Unsupervised Cross-Domain Feature Extraction Using Autoencoders for Analyzing Single-Cell Images from Acute Myeloid Leukemia Patients

Identification and classification of white blood cells in peripheral blood smears is a key step for the diagnosis of hematological malignancies. Different lab procedures, illumination, staining, and microscope settings are some of the factors resulting in domain shifts which hampers the applicability and reusability of machine learning methods when applied to data collected from different sites. In this thesis, we propose an autoencoder to extract unsupervised cross-domain features on three different datasets of single white blood cells. Using a Mask R-CNN architecture as a first step allows the autoencoder to focus on the relevant white blood cell and eliminate artifacts in the image. We use a simple random forest method to classify the extracted features of the single cells as a way to evaluate the quality of the features extracted by the autoencoder. We show that the random forest classifier trained on only one of the datasets, can perform satisfactorily on the unseen datasets thanks to the rich features extracted by the autoencoder, and in the cross-domain task, it outperforms the published oracle models. According to the results, this unsupervised approach can be employed in more complicated diagnosis and prognosis tasks without the need to add expensive expert labels to unobserved datasets and it is expected that it will improve the applicability and reliability of these systems in different centers and hospitals.