key: cord-0850246-0oz4fgm2 authors: Zhang, Chi; Jiang, Hao; Jiang, Hanlin; Xi, Hui; Chen, Baodong; Liu, Yubing; Juhas, Mario; Li, Junyi; Zhang, Yang title: Deep learning for microscopic examination of protozoan parasites date: 2022-02-11 journal: Comput Struct Biotechnol J DOI: 10.1016/j.csbj.2022.02.005 sha: ed8a78e7863a03b8c1d4b08c0f36e33e3a9069f4 doc_id: 850246 cord_uid: 0oz4fgm2 The infectious and parasitic diseases represent a major threat to public health and are among the main causes of morbidity and mortality. The complex and divergent life cycles of parasites present major difficulties associated with the diagnosis of these organisms by microscopic examination. Deep learning has shown extraordinary performance in biomedical image analysis including various parasites diagnosis in the past few years. Here we summarize advances of deep learning in the field of protozoan parasites microscopic examination, focusing on publicly available microscopic image datasets of protozoan parasites. In the end, we summarize the challenges and future trends, which deep learning faces in protozoan parasite diagnosis. Parasitic diseases have far-reaching consequences on public health globally. According to World Health Organization's Global Health Estimates in 2020 [1] , infectious and parasitic diseases are among the main causes of mortality in Africa. There are three main classes of parasites that can cause disease in humans: protozoa, helminths, and ectoparasites. Most parasitic protozoan in humans range less than 50 lm in size [2] . Due to the smaller size, protozoan parasites present a greater challenge to diagnosis from microscopic examination. Best known disease-causing protozoan parasites include Plasmodium, Trypanosome, Babesia, Toxoplasma, Leishmania, and Trichomonad. Malaria is caused by Plasmodium, which destroys the red blood cells, thus causing fatal disease. According to the World Health Organization (WHO) , malaria deaths has increased to an estimated 627,000 (±5%) due to disruptions during the COVID-19 pandemic [3] . Trypanosome is the causative agent of trypanosomiasis, which threatens the lives of people in sub-Saharan African countries and is transmitted by the tsetse fly [4] . Chagas Disease is a tropical parasitic disease in Latin America caused by Trypanosoma cruzi (T. cruzi) that leads to approximately 14,000 annual deaths [5] . Babesiosis, a malaria-like parasitic disease, affects not only a wide range of domestic and wild animals, but occasionally also people [6] . In the past 20-30 years, the United States has reported increasing numbers of blood transfusion-transmitted cases of Babesia, making Babesia infections a major public health problem [7] . Toxoplasma is the causative agent of toxoplasmosis that infects over a third of the world's population [8] . Toxoplasma can reach various parts of the body via the bloodstream, leading to a decreased immunity and various symptoms by destroying the brain, heart, and fundus of the eyes [9] . Most infected people have either no symptoms or mild symptoms, but Toxoplasma gondii (T. gondii) infection during pregnancy can cause severe nerve damage and even fetal death [10] . Leishmania causes leishmaniasis, which is a neglected tropical disease and transmitted by the female phlebotomine sandflies, affecting more than 700,000 people every year [11] . Trichomonad parasites in the genitourinary system, intestines and oral cavity cause human diseases, collectively referred to as trichomoniasis [12] . Trichomoniasis is highly epidemic in women, can cause poor perinatal outcomes and infertility [13] . Each year approximately 3.1 million women are infected with Trichomonas vaginalis in United States [14] . Mass drug administration programs have significantly reduced the disease burden [15] . To administer the correct drugs and to avoid drug abuse requires first a reliable and efficient diagnosis. Microscopy is one of the most commonly used methods for the diagnosing of the parasitic diseases [16, 17] . However, microscopy-based parasite identification and quantification is challenging, time-consuming and labor-intensive, requires microscope and well-trained researchers [19] [20] . Parasite examination in microscopic images is challenging for human experts because of the variations and uncertainties in the shape, density and staining color of the parasites [21] . Human infections caused by parasites usually occur in developing and undeveloped countries, which often suffer from a lack of good medical conditions and well-trained microscope operators [22] . Another disadvantage of the microscopic examination of some parasites is the relatively low sensitivity, particularly at low