Multisequence MRI and clinical data-based deep learning radiomics model for predicting adjacent segment degeneration post-lumbar fusion: a retrospective multicenter study.

BACKGROUND: While adjacent segment degeneration (ASDeg) is a major complication following lumbar fusion, objective tools for preoperative risk prediction remain lacking. This study developed a model integrating clinical data, deep learning, and radiomic features from multisequence magnetic resonance imaging (MRI) to predict high-risk patients for ASDeg after lumbar fusion. METHODS: This study included 305 patients who underwent lumbar MRI before surgery. The patients were randomly divided into a training cohort (n = 192) and an internal validation cohort (n = 83). An external test cohort (n = 30) was recruited from two additional centers. The Vision Transformer 3D model architecture was used for model development. Deep learning-based and handcrafted radiomic features were extracted from multisequence MR images of patients. A logistic regression model incorporating the least absolute shrinkage and selection operator was established, and both radiomic feature and deep learning radiomics (DLR) feature classification models were developed. ASDeg was defined as degeneration observed during radiological follow-up > 6 months after lumbar surgery and served as the reference standard. The performance of 14 machine learning classification models in predicting high-risk patients for ASDeg was evaluated using receiver operating characteristic (ROC) curves. A combined model was developed by integrating clinical baseline variables. RESULTS: After feature selection, 17 handcrafted radiomic features, 12 DLR features, and three clinical features were ultimately selected. The most effective machine learning algorithm for the radiomic feature model was the linear support vector machine, and AdaBoost was the best performing algorithm for the DLR feature model. By combining these features with clinical features, a combined model was developed, with the gradient boosting machine being the most effective machine learning algorithm. The AUCs for training, internal validation, and external test cohorts were 0.959, 0.818, and 0.895, respectively. Compared to the combined predictions of two spinal surgeons, the combined model also demonstrated superior performance. Calibration curves showed better calibration in the validation set. Decision curve analysis demonstrated greater net benefit for patients. CONCLUSIONS: The combined model effectively identified patients at high risk of ASDeg following lumbar fusion based on clinical characteristics and MRI, which may help reduce revision surgeries and alleviate the burden on the public healthcare system. CLINICAL TRIAL NUMBER: Not applicable.