Automatic measurement of mesenteric vascular and portal vein parameters via PE-NET in the diagnosis of Crohn's disease.

OBJECTIVE

Vascular changes are concomitant of the course of Crohn's disease (CD). In this study, we evaluated the value of the parallel encoder network (PE-NET) method for the automated measurement of mesenteric vascular and portal vein parameters and explored the performance of PE-NET combined with support vector machine (SVM) classifier in CD diagnosis.

METHODS

The automatic vascular segmentation model was trained using computed tomography enterography (CTE) imaging data from our hospital based on PE-NET. The segmentation performance of the trained model was evaluated using the sensitivity (SEN), the Dice Similarity Coefficient (DSC), and the Average Hausdorff Distance (AHD). Then, the model was used for the automatic measurement of vascular parameters in the classification set, and machine learning classifier SVM was applied based on selected vessel features. The diagnosis performance of the PE-NET + SVM model was evaluated and compared with that of human radiologists and the clinical biomarkers [C-reactive protein (CRP) and fecal calprotectin (FCP)]. The impact of PE-NET + SVM on the reading time of radiologists was also evaluated.

RESULTS

The segmentation dataset included the CTE data from 54 CD patients and 20 healthy controls. The classification dataset included the CTE data from 40 CD patients and 45 healthy controls. We found that PE-NET performed well in the vascular segmentation of the superior mesenteric artery (SMA), portal vein (PV), and abdominal aorta (AA) in both validation sets and the testing set. Vascular parameters were automatically extracted by PE-NET. We found that the mesenteric artery, portal vein, abdominal aorta, and the ratio of portal vein to superior mesenteric artery or abdominal aorta were increased in the testing set, with no statistical difference between the automatic measurement obtained using PE-NET and the manual evaluation of CTE. Moreover, an support vector machine (SVM) classifier was applied for CD diagnosis based on the vascular parameters. The F1 scores indicated the comparable diagnostic ability of PE-NET + SVM to senior radiologists with over 10 years of experience, and the receiver operating curves (ROCs) revealed that the area under the curve (AUC) of PE-NET + SVM was 0.934, which was higher than those of clinical biomarkers such as FCP (AUC of 0.913) and CRP (AUC of 0.893), suggesting the great potential of PE-NET in aiding CD diagnosis. Additionally, the reading time of a junior radiologist on CTE images was significantly reduced and comparable to that of a senior radiologist with the help of PE-NET.

CONCLUSION

The PE-NET enables the automated measurement of mesenteric vascular and portal vein parameters and potentially assists the efficient diagnosis of CD.