Adaptive Cycle-consistent Adversarial Network for Malaria Blood Cell Image Synthetization

Malaria is a tropical infectious disease that causes massive global deaths. The convolution neural network (CNN) models can theoretically classify the malaria infected blood cells from normal cells, but they are vulnerable to network attacks even with simple uniform noise. A typical drawback of CNN is that the algorithm cannot properly capture the meaningful patterns with clinical significance. We propose a novel adaptive cycle-consistent adversarial network (Ad Cycle GAN) to synthesize malaria significant patterns based on a homogeneous image template with randomness. The Ad Cycle GAN model consists of a pretrained convolutional variational autoencoder (CVAE) and conventional cycle-consistent adversarial network (Cycle GAN). The CVAE model is trained by a large, segmented blood cell dataset with 27,578 images. The model is optimized for 120 epochs. The CVAE is pipelined to a conventional Cycle GAN model with two generator-discriminator combinations. The real malaria positive images are at first sent to the pretrained CVAE to generate template images for the adversarial optimization with the real images. Therefore, the optimization process is to use generator G to convert the CVAE generated images from the synthetic domain (X) to the real malaria positive image domain (Y), then use generator F to convert the real malaria positive images from the real positive image domain (Y) to the CVAE synthetic image domain (X). The total generator loss is composed of adversarial loss, cycle loss, and identity loss, all loss terms are computed by least squared loss. The Ad Cycle GAN architecture is optimized by 150 epochs. When using a pretrained classifier to differentiate the real and synthetic malaria positive image, 99.61% of the real images from the real image set are accurately recognized, compared to 86.6% of the synthetic images are accurately classified. The average score of Frechet Inception Distance (FID) of the generated images by the Ad Cycle GAN is 0.0053 (Std=0.0004). By human eye observation, the Ad Cycle GAN generated images have reasonable fidelity as real blood cells with meaningful pathological patterns that properly mimics real malaria infected blood cells. The proposed Ad Cycle model can generate synthetic malaria infected blood cell images to successfully optimize the deep neural network model for high classification accuracy. We conclude that the new Ad Cycle GAN model can generate high quality malaria infected blood cell images with good diversity.

I. Introduction

Malaria is a tropical infectious disease caused by the infection of plasmodia. The microorganism is usually transmitted by mosquito bites and parasitizes in human blood cells (particularly red blood cells). According to the statistics by WHO, there are 409,000 deaths related to malaria in 2019 and the total death toll is accumulated to 7.6 million since 2000[1]. The standard malaria fast screening diagnosis technique is microscopic malaria infected blood cell counting by medical professionals. This method is inefficient because it is not only time and labor consuming but also highly affected by individual expertise. To solve this problem automatic image processing technology has been applied to malaria diagnosis since 2005[2]. In 2016, our team developed a convolution neural network (CNN) with 6 convolutional layers for classification of malaria infected blood cells. The CNN was trained with a dataset with 27,578 blood cell images (ration: 1:1) and the average accuracy is 97.37%[3]. The following studies also report extremely high classification accuracy[4], [5]. However, these results are all achieved based on a large, well annotated dataset for training the CNN models. In most cases, big annotated medical image datasets are difficult to acquired. If the medical images are annotated by non-medical persons, the quality of the image data is suspicious due to the lack of expertise. Therefore, we should seek for a solution to minimize the human expertise intervention to the deep neural network (DNN) optimization.