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

The detection of tuberculosis lesions in chest X-rays is increasingly critical but remains hindered by sparse annotations and significant class imbalance within medical imaging datasets. This study addresses these limitations by proposing an intrinsic motivation-based exploration framework, leveraging curiosity-driven algorithms to encourage artificial intelligence models to autonomously seek and interpret novel radiological cues of tuberculosis. By introducing learned novelty bonuses as intrinsic rewards, the approach enables deep models to discover rare or unannotated lesion states that would otherwise be overlooked in conventional supervised learning. Experiments conducted on a large-scale chest X-ray dataset with incomplete labels demonstrate significant improvements in exploration coverage, recall of rare lesion regions, intersection over union for lesion detection, and overall F1-score compared to state-of-the-art baselines. Notably, the model achieved higher average precision and substantially reduced annotation gaps without relying on exhaustive manual labeling, indicating robust generalization to challenging and unseen cases. The primary contribution of this work is the advancement of autonomous, curiosity-driven methods that enhance lesion discovery and diagnostic reliability in resource-constrained global healthcare environments.