Automated quantitative computed tomography grading of lumbar facet joint degeneration: influence of segmentation errors on instance-level clinical assessment.
Authors
Affiliations (2)
Affiliations (2)
- Department of Orthopedics, Affiliated Nantong Clinical College of Nantong University, Nantong First People's Hospital, School of Medicine, Nantong University, Nantong, China.
- Department of Orthopedics, Nantong First People's Hospital, Southeast University, Nantong, China.
Abstract
Automated lumbar facet joint degeneration grading requires both accurate anatomical segmentation and clinically meaningful region-level classification. We evaluated whether voxel-level segmentation performance was aligned with downstream degeneration grading. This retrospective study included 145 lumbar computed tomography (CT) examinations and 725 level-specific bilateral facet joint regions of interest (ROIs), including 165 severe and 560 non-severe regions. Weishaupt grades 0-1 were categorized as non-severe and grades 2-3 as severe. Scheme A used binary nnU-Net segmentation followed by 11 quantitative region-level features, training-fold-specific standardization, and logistic regression; an oracle comparison using ground-truth masks was compared with analysis using predicted masks. Scheme B used direct three-class segmentation. Patient-level five-fold cross-validation was performed, with 116 training and 29 validation patients per fold. The results were reported as mean ± standard deviation (SD). Scheme A achieved a binary segmentation Dice coefficient of 0.770±0.008. In the oracle comparison, the region-level area under the receiver operating characteristic (ROC) curve (AUC), recall, F1-score, accuracy, and case-level accuracy were 0.671±0.100, 0.538±0.172, 0.399±0.102, 0.640±0.051, and 0.579±0.096, respectively. Using predicted masks, the corresponding values were 0.650±0.068, 0.603±0.250, 0.375±0.072, 0.563±0.123, and 0.545±0.071. Scheme B achieved a native nnU-Net mean validation Dice of 0.493±0.044 and region-level AUC, recall, F1-score, accuracy, and case-level accuracy of 0.730±0.024, 0.445±0.051, 0.510±0.045, 0.808±0.015, and 0.738±0.058, respectively. Voxel-level segmentation and clinically oriented grading performance were not directly aligned. Scheme A provided stable anatomical segmentation but more variable downstream grading, whereas scheme B achieved better overall region- and case-level classification performance but lower severe-degeneration recall. Both spatial and clinical decision metrics should therefore be considered when evaluating automated facet joint grading systems.