Deep-Learning-Based Acute Leukemia Classification Using Imaging Flow Cytometry and Morphology
Jakkrich Laosai, Kosin Chamnongthai · 2018
This paper presents to classify a blood cell into one type out of 11 subtypes of leukemia, which is important for diagnosis. Originally, the classification requires a huge training dataset collected from their races, which depend upon genes. Due to the recent discovery of imaging flow cytometry, it becomes possible to manually classify acute leukemia among different people and races by using a small training dataset. This paper proposes a method of automatic acute leukemia classification using imaging flow cytometry and morphology. The method utilizes intensity and morphology, which are the features from the imaging flow cytometry and blood smear, respectively, to classify a blood cell into one out of three groups; healthy, ALL, and AML ones in the coarse step, and categorizes it into a subtype in the group by deep learning in the fine step. The evaluation has been performed by some samples between Thai and American people, and the results show better accuracy compared with conventional methods. The computer simulations show the proposed system robustly segments and classifies Acute Leukemia based on complete microscopic blood images. We have obtained an accuracy of 99% which is a 4% improvement compared with the conventional method.