2406 - From Adults to Pediatrics: The Reliability and Reproducibility of AI Auto-Contouring in Pediatric Cancer Treatment

S. Sadeghi¹, A. C. Paulino², and R. M. Howell¹; ¹Department of Radiation Physics, The University of Texas MD Anderson Cancer Center, Houston, TX, ²Department of Radiation Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX

Purpose/Objective(s): This research aims to assess the performance of deep learning-based auto-segmentation tools on pediatric cancer patients. Specifically, we evaluated the reliability of an auto-segmentation software, trained on adult cases (20-60+ years), for the delineation of organs at risk (OARs) in a radiotherapy pediatric and adolescent patients aged 1-18 years. Materials/

Methods: A vender neutral platform (MIM, Cleveland, OH) was utilized to delineate OARs in non-contrast computed tomography scans for of 24 pediatric/adolescent patients, with tumor sites in the brain (10 cases), thorax (8), and abdomen/pelvis (6). We evaluated the auto-segmentation tools for these regions, which together included brain, heart, liver, kidneys, eyes, and lung OARs. Auto-Segmentation performance was quantitatively assessed by comparison with physician-approved manual (ground truth) contours using Dice similarity coefficient (DSC), Jaccard index (JI), average Hausdorff distance (AHD), relative absolute volume difference (RAVD), and Volume Correlation (VC).
Results: The mean ± standard deviation (range) DSC, JI, AHD, RAVD, and VC across all OAR in each region were: [A] brain: 0.93±0.05(0.82-0.99), 0.87±0.08(0.98-0.7), 0.64±0.28(0.13-1.55), 0.02±0.10(0.01-0.4), 0.99±0.01(0.97-0.99); [B] thorax: 0.94±0.02(0.86-0.97), 0.87±0.07(0.76-0.94), 1.55±0.58(0.07-2.74), 0.04±0.12(0.01-0.29), 0.96±0.09 (0.95-0.99); [C] abdomen/pelvis: 0.94±0.01(0.92-0.96), 0.89±0.03(0.85-0.92), 1.38±0.59(0.78-2.18), 0.01±0.03(0.01-0.04), 0.99±0.01(0.98-0.99), respectively. The standard deviation reflects minor variability, suggesting consistency across individual assessments. Across individual organs the ranges for DSC JI, AHD, RAVD, and VC were 0.92-0.99, 0.85-0.98, 0.47-1.63, 0.01-0.06, 0.98-0.99. Across all metrics for all organs the differences between manual and auto-segmentation were within the range of interobserver variability.
Conclusion: These preliminary results demonstrated for the OARs considered here that AI-based auto-segmentation tools trained on adult populations can be adapted with high reproducibility for pediatric cancer patients across various tumor sites. Further validation in a larger pediatric cohort is ongoing.

Evaluation Metric		DSC (%)	JI (%)	AHD (mm)	RAVD (%)	Volume Correlation (%)
Site	Brain	0.93±0.05	0.87±0.08	0.64±0.28	0.02±0.10	0.99±0.01
	Thorax	0.93±0.04	0.87±0.07	1.55±0.58	0.04±0.12	0.96±0.09
	Pelvic & Abdomen	0.94±0.01	0.89±0.03	1.38±0.59	0.01±0.03	0.99±0.01
OARs	Brain	0.99±0.01	0.98±0.01	0.47±0.15	0.02±0.01	0.99±0.01
	Heart	0.92±0.03	0.85±0.05	1.63±0.47	0.05±0.013	0.98±0.02
	Liver	0.95±0.01	0.90±0.03	1.40±0.56	0.02±0.03	0.99±0.01
	Kidneys	0.95±0.01	0.90±0.02	0.76±0.17	0.02±0.06	0.99±0.01
	Eyes	0.91±0.02	0.84±0.03	0.61±0.15	0.03±0.11	0.99±0.01
	Lungs	0.97±0.01	0.93±0.02	1.48±0.58	0.01±0.05	0.99±0.01