香京julia种子在线播放

Returned travelers who have recently visited malaria-endemic countries and are unwell are routinely assessed for possible malaria infection in clinics at The Hospital for Tropical Diseases and Homerton University Hospital, London, UK. Diagnostic evaluation primarily involves examination of patient blood samples by manual light microscopy (using stained peripheral blood smears) to confirm the presence of malaria parasites, determine the species and provide an estimation of parasitemia where P. falciparum or P. knowlesi are present. Supplementary testing using Rapid Diagnostic Tests (RDTs) and PCR may also take place (Tangpukdee et al., 2009; Bailey et al., 2013). Patients who are confirmed positive for malaria are given appropriate antimalarial medication and supportive care. For this study, residual patient blood samples were obtained subsequent to routine laboratory testing from a total of 1202 returned travelers over the age of 18 years. These samples were anonymized and used to prepare study-specific Giemsa-stained thick and thin blood films to facilitate a direct comparison between manual light microscopy and the EasyScan GO and for PCR assessment at The London School of Hygiene and Tropical Medicine, London, UK.

Thick and thin smears were prepared on clean glass slides using surplus patient EDTA blood samples which had been obtained by venepuncture (Warhurst and Williams, 1996; WHO, 2016). Thin smears were fixed in methanol for 1 minute prior to staining. Fully air-dried smears were immersed for 30 minutes in 3% (v/v) Giemsa staining solution diluted in Phosphate-buffered water (pH 7.2), rinsed in tap water to remove stain deposit and allowed to air-dry vertically. Blood films were viewed using a x100 oil-immersion objective and a minimum of 200 and 50 fields of view (FOV) were assessed for thick and thin films respectively. Thick films were initially used for positive confirmation of malaria parasites being present within a blood sample whereas thin films were used for Plasmodium species determination. For subsequent quantitation of parasites, thick films were further examined and parasites counted until a total of 500 white blood cells (WBCs) had been seen and an accurate parasite density (parasites/µL of blood) determined using patient WBC counts obtained from laboratory records.

For each blood sample, there were two independent reads performed. The first result was obtained from the routine diagnostic laboratory after examination by two microscopists within that laboratory and was used to determine if samples were malaria parasite positive or negative for recruitment purposes. The second reading was study-specific, also provided by an expert microscopist, who examined slides produced from the same blood samples after they had been anonymized to confirm positivity and perform accurate quantitation. If there was a discrepancy between the results provided by the study microscopist and the initial diagnostic laboratory microscopy result, a second expert study microscopist was then engaged to perform a third read.

The EasyScan GO is a fully automated end-to-end malaria diagnostic system which includes both hardware and software. An automated bright-field microscopy platform scans Giemsa-stained thick and thin blood films, and malaria detection algorithms process the image sets to give parasite detection, species ID, and parasite quantitation for the patient. Given a blood film, the device automatically scans and processes the slide, and outputs a report that includes estimated diagnosis, species ID, quantitation, WBC count, and a mosaic of thumbnail images of top suspected malaria parasites ( Figure 1 ). The images allow a technician or pathologist to quickly check the device’s findings. Using the current software and for the purposes of this study, a complete slide evaluation including an output report with image thumbnails took ~10 minutes (Horning et al., 2021). Thin films generally took longer at ~18 minutes per slide, as more FOV needed to be scanned.

Thick films are used to (i) confirm whether malaria parasites are present, (ii) obtain an accurate parasite count, and (iii) obtain a species identification of P. falciparum vs non-P. falciparum (with the default for non-P. falciparum cases being P. vivax). Thin films are used for refined Plasmodium species determination (e.g. between non-P. falciparum species). The overall algorithm architecture, as well as the EasyScan GO device, are described in (Horning et al., 2021). Thick film algorithms are fully detailed in (Mehanian et al., 2017) and thin film algorithms, in (Delahunt et al., 2019).

Briefly, a new sample is assessed as follows: In a thick film the algorithm analyses 100 image stacks (each 113 µm x 85 µm x 5 depths), containing an average of 1527 WBCs (std dev 862; 90% of samples have over 720 WBCs). Candidate parasite objects are detected, then culled, by rapid morphological methods, which are tuned for high sensitivity yet still eliminate most of the easier distractor objects. The remaining candidate objects pass through two convolutional neural net (CNN) classifiers to receive labels as ring, late-stage, or distractor. WBCs are detected and counted by a separate module. Diagnosis is based on whether the suspected parasite count per WBC exceeds a noise threshold that has been pre-tuned to aim for 90% patient specificity. While the high blood volume examined theoretically allows for a lower limit of detection, in practice the object false positive rate is a limiting factor (Delahunt et al., 2022). Quantitation is reported based on thick films only, leveraging the large blood volume examined to reduce Poisson variability (Delahunt et al., 2019). Species ID is binary, P. falciparum vs non-P. falciparum, based on comparing ring and late-stage counts and leveraging two facts: sequestration in P. falciparum and relatively low parasitemias in non-P. falciparum infections. For those samples labeled non-P. falciparum, the thin film is processed in a similar fashion to the thick film, but with a single CNN assigning a species (or distractor) label to each object. Final species ID from thin film is determined by weighted majority vote.

