Automated radiographic shoulder balance assessment in scoliosis via deep learning

The objective of this study was to develop an automated deep learning-based method for the assessment of shoulder balance in adolescent idiopathic scoliosis (AIS) patients using X-ray images in order to provide a reliable and efficient alternative to manual measurements. A total of 940 AIS radiographs were screened; 937 cases were included in the model-development cohort after quality control and were annotated for precise identification and segmentation of the T1 vertebra, both clavicles, and both coracoids. A deep learning neural network was used to segment these structures. Landmarks were extracted based on morphological image processing, and shoulder balance parameters including the clavicle angle (CA), coracoid height difference (CHD), clavicle tilt angle difference (CTAD), radiological shoulder height (RSH), and T1 tilting angle (T1TA) were calculated. The accuracy of the automated measurements was validated using an external dataset (n = 70) assessed by three senior spinal surgeons. The deep learning neural network achieved reliable segmentation performance for foreground anatomical structures, with macro-average intersection over union (IoU) values of 0.77 and 0.73 and Dice coefficients of 0.87 and 0.84 in the internal and external validation datasets, respectively. In the external dataset, the automated measurements displayed a high level of agreement with observer-averaged measurements, with intraclass correlation coefficients ranging from 0.964 to 0.994. Bland-Altman analysis revealed small mean biases across the five shoulder balance parameters, and 90.0 to 98.6% of automated measurements were within the range of interobserver variability. The proposed method provides an efficient and reproducible approach for radiographic shoulder balance assessment and may help reduce observer-dependent measurement variability.