Automatic Detection of Rouleau Formations Using Image Processing and Deep Learning

Abstract: Analysis of peripheral blood smears is commonly used to aid physicians in diagnosing diseases, checking for infections and evaluating the function of organs. One object of interest in such analysis is the presence of red blood cell aggregates called rouleaux. CellaVision's product currently automates the traditionally resource-intensive process of blood smear analysis, which leads to improved patient value. As of now, it lacks the function of detecting rouleaux formations. The aim of this thesis project is to investigate the addition of this feature to CellaVision's product. For this aim, we have implemented and compared two object detection architectures, employing image processing and machine learning methods. One of the architectures, called Single Shot MultiBox Detector, has previously been used by CellaVision for similar tasks. The other was tailormade for the task of detecting rouleaux, based on the method of region proposals employed in various other object detection architectures. We have also collected and annotated the data necessary to train and evaluate models using these architectures. Said models have been tested and evaluated on expert-annotated data for different sizes of rouleau formations. We found that the implemented models are capable of detecting rouleaux in blood smear images collected by CellaVision's systems. However, results are not robust enough for use in production, and the defnitive performance of the models requires further investigation by experts. Future work is needed to ensure accurate model evaluation and suffcient performance for use in CellaVision's product.