[Study on residual steps network human keypoint detection technology in scoliosis screening]

To assess the potential of residual steps network (RSN)-based human keypoint detection for early scoliosis screening.

Data from children and adolescents who underwent scoliosis screening at the outpatient clinic of Hangzhou Children's Hospital, Zhejiang Provincial Integrated Traditional Chinese and Western Medicine Hospital, and nearby primary schools between December 2021 and December 2022 were retrospectively analyzed. The dataset comprised 987 back photographs of healthy participants and 2 784 back photographs of patients with scoliosis or poor posture. Twelve dorsal keypoints were manually annotated, and their vertical symmetry was used to determine the presence of poor posture or scoliosis. The software was trained using these annotations to detect the keypoints automatically and make corresponding classifications. Model recall and accuracy were calculated as follows:recall=TP/(TP+FN);accuracy=(TP+TN)/(TP+TN+FP+FN).TP is true positive. FN is false negative. FP is false positive. TN is true negative.

The dataset contained 3 771 images, of which 3 017 were used for training and 754 for validation. The validation set contained 194 positive and 560 negative samples. The model correctly identified 179 scoliosis patients (true positives), missed 15 scoliosis patients (false negatives), correctly identified 439 non-scoliosis patients (true negatives), and misclassified 121 non-scoliosis patients as scoliosis (false positives). The model achieved an accuracy of 82.0% and a recall of 92.3%.

RSN-based human pose estimation technology may be used for early scoliosis screening in children and adolescents.