parasite levels [20] . With the increase of computing power and data availability, artificial intelligence has been successfully employed in many applications, such as face recognition, natural language processing and biomedical image analysis [24, 24, 64] . Biomedical image analysis refers to the characterization, recognition and understanding of biomedical images, with the aid of the prior knowledge in large amount of data and the efficient extraction of data features [25] . As a subfield of machine learning, deep learning uses multi-layer neural networks to extract and characterize the features of input data, which can be divided into supervised, weakly supervised and unsupervised models. As shown in Fig. 1 , the main difference between supervised and unsupervised learning is whether the data provided to the model is labeled. Labeled data is usually difficult to obtain, thus requiring labeling personnel, special equipment and additional expenses. Relatively speaking, unlabeled data is cheap and easy to obtain. Deep learning has been applied in the life and health sciences with the aim to improve the accuracy and efficiency of clinical diagnosis [26] , and to discover novel biomarkers. From the perspective of biomedical image datasets, neural network with convolutional neural network (CNN) can extract the features of the images of these diseases [28, 28] . Furthermore, deep learning has played an important role in the analysis of a wide variety of microorganisms, including bacteria, fungi, parasites, and viruses [29] . Due to its high speed, accuracy, flexibility and low cost [20, 30] , deep learning has been widely applied in various microscopic image analyses, including microscopic examination of protozoan parasites. This analysis includes detection, segmentation, tracking The main difference between supervised learning and unsupervised learning is whether the training dataset is labeled. Datasets should be clean and representative with balanced number of categories and statistically independent. After multiple epochs of training, the loss function converge to the optimal value. Then the best performing model can be used for predicting the infected and uninfected samples from the testing dataset. and classification. Detection is used to obtain the distribution and location of individual parasites, clusters and hosts [32, 33] . Classification is used for multi-parasites identification in the case of mixed infections [35, 36] . Segmentation is used to observe the edges of parasites morphologically [36] . Tracking is a unique task for the dynamic observation of motile parasites [37] . The detection model represented by You Only Look Once (YOLO) series has been improved to detect parasites [38] . CNN-based models are usually used for parasite classification [39] . U-Net-based model is widely used for segmentation [36, 40] . In clinical scenarios of these algorithms, researchers use methods such as transfer learning [20, 34] and GAN-based model [19] to develop weakly supervised learning and unsupervised learning [31, 42] , which reduce the volume of annotation needed. Notably, smartphones play an important role in developing parasite diagnosis algorithms, thus greatly reducing the need for well-trained microscope operators [38] . Plasmodium infects humans through mosquito bites. The malaria parasite Plasmodium goes through two life stages in the human host, a clinically silent liver stage followed by a blood stage with clinical manifestation of the malaria disease [42] . Therefore, the examination of the blood smear is essential for the diagnosis of malaria. [43] developed a two-stage method for detecting and counting red blood cells (RBCs) in thin blood smear microscopy images. In the first stage, U-Net is used to superpixel or cellcluster segmentation. Then, Faster R-CNN is used to detect small cell objects in a connected component cluster. Due to the foreground cell-cluster masks from U-Net adaptively guide the detection stage, this work achieved higher true positive and lower false alarm rates. Inspired by residual and dense networks and the combined attention mechanism, [44] proposed a miniature and effective CNN named Attentive Dense Circular Net (ADCN) for classifying Red Blood Cells in Malaria. It is worth noting that ADCN combines the attention mechanism, which can make CNN focus on critical factors. It has a patient-level accuracy of 97.47% in infected RBCs classification task. Although great achievements have been made in the accuracy of the above algorithms, their application in most of the developing and undeveloped countries with high incidence of Malaria diseases remains difficult, due to their