2409 - Class-Based Automatic Contour Correction Method for MR-Guided Online Adaptive Radiotherapy on a MR-Linac

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

Purpose/Objective(s): MR-guided online adaptive radiotherapy (MRgOART) relies on fast and accurate contouring methods to adapt treatment plans to patient daily anatomies. Deep learning-based auto-segmentation (DLAS) provides fast initial contours but can require time-consuming manual edits. Previously we presented a population-based automatic contour correction (ACC) method that improves initial DLAS contours on unseen patient data but showed limited improvement for the duodenal contours. Utilizing patient-specific information in training DLAS-type models can improve contour accuracy but requires substantial effort offline to train separate models for each patient after each fraction. In this study, we present a class-based ACC (C-ACC) model that leverages the benefits of patient-specific models while reducing the workload for model training. Materials/

Methods: We trained the C-ACC model using first fraction images from a population subset of ten patients treated on a 1.5T MR-Linac with the goal of applying the model to the subsequent fractions of this class of patients to better improve the contours. Prior to training, the images were contoured manually to obtain “ground truth” contours and run through a research DLAS tool to obtain initial contours to be corrected. The images also underwent geometric and image quality augmentation using a previously developed tool that aims to model inter-fractional variations, resulting in 100 additional images for training. The C-ACC model was then applied to subsequent fractions of patients in the training class. We compared the dice similarity coefficient (DSC) and mean distance to agreement (MDA) between the contours from initial DLAS, the ACC model, and the C-ACC model.
Results: The model was tested on duodenal contours for seven fractions from three of the patients used in training the C-ACC. The average DSC for the seven fractions was (0.52, 0.58, 0.66) and the average MDA was (6.4, 4.8, 4.1) mm for the DLAS, ACC, and C-ACC, respectively. The metrics for the individual fractions are shown in Table 1. Five of the seven fractions tested showed highest DSC and lowest MDA using the C-ACC model.
Conclusion: By utilizing a small class of patients, one trained C-ACC model can be applied to all patients in the class, reducing the number of models that need to be trained while still offering the benefits of a patient-specific model. Table 1. Comparison of 3D contour accuracy metrics (DSC and MDA) for duodenal contours produced by DLAS, corrected by the population-based ACC, and corrected by the class-based ACC (C-ACC).

			DSC	MDA (mm)
Patient A	Fx 2	DLAS	0.61	5.0
		ACC	0.64	3.8
		C-ACC	0.57	3.8
Patient B	Fx 2	DLAS	0.48	7.5
		ACC	0.51	6.5
		C-ACC	0.77	2.5
	Fx 3	DLAS	0.68	3.8
		ACC	0.76	2.5
		C-ACC	0.82	1.8
Patient C	Fx 2	DLAS	0.59	4.4
		ACC	0.67	3.2
		C-ACC	0.73	2.8
	Fx 3	DLAS	0.32	9.2
		ACC	0.35	7.7
		C-ACC	0.55	5.8
	Fx 4	DLAS	0.50	8.4
		ACC	0.64	5.1
		C-ACC	0.77	5.0
	Fx 5	DLAS	0.48	6.5
		ACC	0.49	4.9
		C-ACC	0.41	6.7