[Paper Review] DeepDoseNet: A Deep Learning model for 3D Dose Prediction in Radiation Therapy

Mumtaz Hussain Soomro, Victor Gabriel Leandro Alves+2 more|arXiv (Cornell University)|Oct 29, 2021

Advanced Radiotherapy Techniques11 citations

TL;DR

DeepDoseNet is a deep learning model combining ResNet and Dilated DenseNet architectures for 3D dose prediction in head-and-neck intensity-modulated radiation therapy. Using the 2020 AAPM OpenKBP dataset, it achieved state-of-the-art performance with a 60% reduction in 3D dose score and 70% reduction in DVH score using a combined MAE and DVH-based loss function, significantly outperforming baseline models.

ABSTRACT

The DeepDoseNet 3D dose prediction model based on ResNet and Dilated DenseNet is proposed. The 340 head-and-neck datasets from the 2020 AAPM OpenKBP challenge were utilized, with 200 for training, 40 for validation, and 100 for testing. Structures include 56Gy, 63Gy, 70Gy PTVs, and brainstem, spinal cord, right parotid, left parotid, larynx, esophagus, and mandible OARs. Mean squared error (MSE) loss, mean absolute error (MAE) loss, and MAE plus dose-volume histogram (DVH) based loss functions were investigated. Each model's performance was compared using a 3D dose score, $\\bar{S_{D}}$, (mean absolute difference between ground truth and predicted 3D dose distributions) and a DVH score, $\\bar{S_{DVH}}$ (mean absolute difference between ground truth and predicted dose-volume metrics).Furthermore, DVH metrics Mean[Gy] and D0.1cc [Gy] for OARs and D99%, D95%, D1% for PTVs were computed. DeepDoseNet with the MAE plus DVH-based loss function had the best dose score performance of the OpenKBP entries. MAE+DVH model had the lowest prediction error (P<0.0001, Wilcoxon test) on validation and test datasets (validation: $\\bar{S_{D}}$=2.3Gy, $\\bar{S_{DVH}}$=1.9Gy; test: $\\bar{S_{D}}$=2.0Gy, $\\bar{S_{DVH}}$=1.6Gy) followed by the MAE model (validation: $\\bar{S_{D}}$=3.6Gy, $\\bar{S_{DVH}}$=2.4Gy; test: $\\bar{S_{D}}$=3.5Gy, $\\bar{S_{DVH}}$=2.3Gy). The MSE model had the highest prediction error (validation: $\\bar{S_{D}}$=3.7Gy, $\\bar{S_{DVH}}$=3.2Gy; test: $\\bar{S_{D}}$=3.6Gy, $\\bar{S_{DVH}}$=3.0Gy). No significant difference was found among models in terms of Mean [Gy], but the MAE+DVH model significantly outperformed the MAE and MSE models in terms of D0.1cc[Gy], particularly for mandible and parotids on both validation (P<0.01) and test (P<0.0001) datasets. MAE+DVH outperformed (P<0.0001) in terms of D99%, D95%, D1% for targets. MAE+DVH reduced $\\bar{S_{D}}$ by ~60% and $\\bar{S_{DVH}}$ by ~70%.

Motivation & Objective

• To develop a deep learning model for accurate 3D dose prediction in radiation therapy to improve treatment planning efficiency.
• To evaluate the impact of different loss functions—MSE, MAE, and MAE+DVH—on prediction accuracy for both 3D dose and DVH metrics.
• To compare model performance across key clinical structures including PTVs and OARs such as parotids, mandible, and spinal cord.
• To determine whether incorporating DVH-based loss improves clinical relevance of predicted dose distributions.
• To achieve superior performance on the 2020 AAPM OpenKBP challenge using a hybrid ResNet and Dilated DenseNet architecture.

Proposed method

• A hybrid deep learning architecture combining ResNet and Dilated DenseNet was designed to capture multi-scale anatomical features and dose deposition patterns.
• The model was trained on 200 head-and-neck patient datasets from the 2020 AAPM OpenKBP challenge, with 40 for validation and 100 for testing.
• Three loss functions were evaluated: mean squared error (MSE), mean absolute error (MAE), and a combined MAE plus dose-volume histogram (DVH)-based loss.
• The 3D dose score ($\bar{S_{D}}$) and DVH score ($\bar{S_{DVH}}$) were used as primary metrics to compare model performance.
• Clinical metrics including D99%, D95%, D1% for PTVs and Mean[Gy], D0.1cc[Gy] for OARs were computed and compared across models.
• Statistical significance was assessed using the Wilcoxon signed-rank test to compare prediction errors across models.

Experimental results

Research questions

• RQ1 Does combining MAE and DVH-based loss improve 3D dose prediction accuracy compared to using MAE or MSE alone?
• RQ2 How does the DeepDoseNet model compare to other entries in the 2020 AAPM OpenKBP challenge in terms of 3D dose and DVH score performance?
• RQ3 To what extent does the model reduce prediction error for clinically relevant metrics such as D0.1cc[Gy] in critical organs?
• RQ4 Is there a significant difference in prediction accuracy between the MAE+DVH, MAE, and MSE loss variants on both validation and test datasets?
• RQ5 Can the model achieve superior performance in predicting target coverage (D99%, D95%, D1%) and organ-at-risk sparing (Mean[Gy], D0.1cc[Gy])?

Key findings

• The MAE+DVH model achieved the lowest 3D dose score ($\bar{S_{D}}$) of 2.0 Gy on the test set, representing a ~60% reduction compared to the MSE model.
• The MAE+DVH model achieved the lowest DVH score ($\bar{S_{DVH}}$) of 1.6 Gy on the test set, representing a ~70% reduction compared to the MSE model.
• The MAE+DVH model significantly outperformed both MAE and MSE models in D0.1cc[Gy] for mandible and parotids (p<0.0001 on test set) and in PTV metrics D99%, D95%, D1% (p<0.0001).
• No significant difference was found among models for Mean[Gy] across OARs, indicating consistent mean dose prediction.
• The MAE+DVH model had the lowest prediction error on both validation and test sets (p<0.0001, Wilcoxon test), confirming statistical superiority.
• The model demonstrated state-of-the-art performance among all OpenKBP challenge entries, particularly in clinical metric accuracy and dose distribution fidelity.