Automated grading and diagnosis of sacroiliitis on CT images using a 3D convolutional neural network: a multicenter retrospective study

Abstract

Abstract Radiographic grading of the sacroiliac joints plays a critical role in the differential diagnosis of ankylosing spondylitis (AS) and in guiding the treatment. The aim of this study was to develop an automated 3D convolutional neural network (3D CNN) for grading and diagnosing sacroiliitis on CT images to assist clinicians. This study included CT images from 2,144 participants, comprising healthy controls and patients with suspected ankylosing spondylitis (AS). A V-Net based segmentation model was applied, followed by training a three-dimensional DenseNet-based convolutional neural network (3D CNN) for both five-class and three-class classification tasks. Grading by three radiologists according to the New York criteria served as the reference standard. The model’s diagnostic performance was evaluated on an external multicenter validation set and compared with radiologist interpretations. For the five-class task, the model’s area under the receiver operating characteristic curve (AUC) for grades 0–IV were 0.966, 0.937, 0.881, 0.962, and 0.994, respectively. In the simplified three-class task, AUCs for classes 0, 1, and 2 were 0.984, 0.967, and 0.994, respectively. On the external validation set, three-class AUCs were 0.957, 0.934, and 0.992. With AI assistance, two radiologists’ diagnostic accuracy improved by 6.9% and 8.4%, respectively. The proposed segmentation–classification framework enables accurate and reproducible CT grading of sacroiliitis.