Guided Data Augmentation by Transfer Function using Uncertainty Scores for Medical Image Classification

Deep Learning architectures are widely used to deal with different types of unstructured data (images, text, sound, etc.). However, its successful implementation depends upon the availability of large labeled datasets, avoid model over-fitting. To tackle such challenge, a number of regularization approaches have been developed, among the most popular transfer learning and data augmentation. However, in a transfer learning setting, often the distribution of the source dataset can be too different to the distribution of the target dataset. In this work, we propose a simple methodology to alleviate such distribution mismatch, by using a scoring based approach to augment the data. The scoring consists in measuring the likelihood of the target dataset, according to the distribution in the source data. These scores are fed into a transfer function which computes the probability of augmenting each observation in the source dataset. We test four simple different transfer functions in the context of chest X-ray images binary classification.