An ROI-Guided Imaging Informatics Framework for Image-Language Differentiation of Tuberculous and Pyogenic...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology
• Infectious Diseases

Background:

• Differentiating tuberculous spondylitis (TB) from pyogenic spondylitis (PS) is crucial for effective treatment but can be challenging using conventional MRI alone.
• Existing diagnostic methods may lack specificity, necessitating advanced analytical tools for accurate patient-level differentiation.

Purpose of the Study:

• To develop and validate an MRI-based image-language collaborative framework for differentiating TB from PS at the patient level.
• To evaluate the diagnostic performance of various AI models, including supervised classifiers and large language models (LLMs), in distinguishing these spinal infections.

Main Methods:

• A retrospective study of 347 patients with confirmed spinal infections (TB or PS) using sagittal T2-weighted MRI.
• An enhanced YOLO model was employed for localizing infection-related regions of interest (ROIs), serving as standardized inputs for AI models.
• Patient-level diagnostic performance was assessed using ROI-guided supervised classifiers (ConvNeXt, Swin Transformer) and LLM inference (zero-shot, few-shot, ROI-guided).

Main Results:

• Vertebral collapse and marked disc involvement were identified as independent imaging features differentiating TB from PS.
• The Swin Transformer achieved an AUC of 0.804 and 71.4% accuracy on an independent test set.
• ROI-guided 10-shot LLM inference demonstrated superior performance with 82.9% accuracy, 81.8% sensitivity, 84.6% specificity, 90.0% precision, and 85.7% F1 score.

Conclusions:

• ROI-guided constrained LLM reasoning provides stable, interpretable, and accurate patient-level differentiation of spinal infections on MRI.
• The developed framework, linking diagnostic output to traceable ROI-level evidence, has the potential to serve as a radiologist-supervised second-reader tool for complex TB/PS cases.