Yefu Xu, Jiangkai Yu, Yan Zhou, Sangni Liu, Chao Ma, Ying Han, Daojuan Wang, Xinghui Xie, Gaojian Tao
Key factors were paraspinal muscle quality features, including the fatty infiltration and relative cross-sectional area. The DL model achieved 83.6% accuracy internally, compared with XGBoost (76.4%) and the surgeon (76.4%), with an F1 score of 0.769 and κ of 0.644. External validation accuracy was 79.4% and κ was 0.564.
BACKGROUND: Percutaneous transforaminal endoscopic discectomy (PTED) is a common treatment for lumbar disc herniation (LDH), but postoperative recovery varies greatly.
PURPOSE: To develop and validate a deep learning (DL) model integrating MRI and clinical features to estimate 1-year outcomes after PTED.
STUDY TYPE: Retrospective.
POPULATION: 252 LDH patients who underwent single-level PTED, with lumbar spine MRI acquired within 1 month before surgery: training set (n = 129) and internal test set (n = 55) from Center 1; external test set (n = 68) from Center 2.
FIELD STRENGTH/SEQUENCE: T2-weighted spin echo sequence at 1.5 T.
ASSESSMENT: The primary endpoint was achieving the Patient Acceptable Symptom State at 1 year (Oswestry disability index ≤ 22). A front-end deep learning model extracted disc and paraspinal muscle features from MRI data. Key features identified by SHAP analysis were incorporated into a DL framework that fused them with ResNet50-derived imaging features and clinical data via multi-head attention. Model performance was compared with conventional machine learning models (using clinical and radiomic features) and a spine surgeon. Sensitivity analyses were carried out using alternative ODI thresholds (≤ 15 and < 25).
STATISTICAL TESTS: Accuracy with confidence interval (CI), F1 score, and Cohen's kappa (κ) were calculated to assess model performance. SHAP values ranked feature importance. The DeLong test compared areas under the receiver operating characteristic curves (AUCs) across models. Statistical significance was set at p < 0.05.
RESULTS: Key factors were paraspinal muscle quality features, including the fatty infiltration and relative cross-sectional area. The DL model achieved 83.6% accuracy internally, compared with XGBoost (76.4%) and the surgeon (76.4%), with an F1 score of 0.769 and κ of 0.644. External validation accuracy was 79.4% and κ was 0.564.
DATA CONCLUSION: The DL model showed encouraging performance compared with XGBoost and a spine surgeon for estimating 1-year outcomes after PTED, with paraspinal muscle quality as a key factor, though these preliminary findings require larger multi-center validation.
TECHNICAL EFFICACY: Stage 2.