Intrinsic motivation-based exploration for enhancing tuberculosis lesion discovery in sparse annotation chest X-ray...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology and Diagnostic Imaging
• Computational Pathology

Background:

• Tuberculosis lesion detection in chest X-rays is critical but challenged by limited annotations and imbalanced data.
• Existing supervised methods often miss rare or unannotated tuberculosis indicators.

Purpose of the Study:

• To develop an intrinsic motivation-based exploration framework for autonomous discovery of novel radiological cues of tuberculosis.
• To enhance the performance of AI models in detecting tuberculosis lesions, particularly rare or underrepresented ones.

Main Methods:

• Proposed an intrinsic motivation-based exploration framework using curiosity-driven algorithms.
• Introduced learned novelty bonuses as intrinsic rewards for deep learning models.
• Evaluated the framework on a large-scale chest X-ray dataset with incomplete labels.

Main Results:

• Achieved significant improvements in exploration coverage and recall of rare lesion regions.
• Demonstrated enhanced intersection over union for lesion detection and a higher overall F1-score compared to baselines.
• Showcased higher average precision and reduced annotation gaps without extensive manual labeling.

Conclusions:

• The curiosity-driven approach effectively enhances autonomous lesion discovery in chest X-rays.
• This method improves diagnostic reliability for tuberculosis, especially in resource-limited settings.
• The framework shows robust generalization to challenging and unseen cases, reducing reliance on exhaustive manual annotation.