GAN-Assisted Sample Equalization for CNN-Based Human Peripheral Leukocyte Image Recognition and Classification

Morphology analysis and the count of distinct cells in human peripheral blood are dominant methods to diagnose common blood diseases. In recent years, with the purpose of improving the efficiency of these methods, the convolutional neural network (CNN), a widely used model of artificial intelligence technology, has been utilized in computer-aided blood tests to recognize and classify blood cell images. However, due to the nature of a smaller number of leukocytes in one blood sample, the CNN model is generally trained by insufficient leukocyte images and thus shows lower accuracy on leukocytes. To address the above issue, we propose a generative adversarial network (GAN) assisted framework to increase leukocyte images and further enable the CNN model to improve the performance of human peripheral leukocyte image recognition and classification. We validate the effectiveness of the proposed framework through a comparative performance analysis of four CNN models trained by four image datasets deriving from different sample equalization methods. The numerical results show that the GAN-assisted sample equalization method is superior to making the CNN model converge fast and attain high recognition and classification accuracy.

I. Introduction

The count and morphology of human peripheral leukocytes can reflect a person’s state of health. The detection and classification of leukocytes play a vital role in diagnosing malignant hematologic disease [1]. With the rapid development of artificial intelligence (AI) methods in the medical field, machine learning (ML)-assisted diagnosis based on AI has attracted more and more attention. A large number of techniques for automatic classification of leukocytes based on image processing and convolutional neural network (CNN) have been proposed by researchers at present [2]–[5]. The challenge in the practical application of a classifier based on CNN is that the count of samples in large categories is higher than that in minor categories, which is also called category imbalance. Imbalanced datasets have a significant negative effect on the training of the classifier. The model is desirably biased towards the categories with more samples in the training set but poor in those with fewer samples [6]. Specifically, the natural distribution of the five categories of leukocytes in the human periphery is uneven; for example, the proportion of neutrophils in the adult periphery ranges from 40% - 75%, while the eosinophils are only 0 - 1%. The imbalance and scarcity of leukocyte datasets prevent CNN models from performing satisfactorily on small-category samples [7].