inconvenience, complexity and need for data annotation. Supervised learning usually requires extensive annotation and is therefore time-consuming and laborious, microscopic image analysis of parasites requires progress in the development of weakly supervised and unsupervised learning strategies. Unlike other weakly supervised algorithms detecting specific objects by using image-level labels, [41] proposed Multiple Objects Features Fusion (MOFF), which is based on the fusion of the convolutional features of multiple objects. This study successfully diagnosed malaria in one hundred high resolution thick blood film (TBF) fields of view by using sample-level labels associated with multiple image fields. In evaluation of automated malaria diagnosis in blood films, MOFF outperforms the existing supervised detection in TBF dataset. [30] proposed a graph convolutional network (GCN) based model to recognize various stages of malaria parasites in blood smear images. This model based on unsupervised learning is the first application of GCN on recognition of malaria parasite image. It successfully recognizes multiple stages of malaria parasites with a patch-level accuracy of 95.4 ± 0.05% in the P. vivax larger-scale dataset without any image label, which consists of 13,780 testing images with an equal quantity of infected and uninfected images. However, this method has only been evaluated on multiple stages of Plasmodium, and its recognition performance for other multistage parasites has not been validated. In addition, other studies have tried to solve the scarcity of highly skilled microscopists in undeveloped countries and remote areas of developing countries by focusing on the diagnosis of malaria using smartphones [38, 45, 46] . [46] could be among the first to apply deep learning to detect parasites in thick smears and run on smartphones. [45] developed an Android smartphone application for the automated malaria parasites detection method based on CNN architecture. Smartphone is attached to the eyepiece of microscope to obtain the target image. Images are caught with 100x magnification in RGB color space with pixel resolution of 3024 Â 4032. Then, Iterative Global Minimum Screening (IGMS) is used to find the parasite candidates. Finally, the classification and detection of Plasmodium cells is achieved by running the customized CNN, and the parasite counting results are recorded in the patient database and displayed in the user interface of the mobilephone applications. [38] reached the state-of-the-art in the detection of small objects such as Plasmodium by optimizing YOLO3 and YOLO4. YOLO3 and YOLO4 are state-of-art detectors with simple network structure and great generalization capability, but their drawback is that they are not optimized for detecting small objects. [38] modified the architecture of YOLO3 and YOLO4 to increase their applicability in detecting malaria parasites images captured by smartphone camera over the microscope eyepiece. In this study, feature scales were increased and detection layers were added to enhance capability and speed of detecting small objects. They applied the optimized models to detect malaria parasites from 608 Â 608 images, which were captured by smartphone over microscope eyepiece with 1000x magnification. This YOLO based architecture models were shown to have obvious advantages, including high performance of mean average precision (mAP), compared to existing models such as single shot detector (SSD) and two-stage detector Faster R-CNN for P. falciparum detection [38] . In the absence of treatment, T. gondii infection is fatal to human infants and people with the weakened immunity, such as AIDS patients [47] . [19] presented a transfer learning-based method for T. gondii microscopic image recognition without any data annotation. The connection between micro and macro objects in this study established by embedding fuzzy C-means cluster into Cycle-GAN. The algorithm achieved patch-level accuracy of 93.1% and 94.0% in the detection of T. gondii microscopic image dataset with magnification of 400x and 1000x, respectively. [40] applied U-net to segment T. cruzi and detected T. cruzi amstigotes nests from histopathological images, thus decreasing the workload of pathologists. The proposed model achieved a segmentation accuracy of 99.19% and Jaccard index of 49.3%, which outperform the results of SVM [40] . However, this research also showed some drawbacks of U-net, such as the network performance can be improved by increasing the number of images and variability, whose result is heavily dependent on the input data. [36] proposed FCN-based system and developed U-net method to segment and classify Leishmania