AlignPro: a robust deep learning-based prediction of spinal alignments irrespective of image qualities acquired from smartphone photographs of radiographs displayed on PACS

Abstract

Introduction: Original X-rays are not easily accessible for telemedicine and existing deep learning–based automated Cobb angle (CA) predictions are not accurate on suboptimal quality X-rays. Objective: To develop an automated CA prediction system irrespective of image quality: AlignPro, with no restrictions on curve patterns to facilitate clinical practice and telemedicine. Methods: In total, 367 consecutive patients attending a scoliosis clinic were recruited prospectively and their coronal X-rays were re-captured using mobile phones. Five-fold cross-validation was conducted (5 experiments, each with 294 images to train a deep neural network named HRNet for endplates landmarks and end-vertebrae detection, and the remaining 73 images were used to test). The predicted heatmaps of the vertebral landmarks were visualised to enhance interpretability. Per-landmark-absolute-errors and recall of the landmark detection results were calculated to assess the accuracy of the predicted landmarks. Further calculated CAs were quantitatively compared with the spine specialists measured ground truth (GT). Results: The average per-landmark absolute distance error and the recall of the detected endplates landmarks were 2.8 pixels and 0.99, indicating a highly accurate detection. The predicted CAs were all significantly correlated with GT (p<0.01). Compared with GT, the mean error was 3.73-4.15° and standard error of the measurement was 0.8-1.7° for the predicted CAs at different spinal regions. Conclusion: This is the first study using HRNet on non-original X-rays to automatically and accurately predict vertebral landmarks of the scoliotic spine. HRNet’s applicability is evidenced by our thorough cross-validation, which can be used with telemedicine to facilitate fast reliable auto-diagnosis and follow-up.