355 - Toward Advancement of Pediatric Cancer Treatment: A Comparative Study of Pediatric-Specific Models and Adult-Trained Commercial Tools for Radiotherapy Auto-Contouring

Purpose/Objective(s): Accurate & efficient organ-at-risk (OAR) contouring is crucial in pediatric radiotherapy. Auto-contouring tools offer promising solution to reduce time-intensive manual contouring. However, current commercial platforms are predominantly trained on adult datasets, compromising performance in pediatric patients due to distinct anatomical characteristics & smaller organ volumes. Study investigates performance of pediatric-trained deep learning models compared to adult-trained commercial platforms for radiotherapy auto-contouring in pediatric CT scans.

Materials/Methods: Total 276 non-contrast pediatric CT scans (ages 1–20 years, 151 male, 125 female) from patients treated btw 2000-2022 collected as part of multi-institutional Childhood Cancer Survivor Study (CCSS) pilot study. Datasets included scans of abdomen/pelvis (n=70), thorax (n=92), brain (n=114). 3 deep learning models (nnUNet, SwinUNeTr, 3DSegResNet) were trained on this pediatric cohort using 5-fold cross-validation strategy. Performance evaluated against 2 adult-trained commercial platforms (MiM and treatment planning system) for structure-specific contouring accuracy, utilizing Dice Similarity Coefficient (DSC) and Average Hausdorff Distance (AHD).

Results: Pediatric models demonstrated superior auto-contouring performance compared to adult-trained platforms across all structures (mean DSC values: nnUNet 89.53% (±3.15%), SwinUNeTr 87.14% (±2.68%), and 3DSegResNet 88.58% (±3.79%), compared to MiM 75.66% (±17.23%) and treatment planning system 77.20% (±17.19%)). Mean AHD values were lower for pediatric models: nnUNet 0.83 mm (±0.20 mm), SwinUNeTr 1.05 mm (±0.51 mm), and 3DSegResNet 1.14 mm (±0.59 mm), versus MiM 2.83 mm (±5.69 mm) and a treatment planning system 1.01 mm (±0.23 mm). Pediatric models yielded consistence performance across different anatomical sites. In the Abdomen/Pelvis region, pediatric models (nnUNet DSC: 89.46% ±3.23%; SwinUNeTr DSC: 87.29% ±2.48%; 3DSegResNet DSC: 88.61% ±3.37%) outperformed adult-trained platforms (MiM DSC: 79.41% ±15.36%; a treatment planning system DSC: 81.07% ±15.38%).

Conclusion: Deep learning models trained on pediatric cohort improved auto-contouring accuracy compared to adult-trained commercial tools across 23 anatomical structures. Findings underscore critical need for implementing pediatric-specific AI models in clinical workflows to enhance precision & efficiency of pediatric radiotherapy planning.