Deep Learning Based Automated Classification And Enumeration Of Blood Cells

Abstract

For blood analysis, precise cell identification is key to successful disease diagnosis and subsequent treatment. In this project, we are developing deep learning models to count and classify blood cells in microscopic images. Our dataset will be sourced from publicly available datasets, collaboration with hematology clinics, and our own image acquisitions. Each image was preprocessed to normalize the staining process, quality check the image, and maintain metadata. Our most recent advances in cell detection have been achieved using the YOLOv10 object detection model. These system modules collectively help in creating a workflow that helps in image processing of blood cells through the GAN and then providing features as inputs to the convolution layers of YOLOv10. Features obtained in this way are used by the detection head in order to produce the bounding boxes and the class probability values for the detected blood cells. The loss function that is used for training the model is the combined loss of binary cross-entropy and bounding box regression loss and is calculated during backpropagation. There are four metrics that would be used for evaluation of the proposed method, namely, precision, recall, F1-score, and mAP. In terms of investigation, the advantage of deep learning is obvious since there is a measurable benefit in all criteria including accuracy and reliability as compared to the older traditional machine learning. More testing experiments are being carried out in order to have models which will be generalizable in larger databases whereby each model will be trained in all classes of the database (majority classification of any one particular cell type using all data).

KEYWORDS: Blood cell classification, deep learning, YOLOv10, microscopic image analysis, object detection, automated hematology, medical image processing.