LDH-HybridNet: a hybrid ConvNeXt-Swin Transformer for nine-class CT-based diagnosis of lumbar disc herniation

Lumbar disc herniation (LDH) is a common cause of low-back pain and radiculopathy. Magnetic resonance imaging remains the reference standard, but computed tomography (CT) is widely used in routine and emergency practice because it is fast, accessible, and suitable for patients who cannot undergo MRI. Accurate CT-based diagnosis remains difficult when both the involved disc level and the type of disc abnormality must be identified. In this two-campus retrospective diagnostic study (both campuses within a single regional healthcare system), 11,928 axial lumbar CT images from 1,247 patients were used to develop and evaluate LDH-HybridNet. The reference labels were assigned by MRI consensus at the patient-disc level. The model combines a ConvNeXt-Tiny stem for local feature extraction with a Swin-Tiny trunk for global anatomical context and uses a multi-task head for nine-class LDH diagnosis. Evaluation included patient-level five-fold stratified cross-validation, comparison with nine reference models (seven baseline architectures, an adapted CT-ViT baseline, and a 2.5D multi-slice variant), single-variable ablation experiments, and Grad-CAM-based interpretability analysis. At the patient-disc level, LDH-HybridNet achieved 95.0% nine-class accuracy (345/363; 95% CI 92.6–97.4) and 94.7% (analytical 95% range 91.8–97.6) macro-F1 on the independent Site B test set. Examination exact-match accuracy was 86.8% (105/121; 95% CI 79.6–91.7). At the slice level, Site B accuracy was 91.3%, macro-F1 was 91.2%, disc-level identification accuracy was 95.4%, and abnormality-type accuracy was 93.5%. On the Site A development set, five-fold cross-validation yielded 93.3 ± 0.5% accuracy, 92.2 ± 0.5% macro-F1, 98.6 ± 0.1% macro-AUROC, and Cohen’s kappa of 0.920 ± 0.004. The observed Site B macro-F1 exceeded that of all nine comparison models. Ablations showed lower macro-F1 after removal of each evaluated component. LDH-HybridNet provided accurate fine-grained CT-based LDH classification by combining local convolutional features, hierarchical Transformer attention, and multi-task learning. Performance was validated on data from a second campus within the same regional healthcare system; external validation across regions and healthcare systems, as well as prospective and volume-level validation, will be needed before clinical use. Not applicable; this was a retrospective diagnostic model development and validation study and did not involve a healthcare intervention.