This retrospective study was conducted at our hospital. The PE-NET model, established as previously described, was applied in this study (12).

To train the model, this study collected the imaging data and other relevant clinical data of 54 CD patients admitted to our hospital from January 2014 to December 2024 and 20 healthy control individuals without intestinal inflammation (validated by endoscopic examination). The training/validation set included 63 cases, and the testing set included 11 patients. The external validation dataset included 20 abdominal images of CD patients and 20 abdominal images of healthy controls obtained from the open public database.¹ For the datasets used for the classification model, a total of 85 cases were included, with 30 cases in the training/validation set and 55 cases in the testing set. The study was approved by the Ethics Committee of our hospital and conforms to the principles of the Declaration of Helsinki. The inclusion criteria were as follows: (1) CD patients confirmed by histological examination, (2) those with CTE imaging data, (3) those who underwent endoscopic examination before or after imaging examination; and (4) those with comprehensive clinical data. The exclusion criteria were as follows: (1) patients who received abdominal surgery before the examination; (2) patients with scanning images of poor quality; and (3) patients with comorbidities such as cancer, coronary heart disease, and other gastrointestinal diseases.

The CT scanning was conducted using a FORCE CT scanner (Siemens, Berlin, Germany). Patients were instructed to undergo standard bowel preparation before the CT examination, including a low-slag diet 2 days before the examination, a liquid diet 1 day before the examination, and fasting for 4–8 h before the examination. The scan area was from the top of the diaphragm to the lower margin of the pubis, and patients were in the supine position. The scan parameters were set as follows: tube voltage at A: 90 kV, B: Sn150 kV, tube current at A: 144 mass, B: 90 mAs, pitch of 1.0, speed of 0.5 s, collimation of 2 × 192 × 0.6 mm, slice thickness of 1 mm, and slice gap of 1 mm. Patients were first injected with 75 mL of iopromide (370 mgI/ml) at 3.5 mL/s in the antecubital vein and then with 20–30 mL of normal saline. The monitoring methods were applied based on a three-stage enhanced scanning automatic tracking technique. When the aortic monitoring threshold reached 100 HU, the arterial phase scan was triggered. The venous phase scan was performed 40 s later, followed by a delayed phase scan conducted 80 s after completion of the venous phase scan. The images were sent to a post-processing workstation (syngo via, VB20, Siemens Medical Solutions, Forchheim, Germany) for further processing.

The 120 kVP images with a slice thickness of 1 mm were obtained after inputting the raw data into the post-processing workstation. The virtual monoenergetic images with computational fusion were also obtained under the dual energy Mono+ mode (40–90 KeV) with the slice thickness of 1 mm. The cross-sectional images with a slice thickness of 3 mm and a slice gap of 3 mm were reconstructed for the above images. Two radiologists, each with over 10 years of experience and blinded to patient information, delineated the vessels. This was reviewed by a specialist with over 20 years of experience in abdominal imaging diagnosis. The robustness assessment was conducted in cases with anatomical variability or disease-induced deformation by evaluating the ability of the model in visualizing major vascular variants compared to the expected anatomical course based on CT images. The vessels delineated in consensus by radiologists were referred to as the ground truth. Figure 1 shows the example of CT images from a CD patient and a healthy control individual.

PE-NET was established as previously described (12, 13). It is a novel parallel network combining a transformer with CNN and is composed of a contracting path (encoder), an expanding path (decoder), and skip connections.

The features of vessel parameters were fused and input into a support vector machine (SVM) for diagnosis, with the parameters of C = 1, gamma = ‘scale’, kernel = ‘rbf’, and degree = 3.

(1) To validate the performance of the model, voxel-based metrics including the sensitivity (SEN), Dice Similarity Coefficient (DSC), and Average Hausdorff Distance (AHD) were evaluated. Considering vascular connectivity, the evaluation of AHD is of top priority, followed by the DSC and SEN values.

(2) The vascular morphological characteristics, such as the diameter of superior and inferior mesenteric artery, the portal vein, and the abdominal aorta, were measured. The ratio of the portal vein/aorta or the portal vein/superior mesenteric artery diameter was evaluated to minimize the effect of patient size.

(3) Observer studies: Four radiologists of varying experience were included in the observation studies. The diagnostic performance of PE-NET + SVM was compared with that of human radiologists using F1 scores. The inter-rater consistency among human radiologists was evaluated using Cohen’s kappa scores.

(4) Traditional imaging indicators obtained from CT imaging were evaluated, including multisegmental bowel involvement, bowel wall thickening, mural stratification, mesenteric vessel engorgement, lymph node enlargement, and increased mesenteric fat density.

SPSS 23.0 software and GraphPad Prism 8.0 software were used for data analysis. The Shapiro–Wilk test was used to evaluate the normal distribution of the data. For data that conformed to normal distribution, the measurement data were shown as the mean ± SD and evaluated using Student’s t-test. Categorical data were shown as the frequency and percentage (%), and comparisons between groups were conducted using the chi-square test. For measurement data that did not conform to normal distribution, data were shown as the median (Q25 and Q75) and compared with the Mann–Whitney U test. A p-value of <0.05 was regarded as a statistically significant difference.

