Validating Noisy Label Learning in Noisy Clinical Datasets

André Aquilina, Emmanouil Papagiannakis

Keywords: colposcopy, learning with noisy labels, medical imaging, deep learning, cervical cancer, histology classification

Abstract: Noise in ground truth labels limits model performance; in response, learning with noisy labels (LNL) has received much attention in recent years. However, most research has been applied to competition datasets where a clean (noiseless) test set is available. A gap exists in applying LNL to practical datasets such as colposcopy, where such clean sets are not available. By synthesizing additional noise, targeted to mimic real-world errors, to the training labels, and using an imperfect test set, we demonstrate that LNL methods outperform traditional learning, thus bridging this gap.

Submission Number: 2