Automated Diagnosis and Severity Quantification of Diabetes on Abdominal CT Scans

Purpose: Diabetes affects millions of individuals in the United States each year. However, it is estimated that 17.8% of people who have diabetes are undiagnosed. Early diagnosis of diabetes is key to preventing complications that are associated with billions in healthcare expenditure and increased mortality. Deep learning methodologies have enabled scans to be used for detecting diagnoses, and prior work has shown how algorithms can diagnose diabetes from Computed Tomography (CT) images. However, no work has estimated Hemoglobin A1c (proxy for severity of diabetes) from scans. The purpose of this study is to use deep learning on abdominal CT scans to estimate A1c therefore allowing both diagnosis of diabetes and quantification of severity.

Methods/Materials: We examined a cohort of patients (2005-2023) who had a CT abdomen and an A1c result within 3 months of the scan (822 scans, 580 patients, 68.6% female, A1c 8.2±2.5(±1SD)%). A convolutional neural network (DenseNet121 architecture) was trained to predict A1c using segmentations of the pancreas and liver (automatically generated from the TotalSegmentator algorithm) as input. Learned embeddings from the network were combined with 13 imaging biomarkers automatically extracted from the CT scan (derived from liver, pancreas, muscle, and atherosclerotic plaque) via a Light Gradient Boosting machine learning model to predict a final A1c. Figure 1 shows an overview of the architecture. The algorithm was tested on scans from 290 patients (1 scan/patient).

Results: The model yielded a mean absolute error in A1c of 0.49±0.40(±1SD)% with a Pearson's R of 0.74 (p <.001, Figure 2). The Bland-Altman plot showed a mean difference of 0.25±1.14(1.96SD). Area Under the Receiver Operator Curve was 0.84±0.12.

Conclusions: Deep learning on abdominal CT scans can accurately detect and quantify the severity of diabetes on prior imaging, enabling clinicians to screen individuals for an A1c blood test and formal diabetes diagnosis. Further testing on a larger patient population is needed to determine this algorithm’s performance across all A1c ranges as our dataset contained individuals with higher A1cs.