Guided data augmentation by transfer function (GUIDATFUN)

Abstract

Deep Learning models are used in a wide variety of contexts, one of which is the classification of medical images for the diagnosis or detection of deceases. For the models to perform adequately great amounts of data to train them are needed, nonetheless the lack of labeled data in the medical field is noticeable due to the scarcity of medical professionals. To solve this other approaches lean on transfer learning to gather data from different sources but often the distribution between the clusters of data is too different causing accuracy issues for the models. To solve the distribution mismatch this study proposes a scoring base data augmentation policy called GUIDATFUN that measures the relatedness between the source and the target datasets and then a transfer function assigns an augmentation probability to the source images. The approach was tested with four different transfer functions in the context of chest X-ray images binary classification, the results showed that a supervised deep learning model trained with the data generated employing the GUIDATFUN method measured with statistical significance with a higher accuracy in comparison to trained with regular data in the context of domain adaptation for medical images.