Application of Multitask Deep Learning Model in Grading the Severity of Spinal Facet Joint Degeneration

This project is a retrospective clinical study. From 2020 to 2022, DICOM-format images and basic information of X-ray, CT, and magnetic resonance (MR) images were collected from outpatients and inpatients with suspected low back pain at the Fifth Affiliated Hospital of Sun Yat-sen University. After obtaining the DICOM image mode, data were exported from the information module upon the successful submission of OA batches; basic patient information was collected from inpatient medical records. This study plans to include 1,132 patients from a single center, who will be randomly divided into a training set, a validation set, and a test set according to the proportion for automatic diagnosis by the computer deep learning model, aiming to test the stability and reliability of the model. Among these 1,132 patients, two doctors separately conducted graded image reading for joint stenosis, hypertrophy, osteophytes, articular surface erosion, and subchondral cysts. Controversial results were determined by another more experienced doctor, and results of the reference standard group were confirmed by the senior doctor group. The data analysis methods for other centers were consistent with those described above. By comparing the diagnostic results of clinicians and the model, the performance and clinical feasibility of the deep learning model for the automatic diagnosis of lumbar facet joint degeneration were evaluated. The doctors' judgment results were compared with the model's prediction results, and statistical analysis was performed on the performance of the model's automatic diagnosis. Performance evaluation indicators included accuracy, sensitivity, specificity, precision, recall, F1 score, and AUC value. Among them, F1 score and AUC value are the main indicators for the comprehensive evaluation of model performance; the higher their values, the stronger the model performance.