AI Radiomic Segmentation Methods Showed Variable Accuracy
Research suggests automated segmentation models currently struggle to match manual performance in pediatric abdominal imaging.
Updated on Oct. 14, 2026 in Artificial Intelligence
Live Poll
Do you trust automated medical diagnostic tools as much as manual procedures performed by professionals?
A study published on October 14, 2026, in the American Journal of Roentgenology found that radiomic features extracted from deep-learning segmentation models exhibit limited agreement with manual methods. Researchers evaluated four common models, observing consistent biases that affect the reliability of automated analysis in clinical imaging.
Why it matters
Understanding how segmentation methodologies impact radiomics is critical for ensuring the generalizability of AI-driven diagnostic tools. Discrepancies in feature extraction could lead to inconsistent clinical outcomes if automated models are used as direct replacements for manual review.
Automated models displayed significant feature divergence, with MRSegmentator and TotalSegmentator yielding 30 and 32 features showing poor agreement, respectively. Mean accuracy differences reached 0.12 for PaNSegNet compared to manual results.
The players
Cincinnati Children's Hospital Medical Center
A pediatric academic medical center in Ohio that conducts research on medical imaging and AI applications.
The details
The research team at Cincinnati Children's Hospital Medical Center compared manual segmentations—the process of outlining regions of interest in an image—against outputs from MRSegmentator, TotalSegmentator, PaNSegNet, and NV-Segment-CTMR. Radiomic features, or quantifiable mathematical data extracted from medical images, were compared across both pediatric CT and MRI datasets to detect algorithmic bias. The study found that automated models, which replace human-led contouring with deep-learning predictions, introduced systematic variations in output values.
Timeline
October 14, 2026: The study was published in the American Journal of Roentgenology.
The Tech Race
This research follows a growing trend of validation studies seeking to standardize AI-driven diagnostic tools in radiology. It marks a departure from purely performance-focused development, emphasizing the need for feature-level agreement across competing models like TotalSegmentator and PaNSegNet.
Clinicians and imaging departments should remain cautious when adopting deep-learning segmentation for diagnostic radiomics. These findings suggest that manual validation is still required for high-stakes interpretations until model biases are addressed in future software updates.
The takeaway
Reliability in radiomics requires human oversight until automated models can mitigate their existing biases. Readers should watch for future model iterations that attempt to align feature extraction with standardized manual segmentation benchmarks.
Further reading
For broader context on how research teams in the region are evaluating diagnostic algorithms, see Artificial Intelligence.
Source note: This article includes information reported by AuntMinnie.
Live Poll
Do you trust automated medical diagnostic tools as much as manual procedures performed by professionals?
