Accurate screening of patients with diabetes based on physical examination: pancreas imaging features from chest CT and laboratory data

Abstract

Background: Fasting plasma glucose (FBG) was used in the large-scale primary screening of diabetes mellitus (DM). However, some people who actually have DM have normal fasting glucose (NFG), which shows a high false negative rate.

Objective: To investigate the application value of pancreatic radiomics features and laboratory data from chest CT scan in screening DM patients.

Methods: Patients with DM were diagnosed according to HbA1c≥48 mmol/mol(6.5%). The radiomics features of pancreas in lung window (L) and soft tissue window (S) of chest CT and laboratory data of 3587 patients from D1 (model training and testing) and D2/D3/D4/D5 (external validation) were retrospectively analyzed to construct a diagnostic model for DM screening.

Results: The AUC of the lung window-laboratory models (L-Lab-LR and L-Lab-SVM) in the test set were 0.958/0.965. The AUC of soft tissue windows-clinical models (S-Lab-LR and S-Lab-SVM) were 0.958/0.969, 0.875/0.881, 0.935/0.959, 0.905/0.919, respectively.

Conclusions: The diagnostic model based on chest CT pancreatic radiomics features and laboratory data has a very high diagnostic efficiency, which is accurate and reliable for the initial screening of DM patients. This method will facilitate early diagnosis of those individuals with DM who may have been missed.

Keywords: CT; diabetes mellitus; diagnostic model; fasting blood glucose; pancreas.

Figure 1

Flowchart of data screening.

Figure 2

(A–C) show the ROI of the pancreatic head-neck, neck-body, and body-tail delineated by the soft tissue window of the same subject, respectively. (D–F) are the ROI of the lung window registered by the automatic registration function of the software, respectively. A total of six samples were obtained for one subject.

Figure 3

Simple process of data analysis and model construction: First, extract the image features from the CT image. Then, screen out the optimal model features. Build the model based on the selected optimal features. Finally, evaluate the diagnostic efficacy of the model.

Figure 4

The importance map of the best radiomics features. (A) shows the best clinical feature from the clinic laboratory. (B, C) respectively display the optimal radiomics features of the pancreas in the lung window and soft tissue window of chest CT. (D, E) respectively present the joint radiomics features of the pancreas in the lung window and soft tissue window of chest CT.

Figure 5

(A, B) are correlation graphs of predict_score and optimal laboratory features of the lung window and the soft tissue window, respectively. The correlation between different features ranged from 0.09 to 0.56 and from 0.14 to 0.56. This indicates that there is no collinearity or only weak collinearity between the features.

Figure 6

(A) stands for single-laboratory clinical model. (B, C) are ROC curves of the training set and test set of the four joint diagnostic models (L-Lab-LR with clinic, L-Lab-SVM with clinic, S-Lab-LR with clinic, and S-Lab-SVM with clinic) trained by D1(medical A), respectively.

Figure 7

(A–D) are calibration curves of training sets and test sets of four different diagnostic models, L-Lab-LR, L-Lab-SVM, S-Lab-LR and S-Lab-SVM, respectively. The closer the fit of the calibration curve of the diagnostic model to the true curve (dashed line), the better the prediction performance. It can be seen that the L-Lab-LR and S-Lab-LR models have better predictive performance, while the L-Lab-SVM and S-Lab-SVM models have slightly poorer predictive performance.

Figure 8

Cluster heatmap [(A, B) show CT lung window-laboratory characteristics and CT soft tissue window-laboratory characteristics respectively] uses different colors to reflect the value of each one-dimensional feature (row) of the sample (column). The top color bar expresses the true class of the samples, the left color bar expresses the class of the features, and the dendrogram shows the results of the hierarchical clustering. (1=DM, 0=non-DM).

Figure 9

(A, B) are nomograms constructed by L-Lab-LR and S-Lab-LR diagnostic models, respectively. L-Lab-LR Rad Score = -1.964×clinic_fibronectin -1.613×clinic_albumin +0.975×predict_score -0.742×clinic_Lymphocytes2 + 0.474. S-Lab-LR Rad Score = -1.942×clinic_fibronectin -1.649×clinic_albumin +0.745×predict_score -0.741×clinic_Lymphocytes2 + 0.393.