Automated radiomics pipeline integrating TotalSpineSeg for Pfirrmann grading of lumbar disc degeneration with SHAP-based interpretability.
Authors
Affiliations (7)
Affiliations (7)
- Department of Pathology, Wangjing Hospital, China Academy of Chinese Medical Sciences, Beijing, China.
- Tianjin University of Traditional Chinese Medicine, Tianjin, China.
- Institute of Chinese Materia Medica, China Academy of Chinese Medical Sciences, Beijing, China.
- Institute of Orthopaedics and Traumatology, China Academy of Chinese Medical Sciences, Beijing, China.
- Department of Sports Medicine, Wangjing Hospital, China Academy of Chinese Medical Sciences, Beijing, China.
- Department of Radiology, Wangjing Hospital, China Academy of Chinese Medical Sciences, Beijing, China.
- Department of Radiology, The Affiliated Hospital of Guizhou Medical University, Guiyang, Guizhou, China.
Abstract
Pfirrmann grading of lumbar disc degeneration (LDD) is constrained by inter-observer variability. Although radiomics enables objective quantification, the manual segmentation required by existing models is time-consuming and undermines feature reproducibility. TotalSpineSeg, an nnU-Net-based tool, allows fully automated segmentation, yet its integration with radiomics for Pfirrmann grading remains unexplored. This study aimed to develop an automated radiomics pipeline incorporating TotalSpineSeg for Pfirrmann grading of LDD, and to assess whether three-dimensional features contributed discriminative information beyond mid-sagittal two-dimensional radiomics, simple measurements, and frozen ResNet50 embeddings (ImageNet and RadImageNet). Interpretability analysis was applied to identify candidate imaging biomarkers. Magnetic resonance imaging (MRI) data from 210 patients with LDD yielded 630 disc samples (L3-S1, Pfirrmann Grades I-V). Segmentation was performed automatically with TotalSpineSeg. Of 1,774 extracted radiomics features, 517 were retained through a multi-step selection process. Ten machine learning algorithms were evaluated using a patient-level 80:20 train-test split with 5-fold cross-validation, and SHapley Additive exPlanations (SHAP) analysis was applied for feature interpretation. Automated segmentation averaged 6.9 s per case. The CatBoost model achieved a macro-averaged area under the curve (AUC) of 0.936 (95% confidence interval [CI]: 0.913-0.958) for three-dimensional radiomics, followed by mid-sagittal two-dimensional radiomics (0.916), frozen convolutional neural network (CNN) embeddings (RadImageNet, 0.874; ImageNet, 0.864), and the strongest simple measurement (0.760). For the three-dimensional model, sensitivity and specificity were 63.9% and 93.1%, grade-specific AUCs ranged from 0.876 to 0.978, and SHAP analysis identified the Laplacian of Gaussian first-order minimum, Flatness, and gray-level co-occurrence matrix (GLCM) Imc2 as candidate biomarkers. Integrating TotalSpineSeg with radiomics offers an efficient and interpretable approach to Pfirrmann grading. Three-dimensional features contributed the greatest discriminative information, with a modest increment over two-dimensional radiomics and a significant advantage over simple measurements and frozen ResNet50 embeddings. The candidate biomarkers identified are pathophysiologically meaningful features for LDD assessment.