Construction of segmentation and classification models for spinal tuberculosis lesions Based on CT: A multi-center study

Abstract Purpose To establish a segmentation model and classification model of tuberculosis lesions based on CT images to improve the accuracy of early diagnosis of spinal tuberculosis. Methods This study adopted multicenter retrospective data (n = 1025). Firstly, the vertebral body region of the spine was extracted through the U-Net segmentation model. Then, the segmented images were input into the improved ResNet50 network. Combined with the CT bone window gradient attention mechanism, an end-to-end deep learning diagnostic model was constructed. Results In the internal validation datasets, the model achieved an AUC of 0.920, accuracy of 0.874, and sensitivity of 0.876. For External test datasets 1, the AUC was 0.867, accuracy 0.801, and sensitivity 0.794; for External test datasets 2, the AUC was 0.866, accuracy 0.769, and sensitivity 0.883; and for External test datasets 3, the AUC was 0.941, accuracy 0.843, and sensitivity 0.790. Conclusion The multi-center study built up a deep learning model for spinal tuberculosis diagnosis with the assist of the CT bone window gradient attention mechanism. The model achieved a good internal verification ability (AUC = 0.920, accuracy rate = 0.874) and external verification ability (AUC = 0.866–0.941,accuracy rate = 0.769–0.843) which showed the wide applicability of the model to different medical institutions.The main developments of this work are good performances for features that extract relevant information about trabecular micro-fractures and calcification contours’ gradients.