Ebba Blomdahl och Simon Moor presenterar sitt examensarbete "White Blood Cell Classification with Point Light Source Images"

Abstract:

Identification and classification of white blood cells is a well-studied area and a vital tool for diagnosing different diseases. Classical microscopy imaging only captures a two-dimensional snap-shot of the cells. Resulting, the images might lack important aspects useful for classification, for example spacial information. Point Light Source (PLS) is a microscopical technique which captures a sequence of images from different angles of a blood sample. In this thesis classification of white blood cells using PLS images was investigated.

Images of cells were collected using both traditional microscopy and a PLS setup. The resulting dataset contained images and classes. Since this is a rather limited dataset experiments were carried out to try to handle both the class imbalance and few datapoints.

Two final models using the most promising hyper-parameters were trained for the PLS images, the best model obtained an average accuracy on the external test set of 80.4% and an F-score of 0.652. The PLS model was compared to the best model for classical images with x resolution, to investigate if using PLS images for classification could improve performance. The average accuracy of our best model using classical x images was 83.7% and the F-score was 0.678. The findings in this thesis show that dataset size is instrumental to classification performance for difficult classes and indicate that classification using PLS images may perform at a similar level as classical x images, however further research using larger datasets is necessary in order to reach a conclusion.