[Paper Review] Dose Prediction with U-net: A Feasibility Study for Predicting Dose Distributions from Contours using Deep Learning on Prostate IMRT Patients.

Dan Nguyen, Troy Long+5 more|arXiv (Cornell University)|Sep 26, 2017

Advanced Radiotherapy Techniques35 references45 citations

TL;DR

This study proposes a modified U-net deep learning model to predict intensity-modulated radiation therapy (IMRT) dose distributions directly from patient contours of the planning target volume (PTV) and organs at risk (OAR). Using only a small dataset, the model achieves high accuracy, with a mean Dice similarity coefficient of 0.91 for isodose volumes and absolute dose differences below 5% of prescription dose across all structures, demonstrating feasibility for accelerating treatment planning.

ABSTRACT

With the advancement of treatment modalities in radiation therapy for cancer patients, outcomes have improved, but at the cost of increased treatment plan complexity and planning time. The accurate prediction of dose distributions would alleviate this issue by guiding clinical plan optimization to save time and maintain high quality plans. We have modified a convolutional deep network model, U-net (originally designed for segmentation purposes), for predicting dose from patient image contours of the planning target volume (PTV) and organs at risk (OAR). We show that, as an example, we are able to accurately predict the dose of intensity-modulated radiation therapy (IMRT) for prostate cancer patients, where the average Dice similarity coefficient is 0.91 when comparing the predicted vs. true isodose volumes between 0% and 100% of the prescription dose. The average value of the absolute differences in [max, mean] dose is found to be under 5% of the prescription dose, specifically for each structure is [1.80%, 1.03%](PTV), [1.94%, 4.22%](Bladder), [1.80%, 0.48%](Body), [3.87%, 1.79%](L Femoral Head), [5.07%, 2.55%](R Femoral Head), and [1.26%, 1.62%](Rectum) of the prescription dose. We thus managed to map a desired radiation dose distribution from a patient's PTV and OAR contours. As an additional advantage, relatively little data was used in the techniques and models described in this paper.

Motivation & Objective

• To reduce treatment planning time in IMRT by enabling rapid dose prediction from patient anatomy and contours.
• To address the growing complexity of IMRT plans that increases planning workload and time.
• To explore whether deep learning can accurately predict dose distributions using only contour data and limited training examples.
• To evaluate the feasibility of adapting a segmentation-based U-net model for dose prediction in clinical radiotherapy.

Proposed method

• A U-net architecture, originally designed for medical image segmentation, was retrained for dose prediction using patient contours as input.
• Input data consisted of delineated contours for the PTV and OARs, without raw CT images or Hounsfield units.
• The model was trained to predict the full 3D dose distribution across the patient volume, with loss functions optimized for dose accuracy.
• Training used a relatively small dataset of prostate IMRT cases, emphasizing data efficiency.
• The network employed skip connections and encoder-decoder architecture to preserve spatial resolution and anatomical detail.
• Model performance was evaluated using Dice similarity coefficient and absolute dose differences for key structures.

Experimental results

Research questions

• RQ1 Can a U-net-based deep learning model accurately predict IMRT dose distributions using only patient contour data?
• RQ2 How does the model's performance compare to clinical plans in terms of dose distribution similarity and absolute dose differences?
• RQ3 To what extent can dose prediction be achieved with minimal training data in a clinical setting?
• RQ4 How well does the model generalize across different OARs and PTVs in prostate cancer patients?
• RQ5 Can a segmentation-focused network like U-net be effectively repurposed for dose prediction with acceptable accuracy?

Key findings

• The model achieved a mean Dice similarity coefficient of 0.91 between predicted and actual isodose volumes across 0–100% of the prescription dose.
• For the PTV, the absolute mean dose difference was 1.03% of prescription dose, with a maximum difference of 1.80%.
• The bladder showed a mean dose difference of 4.22% and a maximum difference of 1.94% of prescription dose.
• The body structure had a mean dose difference of 0.48% and a maximum difference of 1.80% of prescription dose.
• The left and right femoral heads had mean dose differences of 1.79% and 2.55%, respectively, with maximum differences of 3.87% and 5.07%.
• The rectum showed a mean dose difference of 1.62% and a maximum difference of 1.26% of prescription dose.