Performance of Adaptive Deep Learning Models for Dose Predictions on High-Quality Cone-Beam Computed Tomography Images.

Researchers

Journal

Modalities

Models

Abstract

Online plan generation remains a patient-specific and time-consuming process that can place a significant burden on clinics strained with staffing shortages. As previous research show that dose-volume histogram (DVH) prediction plays a crucial role in automatic treatment planning, the objective of this study is to assess the capability of adaptive deep learning models in predicting dose information in volumetric modulation radiotherapy plans using the high-quality CBCT images and contour information of organs-at-risk (OARs).The relationship between dose-volume histograms (DVHs) in radiotherapy plans and the geometric information of organs-at-risk (OAR) and planning target volume (PTV) has been well established. To evaluate the performance of the current state-of-the-art convolutional neural network (CNN) models including VIT3D and Unet3D, and intuitive machine learning methods (i.e., SVM and MLP), we implemented those models for dose prediction and conducted a comprehensive analysis with treatment plans created from images acquired from patients who consented to participate an IRB-approved imaging study designed to evaluate the imaging performance of the system. In total, 20 plans created by certified medical dosimetrists were employed in this study, with 15 used for training the machine-learning models and the remaining 5 used for performance testing. Two evaluation metrics were used: 1) root mean square error (RMSE) of the predicted dose and true dose and 2) time spent on dose prediction.The results of the analysis showed that the ViT-3D (Transformer) model had the lowest RMSE of 3.682 ±0.010, followed by the Unet-3D (CNN) model with an RMSE of. 3.973 ±0.021 The MLP model had an RMSE of 8.007 ±0.019 while the SVM model had the highest RMSE of 9.156 ±0.032. For a fair comparison, we use 4-fold cross validation (each has 15 training plans and 5 testing plans), and report the mean value with standard deviation. All models are optimized with Adam optimizer of a learning rate 0.01, and the training process is stopped after 100 epochs. These findings indicate that the ViT-3D (Transformer) model performed the best in terms of predicting the dose information in volumetric modulation radiotherapy plans based on the CBCT images and contour information of OARs. For tested plan which contains 81 CT images (512 × 512 resolution), the inference time to predict dose information with a general CPU machine (6-Core Intel Core i7) is about 1.5 minutes. With GPU resources, such as NVIDIA A100, the inference process can be finished within seconds.The study demonstrated that current state-of-the-art machine-learning models can achieve promising accuracy in dose prediction using high-quality CBCT images. A well-trained machine-learning model could offer clinicians a quick and reliable prediction of the true dose to patients in the case of significant anatomical changes or provide patient-specific optimization objectives if replanning is warranted.Copyright © 2023. Published by Elsevier Inc.