关键词: 布鲁氏菌,脊柱炎, 结核,脊柱, 深度学习, 图像分割, 诊断,鉴别

Abstract:

Objective: To develop and evaluate a deep learning-based cascaded ensemble system that integrates lesion segmentation and classification models for the intelligent differentiation of tuberculous spondylitis (TS) and brucellar spondylitis (BS), aiming to improve diagnostic accuracy and efficiency in clinical practice. Methods: In this retrospective study, spinal magnetic resonance imaging (MRI) data were collected from 202 patients with pathologically or microbiologically confirmed spondylitis treated at the First People’s Hospital of Kashi Prefecture between January 2021 and January 2025, including 113 TS and 89 BS cases. All patients underwent MRI scans incorporating fat-suppressed T2-weighted imaging (T₂WI-FS) sequences. The proposed end-to-end diagnostic framework combined a U-Net-based lesion segmentation model with ImageNet-pretrained ResNet50 or EfficientNet classification models in a cascade, using both soft and hard voting strategies. Segmentation performance was assessed with Dice coefficient, intersection over union (IoU), sensitivity, specificity, precision, and accuracy on validation and independent test sets. Classification performance was evaluated using accuracy, F1-score, precision, recall, and the area under the receiver operating characteristic curve (AUC). Results: For the lesion segmentation model based on U-Net, on the validation set, the Dice coefficient was 0.851±0.057, IoU was 0.744±0.081, sensitivity was (87.4±8.1) %, specificity was (99.5±0.3) %, precision was (83.8±8.2) %, and accuracy was (99.1±0.4) %. On the test set, the Dice coefficient was 0.835±0.085, IoU was 0.725±0.115, sensitivity was (83.9±10.4) %, specificity was (99.6±0.2) %, precision was (83.8±9.0) %, and accuracy was (99.1±0.4) %. For lesion classification, the ResNet50 model achieved an accuracy of 79.6%, F1-score of 83.8%, precision of 85.3%, recall of 82.5%, and AUC of 0.855 on the validation set; on the test set, it achieved an accuracy of 75.2%, F1-score of 78.6%, precision of 75.7%, recall of 81.7%, and AUC of 0.822. The EfficientNet model showed an accuracy of 79.0%, F1-score of 84.4%, precision of 80.7%, recall of 88.5%, and AUC of 0.852 on the validation set; on the test set, it had an accuracy of 73.2%, F1-score of 78.2%, precision of 71.7%, recall of 86.0%, and AUC of 0.800. In the cascade ensemble system, the ResNet50-based model with soft voting achieved the optimal diagnostic performance on the test set, with an accuracy of 80.4%, F1-score of 83.1%, precision of 78.3%, recall of 88.5%, and AUC of 0.853. Conclusion: The proposed cascaded deep learning system provides an effective solution for differentiating TS from BS. By integrating multimodal MRI radiomic features, it captures subtle microstructural and pathological differences between the two diseases, significantly enhancing diagnostic performance and offering a promising auxiliary tool for clinical decision-making.

Key words: Brucellar, spinal, Tuberculosis, spinal, Deep learning, Image segmentation, Diagnosis, differential