A deep-learning framework for the prediction of the type of adaptive strategy of MR-guided prostate radiotherapy

Abstract

Background and purpose: In MR-guided adaptive radiotherapy (MRgART), adaptive strategies are currently mainly determined through subjective review of anatomical changes. Machine learning (ML) models based on deformable image registration (DIR) have been developed to predict type of adaptive strategy: adapt to position (ATP) versus adapt to shape (ATS). However, subjective review may result in sub-optimal plans and DIR processing for ML models can be time-consuming. This study aims to develop a deep learning (DL) model that uses images as input data for fast and accurate adaptive strategy selection. Methods: Data from 180 fractions of 36 prostate cancer patients were used retrospectively for this study. The optimal adaptive strategy was determined between ATP and ATS according to dosimetric evaluation. A multi-stage network method was proposed and used for adaptive strategy prediction. A DL-based image registration (DLIR) network was first trained to register the reference image to the daily image. Then, a DL-based adaptive strategy prediction (DLSP) model was constructed and trained using the encoder section of the trained DLIR network. Data from 24 patients were used for training, while the data from the remaining 12 patients formed the independent test set. Results: The DLSP model demonstrated high performance with an area under the curve (AUC) value of 0.861, and corresponding accuracy (ACC), sensitivity (SEN), and specificity (SPC) were 0.867, 0.898, and 0.727, respectively. The DL prediction process required approximately 2.5 min, representing a 5-fold improvement in efficiency over the existing ML method. Conclusions: The DL-based model could provide fast and accurate adaptive strategy selection, which further improves the efficiency of the MRgART process.