Multimodal-Imaging-Based Interpretable Deep Learning Framework for Distinguishing Brucella from Tuberculosis...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Brucella spondylitis (BS) and tuberculosis spondylitis (TS) present similar clinical and imaging features, making differentiation challenging.
• Accurate early diagnosis is crucial as treatment protocols for BS and TS differ significantly.

Purpose of the Study:

• To develop and evaluate a deep learning framework utilizing multimodal computed tomography (CT) and magnetic resonance imaging (MRI) data for distinguishing between BS and TS.
• To improve diagnostic accuracy and efficiency in differentiating these two conditions.

Main Methods:

• Multimodal imaging data (CT, T1WI, T2WI, T2WI FSE) were collected from two centers.
• Image preprocessing involved ROI segmentation, normalization, and augmentation.
• A GoogleNet-based deep learning model was trained and validated against human radiologists using accuracy, sensitivity, and AUC metrics.

Main Results:

• The GoogleNet model achieved high AUC values (95.97% training, 91.24% test, 81.25% external validation), outperforming other deep learning architectures.
• GoogleNet demonstrated superior diagnostic accuracy and speed compared to radiologists (AUC 88.01%).
• Grad-Cam visualization effectively localized lesions, enhancing model interpretability.

Conclusions:

• A multimodal imaging deep learning model can effectively differentiate between tuberculosis spondylitis and Brucella spondylitis.
• Deep learning eliminates the need for manual feature engineering, offering significant potential for clinical application.
• AI-driven diagnostic tools can enhance accuracy and efficiency in clinical practice.