aims to investigate hybrid approaches combining:
- deep learning,
- geometric anatomical priors,
- and physics-guided optimization strategies.
For each patient, the model uses four aligned volumetric input channels:
| Channel | Description |
|---|---|
| 0000 | CT Scan |
| 0001 | Planning Target Volume (PTV) |
| 0002 | Bladder Signed Distance Map (SDM) |
| 0003 | Anorectum Signed Distance Map (SDM) |
The target output is the clinically delivered 3D radiation dose distribution.
The preprocessing pipeline performs:
- DICOM/NIfTI loading,
- channel formatting,
- voxel spacing resampling,
- CT intensity normalization,
- multi-channel tensor concatenation,
- and volumetric patch extraction.
All patient volumes are resampled to a uniform voxel spacing:
TARGET_SPACING = (1.27, 1.27, 2.5)Training is performed using random 3D patches:
PATCH_SIZE = (96, 96, 96)The current implementation uses a MONAI-based 3D U-Net architecture:
UNet(
spatial_dims=3,
in_channels=4,
out_channels=1,
channels=(16, 32, 64, 128, 256),
strides=(2, 2, 2, 2),
num_res_units=2,
)The network predicts a continuous-valued radiation dose distribution.
The baseline implementation currently uses voxel-wise Mean Squared Error (MSE):
loss = MSE(predicted_dose, true_dose)Future work aims to extend this toward physics-guided losses incorporating:
- PTV coverage constraints,
- organ-at-risk penalties,
- smoothness regularization,
- distance-aware weighting,
- and DVH-based optimization.
Inference is performed using MONAI sliding-window inference to handle large 3D patient volumes efficiently.
The workflow:
- Loads the trained model,
- Applies preprocessing transforms,
- Performs patch-wise volumetric inference,
- Reconstructs the full dose distribution,
- Saves predicted dose volumes in NIfTI format.
Current validation metrics include:
- Validation Mean Squared Error (MSE),
- PTV D95,
- Mean Bladder Dose,
- Mean Rectum Dose.
These metrics provide clinically interpretable measures of:
- target coverage,
- and organ-at-risk sparing.
Planned future developments include:
- Physics-guided loss formulations,
- Dose Volume Histogram (DVH)-based optimization,
- Beam geometry integration,
- Attention-based architectures,
- Improved volumetric sampling strategies,
- and clinically informed evaluation pipelines.
This project currently uses:
- PyTorch
- MONAI
- NumPy
- SimpleITK
This work is being developed as part of a research internship in deep learning-assisted radiation treatment planning.
The project lies at the intersection of:
- medical physics,
- deep learning,
- optimization,
- and computational radiotherapy.
This project is currently intended for research and educational purposes only. It is not validated for clinical deployment or patient treatment use.
Shield:
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License.
