Semantic Segmentation Based Method for Platelet Quantification in Dense​ Aggregates​

Abstract

Identifying different blood components such as red blood cells, different types of white blood cells, and platelets is an important tool for health practitioners. A local company has developed a lens-less near-field microscope that can take large images of blood samples. These images are analysed with a segmentation and classification neural network to count the different components. One of the most challenging components is thereby the class of platelets, which are small components in our blood that form clots and prevents bleeding. We re-implemented the current production system to be able to experiment with different solutions and tested recent architectures such as SegFormer and MAResUNet.

Furthermore, in this thesis we report on the development of methods to improve platelet counts. In the images of blood samples, platelets are either appearing as small single entities or as aggregates where several platelets are bound together. The approach we developed uses a two-stage process where at first we train a network to classify single platelets and platelet aggregates as separate classes. We then tested several methods to separately estimate the number of platelets in the aggregate state. This included taking the number of pixels and the average size of platelets in pixel units into account. Since platelets have a characteristic signature with higher intensity in the center than the background, we also count the number of intensity peaks. We find that treating singular platelets and platelet aggregates separately better captures overall platelet counts, improving base system.