Automatic Localization and Classification of Crohn's Disease Activity in Computed Tomography Enterography Images Using Deep Learning.

Rapid and accurate localization and activity grading of Crohn's disease (CD) lesions on computed tomography enterography (CTE) images enhance the diagnostic efficiency of radiologists. We developed a one-stage model (called CD-YOLO) based on YOLOv5 for the automatic detection and classification of CD lesions. To tackle the challenges of variable sizes and shapes, and boundary blurring in CD lesions, we integrated a suite of specialized modules that not only enhanced the model's feature extraction capability but also enabled the effective fusion of multi-scale features. We retrospectively collected CTE images from 233 patients with pathologically and endoscopically confirmed CD. Patients were divided into training, validation, and testing cohorts at the patient level. The model performance was evaluated using the mAP@0.5, precision, and recall. Gradient-weighted class activation mapping (Grad-CAM) was employed to interpret the model predictions via visualization. The mAP@0.5 of CD-YOLO for detecting active and remission CD was 92.1% and 84.2%, respectively. Compared to YOLOv5s, the mAP@0.5 improved by 2.0% and 4.9% in detecting active and remission CD, respectively. Meanwhile, the missed detection and misdetection lesion rates decreased by 4.9% and 4.18%, respectively. The proposed CD-YOLO framework demonstrated superior performance in CD lesion localization and activity classification and may serve as an effective computer-aided diagnostic tool to support radiologists and gastroenterologists, as well as facilitate clinical teaching.