into promastigotes, amastigotes and adhered parasites. In order to overcome labor-intensive diagnostic procedures and false negative scores caused by straightforward nature of sample preparation, [4] proposed a ResNet18 based method to detect T. cruzi automatically in unstained microscope images. This method can overcome sensitivity limitations related to microscopic examination. [48] proposed an efficient artificial neural network named MCellNet to detect Giardia and Cryptosporidium from drinking water. In this study, 346 frames per second were captured and processed by the system of MCellNet combined with imaging flow cytometry to increase the high-throughput and efficiency of deep learning. This model achieved a classification accuracy of greater than 99.6%, but could not handle the data that model did not see. [49] automatically and quantitatively analyzed the Babesiainfected Erythrocytes by applying DenseNet, thus avoiding the quantitative analysis error generated by the manual microscopy. Integrated gradient used as an interpretability tool of model revealed that cell boundaries, precipitate and rouleaux formations are the direct factors of false positives. In clinical diagnosis scenarios, it is very important to identify the parasite species and presence of potentially mixed infections [35] . The multi-parasite algorithm has better practicability and generalization. [37] built a cost-effective mobile instrument to quickly detect the moving parasites of optically dense bodily fluids. The CNN model was used to automatically detect the signals created by motile Trypanosomes. This study showed that this instrument can detect other motile parasites by imaging Trichomonads. [32] proposed a deep-learning approach of geometry-aware and applied geometric-feature spectrum to detect host nucleus, Toxoplasma, Trypanosome, and Babesia by ExtremeNet based architecture without misdiagnosis. Different from ordinary unsupervised learning, [34] proposed a deep cycle transfer learning (DCTL) method for parasite detection. DCTL can avoid tedious data annotation by transferring macro shape knowledge from human experts to replace micro shape knowledge of parasites. DCTL successfully uses the shape information of macro objects that are similar with parasites' shape such as ring, pear and banana, to detect Plasmodium, Babesia and Toxoplasma. The drawback of DCTL is that highly depend on shape information of macro objects. To facilitate the development of deep learning approaches to the diagnostic of parasites, we summarize and describe the publicly available protozoan parasite datasets in Table1. As shown in the Table 1 , most of them are Plasmodium datasets, because malaria diagnosis still remains the main focus of deep learning based parasite research [21] . Studies containing datasets of other parasites, such as Toxoplasma, Babesia, Leishmania, Trypanosome and Trichomonad [20, 29, 34] , are precious because there are only few publicly available datasets of these parasites. Most of the datasets are used for the detection and classification with only few datasets focusing on the segmentation [50] . The detailed parameters of representative datasets are displayed in Table 1 . Three representative datasets were displayed [51] has attracted the attention of many researchers [39, 52, 44, 53] . This dataset was collected from the Chittagong Medical College Hospital, Bangladesh, and includes the Giemsa stains of 150P. falciparum-infected patients and 50 healthy patients. The uninfected cases did not contain infected cells but contained other interferences, such as staining interferences or dust impurities. These images were obtained by imaging of thin blood smear slides on a traditional optical microscope using the mobile phone's built-in camera. The focus of the available public datasets on the Plasmodium poses a difficulty for researchers who want to study other parasites. [29] collected 6 types of protozoan parasites' images, and made them publicly available. Images in this dataset are collected using a bright-field light microscope with either 400x or 1000x magnification. 