BoMaCNet: A convolutional neural network model to detect bone marrow cell cytology

Abstract

Bone Marrow is responsible for the creation of all the different types of blood cells in the human body and helps us to identify various types of bone marrow cell disorders. Therefore it is necessary to correctly identify and classify the different types of cells. Conducting different pathological and blood tests may take some time. Applying a Deep Neural Network (DNN) for blood cell detection allows us to quickly classify the call types, which further enables us to identify multiple types of blood cells simultaneously from the same sample. Not only does this save us the time needed for cell classification but also removes the possibility of human error as an automated system can deliver more precise and instantaneous results than a hematologist or pathologist. Machine Learning algorithms are capable of solving these problems quite easily. With that in mind, we propose a CNN-based architecture named BoMaCNet, which is capable of detecting and classifying bone marrow cell images quickly and accurately. Our CNN model takes 96000 images in total, which are then split into training, testing, and validation. Six common types of bone marrow cells (Artefact, Blast, Erythroblast, Lymphocyte, Segmented Neutrophil and Promyelocyte) are chosen for this research. Our entire data set was split into three parts 80% was kept for training, 10% was kept for validation and 10% was used for testing. For testing, 1600 instances of each label were used. Our model was able to produce the highest by far results on the used dataset by achieving an overall accuracy of 95.71%. With 95.71% accuracy in training and 93.06% accuracy in validation along with achieving an impressive mean average F-1 score of 0.93, we were able to achieve exceptional results.