Classification of spinal tuberculous infection, pyogenic infection and spinal metastasis from magnetic resonance...

Differentiating between spinal tuberculosis, pyogenic (bacterial) spondylitis and spinal metastasis remains a major diagnostic challenge because their radiological features often overlap. Delayed or incorrect diagnosis may lead to inappropriate treatment, permanent disability or death.

To develop and evaluate deep learning models for automated classification of spinal tuberculosis, pyogenic infection, and spinal metastasis using magnetic resonance imaging (MRI).

T2-weighted sagittal MRI scans from 120 patients (40 per disease group) with pathologically or microbiologically confirmed diagnoses between 2014 and 2019 were retrospectively analyzed. Lesion regions were manually annotated by radiologists, and data were split into 80% training and 20% testing sets at the patient level. Extensive data augmentation (rotation ± 5°, zoom 1.1-1.2×, shearing ± 5°, grid distortion 2 × 2) was applied to mitigate overfitting. Three models were trained and compared: (1) a single-layer perceptron baseline, (2) a custom dense neural network (2 × 1024 neurons), and (3) pre-trained convolutional neural networks (ResNet50, VGG16, InceptionV3). Model performance was evaluated using accuracy, precision, recall, and F1-score on both whole and segmented images.

After augmentation, 1,000 synthetic samples were generated per class. The baseline model achieved 27-33% accuracy, whereas the dense and pre-trained models achieved 98-100% accuracy on the test set. Although pre-trained networks demonstrated marginally higher performance, the difference compared with the dense model was not statistically significant. Activation heatmaps revealed inconsistent localization of attention regions, suggesting potential overfitting and limitations in visualization interpretability.

Deep learning models demonstrated strong potential in distinguishing between spinal tuberculosis, bacterial spondylitis, and spinal metastasis on MRI. However, the near-perfect performance likely reflects dataset homogeneity and augmentation effects rather than full generalization. External, multi-center validation and improved interpretability methods (e.g., Grad-CAM) are warranted to confirm clinical applicability and ensure reliable decision support for radiologists.