2419 - Auto-Contouring of Normal Organs Using AI and Intensity-Based Deformable Image Registration Algorithms: Opportunities and Caveats

B. Stiehl, J. L. Dolan, E. McKenzie, J. Steers, J. DeMarco, and I. J. Chetty; Department of Radiation Oncology, Cedars-Sinai Medical Center, Los Angeles, CA

Purpose/Objective(s): Deformable image registration (DIR) algorithms are commonly employed to derive the deformation vector fields (DVFs) for contour propagation. AI-based methods have recently been implemented for the auto-contouring of critical structures. Due to the reliance of DIR-based methods on the accuracy of registration, we hypothesized that AI contouring may be less prone to error in scenarios where DIR has traditionally underperformed. The goal was to assess the geometric accuracy of organ-at-risk (OAR) contours from DIR-based propagation methods against a deep-learning auto-contouring algorithm for multiple anatomies and to evaluate caveats for each algorithm type. Materials/

Methods: Twelve patients, 4 each in the abdomen, pelvis, and head/neck, having CT-simulation and PET/CT image datasets were identified, retrospectively. DIR was performed between CT-sim and PET/CT volumes using Demons and multi-resolution, B-spline algorithms, and PET/CT-defined contours were deformed according to the resulting DVFs. Each OAR was then automatically contoured on CT-sim images using a deep-learning algorithm. Quantitative analysis of geometric accuracy compared AI-generated and DIR-propagated contours to ground truth, manual, contours. Similarity metrics calculated included the Dice Similarity Coefficient (DSC) and mean distance-to-agreement (MDA).
Results: For DIR-based contour propagation, using Demons and B-spline algorithms, mean DSC values were 0.71 ± 0.20 and 0.70 ± 0.17 and mean MDA values were 4.55 ± 1.63 and 4.56 ± 1.45 mm, respectively. For AI-generated contours, the mean DSC value was 0.77 ± 0.16 and the mean MDA value was 4.23 ± 1.51 mm. The AI algorithm performed better than DIR-based propagation on average, and especially in scenarios where the deformation magnitude was large, CT datasets were truncated, or soft tissue contrast was low. For patients who received contrast during imaging, all algorithms performed poorly. This is likely due to the limited number of contrast datasets in the AI training data.
Conclusion: Contours generated by an AI application showed good agreement with ground-truth contours and outperformed DIR algorithms in a quantitative analysis. AI-based auto-contouring may offer advantages over DIR-based contour propagation in cases where DIR is limited, however, AI algorithms must incorporate robust training datasets to perform accurately under the spectrum of clinical scenarios. Table 1: Mean similarity metric results for DIR-propagated and AI-generated contours

	DSC			MDA (mm)
Abdomen
Structure	B-spline	Demons	AI	B-spline	Demons	AI
Liver	0.93	0.90	0.96	7.54	7.32	7.41
Stomach	0.72	0.77	0.86	4.47	5.82	6.04
Duodenum	0.57	0.58	0.48	5.34	5.51	4.27
Pelvis
Bladder	0.65	0.75	0.88	5.16	4.92	4.70
Rectum	0.60	0.64	0.80	3.93	3.92	3.26
Sigmoid Colon	0.45	0.52	0.62	3.83	3.95	3.76
Head & Neck
Lt Submandibular	0.70	0.67	0.81	3.14	3.10	3.05
Rt Submandibular	0.73	0.65	0.86	2.44	2.55	2.25
Larynx	0.80	0.85	0.78	4.19	4.18	3.93
Mandible	0.84	0.77	0.89	4.20	4.21	3.62