New research demonstrates that attenuation-based biomarkers from computed tomography (CT) scans have a high capacity for predicting diabetes across multiple pancreatic segmentation algorithms.
For the retrospective study, recently published in Academic Radiology, researchers reviewed data from CT scans and HbA1c tests for 9,772 patients (average age of 56.1) to assess the prognostic capability of pancreatic imaging biomarkers — ranging from average attenuation to pancreatic volume — across three different pancreatic segmentation algorithms (TotalSegmentator, nnU-Net and DM-UNet).
The study authors found that the three algorithms had an average area under the receiver operating characteristic curve (AUC) of 87 percent for predicting diabetes. The algorithms also had average negative predictive value (NPV) of 92 percent and average specificity of 98 percent, according to the researchers.
The study findings also revealed that attenuation-based biomarkers on CT had a 93 percent interclass correlation coefficient (ICC) agreement across the pancreatic segmentation algorithms.
“Overall, we found that segmentation algorithms agreed well with respect to calculating imaging biomarkers that are dependent on attenuation measures on CT rather than shape. Furthermore, we found that diabetes prediction models trained on imaging biomarkers derived from the segmentation algorithms retained excellent overall agreement for classifying patients by diabetes status,” wrote lead study author Abhinav Suri, M.P.H., who is affiliated with the David Geffen School of Medicine at the University of California, Los Angeles, and the Imaging Biomarkers and Computer-Aided Diagnosis Laboratory at the National Institutes of Health in Bethesda, M.D., and colleagues.
In assessing the impact of contrast on the predictive capacity of the segmentation algorithms, the researchers noted a significant difference in AUC for the nnU-Net algorithm (73 percent AUC for contrast-enhanced CT scans vs. 62 percent AUC for non-contrast CT scans).
However, the study authors noted comparative AUCs for the TotalSegmentator (73 percent AUC for contrast CT scans vs. 71 percent for non-contrast CT scans) and DM-UNet models (80 percent AUC for contrast and non-contrast scans).
“ … We found that imaging biomarkers that were predictive of diabetes on non-contrast scans retained their predictive utility in the setting of contrast scans (at a different institution), highlighting that these biomarkers may be invariant to imaging characteristics,” added Suri and colleagues.
Three Key Takeaways
1. Strong predictive capability of segmentation algorithms.The three pancreatic segmentation algorithms demonstrated high predictive capability for diabetes, achieving an average AUC of 87 percent, specificity of 98 percent, and NPV of 92 percent.
2. Robust agreement for attenuation-based biomarkers. The study found a 93 percent interclass correlation coefficient (ICC) agreement for attenuation-based biomarkers across segmentation algorithms, whereas morphological measures like pancreatic volume and 3D fractal dimension showed significantly lower agreement (50 percent and 15 percent ICC, respectively).
3. Contrast Impact on Predictive Performance: Contrast-enhanced CT scans improved predictive performance for some models (e.g., nnU-Net had an AUC of 73 percent vs. 62 percent for non-contrast), but overall, imaging biomarkers retained predictive utility across different imaging settings, suggesting potential robustness to contrast variations.
The researchers acknowledged that sensitivity rates for the segmentation algorithms ranged between 43 to 59 percent.
In contrast to the robust agreement between the segmentation algorithms with respect to attenuation-based CT biomarkers, the researchers also noted significantly lower agreement for morphological measures. While pointing out that 3D fractal dimension was a key predictive factor for models originating from the three segmentation algorithms, the study authors found it only had a 15 percent ICC agreement across the algorithms. The researchers also pointed out a 50 percent ICC agreement between the algorithms on pancreatic volume.
(Editor’s note: For related content, see “What a Large CT Study Reveals About Potential Kidney Injury, Diabetes and Risk Stratification,” “Assessing MACE Risk in Women: Can an Emerging Model with SPECT MPI Imaging Have an Impact?” and “Can AI Enhance CT Detection of Incidental Extrapulmonary Abnormalities and Prediction of Mortality?”)
In regard to study limitations, the authors conceded that the absence of ground truth diagnoses for one of the study’s participating centers prohibited definitive confirmation of the date of diagnosis for type 2 diabetes in patients from this center. The researchers also acknowledged that contrast phase changes may lead to variation with chosen biomarkers utilized for contrast scans.
