A multi-task deep learning framework for simultaneous classification of scoliosis severity and sagittal parameters from full-spine radiographs.
Authors
Affiliations (5)
Affiliations (5)
- Fujian University of Traditional Chinese Medicine, Fuzhou, Fujian, 350122, China.
- Xiamen University, Xiamen, Fujian, 361005, China.
- Hangzhou First People's Hospital Xiasha Campus, Hangzhou Rehabilitation Hospital, Hangzhou, Zhejiang, 310018, China.
- Xiamen University, Xiamen, Fujian, 361005, China. [email protected].
- Xiamen University, Xiamen, Fujian, 361005, China. [email protected].
Abstract
Scoliosis is commonly diagnosed and assessed from standing radiographs, but manual measurement is time-consuming and subject to inter-observer variability. Most existing artificial intelligence approaches focus on coronal Cobb angle estimation alone and therefore provide limited support for multi-dimensional clinical assessment. This study developed a multi-task classification framework for simultaneous assessment of scoliosis severity, curve type, and sagittal parameter categories from a single anteroposterior (AP) full-spine radiograph. A total of 239 participants contributed 239 AP full-spine radiographs, with one AP radiograph per participant. The development cohort comprised 192 participants and was evaluated using patient-level stratified five-fold cross-validation, while 47 consecutive later cases formed an independent temporal validation cohort. The network received only one AP radiograph as input and classified Cobb angle severity, curve type, thoracic kyphosis (TK) category, lumbar lordosis (LL) category, and sacral slope (SS) category. Lateral radiographs were never used as network inputs during training or inference; orthopedic experts used them solely to measure TK, LL, and SS, after which the measurements were converted into predefined categorical labels. Over three independent experiments using the same cross-validation protocol and different random initializations, the proposed ResNet-based framework achieved a mean Cobb angle classification accuracy of 93.68% ± 1.82% and a Macro-F1 of 93.15% ± 1.80%. Cobb classification accuracy was 91.23% on the independent temporal validation cohort and should be interpreted cautiously given its limited size. Ablation experiments and training-validation curves supported the contributions of multi-task coupling, multi-scale attention, and dynamic task weighting and showed stable convergence without a progressively widening train-validation gap. The proposed AP-only framework simultaneously provides five clinically relevant categorical outputs. The sagittal outputs should be interpreted as auxiliary classification or sc reening information and not as recovery of exact sagittal angles or a replacement for standard lateral-radiograph measurement. Further validation in larger, multi-center cohorts is required before clinical deployment.