Development and validation of the MRI-based deep learning classifier for distinguishing perianal fistulizing Crohn's disease from cryptoglandular fistula: a multicenter cohort study.

BACKGROUND

A singular reliable modality for early distinguishing perianal fistulizing Crohn's disease (PFCD) from cryptoglandular fistula (CGF) is currently lacking. We aimed to develop and validate an MRI-based deep learning classifier to effectively discriminate between them.

METHODS

The present study retrospectively enrolled 1054 patients with PFCD or CGF from three Chinese tertiary referral hospitals between January 1, 2015, and December 31, 2021. The patients were divided into four cohorts: training cohort (n = 800), validation cohort (n = 100), internal test cohort (n = 100) and external test cohort (n = 54). Two deep convolutional neural networks (DCNN), namely MobileNetV2 and ResNet50, were respectively trained using the transfer learning strategy on a dataset consisting of 44871 MR images. The performance of the DCNN models was compared to that of radiologists using various metrics, including receiver operating characteristic curve (ROC) analysis, accuracy, sensitivity, and specificity. Delong testing was employed for comparing the area under curves (AUCs). Univariate and multivariate analyses were conducted to explore potential factors associated with classifier performance.

FINDINGS

A total of 532 PFCD and 522 CGF patients were included. Both pre-trained DCNN classifiers achieved encouraging performances in the internal test cohort (MobileNetV2 AUC: 0.962, 95% CI 0.903-0.990; ResNet50 AUC: 0.963, 95% CI 0.905-0.990), as well as external test cohort (MobileNetV2 AUC: 0.885, 95% CI 0.769-0.956; ResNet50 AUC: 0.874, 95% CI 0.756-0.949). They had greater AUCs than the radiologists (all p ≤ 0.001), while had comparable AUCs to each other (p = 0.83 and p = 0.60 in the two test cohorts). None of the potential characteristics had a significant impact on the performance of pre-trained MobileNetV2 classifier in etiologic diagnosis. Previous fistula surgery influenced the performance of the pre-trained ResNet50 classifier in the internal test cohort (OR 0.157, 95% CI 0.025-0.997, p = 0.05).

INTERPRETATION

The developed DCNN classifiers exhibited superior robustness in distinguishing PFCD from CGF compared to artificial visual assessment, showing their potential for assisting in early detection of PFCD. Our findings highlight the promising generalized performance of MobileNetV2 over ResNet50, rendering it suitable for deployment on mobile terminals.

FUNDING

National Natural Science Foundation of China.