Development and evaluation of automated organ segmentation models for abdominal ultrasound in injured children

Background: Accurate organ segmentation in pediatric focused assessment with sonography in trauma (FAST) is an important prerequisite for automated image analysis and image quality assessment, but manual segmentation is time-intensive, subject to inter- and intra-observer variability, and limits the scalability of deep learning dataset development. Although deep learning has shown promise for ultrasound organ segmentation, its application to pediatric FAST has not been evaluated. Objective: To describe the development and evaluation of deep learning models to automatically localize and segment six abdominal organs in pediatric FAST. Methods: We conducted a cross-sectional study using a convenience sample of archived FAST examinations from 200 injured children evaluated at a Level I pediatric trauma center. Emergency ultrasound fellowship-trained physicians manually segmented six key abdominal organs. Annotated image-mask pairs were split into training and test sets at 80:20. We developed or fine-tuned U-Net, EVIT U-Net, UltraSAM, and nnU-Net models to automatically segment the bladder, kidney, liver, spleen, spine, and diaphragm. Segmentation accuracy was assessed using the Dice-Sørensen Coefficient (DSC) compared with manual annotations. Pediatric emergency ultrasound experts reviewed previously unlabeled FAST clips and rated the acceptability of model-generated organ masks. Results: The dataset included 1,807 manually annotated image-mask pairs from 200 pediatric FAST clips. Using an 80:20 split with no overlap between video clips, 1,446 image-mask pairs were used for training and 361 for testing. Across all models and organs, mean DSC was 0.79 (95% CI: 0.75-0.84). Mean DSC was highest for the kidney and bladder, each 0.91 (95% CI: 0.87-0.95), and lowest for the diaphragm at 0.63 (95% CI: 0.60-0.67). EVIT U-Net achieved the highest DSC for kidney, liver, and spleen, performed comparably to the highest-performing models for bladder, and had a mean inference time of 54 milliseconds per image. In previously unlabeled pediatric FAST clips, experts rated model-generated organ masks as acceptable for localization in 73%-95% of clips across organs. Conclusion: We describe the automatic segmentation of key abdominal organs in pediatric FAST with limited annotated data. This represents an important advancement in our image processing pipeline and can accelerate the development of large, robust datasets with segmentation for anatomy and pathology. Experts found that model-generated masks provided acceptable localization in most previously unlabeled pediatric clips, supporting the potential use of automated segmentation in pediatric FAST image analysis.