34,298 microscopic images of multiple parasites and host cells were obtained, including Plasmodium, Toxoplasma, Babesia, Leishmania, Trypanosome, Trichomonad, Red blood cells and Leukocytes. It is worth noting that the current publicly available Leishmania datasets are very rare, and the flagellated promastigote stage present only in the midgut of insect vectors (sandflies) [54] . The images of Leishmania collected and presented here with promastigote morphology [29] might be not useful for medical diagnostic, but this dataset is still very useful for the study the life cycle of the Leishmania parasite, and would be also useful for deep learning training and model development. [29] . This dataset contains six types of protozoan parasites (Plasmodium, Toxoplasma, Babesia, Leishmania, Trypanosome, Trichomonad) and two types of host cells (RBCs and Leukocyte) presented in the cropped image patches. C: Plasmodium falciparum dataset [51] . The dataset contains cropped image patches from parasitized and uninfected RBCs. To aid parasite classification, detection and segmentation of deep learning, we summarized the algorithms-related formulas for standard metrics in Table 2 . The evaluation metrics of classification mainly include Accuracy, Precision, Recall, Specificity, F1 score and Confusion Matrix. The evaluation indicators of object detection mainly include precision-recall curve, average precision (AP), mean average precision (mAP), Intersection-over-Union (IoU), receiver operating characteristic curve (ROC curve) and area under curve (AUC). In order to measure the segmentation task, Pixel Accuracy (PA), Dice, volumetric overlap error (VOE) and average symmetric surface distance (ASD) are usually calculated in medical image analysis. Deep learning assists in the analysis of parasites and their hosts. In this paper, we systematically and comprehensively summarized the development and the application of deep learning in microscopic examination of protozoan parasites. Furthermore, we also collected and described the available datasets and model evaluation metrics used in protozoan parasites. There is a number of methods used for protozoan parasite examination. For example, the main methods of malaria diagnosis include rapid diagnostic tests (RDTs) and microscopy [3] . Compared to microscopic examination by human experts, RDTs are quite convenient, effective and fast. However, RDTs is not always available and detectable in all regions with the largest disease burden [3] . According to world malaria report 2021, Nicaragua suffered undersupply of RDTs during a malaria outbreak [3] ; between 2020 and 2021, 17 publications from 13 countries have recorded the pfhrp2/3 deletions, which made plasmodium undetectable by RDTs [3] . Hence, microscopic examination and other molecular or antigenic diagnostic tests are more complementing than replacing each other. Microscopic examination remains the gold standard for laboratory diagnosis [17] , and deep learning represents promising technology for microscopic parasite examination [18] . Compared to parasite microscopic examination by the human experts, microscopic parasite diagnosis employing deep learning is fast, automatable and accurate, while traditional microscopic examination by human experts is usually tedious, timeconsuming and labor-intensive, which can lead to misdiagnosis and missed diagnosis [55, 56] . Deep learning-based microscopic parasite diagnosis can save a lot of time for doctors, thus reducing the rate of misdiagnosis, missed diagnosis and drug abuse. Using deep learning-based microscopic parasite diagnosis clinicians can spend more time on advanced decision-making tasks. Besides, deep learning has advantages in multi-parasite diagnosis, multistage parasite recognition and parasite quantification in realtime, high-throughput and at low parasite density [31, 33, 48, 58] . Although deep learning has made progress in protozoan parasite diagnosis, several challenges remain. As our article highlights, dataset is an important factor restricting the wide application of deep learning for parasite diagnosis. Compared with other fields, the scarcity of publicly available datasets in the field of parasite diagnosis is more prominent, especially in other protozoan parasites other than Plasmodium. In some protozoan parasite examination studies, datasets are too small to give a convincing evaluation about model's performance. Therefore, collecting existing publicly available protozoan parasite datasets is very important in training model and distinguishing model's over-fitting, robustness and good generalization [40] . In supervised learning, the lack of labeled data is an important factor hindering the application of deep learning in parasitology. Collecting and labeling large datasets for training models is costly and time-consuming, especially in the pixel-level labeling for segmentation tasks [36, 40] . Moreover, some datasets currently use only parts of samples in the microscopy slide, while whole slide image needs to be analyzed [43, 45] . This requires additional time, computing resources and more robust models for small objects. Inference speed, model size and accuracy of model need to be compromised in clinics. Parameters of dataset, such