Impact of deep learning on CT-based organ-at-risk delineation for flank irradiation in paediatric renal tumours: a SIOP-RTSG radiotherapy committee study

Abstract

Background and purpose: Integrating deep learning (DL) for auto-contouring has significantly improved organ-at-risk (OAR) delineation in adult radiotherapy. However, its application in paediatric radiotherapy remains limited. This study evaluates DL-based auto-contouring of OARs, followed by manual revisions, for paediatric flank irradiation, focusing on delineation time, accuracy, and inter-observer variability (IOV). Materials and methods: Twelve paediatric radiation oncologists from nine countries affiliated with the SIOP Renal Tumour Study Group participated in a two-day workshop. Participants were randomly divided into two groups: one performed manual delineation first, followed by DL-based revision, while the other group performed in reverse order. Eight thoracoabdominal OARs were delineated on non-contrast CTs of renal tumour patients (ages 1–6). DL-based contours were generated using a model for paediatric abdominal cases. Delineation time was recorded, accuracy and IOV were assessed using the Dice similarity coefficient (DSC), 95th percentile Hausdorff distance, mean surface distance against a STAPLE consensus (threshold = 0.95), and an expert reference. Results: In total, 122 manual delineations and 254 DL-based revisions were collected. DL-based auto-contouring reduced delineation time by 59 %, from 25.5 to 10.2 min. The mean DSC of all eight OARs improved from 0.91 to 0.97 using STAPLE reference and from 0.89 to 0.93 using expert reference. The pancreas exhibited the largest gain, with mean DSC increases ranging from 0.18 to 0.25. Delineation accuracy was significantly improved for seven OARs (p < 0.05), while IOV significantly decreased for the pancreas and heart in both references (p < 0.05). Conclusion: Manually revising DL-based auto-contouring reduces delineation time, enhances accuracy, and reduces inter-observer variability in paediatric CT-based OAR delineation.