Summary: A new study showed that transfer learning models, trained on diverse public datasets and fine-tuned with institutional pediatric data, outperformed both internally trained models and TotalSegmentator.

Key Takeaways:

  • Transfer learning (TL) models, fine-tuned with institutional pediatric data after training on diverse public datasets, showed superior performance compared to internally trained models and TotalSegmentator (TS) in both internal and external pediatric test data.
  • The best-performing organ segmentation model, validated internally and externally, is now accessible as an open-source tool for further research and potential clinical implementation.
  • The study encompassed the development and validation of deep-learning models for liver, spleen, and pancreas segmentation using a substantial dataset of CT examinations, drawn from both internal institutional pediatric datasets and public datasets with various pathologies.

__________________________________________________________________________________________________________________________________

A recent study in the American Journal of Roentgenology found that transfer learning (TL) models, trained on diverse public datasets and fine-tuned with institutional pediatric data, surpassed both internally trained (NT) models and TotalSegmentator (TS) in performance on both internal and external pediatric test data. 

“The best-performing organ segmentation model, based on internal and external validation, is now available as an open-source application for further research and potential clinical deployment,” says first author Elanchezhian Somasundaram, PhD, a faculty member in the division of radiology at Cincinnati Children’s Hospital Medical Center.

In their AJR-accepted manuscript, Somasundaram and his colleagues developed and validated deep-learning models for liver, spleen, and pancreas segmentation using 1,731 CT examinations (1,504 training; 221 testing), derived from three internal institutional pediatric (age ≤ 18) datasets (n = 483) and three public datasets comprising pediatric and adult examinations with various pathologies (n = 1,248). Three deep-learning model architectures (SegResNet, DynUNet, SwinUNETR) from the Medical Open Network for AI (MONAI) framework underwent training using NT, relying solely on institutional datasets, and TL, incorporating pre-training on public datasets. For comparison, TS, a publicly available segmentation model, was applied to test data without further training.

Ultimately, the AJR authors’ optimal model outperformed models trained using only internal pediatric data, as well as a publicly available model. Acknowledging that segmentation performance was better in liver and spleen than in pancreas, “the selected model may be used for various volumetry applications in pediatric imaging,” Somasundaram and his colleagues conclude.