2457 - Multiplexed Deformable Contour Propagation and Deep Learning Auto Segmentation Using Automatic Contour Quality Assurance for Abdominal MR-Guided Online Adaptive Radiotherapy

M. Zarenia, Y. Zhang, R. Conlin, C. Sarosiek, A. Amjad, and E. S. Paulson; Department of Radiation Oncology, Medical College of Wisconsin, Milwaukee, WI

Purpose/Objective(s): Fast and accurate auto-segmentation is crucial for MR-guided online adaptive radiation therapy (MRgOART). While Deep Learning Auto-Segmentation (DLAS) methods are promising for MRgOART, current DLAS algorithms do not always generate clinically acceptable contours. Additionally, for specific abdominal organs at risk (OARs), deformable contour propagation (DCP) from prior fractions using deformable image registration (DIR) may yield contours comparable to DLAS. This study aims to combine both DCP and DLAS approaches by evaluating them using a newly developed organ-specific 3D Deep Learning-based Automatic Contour Quality Assurance (ACQA) model to improve abdominal OAR contours for MRgOART. Materials/

Methods: The ACQA models, based on a 3D convolutional neural network (CNN), were trained for duodenum and stomach contours obtained from a research DLAS tool. DLAS was performed on abdominal MRIs acquired during routine MRgOART on a 1.5T MR-Linac using either turbo field-echo or balanced turbo field-echo sequences. The DCP contours were propagated from a different daily image, using an intra-patient DIR contour propagation research tool. The training dataset contained abdominal MR images, DLAS contours, and their corresponding quality ratings, from 103 datasets. The quality of DLAS and DCP contours was determined using an in-house contour classification pipeline, which categorized contours as acceptable or edit-required based on the expected editing effort. Through the DCP/DLAS/ACQA workflow, a DCP contour was replaced by the corresponding DLAS contour whenever the DLAS contour received an "acceptable" prediction from the ACQA model. The performance of the workflow was tested on an independent dataset consisting of 17 sets of abdominal MRIs.
Results: By multiplexing both the DCP and DLAS contours using the ACQA model, the accepted DLAS and DCP contours were improved from 32% and 35% to 40% for stomach. For duodenum, the accepted DLAS and DCP contours were improved from 22% and 27% to 35%, after the DCP/DLAS/ACQA workflow. The developed ACQA models predicted acceptable and edit-required contours at 89.5% and 63% accuracy for duodenum and at 60.5% and 64.5% for stomach contours, respectively.
Conclusion: Multiplexing DCP and DLAS contours using the 3D DL-based ACQA models can improve the quality of auto-segmented duodenum and stomach contours, minimizing subsequent manual editing time required for MRgOART. In addition, the developed ACQA models can be integrated into automatic contour correction processes and can also be extended for evaluation of DLAS contours generated by various vendors or different contour propagation techniques.