Training slides included over 500 imaged blood slides from 12 countries, encompassing a wide range of different staining presentations and containing a variety of artefacts. The collection included large numbers of P. falciparum-positive, P. vivax-positive, and Plasmodium spp.-negative slides, and much smaller numbers of P. ovale and P. malariae-positive slides, as these species are much less commonly encountered. Slides used for training were annotated by expert microscopists specializing in the diagnosis of malaria. Details about slide collections are found in (Mehanian et al., 2017; Delahunt et al., 2019; Horning et al., 2021). Annotation methods are fully described in (Mehanian et al., 2017).

The WHO international standard for P. falciparum DNA for nucleic acid amplification techniques (Padley et al., 2008) was used as a positive control for P. falciparum. Lyophilized blood samples derived from patients infected with P. vivax, P. ovale and P. malariae, used as part of the WHO external quality assessment scheme for malaria nucleic acid amplification testing (WHO, 2018b) were used as non-P. falciparum positive controls. Negative controls consisted of negative extraction controls (whole uninfected blood) and negative assay controls (nuclease-free water). Parasite DNA was extracted from samples using the PureLink™ Pro 96 Genomic DNA Purification Kit (Invitrogen, US).

For real-time PCR, in the first instance, each sample was amplified in a multiplex reaction targeting the conserved region of the Plasmodium 18S rRNA gene and the human beta-2 microglobulin (β2 M) gene (a DNA extraction control). After genus-specific amplification, positive samples were subsequently tested for P. falciparum, P. vivax, P. ovale and P. malariae in a multiplex real-time PCR reaction. Primers used for these amplifications are shown in supplemental table 1 and are modified slightly from those described previously (Shokoples et al., 2009). The ABI 7500 FAST System (Applied Biosystems, US) was used for all amplification reactions. The conditions consisted of an initial activation of DNA Polymerase at 95°C for 30 secs followed by 40 cycles of amplification comprising denaturation at 95°C for 3 secs, annealing and extension at 60°C for 30 secs.

Nested PCR (nPCR) is widely regarded as the gold standard nucleic acid amplification (NAA) method for detection of malaria parasites in very low density samples (Cordray and Richards-Kortum, 2012; Vasoo and Pritt, 2013) and was performed as a confirmatory test where there was discordance between results reported for manual light microscopy, EasyScan GO and real-time PCR. The species-specific nucleotide sequences of the 18S rRNA gene of P. falciparum, P. vivax, P. malariae and P. ovale were amplified as described previously (Snounou et al., 1993; Snounou and Singh, 2002; Padley et al., 2003; Singh et al., 2004; Calderaro et al., 2007) with slight modifications, and primers used are shown in supplemental table 1 . Assays were performed using a PxE thermal cycler (ThermoFisher Scientific, US) and a DNA Engine Tetrad® 2 cycler (Bio-Rad, US). Thermal cycling parameters used are described previously (Snounou and Singh, 2002) with the only adaptation being that Nest 1 and Nest 2 reactions were given 30 and 25 cycles of annealing, extension and denaturation respectively.

The required sample size of 104 malaria positive slides and 1125 malaria negative slides was calculated as a non-inferiority study to be able to jointly test that the sensitivity of EasyScan GO was not decreased by more than 80% compared to expert microscopy, and that the false positive fraction (1-specificity) was not increased by more than 50% compared to expert microscopy, with 5% significance and 80% power. The malaria positive slides were expected to be derived from symptomatic clinical cases, plus follow-up slides from these same patients following treatment (i.e. with lower parasitemia). The sensitivity and specificity for expert microscopy with this slide composition was assumed to be 80% and 90%, respectively.

A total of 1202 patient samples were collected and the same sets of Giemsa-stained thick and thin slides evaluated concurrently by manual light microscopy and the EasyScan GO. By light microscopy, 113 of the samples were confirmed as malaria parasite positive and 1089 were negative. When compared to light microscopy, the EasyScan GO exhibited a sensitivity of 88% (95% CI: 80-93%) and a specificity of 89% (95% CI: 87-91%). The EasyScan GO correctly identified 99 of the 113 light microscopy positives but also incorrectly reported a positive result for 122 samples that were identified as parasite negative (1089) by light microscopy ( Table 1 ).

Of the 99 samples labelled positive by both light microscopy and the EasyScan GO, manual microscopy identified 87 as P. falciparum and 12 as non-P. falciparum comprising 6 P. vivax, 5 P. ovale and 1 P. malariae. The EasyScan GO correctly reported P. falciparum for 86 of the 87 P. falciparum samples and reported non- P. falciparum for 11 out of 12 non-P. falciparum samples (Kappa = 0.905). However, it failed to distinguish between the non-falciparum species, reporting all non-P. falciparum samples as P. vivax.