as the different staining methods, preparation of samples, color differences, standards of images and digitization devices, need to be unified and be more specific to determine their impact on the performance of deep learning models [58] . To solve the limitation of available labeled datasets, some transfer learning-based methods have been created [20, 34] . These methods rely heavily on the selection of specific macro objects to match the parasites, and might be only effectively against specific types of protozoan parasites tested, such as Toxoplasma, Plasmodium and Babesia. These methods might not be able to apply to other parasites which might have no corresponding macro objects, such as Leishmania, Trichmonad and Trypanosoma. Although deep learn- Table 2 Standard evaluation metrics for parasite analysis based on deep learning. A is the predicted result and B is the Ground truth. TP: true positive, TN: true negative, FP: false positive, FN: false negative. Accuracy ing model performs well in the analysis of microscopic parasite images, deep learning is like an opaque black box, which lack convincing interpretability [59] . The complete interpretability of deep learning model would be conducive to the promotion of artificial intelligence in the protozoan parasites diagnosis. The deep learning model has a large number of hyperparameters which usually depends on empirical settings, such as learning rate, batch size and epochs, which are also unexplainable. Although there are currently some automated tuning tools such as Neural Network Intelligence and AutoML [60] , the long training time and vast parameter verifications make those newly proposed state-of-the-art models require a large amount of time and computing resources. Deep learning-based protozoan parasite researches should not only focus on improving model accuracy, but also pay more attention to the biological view. The complex and divergent life cycles of the protozoan parasites make lots of difficulties associated with the diagnosis of these organisms by deep learning. Therefore, deep learning-based protozoan parasite researches need to consider the different growth stages and distinct morphological characteristics of the parasites. Protozoan parasites accomplish their life cycle in different hosts where they adopt completely different morphological stages, which needs to be considered for accurate parasite microscopic examination. Besides Plasmodium [30] , it is also true for other protozoan parasites such as T.gondii and Leishmania. T.gondii has three infective stages: tachyzoite, bradyzoite and sporozoite [10] . However, Plasmodium is perhaps the simplest case because it infects non-nucleated cells (RBCs) during blood stage. This is not the case for all protozoan parasites, care must be taken before generalizing the benefits of such deep learning approaches to other protozoa and in the context of medical decision making. Presenting and distinguishing their different morphological stages and divergent life cycles can help us to better understand the difficulties associated with the diagnosis of these organisms by microscopic examination and deep learning. African malaria deaths set to dwarf covid-19 fatalities as pandemic hits control efforts, WHO warns Protozoa: structure, classification, growth, and development World malaria report 2021. World Health Organization Automatic detection of Trypanosomosis in thick blood smears using image pre-processing and deep learning. Intelligent Human Computer Interaction Geographical origin of chronic Chagas disease patients in Brazil impacts the performance of commercial tests for anti-T. cruzi IgG. Memórias do Instituto Oswaldo Cruz The Clinical Assessment, Treatment, and Prevention of Lyme Disease, Human Granulocytic Anaplasmosis, and Babesiosis: Clinical Practice Guidelines by the Infectious Diseases Society of America. Clinical infectious diseases : an official publication of the Infectious Diseases Society of Treatment of Human Babesiosis: Then and Now Brains and Brawn: Toxoplasma Infections of the Central Nervous System and Skeletal Muscle Human immunity to Toxoplasma gondii The life-cycle of Toxoplasma gondii reviewed using animations Macrophage Polarization in Leishmaniasis: broadening horizons Trichomonads parasitic in humans Trichomonas vaginalis: pathogenicity and potential role in human reproductive failure Neglected Tropical Diseases-North America Mass drug administration can be a valuable addition to the malaria elimination toolbox Comparison of three diagnostic methods (microscopy, RDT, and PCR) for the detection of malaria parasites in representative samples from Equatorial Guinea Challenges in laboratory