Fig. 5: Tuberculosis diagnosis results with real-world data collection and label corruption.
From: Self-evolving vision transformer for chest X-ray diagnosis through knowledge distillation
a Even after adding unseen class data that are commonly encountered in clinics, the performance was stably improved with increasing time T, even though these other class data were not included for the training of the initial model. b In the simulation for label corruption, the model trained with the proposed framework was not compromised, while the model trained with supervised learning using corrupted labels showed significant deterioration in performance. Data are presented with calculated area under the receiver operating characteristics curves (AUCs) in the study population (center lines) ±95% confidence intervals calculated with the DeLong's method (shaded areas). The AUCs of the proposed method and the supervised learning method with and without corruptions were compared at each time point T with the DeLong test to evaluate statistical significance, except for the T = initial where all methods start from the same baseline. * denotes statistically significant (p < 0.050) superiority of the proposed framework. All statistical tests were two-sided.
