DeepFake for Life Sciences

Abstract

Spheroids are three-dimensional cellular aggregates and one of the most common and versatile way to culture cells in 3D. In order to scale laboratory tests, automated processes are needed, including robust classification. The spheroids need to be sorted with a high accuracy as misclassification can pollute the entire batch of healthy spheroids. The automation often uses deep learning algorithms to sort healthy and unhealthy spheroids. Nevertheless, the spheroids grown under normal conditions exhibit a high-class imbalance. It is estimated that only 1-10% of unhealthy spheroids are contained in one cell culture batch. This imbalance can bias the classification algorithm. Classical data warping augmentations and weighted loss are state-of-the-art methods when dealing with class imbalance. However, they are limited to the extent they can increase the performance of the classifier. Thus, the underrepresented class is oversampled with synthetic images generated with Generative Adversarial Networks (GAN). The results show significant improvement of the classification performance of an imbalanced dataset with this novel method of data augmentation.