diagnosis of Malaria in a low resource country among cases of acute febrile illness at tertiary care hospital in eastern Nepal: Comparative study on Conventional Vs Molecular approach Label-free imaging and classification of live P. falciparum enables high performance parasitemia quantification without fixation or staining Transfer learning for toxoplasma gondii recognition Malaria diagnosis: a brief review Applying Faster R-CNN for Object Detection on Malaria Images Systematic review and meta-analysis of diagnostic accuracy of loop-mediated isothermal amplification (LAMP) methods compared with microscopy, polymerase chain reaction and rapid diagnostic tests for malaria diagnosis Deep learning in medical imaging: general overview Challenges related to artificial intelligence research in medical imaging and the importance of image analysis competitions Artificial intelligence in lung cancer pathology image analysis Deep neural networks improve radiologists' performance in breast cancer screening Integrating deep learning CT-scan model, biological and clinical variables to predict severity of COVID-19 patients Deep neural network models for computational histopathology: A survey Deep Learning for Imaging and Detection of Microorganisms Multi-stage malaria parasite recognition by deep learning Automated blood smear analysis for mobile malaria diagnosis. Mobile Point-of-Care Monitors and Diagnostic Device Geometry-aware cell detection with deep learning A novel stacked CNN for malarial parasite detection in thin blood smear images Parasitologist-level classification of apicomplexan parasites and host cell with deep cycle transfer learning (DCTL) Image analysis and machine learning for detecting malaria Leishmaniasis Parasite Segmentation and Classification Using Deep Learning. International Conference on Articulated Motion and Deformable Motility-based label-free detection of parasites in bodily fluids using holographic speckle analysis and deep learning Malaria parasite detection in thick blood smear microscopic images using modified YOLOV3 and YOLOV4 models Malaria Detection on Giemsa-Stained Blood Smears Using Deep Learning and Feature Extraction Convolutional Neural Networks for Chagas' Parasite Detection in Histopathological Images A Weakly Supervised Deep Learning Approach for Detecting Malaria and Sickle Cells in Blood Films Protective immunity to liver-stage malaria Clustering-Based Dual Deep Learning Architecture for Detecting Red Blood Cells in Malaria Diagnostic Smears An Effective Convolutional Neural Network for Classifying Red Blood Cells in Malaria Diseases Deep Learning for Smartphone-Based Malaria Parasite Detection in Thick Blood Smears Smartphone-supported malaria diagnosis based on deep learning Human toxoplasmosis in Mozambique: gaps in knowledge and research opportunities. Parasites Vectors Deep learning-enabled imaging flow cytometry for high-speed Cryptosporidium and Giardia detection Applications of Digital Microscopy and Densely Connected Convolutional Neural Networks for Automated Quantification of Babesia-Infected Erythrocytes MP-IDB: the malaria parasite image database for image processing and analysis Pre-trained convolutional neural networks as feature extractors toward improved malaria parasite detection in thin blood smear images Medical Use of Deep Learning: Malaria Testing Using Pretrained ResNet. International Conference on Advanced Intelligent Systems for Sustainable Development Performance evaluation of deep neural ensembles toward malaria parasite detection in thin-blood smear images Developmental differentiation in Leishmania lifecycle progression: post-transcriptional control conducts the orchestra Misdiagnosis of babesiosis as malaria, Equatorial Guinea Comparison of Three PCR-Based Methods to Detect Loa loa and Mansonella perstans in Long-Term Frozen Storage Dried Blood Spots A novel deep learning-based point-of-care diagnostic method for detecting Plasmodium falciparum with fluorescence digital microscopy Convolutional Neural Networks for Chagas' Parasite Detection in Histopathological Images Interpretable deep learning systems for multi-class segmentation and classification of non-melanoma skin cancer Taking human out of learning applications: A survey on automated machine learning Automated blood smear analysis for mobile malaria diagnosis. Mobile Point-of-Care Monitors and Diagnostic Device Design Deep convolutional neural networks for microscopy-based point of care diagnostics Annotated high-throughput microscopy image sets for validation A machine learning-based system for detecting leishmaniasis in microscopic images The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.