This study included 54 CD patients and healthy control individuals. The mean age was 25.32 ± 4.36 for CD patients and 26.24 ± 3.96 for healthy controls. Among the CD patients, 17 were women and 37 were men. There were 8 women and 12 men in the healthy control group. There was no statistical significance between the baseline characteristics, including age and sex of CD patients and healthy controls (p > 0.05), while the CRP and ESR levels of CD patients were significantly higher than those of the healthy controls, and the BMI of CD patients was significantly lower relative to healthy controls (p < 0.05). Moreover, the CTE showed that, among CD patients, 26 had multisegmental bowel involvement, 31 had mural stratification, 40 showed comb signs, 12 had lymph node enlargement, 54 showed bowel wall thickening and mesenteric fat thickness of 0.85 ± 0.42, which were all significantly higher compared with the healthy controls (p < 0.05, Table 1).

Data from 74 participants (54 CD patients and 20 healthy controls) were used for modeling, of which data from 63 cases were used for the training and validation sets and those from 11 cases were used for the testing set. The automated segmentation performance of the model was evaluated with the five-fold cross-validation and compared against the ground truth expert-labeled segmentations. The data of 20 CD cases and 20 healthy controls from a public dataset were used for external validation. The segmentation performance of PE-NET of each type of vessel (SMA, PV, and AA) was tested and evaluated as shown in Table 2. The model showed comparable SEN, DSC, and AHD across the three cohorts, although the SEN and DSC values in the external validation group were slightly slower, and the AHD were slightly higher for PV segmentation compared with the internal validation group.

The accuracy of the automatic segmentation by PE-NET was evaluated in the testing set. As measured by PE-NET or CTE evaluation, the diameter of the superior mesenteric artery, the inferior mesenteric artery (IMA), the portal vein, and the abdominal aorta (AA), as well as the ratio of portal vein (PV)/superior mesenteric artery (SMA) and PV/AA diameter was significantly increased in CD patients compared with the control group (Figures 2A–F). Moreover, we found that the bias of mesenteric vascular, portal vein, and abdominal aorta parameters between PE-NET and CTE evaluation showed no significant difference, which indicated minimal systemic error. Overall, these data revealed that PE-NET performed well in the automated measurement of mesenteric vascular and portal vein parameters as well as in the calculation of PV/SMA and PV/AA diameter.

For the evaluation of vascular variants (Table 3), the successful visualization rate of the model was 50% (1 of 2) and 66.67% (2 of 3) in the internal validation and testing groups, respectively.

The SVM classifier was applied based on the vessel parameters automatically measured by PE-NET in the classification set. The classifying accuracy reached 81.82% (Sensitivity: 75.86%; Specificity: 88.46%), and the AUC value of the SVM model was 0.934 in the testing set (Table 4). Figure 3 shows the importance of the permutation feature for the SVM model and highlights the contribution of the selected features to the diagnostic efficiency. The results indicated the top importance of the PV diameter to the SVM model.

The diagnosis performance of the PE-NET + SVM model alone was compared with that of human radiologists with varying expertise using F1 scores. The results showed that the model exhibited better performance compared with junior human radiologists with over 5 years of experience and nearly comparable performance to the senior human radiologists with over 10 years of experience (Table 5). When combined with the PE-NET + SVM model, the junior human radiologists showed improved F1 scores that were comparable to the senior human radiologists without PE-NET + SVM model assistance. We also found that the F1 scores of senior human radiologists were increased with the aid of the PE-NET + SVM model, suggesting the model as a promising tool in aiding the clinical diagnosis of radiologists. Additionally, the human expert variability was evaluated by Cohen’s kappa scores, and inconsistencies were observed between the junior and senior radiologists. The results suggested that applying the PE-NET + SVM model might exert improved stability in clinical diagnosis (Table 6).

Moreover, we also compared the diagnostic ability of the PE-NET + SVM model with two clinical biomarkers (FCP and CRP) of CD. The results of the ROC analysis showed that the AUC value was 0.934 for the model, 0.913 for FCP, and 0.893 for CRP in the testing set, which indicated the good performance of the PE-NET + SVM model in CD diagnosis (Figure 4).

Moreover, we evaluated the potential of PE-NET in improving the diagnostic efficiency of CD. The results showed that the average reading time and total reading time for a junior radiologist (radiologist 1) were significantly reduced with the aid of PE-NET + SVM model and comparable to that of a senior radiologist (radiologist 3), while the average reading time for a senior radiologist (radiologist 3) was not evidently decreased by the application of PE-NET + SVM. The results suggested that the diagnostic efficiency of junior radiologists was improved with the help of the PE-NET + SVM model. Though the reading time was not significantly decreased for a senior radiologist, the application of the model might potentially reduce their burden in correcting the mistakes of junior radiologists and decrease the misdiagnosis or missed diagnosis in the clinic (Table 7).