Fig. 3. Tuberculosis diagnosis results with gradually increasing unlabeled data.

a Gradually evolving performance with the proposed framework under an increasing amount of data. Compared with the supervised model using the same amount of data with labels, the model trained with the proposed framework without labels showed even better performances. b Gradual attention change of the evolving model for a tuberculosis case. For the exemplified tuberculosis case, the attention of the Vision Transformer (ViT) model gets gradually refined to better catch the target lesions and semantic structures as the model evolves with increasing time T. Data are presented with calculated area under the receiver operating characteristics curves (AUCs) in the study population (center lines) ±95% confidence intervals (CIs) calculated with the DeLong's method (shaded areas). The AUCs of the proposed method and the supervised learning method were compared at each time point T with the DeLong test to evaluate statistical significance, except for the T = initial where the two methods start from the same baseline. * denotes statistically significant (p < 0.050) superiority of the proposed framework. All statistical tests were two-sided. CXR, chest X-ray.