https://atlas.rsna.org/schemas/2025-11/model.json

Deep Learning Models for Abdominal CT Organ Segmentation in Children: Development and Validation in Internal and Heterogeneou...

https://github.com/cchmc-dll/pediatric_abdominal_segmentation_bundle.git

Elanchezhian.Somasundaram@cchmc.org

This HIPAA-compliant retrospective study was approved by the institutional review board. The requirement for participant written informed consent was waived.

This retrospective study developed and validated deep learning models for liver, spleen, and pancreas segmentation on pediatric CT examinations. The DynUNet TL model was selected as the best-performing model and made available as an open-source MONAI bundle.

Three deep learning model architectures (SegResNet, DynUNet, and SwinUNETR) from the Medical Open Network for Artificial Intelligence (MONAI) framework were trained. TotalSegmentator, based on nnU-Net architecture, was used for comparison. DynUNet TL model was selected as the best-performing.

Available as an open-source MONAI bundle on GitHub (https://github.com/cchmc-dll/pediatric_abdominal_segmentation_bundle.git).

The selected model may be used for various volumetry applications in pediatric imaging. Organ volumes are potential biomarkers of disease, objectively better quantitative representations of organ sizes, and may be more sensitive to change. It can facilitate image analysis for biomarker application and validation, disease monitoring, opportunistic screening, patient-specific dosimetry, nomogram generation, and input to computer-aided diagnostic models.

Fully automated segmentation of liver, spleen, and pancreas on CT images, significantly reducing time compared to manual methods (2.9 seconds vs. at least 15 minutes per organ).

Segmentation of liver, spleen, and pancreas on pediatric abdominal CT examinations for volumetry applications, disease monitoring, opportunistic screening, patient-specific dosimetry, and nomogram generation.

Abdominopelvic CT examinations, contrast-enhanced or non-contrast.

The model is available as an open-source MONAI bundle. Institutions applying the model should establish AI governance and monitoring infrastructure before deployment.

Internal pediatric test data did not include examinations with abnormalities of the liver or spleen. Exact distributions of pathologic conditions in public datasets were unknown. Models were compared only in terms of segmentation performance; further investigations are warranted to assess potential benefits or cost implications in real-world clinical settings.

The DynUNet TL model is available as an open-source MONAI bundle on GitHub, facilitating future research and potential clinical deployment.