[Development and validation of an automatic diagnostic tool for lumbar stability based on deep learning].

OBJECTIVE

To develop an automatic diagnostic tool based on deep learning for lumbar spine stability and validate diagnostic accuracy.

METHODS

Preoperative lumbar hyper-flexion and hyper-extension X-ray films were collected from 153 patients with lumbar disease. The following 5 key points were marked by 3 orthopedic surgeons: L posteroinferior, anterior inferior angles as well as L posterosuperior, anterior superior, and posterior inferior angles. The labeling results of each surgeon were preserved independently, and a total of three sets of labeling results were obtained. A total of 306 lumbar X-ray films were randomly divided into training (=156), validation (=50), and test (=100) sets in a ratio of 3∶1∶2. A new neural network architecture, Swin-PGNet was proposed, which was trained using annotated radiograph images to automatically locate the lumbar vertebral key points and calculate L intervertebral Cobb angle and L lumbar sliding distance through the predicted key points. The mean error and intra-class correlation coefficient () were used as an evaluation index, to compare the differences between surgeons' annotations and Swin-PGNet on the three tasks (key point positioning, Cobb angle measurement, and lumbar sliding distance measurement). Meanwhile, the change of Cobb angle more than 11° was taken as the criterion of lumbar instability, and the lumbar sliding distance more than 3 mm was taken as the criterion of lumbar spondylolisthesis. The accuracy of surgeon annotation and Swin-PGNet in judging lumbar instability was compared.

RESULTS

① Key point: The mean error of key point location by Swin-PGNet was (1.407±0.939) mm, and by different surgeons was (3.034±2.612) mm. ② Cobb angle: The mean error of Swin-PGNet was (2.062±1.352)° and the mean error of surgeons was (3.580±2.338)°. There was no significant difference between Swin-PGNet and surgeons (>0.05), but there was a significant difference between different surgeons (<0.05). ③ Lumbar sliding distance: The mean error of Swin-PGNet was (1.656±0.878) mm and the mean error of surgeons was (1.884±1.612) mm. There was no significant difference between Swin-PGNet and surgeons and between different surgeons (>0.05). The accuracy of lumbar instability diagnosed by surgeons and Swin-PGNet was 75.3% and 84.0%, respectively. The accuracy of lumbar spondylolisthesis diagnosed by surgeons and Swin-PGNet was 70.7% and 71.3%, respectively. There was no significant difference between Swin-PGNet and surgeons, as well as between different surgeons (>0.05). ④ Consistency of lumbar stability diagnosis: The of Cobb angle among different surgeons was 0.913 [95% (0.898, 0.934)] (<0.05), and the of lumbar sliding distance was 0.741 [95% (0.729, 0.796)] (<0.05). The result showed that the annotating of the three surgeons were consistent. The of Cobb angle between Swin-PGNet and surgeons was 0.922 [95% (0.891, 0.938)] (<0.05), and the of lumbar sliding distance was 0.748 [95%(0.726, 0.783)] (<0.05). The result showed that the annotating of Swin-PGNet were consistent with those of surgeons.

CONCLUSION

The automatic diagnostic tool for lumbar instability constructed based on deep learning can realize the automatic identification of lumbar instability and spondylolisthesis accurately and conveniently, which can effectively assist clinical diagnosis.