Blood Cell Data Augmentation using Deep Learning Methods

Sammanfattning: In this thesis we aim to improve classification performance on blood cell imagesby using deep learning techniques to augment data. The thesis was conductedat CellaVision, a company providing digital solutions for medical microscopy inthe field of hematology. The goal of CellaVision’s technology is to replace man-ual microscopes used for cell differentials in blood tests with digital microscopesthat perform cell differentials automatically. Classyfying white blood cells is animportant part of this technology and is achieved by using an artificial neuralnetwork. This classifier network requires a great amount of training data inorder to perform well.With the objective to improve the performance of the classifier, we augmenttraining data consisting of blood cell images by generating synthetic data usinga Generative Adversarial Network (GAN). Our goal is to generate images withclose to equal quality of the real images and to use the generated images forclassifier improvement. The results show that the GAN is able to generate im-ages that, apart from some small artefacts, very much resemble the real images,so much that a medical technologist struggled to differentiate them from realimages.In order to generate class specific blood cell images, we implement a versionof the Auxiliary Classifier GAN (AC-GAN), where we use a pre-trained gener-ator and discriminator from a GAN able to produce high quality images. Thegenerator and discriminator are freezed and connected to fully connected lay-ers to be trained. By augmenting the training data with the generated imagesfrom this AC-GAN, classifier performance improved for the majority of classesresulting in an increased F1-score. This leads us to believe that augmentingblood cell image data by using synthetic images is a viable method for classifierperformance improvement.