A Joint Training and Confidence Calibration Procedure That is Robust to Label Noise
Coby Penso, Jacob Goldberger · 2024
Manually annotated medical imaging data tend to have unreliable labels due to the complexity of the medical data and the considerable variability across experts. Noisy data can pose a significant challenge when it comes to learning the model’s parameters and calibrating its predictive confidence. This study presents a joint training and confidence calibration procedure that is robust to label noise. The method is based on estimating the noise level as part of a noise-robust training procedure. The estimated noise level is then used to modify the computed network accuracy on the noisy validation set which is required by the calibration procedure. We demonstrate that, despite the unreliable labels, we can still achieve calibration results similar to those obtained by a procedure using data with noise-free labels.