Streamlining tuberculosis detection with foundation model-based weakly supervised transformer.

Area of Science:

• Medical Imaging
• Computational Pathology
• Infectious Disease Diagnostics

Background:

• Tuberculosis (TB) poses a significant global health burden, especially in resource-limited regions.
• Current diagnostic methods, including microscopy and existing deep learning models, face challenges in scalability, cost, and annotation requirements.
• Automated detection of Mycobacterium tuberculosis (MTB) is crucial for timely diagnosis and treatment.

Purpose of the Study:

• To develop a scalable, weakly supervised deep learning framework for MTB detection in microscopy images.
• To leverage a foundation model (UNI) for cross-domain transfer learning in TB diagnostics.
• To reduce the reliance on extensive expert annotations and intensive preprocessing for automated TB detection.

Main Methods:

• Utilized UNI, a foundation model pretrained on pathology images, to encode microscopy images into patch-level embeddings.
• Employed a Transformer encoder for image classification using only image-level labels, eliminating the need for detailed annotations.
• Trained and validated the model on large, diverse datasets to ensure robustness and generalizability.

Main Results:

• Achieved high Precision-Recall Area Under the Curve (PR-AUC) scores ranging from 0.943 to 0.974.
• Demonstrated strong performance and robustness in detecting MTB.
• Successfully minimized preprocessing steps and annotation costs, enhancing scalability.

Conclusions:

• The proposed weakly supervised approach, leveraging foundation models and image-level labels, significantly reduces annotation burden for TB diagnostics.
• This method shows great potential for streamlining TB detection workflows, particularly in resource-limited settings.
• Highlights the feasibility of foundation models for automated TB detection and broader medical imaging applications.