Comparative analysis of neural network Models for peripheral blood cell image classification

Abstract

It’s quite difficult to fathom that the future of medicine and diagnosis is more dependent on, and much more likely to be dictated by the growth of technology than the quality of doctors. One of the deadliest diseases today that is ubiquitous all around the world is Leukemia. As deadly as it is, it is one of the most difficult diseases to diagnose. One of the biggest challenges is to identify cells as being affected by this condition and this requires highly trained medical professionals to accomplish such tasks. In this paper we have trained four different image processing models to recognize and identify such cancerous cells. We have used more than 9000 images to do so. After the training processes were over, we evaluated the success of these individual models to assess the difference in their final accuracies, we should bear in mind that these images are rather different than what a usual image-based dataset would look like in that the images are quite similar despite being of different classes. We have used the following models: YOLOv5 (precision = 0.82), CNN (precision = 0.74), YOLOv7 (precision = 0.52), EfficientNet (accuracy = 0.89). From this we can clearly agree upon the dominance of EfficientNet over all the other models.