ARRS 2022 Abstracts


1664. Radiomics-Based Machine Learning (ML) Classifier for Detection of Type-2 Diabetes Mellitus (T2DM) on Standard-of-Care Abdomen CTs
Authors * Denotes Presenting Author
  1. Darryl Wright *; Mayo Clinic
  2. Anurima Patra; Mayo Clinic
  3. Garima Suman; Mayo Clinic
  4. Hala Khasawneh; Mayo Clinic
  5. Panagiotis Korfiatis; Mayo Clinic
  6. Timothy Kline; Mayo Clinic
  7. Ajit Goenka; Mayo Clinic
Globally, around 185 million people living with T2DM have silent disease. Our hypothesis was that T2DM-associated pancreatic imaging phenotype could be characterized with quantitative radiomics. Our purpose was to determine if CT radiomics-based ML classifier could differentiate patients with T2DM versus control subjects.

Materials and Methods:
Volumetric pancreas segmentation using 3D Slicer was done on a dataset [portal venous CTs, slice thickness (ST) = 5mm] with morphologically normal pancreas in 422 patients with T2DM (mean age=63 years; male: female=0.8:1) and 456 controls (mean age=64 years; male: female=1.1:1). Data were randomly split into training (n=600; 300 each of T2DM and controls) and test sets (n=278; 122 T2DM, 156 controls). Pre-processing included: modified soft tissue window (level 50 HU, width 500 HU); rescaling to range 0-255; resampling to 0.75 x 0.75 x 2-mm (in-plane x ST). Using a bin width of 25,107 features (shape, first order, and texture features) were extracted using PyRadiomics. Multiple feature selection methods were tested. Four classifiers: logistic regression, SVMs, random forests, and XGBoost - were built using selected features. Parameters were chosen through a 3-fold cross-validation grid search on the training set. Model with the highest performance was evaluated on the independent test set.

Mean (range) age of T2DM and control cohorts in the test set was 63 (32-90) and 64 (35-89) years, respectively. Mean (range) T2DM duration was 4.6 (0-15.7) years. Patients with T2DM were on insulin (n=16, 13%), oral antidiabetics (n=25, 20.5%), both (n=49, 40%), or no medications (n=32, 26%). In cross validation, XGBoost model (n_estimators=100, gamma=0.1, max_depth=4, lambda=10, learning_rate=0.01), which was based on the top 10 features selected according to the ANOVA F-value, had the highest performance [accuracy: 0.673 (0.012), area under the curve (AUC) 0.691 (0.025)]. On the test set, the model correctly classified 73 (60%) T2DM CT scans and 97 (62.2%) controls, which translated to accuracy and AUC of 0.61 (0.02) and 0.65 (0.014), respectively. Model’s performance was similar across genders (p=0.6), CT STs (accuracy 58% on ST <3 mm vs 64% on ST >3 mm; p=0.4) and vendors (accuracy 59.2% on CTs from Siemens vs 76% on CTs from others; p=0.1). Mean (range) T2DM duration (in years) was similar for correctly classified [4.5 (0-15.4)] vs misclassified CTs [4.8 (0-15.7)] (p=0.8). There was no difference in the distribution of correctly classified vs. misclassified T2DM CT scans with regards to type of antidiabetic medication: insulin (22% vs. 8%), oral antidiabetics (9% vs 18%), both (41% vs 39%) (p=0.6). Mean (range) duration (in years) of antidiabetic medications was also similar for correctly classified [5.4 (0-15)] vs misclassified CTs [5 (0-13)] (p=0.4).

ML classifier based on radiomics extracted from a morphologically normal pancreas could differentiate patients with T2DM vs unaffected subjects. Prospective validation on enriched multicenter datasets could offer the prospect of opportunistic CT detection of undiagnosed or unsuspected T2DM.