Machine learning prediction of Dice similarity coefficient for validation of deformable image registration

• Main Authors: Wong, Yun Ming, Yeap, Ping Lin, Ong, Ashley Li Kuan, Tuan, Jeffrey Kit Loong, Lew, Wen Siang, Lee, James Cheow Lei, Tan, Hong Qi
• Other Authors: School of Physical and Mathematical Sciences
• Format: Article
• Language: English
• Published: 2024
• Subjects: Medicine, Health and Life Sciences; Physics; Deformable image registration; Adaptive radiotherapy
• Online Access: https://hdl.handle.net/10356/181833
• Institution: Nanyang Technological University

Introduction: Deformable image registration (DIR) plays a vital role in adaptive radiotherapy (ART). For the clinical implementation of DIR, evaluation of deformation accuracy is a critical step. While contour-based metrics, for example Dice similarity coefficient (DSC), are widely implemented for DIR validation, they require delineation of contours which is time-consuming and would cause hold-ups in an ART workflow. Therefore, this work aims to accomplish the prediction of DSC using various metrics based on deformation vector field (DVF) by applying machine learning (ML), in order to provide an efficient means of DIR validation with minimised human intervention. Methods: Planning CT image was deformed to the cone-beam CT images for 20 prostate cancer patients. Various DVF-based metrics and DSC were calculated, and the former was used as input features to predict the latter using three ML models, namely linear regression (LR), Nu Support Vector Regression (NuSVR) and Random Forest Regressor (RFR). Four datasets were used for analysis: 1) prostate, 2) bladder, 3) rectum and 4) all the organs combined. Average mean absolute error (MAE) was computed to evaluate the model performance. The classification performance of the best-performing model was further evaluated, and the prediction interval and feature importance were calculated. Results: Overall, RFR achieved the lowest average MAE, ranging between 0.045 and 0.069 for the four datasets, while LR and NuSVR had slightly poorer performances. Analysis on the results of best-performing model showed that sensitivity and specificity of 0.86 and 0.51, respectively, were obtained when a prediction threshold of 0.85 was used to classify the fourth dataset. Jacobian determinant was found to be a significant contributor to the predictions of all four datasets using this model. Conclusion: This study demonstrated the potential of several ML models, especially RFR, to be applied for prediction of DSC to speed up the DIR validation process.