Deep learning using CT to identify high-risk patients for adjacent segment disease: Model development and validation in a multicenter study

Objective: The prevalence of adjacent segment disease (ASD) following lumbar surgery is strongly linked to posterior lumbar interbody fusion (PLIF). The goal of this study was to create and validate a composite deep learning model for predicting the development of ASD following PLIF. Methods: We retrospectively collected preoperative lumbar CT data from 331 patients who had PLIF for lumbar degenerative disorders between January 2016 and June 2023 at our center and two other research centers. The 3D UNet model was used to precisely segment the spine, and the 3D ResNet model assessed these segmented pictures to predict postoperative ASD incidence. The internal dataset was separated into three sets: training, testing, and validation, with a 70:15:15 ratio. Immediate data augmentation and cross-validation examined model generalization, which was then validated externally. Gradcam was utilized to visually represent the network's prediction base and to investigate accurately predicted images. Results: The integrated deep learning models revealed great segmentation accuracy and predictive capability for ASD, as well as significant discriminatory ability. The ResNet50 model predicted with 89 % accuracy, 75 % sensitivity, and 95 % specificity. The system outperformed two spine surgeons’ combined forecasts in terms of accuracy and specificity, as well as performance in the external validation set. Conclusion: This study successfully built a deep learning-based composite model that accurately predicts postoperative ASD occurrence in patients undergoing PLIF using preoperative lumbar CT scans. This strategy has the potential to reduce the number of secondary procedures required for ASD, reducing the burden on the public health system.