Detecting red blood cells and platelets in blood smears using a single multi-class object detector

Sammanfattning

Blood analysis is an integral part of diagnostic medicine and used in most medical fields. The concentrations of red blood cells and platelets, and ratio between these, are used for diagnosing several diseases. CellaVision develops machines and software for automatically capturing images of blood sample smears and detecting its cellular contents. The company currently has separate algorithms for detecting red blood cells and platelets. The aim of this master’s thesis is to develop an object detection model that simultaneously detects these blood cell types, with suciently high accuracy and speed for use on CellaVision systems. The object detection model YOLOv5 was selected to develop the detector. Several model parameters, hyper parameters and improvement techniques were evaluated, and the ones maximising the performance were selected. Image augmentations proved to be the most important improvement technique added during development in terms of detection performance. Pseudo labelling was successfully used for creating a large training data set. The results obtained show that it is possible to combine red blood cell and platelet detections in a single object detector with higher speed than when using separate algorithms. Comparing performance with the current individual algorithms, platelet detection was almost as good and red blood cell counting significantly better when using the detector developed during this thesis.