Classifying Parasitized and Uninfected Malaria Red Blood Cells Using Convolutional-Recurrent Neural Networks

This work aims to classify malaria infected red blood cells from those uninfected using two deep learning approaches. Plasmodium parasite transmitted by a female anopheles’s mosquitoes bite is the main cause of malaria. Commonly, Microbiological analyses by a microscope allows detecting cells infected from a blood sample, followed by a specialist interpretation of results to conclude the diagnosis process. Taking advantage of efficient deep learning approaches applied in computer vision field, the present framework proposes two deep learning architecture based on Convolutional-Recurrent neural Networks to detect accurately malaria infected cells. The first one implements a Convolutional Long Short-Term Memory while the second uses a Convolutional Bidirectional Long Short-Term Memory architecture. A malaria’s public dataset consisting of parasitized and uninfected red blood cell images was used for training and testing the proposed models. The methods developed in this work achieved an accuracy of 99.89% in the detection of malaria-infected red blood cells, without preprocessing data.


I. INTRODUCTION
deaths per day, of which 96% lived in Africa [1]. The most 20 affected by malaria are children under 5 years old, whom 21 represents an 80 % mortality rate in the region according 22 to [1]. 23 Among other malaria pathogens, the Plasmodium falci- 24 parum and Plasmodium vivax attack human red blood cells 25 The associate editor coordinating the review of this manuscript and approving it for publication was Fahmi Khalifa .
(RBCs) causing high fever, cough, chills, headache, nausea, 26 and vomiting, muscle pain and fatigue, transpiration, abdom-27 inal and chest pain, and the death [2], [3]. Untreated malaria 28 patients may develop long-term pneumonia, anemia, yellow 29 fever, respiratory or brain disorders (Cerebral Malaria) [4]. 30 The World Health Organization (WHO) is deploying 31 strategies for the prevention, treatment, elimination, and 32 surveillance of Malaria to face this global pandemic [1], start-33 ing with the diagnosis of the disease. Basically, microscopic 34 diagnosis method detecting Malaria consist in a microbiolog-35 ical analysis using peripheral blood slides [5], [6]. Collection 36 of blood smears for manual microscopic analysis remains 37 an effective method in the diagnosis of malaria, compared 38 to other methods such as polymerase chain reaction (PCR) 39 and in-house tests [7]. Nevertheless, analyses by standard 40 frame and the process computational cost. In this sense, 97 various recent works are focused on the data preprocessing 98 step to perform the malaria detection [20], [21]. 99 To deal with the cell detection accuracy and the diag-100 nostic time frame challenges, this work aims to classify 101 malaria red blood cells without a preprocessing step, using 102 two convolutional-recurrent neural networks and the public 103 malaria dataset from the National Library of Medicine [22]. 104 The main contributions of this article are summarized as 105 follows, 106 1) Because of the implemented learning acceleration, the 107 developed models all converge with only 100 training 108 iterations.  3) Comparative results of malaria red blood cells classi-112 fication achieved using convolutional-recurrent neural 113 networks.

114
As novelty, this work investigates on recurrent neural net-115 works contribution detecting malaria red blood cells, empha-116 sizing on data splitting, and images sizing to evaluate the 117 processing time, since various related works are based on the 118 basic CNN or its variants.

119
The paper is organized as follows. Section II presents 120 the referred malaria dataset, artificial neural architectures 121 and methods developed in this framework while the results 122 encountered are reported and discussed in Section III. Finally, 123 Section IV gives the paper conclusion and projections for the 124 future works.

126
In making the binary-class classification of the parasitized 127 and uninfected cells in this paper, two neural network archi-128 tectures are implemented:  Figure 1 illustrates the step-by-step diagram of methods 140 implemented with the same public dataset in this paper. The 141 procedure's main purpose is to evaluate results obtained with 142 different network architectures, classifying malaria unin-143 fected and parasitized cells.   training and test subsets, as illustrated in Table 1. number of memory unit blocks, the number of memory units 175 per block, and how the weights are initialized.
176 Figure 3 illustrates the LSTM memory unit structure, cen-177 tered around the input, forget, and output gates. At time t, the input gate upshot is

186
According to the input x t , the precedent state h t−1 , and the 187 cell state c t−1 , the output of the output gate is, 204 205 where the denominator represents the normalization expres-206 sion for the probability distribution and x i the feature map to 207 be classified in its respective class.

208
All the convolutional layers are activated by the Rec-209 tifed Linear Activation Function (ReLU), offering gradients 210 between 0 and 1 and defined as, The second approach developed in this work contemplates 218 processing data with a CNN-BiLSTM network, as illustrated 219 in Figure 5. A BiLSTM network is constituted by two LSTM 220 blocks working each one in the forward and backward direc-221 tion [26]. Thus, the BiLSTM architecture results in faster 222 convergence and better classification accuracy than the basic 223 LSTM [27]. For a given data processing t th time, the BiLSTM 224 unit considers the past and the future unit state, as shown in 225 Figure 6.
226 Therefore, the t th BiLSTM output merges forward and 227 backward outputs as, where − − → For t and ← −− − Back t are the t th LSTM memory block out-230 puts, and ⊕ is the element-wise sum.

