The Development and External Testing of Deep Learning Model for Automated Classification of Intervertebral Disc Degeneration on CT Images.
Authors
Affiliations (11)
Affiliations (11)
- Laboratory of Key Technology and Materials in Minimally Invasive Spine Surgery, Tongren Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, 200336, China.
- Center for Spinal Minimally Invasive Research, Shanghai Jiao Tong University, Shanghai, 200336, China.
- Department of Imaging, Tongren Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, 200336, China.
- Shanghai Key Laboratory of Flexible Medical Robotics, Institute of Medical Robotics, Tongren Hospital, Shanghai Jiao Tong University, Shanghai, 200336, China.
- Department of Radiology, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, 200025, China.
- Shanghai Haohua Technology Co., Ltd., Shanghai, 201108, China.
- Department of Radiology, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, 200025, China. [email protected].
- Laboratory of Key Technology and Materials in Minimally Invasive Spine Surgery, Tongren Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, 200336, China. [email protected].
- Center for Spinal Minimally Invasive Research, Shanghai Jiao Tong University, Shanghai, 200336, China. [email protected].
- Department of Imaging, Tongren Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, 200336, China. [email protected].
- Shanghai Key Laboratory of Flexible Medical Robotics, Institute of Medical Robotics, Tongren Hospital, Shanghai Jiao Tong University, Shanghai, 200336, China. [email protected].
Abstract
The study aims to develop and externally test a deep learning (DL) model using CT images for automatic classification of intervertebral disc degeneration. This retrospective study enrolled patients aged ≥ 18 years who were diagnosed with intervertebral disc herniation or bulge, covering the cervical, thoracic, and lumbar spine, between January 2018 and July 2024 from two institutions. The disc lesions on CT were labeled by three radiologists with 3 to 5 years of experience as reference standard. A 3D U-Net model was developed to segment the disc lesions area. The lesion type (herniation, bulge), zone (Central canal zone, Subarticular zone, Foraminal zone, Extraforaminal zone), and grade (grade 1, 2, 3) of disc lesions were determined based on key points using the EfficientNet-B4 network architecture. Dice score, accuracy, recall, and precision were calculated to evaluate the model performance. The study randomly divided 140 and 36 cases from one institution into training and internal testing datasets, and 97 cases from another institution were used as external testing dataset. The mean Dice scores of segmentation were 0.735 for external testing. The average accuracy, recall, and precision of external testing were 0.938, 0.938, and 0.944 for lesion type, and 0.919, 0.986, and 0.976 for lesion zone, respectively. The grade metrics for external testing was 0.979. The DL model automatically classified intervertebral disc degeneration on CT images by type, zone and grading, demonstrating good peformance.