A transformer-based deep learning algorithm for diagnosing spinal infections on axial non-contrast computed...

Spinal infections are rare but serious conditions requiring timely diagnosis. Non-contrast computed tomography (CT) is widely used and may incidentally reveal spinal abnormalities; however, subtle infectious findings are often missed, especially on axial images without sagittal reconstruction.

To investigate a deep learning approach for diagnosing primary spinal infections using non-contrast CT images.

This retrospective dual-center study included 157 patients with primary spinal infection. A Swin Transformer model was developed using non-contrast CT slices. Patients from the primary center (n = 127) were randomly split 7:3 into training and internal validation sets. An independent external cohort (n = 30) from a second center was used for external validation. Per-slice diagnostic performance was assessed using the area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (AUPRC), sensitivity, specificity, accuracy, positive predictive value (PPV), negative predictive value (NPV), and F1-score, with a probability threshold ≥0.5 determined by the Youden index. Model performance was compared with that of two radiologists.

The Swin Transformer model demonstrated excellent per-slice diagnostic performance. In the internal validation set, the model achieved an AUC of 0.979, sensitivity of 96.2%, specificity of 90.5%, and accuracy of 89.4%. In the external cohort, similar results were obtained: AUC 0.989, sensitivity 98.2%, specificity 98.3%, and accuracy 98.3%. The deep learning model significantly outperformed both radiologists in AUC and sensitivity across cohorts (all P < 0.05). With artificial intelligence (AI) assistance, both radiologists showed substantial improvements in diagnostic performance and efficiency in both internal and external validation (all P < 0.05). The model's reading time was markedly shorter than that of unassisted radiologists, and AI assistance reduced radiologists' reading time (P < 0.001). Presence of spinal epidural abscess and pathogen type significantly influenced diagnostic outcomes (P < 0.001).

This Swin Transformer-based deep learning model achieves high diagnostic accuracy for detecting spinal infections on axial non-contrast CT images, with performance comparable to or exceeding that of musculoskeletal radiologists. By enhancing radiologists' sensitivity and reducing reading time, the model shows promise as a clinical decision support tool to reduce missed diagnoses, particularly in emergency or resource-limited settings where magnetic resonance imaging (MRI) is unavailable. Its robust performance on external validation supports generalizability and lays the foundation for future multicenter prospective studies and extension to other spinal pathologies.