AI Segmentation Models Show Radiomic Feature Biases

New research revealed significant discrepancies between manual and automated segmentation in pediatric medical imaging.

Updated on Oct. 14, 2026 in Artificial Intelligence

AI Segmentation Models Show Radiomic Feature Biases

Live Poll

Do you trust automated medical diagnostic tools as much as manual procedures performed by professionals?

Researchers at the Cincinnati Children's Hospital Medical Center found that radiomic features derived from deep-learning segmentations often lack agreement with manual methods. This analysis of pediatric abdominal scans suggests that current automated tools may introduce bias into diagnostic workflows.

Why it matters

The findings highlight the challenge of achieving consistency in medical AI, as different segmentation models can significantly alter the radiomic data used for patient assessment. Understanding these variations is essential for ensuring that automated diagnostic tools remain reliable and generalizable across clinical settings.

The study evaluated MRSegmentator, TotalSegmentator, PaNSegNet, and NV-Segment-CTMR models across 203 total imaging sequences. Researchers observed poor agreement in up to 32 features, with mean accuracy differences reaching 0.12 for specific models.

The players

Cincinnati Children's Hospital Medical Center

This is a pediatric research and clinical institution based in Ohio where the study on radiomic segmentation was conducted.

The details

The team compared radiomic features from 139 pediatric abdominal CT exams and 64 pediatric and young adult MRI sequences. By contrasting manual segmentations against three automated deep-learning models, researchers identified consistent biases and limited feature agreement across the tested software.

The research was published in the American Journal of Roentgenology on October 14, 2026.

Elections
Sept. 28
Fairshake PAC Targeted Sherrod Brown Senate Campaign

The Tech Race

This study underscores the push for standardized validation in AI-driven diagnostics as the healthcare industry pivots toward automated image analysis. It highlights how these new models must clear significant hurdles in generalizability before replacing traditional manual segmentation workflows.

For patients and clinicians, these findings emphasize that AI tools are not yet perfectly interchangeable and require careful validation before routine clinical use. Users of diagnostic software should be aware that automated segmentations may produce different feature outputs than human experts.

The takeaway

Diagnostic accuracy depends heavily on the specific segmentation methodology chosen for medical imaging tasks. Clinicians should maintain oversight of automated outputs until these tools demonstrate greater consistency and lower bias across diverse pediatric datasets.