2990 - Characterization of Commercial AI-Assisted Auto-Segmentation for Delineating Visceral Organs In Pediatric Patients with CNS Malignancies Treated with Craniospinal Irradiation
Presenter(s)
R. P. Ermoian1, D. Maes2, D. Melancon1, J. Kang3, S. Cui3, J. Meyer1, F. Lew4, and S. R. Bowen3; 1Department of Radiation Oncology, University of Washington, Seattle, WA, 2Department of Radiation Oncology, University of Washington/ Fred Hutchinson Cancer Center, Seattle, WA, 3Department of Radiation Oncology, University of Washington/Fred Hutchinson Cancer Center, Seattle, WA, 4University of Washington, Seattle, WA
Purpose/Objective(s): Commercial deep learning-based (AI-assisted) auto-segmentation solutions have achieved promising accuracy and time-saving efficiency in adult patients. However, their clinical integration to pediatric populations can present unique challenges related to skeletal maturity and larger variability in anatomy.
Materials/Methods: This retrospective review included 57 pediatric patients at median age 8 years (1.6-17.7 years) with CNS disease without visceral organ disease or anatomic anomalies. Commercial AI auto-segmentation of planning CTs generated 28 visceral organ contours and a total of 1509 unique contours with median volume 7.3 cc (<0.03cc - 4016cc). AI contours were reviewed and manually modified to create clinically approved contours from a pediatric radiation oncologist. Dice similarity coefficients (DSCs) were computed to evaluate the agreement between AI contours and clinical contours. Spearman rank correlation between patient age and DSC, as well as between contour volume and DSC for various age groups (0-5 years [n=19], 5-10 years [n=14], 10-15 years [n=17], 15+ years [n=7]) was estimated.
Results: Distinct patterns of AI auto-segmentation accuracy emerged among various pediatric visceral organs: (i) strong agreement with clinical contours regardless of patient age, (ii) dissimilarity below a discrete threshold age, and (iii) dissimilarity correlating with decreasing age. High mean similarity between AI and clinical contours was noted for bowel bag (DSC 0.97), eyes (DSC 0.99), liver (DSC 0.99), and lungs (DSC 0.97-0.98). AI contouring performance demonstrated high variability among patients aged 1-5 years (DSC IQR 0.83-1.00), with comparatively lower variability observed in patients 15+ years (DSC IQR 0.99-1.00). Dissimilarity occurred for esophagus below a threshold age of 10 years (DSC > 0.92 vs. DSC < 0.56) and kidneys below 5 years (DSC > 0.89 vs. DSC < 0.75). AI contouring accuracy was strongly correlated with patient age for esophagus (R 0.84) and kidneys (R>0.67), while not correlated with age for liver, parotids, and corneas (R<0.1). AI contouring accuracy was strongly correlated with visceral organ volume for patients aged 5-10 years (R>0.8) compared to other age groups. Outlier analysis revealed AI contour limitations of esophagus systematically truncated superiorly / inferiorly compared to clinical contours, kidneys truncated superiorly, bowel bag extending into bony pelvis in young patients, spleens undercontoured cranially and medially, and challenges with contouring empty bladders.
Conclusion: Performance of commercial AI-contouring of visceral organs for pediatric patients with CNS malignancy was age- and organ-dependent, with distinct patterns that can guide clinical workflows. Future studies to validate adult versus pediatric-trained auto-contouring models are warranted.