2786 - Auto-Contouring and Observer Variations in MR-Guided Brachytherapy for Cervical Cancer

10:45am - 12:00pm PT

Hall F

Screen: 3

POSTER

Presenter(s)

Chun Him Frederick Law, MS - Hospital Authority, Hong Kong, Hong Kong

N. T. C. Fung, C. H. F. Law, E. S. N. Cheung, and P. Y. Wu; Pamela Youde Nethersole Eastern Hospital, Chai Wan, Hong Kong

Purpose/Objective(s): To evaluate the time required for target and OAR contouring with or without auto-contouring assistance and quantify geometric and dosimetric impacts of auto-contour errors (A-E), intra-/interobserver variations (Intra-V/Inter-V) for MR-guided brachytherapy (IGBT) in cervical cancer.

Materials/Methods: A self-trained commercial deep learning model was evaluated on 14 clinical retrospective cases. Targets and OARs were re-contoured following the EMBRACE-II protocol, generating auto-contours with manual edits (E-C) and all-manual contours (M-C, reference standard). E-C and M-C were created by the same oncologist, different from the original clinical contours (C-C). Geometric accuracy was assessed using DSC, (1) A-E: auto-contours (A-C) vs. M-C, (2) Intra-V: E-C vs. M-C, and (3) Inter-V: C-C vs. M-C. Dosimetric impacts were evaluated by optimizing dose on A-C, E-C, and C-C, then comparing the planned dose to the same dose distribution on M-C for A-E, Intra-V, and Inter-V. Contouring times for A-C, E-C, and M-C were recorded. Statistical analyses used the Wilcoxon signed-rank test (p < 0.05).

Results: The auto-contouring model performed well, with DSC for A-E comparable to inter-V for most structures except the sigmoid and small bowel. Dosimetric evaluation demonstrated small differences for bladder D2cc, and no significant difference for other structures. A-E, intra-V, and inter-V tended to underdose targets (highest for GTV, followed by HRCTV) and overdose OARs. Intra-V variations were minor, and significantly less than A-E and inter-V for most structures. Nevertheless, dose impact on GTVD98% appears considerable despite a DSC of 0.89. Median contouring times for A-C, E-C, and M-C were 0.6, 37, and 46 min, respectively. E-C significantly reduced contouring time compared to M-C (p = 0.002), with 50% of patients saving >30% time.

Conclusion: Geometric and dosimetric impact of A-E were comparable to inter-V but must be interpreted cautiously. Careful editing of auto-contours is necessary, especially for targets, as clinical findings and initial disease extent are critical. Even with edits, auto-contouring could improve the efficiency of IGBT workflows. *Statistically significant; ?D +ve = Planned dose < Dose on M-C

Abstract 2786 - Table 1

Structure	Parameter	A-E		Intra-V		Inter-V		A-E vs Intra-V	A-E vs Inter-V	Intra-V vs Inter-V
Structure	Parameter	Median	SD	Median	SD	Median	SD	A-E vs Intra-V	A-E vs Inter-V	Intra-V vs Inter-V
GTV	DSC	0.76	0.13	0.89	0.07	0.77	0.19	*		*
GTV	?D98%(Gy)	-0.9	1.4	-0.4	0.6	-0.8	2.7
HRCTV	DSC	0.87	0.05	0.95	0.03	0.89	0.04	*		*
	?D90%(Gy)	-0.3	0.7	0.1	0.3	-0.2	0.5	*		*
	?D98%(Gy)	-0.5	0.7	0.0	0.5	-0.3	0.6	*		*
IRCTV	DSC	0.88	0.04	0.96	0.04	0.87	0.04	*		*
IRCTV	?D98%(Gy)	-0.5	0.5	0.0	0.2	0.0	0.4	*
Bladder	DSC	0.94	0.03	0.98	0.02	0.95	0.02	*		*
Bladder	?D2cc(Gy)	0.4	0.3	0.0	0.2	0.2	0.4	*	*
Rectum	DSC	0.86	0.07	0.93	0.06	0.85	0.06	*		*
Rectum	?D2cc(Gy)	0.1	0.2	0.0	0.1	0.1	0.1	*		*
Sigmoid	DSC	0.69	0.10	0.94	0.04	0.81	0.10	*	*	*
Sigmoid	?D2cc(Gy)	0.0	0.6	0.0	0.2	0.1	0.2
Small bowel	DSC	0.65	0.13	0.87	0.07	0.80	0.10	*	*	*
Small bowel	?D2cc(Gy)	-0.1	1.0	0.0	0.2	0.0	0.4