231
The implemented BiLSTM network uses two LSTM 232 blocks each one configured with 64 cell units, followed by 233 two fully-connected layers of 512 and 2 neurons, respec-234 tively (see Figure 5). 4751872 and 98816 parameters are used 235 VOLUME 10, 2022              Tables 3 and 4 support the observation 329 that the large-sized images produced better performance in 330 their processing than the small-sized. This is understood by 331 the large quantity of pixels necessary for the features learning 332 by the CNN block. Splitting the database into training and 333 testing data on top carries out a significant role. The same 334 tables report performances for the split-1 which considers 335 90 % images for training and 10 % for testing than for the 336 other partitions. The split-5 partition produced the lowest 337 results of all splits (98 % and 97.93 %), despite the considered 338 image sizes.

339
Finally, the processing time that the classifier considers to 340 decide whether a selected sample is infected or not, should be 341 added to these observations. Large-sized images take longer 342 (despite < 1 s) than small-sized, as revealed in Table 5. 343 Therefore, comparing the two proposed architecture models 344 for the 96 × 96 image size, The CNN-BiLSTM model took 345 longer than the CNN-LSTM classifier to detect a data sample, 346 justifiable by the data twice-processing operated by the BiL-347 STM. Obviously, the results presented in Table 5 correspond 348 to the used computer resources and can be improved using a 349 more powerful computer. Therefore, the reasonable compro-350 mise should be negotiated amidst these three observations for 351 a usage application, give-and-take between the size of images 352 -data partition -processing time.

353
The literature is flourishing with scientific research relat-354 ing to the detection of cells infected and uninfected by malaria 355 employing the NLM database and various deep learning 356 approaches. Among all these specific techniques efficiently 357 deployed, the present work contributes in meaningfully com-358 paring results of two combined architectures of convolutional 359 and recurrent neural networks, without preprocessing data. 360 Table 6 presents the comparison between the results of recent 361 state-of-the-art works and those achieved in this framework. 362 Recently, [29] implemented a customized CNN to detect 363 peripheral malarial parasites in blood Smears, achiev-364 ing an accuracy of 98 % with 27500 images resized at 365 128 × 128 pixel size. Next, their used morphological filters 366 and a a fine-tuned pretrained CNN model to classify par-367 asitized and uninfected malaria cells, obtaining an average 368 accuracy of 98.34 ± 0.51%.

369
Likewise, [20] proposed a model based on Barnacles Mat-370 ing Optimizer with Deep Transfer Learning (BMODTL) to 371 detect and classify malaria parasite. Concretely, they pre-372 processed data using the Gaussian filter (GF) and Graph 373 cuts (GC) segmentation technique to locate the contami-374 nated areas in the blood smear images, before classifying 375 malaria parasites. An accuracy of 99.04 % was achieved 376 with their model. In another work, the malaria detection 377 based on offline and Web application was implemented using 378 various approaches contemplating distillation, morphology, 379 Autoencoder, CNN -Support Vector Machine (SVM) or 380 K-Nearest Neighbors (KNN) [12]. They reported an accu-381 racy of 99.23 processing malaria RBC images of 28 × 28, 382 32 × 32, and 64 × 64 pixel size. Splitting data in 80 % 383 97356 VOLUME 10, 2022    Democratic Republic of Congo, in 2008, and the 535 master's degree in electrical engineering (instru-536 mentation and digital systems) and the Ph.D. 537 degree in electrical engineering from the Univer-538 sity of Guanajuato, in 2018 and 2022, respectively. 539 His research interests include biological signal processing, robotics, arti-540 ficial intelligence, embedded systems, image processing, brain-computer 541 interfaces, deep learning, telecommunications, and bio-inspired systems. 542 Currently, he is collaborating with the Telematic and Digital Signal Process-543 ing Research Groups, Department of Electronics Engineering, University of 544 Guanajuato, Mexico. 545 CARLOS HUGO GARCÍA-CAPULÍN received 546 the B.S. degree in electronics engineering from the 547 Instituto Tecnológico de Ciudad Madero, in 1998, 548 the master's degree in electrical engineering from 549 the Universidad de Guanajuato, in 2004, and the 550 Ph.D. degree in optics from the Centro de Inves-551 tigaciones en Óptica, A. C., in 2014. He is cur-552 rently a full-time Professor at the Department of 553 Electronics, Universidad de Guanajuato. He has 554 authored several journals and conferences pro-555 ceedings papers. His research interests include computational intelligence, 556 evolutionary algorithms, bio-inspired computing and robotics, reconfig-557 urable electronics, and parallel computing.