Trade-off of different deep learning-based auto-segmentation approaches for treatment planning of pediatric craniospinal irradiation autocontouring of OARs for pediatric CSI.

Background: As auto-segmentation tools become integral to radiotherapy, more commercial products emerge. However, they may not always suit our needs. One notable example is the use of adult-trained commercial software for the contouring of organs at risk (OARs) of pediatric patients.

Purpose: This study aimed to compare three auto-segmentation approaches in the context of pediatric craniospinal irradiation (CSI): commercial, out-of-the-box, and in-house.

Methods: CT scans from 142 pediatric patients undergoing CSI were obtained from St. Jude Children's Research Hospital (training: 115; validation: 27). A test dataset comprising 16 CT scans was collected from the McGill University Health Centre. All images underwent manual delineation of 18 OARs. LimbusAI v1.7 served as the commercial product, while nnU-Net was trained for benchmarking. Additionally, a two-step in-house approach was pursued where smaller 3D CT scans containing the OAR of interest were first recovered and then used as input to train organ-specific models. Three variants of the U-Net architecture were explored: a basic U-Net, an attention U-Net, and a 2.5D U-Net. The dice similarity coefficient (DSC) assessed segmentation accuracy, and the DSC trend with age was investigated (Mann-Kendall test). A radiation oncologist determined the clinical acceptability of all contours using a five-point Likert scale.

Results: Differences in the contours between the validation and test datasets reflected the distinct institutional standards. The lungs and left kidney displayed an increasing age-related trend of the DSC values with LimbusAI on the validation and test datasets. LimbusAI contours of the esophagus were often truncated distally and mistaken for the trachea for younger patients, resulting in a DSC score of less than 0.5 on both datasets. Additionally, the kidneys frequently exhibited false negatives, leading to mean DSC values that were up to 0.11 lower on the validation set and 0.07 on the test set compared to the other models. Overall, nnU-Net achieved good performance for body organs but exhibited difficulty differentiating the laterality of head structures, resulting in a large variation of DSC values with the standard deviation reaching 0.35 for the lenses. All in-house models generally had similar DSC values when compared against each other and nnU-Net. Inference time on the test data was between 47-55 min on a Central Processing Unit (CPU) for the in-house models, while it was 1h 21m with a V100 Graphics Processing Unit (GPU) for nnU-Net.

Conclusions: LimbusAI could not adapt well to pediatric anatomy for the esophagus and the kidneys. When commercial products do not suit the study population, the nnU-Net is a viable option but requires adjustments. In resource-constrained settings, the in-house model provides an alternative. Implementing an automated segmentation tool requires careful monitoring and quality assurance regardless of the approach.