Deep Learning Based Diagnosis of Acute Lymphoblastic Leukemia (B-ALL) using Image Flow Cytometry

Abstract

Acute Lymphoblastic Leukemia (ALL) is among the world’s most fatal diseases which results in the death of thousands of people suffering from ALL annually. Most of the time, procedural delay and delays in diagnosing ALL have a higher chance of an increase in suffering and pain and sometimes, even death can occur. A rapid and effective system of diagnosing ALL is used in this study to help doctors in the early and quick management of patients. The pipeline is composed of different steps and each step uses a deep learning model to classify the blood smears taken from Image Flow Cytometry (IFC). A high-definition image taken from IFC has more details as compared to normally available microscopic images and hence the detection of change in cells can be monitored efficiently. After learning from high-definition images, models like ResNet, a combination of U-Net and ResNet, Vision Transformer (ViT), YOLOv5, and DenseNet have shown exceptional accuracies over a dataset of 3242 images with an overall accuracy of 96.8% is shown during the test run of the pipeline. It also takes a quick runtime ranging from 9 seconds to 50 seconds depending upon the information a cell has to provide. Achieved results show the efficiency of the pipeline as compared to the previous studies done by using either a single algorithm or two algorithms. Furthermore, this pipeline can be trained for other blood diseases as well as for the genomic and phenomic data of cells as It will aid in complementing the results generated by the pipeline.