To achieve level 1 (expert) competency in malaria microscopy, the WHO requires that 50% of samples containing malaria parasites with densities between 200 and 2000 p/uL be quantified within +/-25% of the “true count” (WHO, 2016). Quantitation of low parasitemia (e.g. under 200 p/µL) samples is intrinsically noisy due to Poisson variability in the number of parasites present in the examined blood (Delahunt et al., 2019). In this study, EasyScan GO quantitation was within +/-25% of true count for 33% (6/18) of samples with parasitemias between 200 and 2000 p/µL; and within +/-25% of true count for 30% (24/79) of samples with parasitemias above 200 p/µL ( Figure 2 ). It should be noted that the WHO standard assumes that the ground-truth quantitation is an average of several counts provided by multiple expert-level microscopists, while here we compare to a ground-truth given by a single expert-level microscopist.

In this study, diagnostic accuracy of the EasyScan GO was similar to that of expert Light Microscopy. The principal difference was the EasyScan GO’s lower specificity (89% vs 100%, PCR as reference). From a clinical perspective, false positives could mean patients receiving drug treatment that is not required, therefore the device would not be used as a standalone diagnostic tool without additional input from other laboratory staff needing to screen further for potential false positives. However, the mosaic of thumbnails of suspected parasites (e.g. as shown in Figure 1 and supplementary figures 1 and 2 ), included in every report outputted by the device, makes this check relatively easy and screening could be done remotely if so required. Sensitivity of the EasyScan GO was very similar to light microscopy, meaning both techniques are able to positively identify parasites at comparable densities and become limited below similar parasite densities. In terms of clinical care, this is reassuring since similar numbers of positive malaria infections could be identified and subsequently treated. The low sensitivity of light microscopy vs PCR was likely due to the inclusion of a number of low-parasitemia samples (under 50 parasites/µL), which would often be undetected by most non-reference microscopists and could therefore be missed in a routine laboratory. In addition, some of the patients sampled may in fact be 1-2 weeks post treatment for malaria and the positive PCR is residual circulating Plasmodial DNA, rather than an active infection with live parasites. The low sensitivity of the EasyScan GO vs PCR was mostly observed in the same low-parasitemia samples, but the device did appear to miss two samples with higher parasitemia (see Figure 2 ). The sample with a very high parasitemia (~35,000 p/µL) was possibly post-treatment, since the parasites appeared to contain no cytoplasm, which could potentially lead to the software algorithm having much more difficulty identifying them as true parasites. The EasyScan GO did detect a few low-parasitemia samples that were missed by light microscopy but picked up by PCR. However, given the high estimated parasitemia of these samples compared with the PCR result, it suggests they may have been “right for the wrong reason”, with the positive diagnoses triggered by incorrectly classified artefacts rather than correctly classified parasites. If so, they are perhaps most similar to false positive negative samples. One of the most common limitations with light microscopy that will also impact any digital imaging device such as the EasyScan GO is variation in quality of blood films (Das et al., 2022). The software was trained with a large number of training slides encompassing a wide range of slide backgrounds containing a variety of different artefacts, but undoubtedly the device will improve further as it learns from a larger range of sample sets in future.

The Easy Scan GO accurately distinguished P. falciparum from non-P. falciparum, using the thick film. However, it failed to distinguish between the various non-P. falciparum species, defaulting to P. vivax. This was likely a reflection of insufficient P. ovale and P. malariae training samples: CNNs require vast amounts of training data, and due to natural distributions of the malaria species the algorithms’ training sets were highly imbalanced, containing ample P. vivax (and P. falciparum) but very few P. ovale and P. malariae blood films. In addition, to determine species, the algorithm applies a logical decision tree to findings from thick and thin films, where thick film decides between P. falciparum and non-P. falciparum, and thin film can modify a non-falciparum finding. Due to prevalence rates and training data imbalances, this logical tree favors P. vivax over P. ovale and P, malariae. That is, P. vivax is the default non-P. falciparum choice. This weakness might be mitigated in two ways: By substantially increasing the number of P. ovale and P. malariae samples in the training set (in practice, a difficult task); or by applying local geographical priors, e.g. the rareness of P. vivax relative to P. ovale in West Africa (Howes et al., 2011) (as in (WHO, 2000; Bailey et al., 2013; WHO, 2016)). Since the EasyScan GO has an implicit bias towards P. vivax (vs. ovale, malariae, and knowlesi), it is more suitable for geographic regions with this same predominance, for example India or Peru. This concern does not apply to diagnosis of P. falciparum.

Evolving drug resistance (Tilley et al., 2016; Balikagala et al., 2021) makes drug efficacy trials and drug resistance sentinel sites a potential use-scenario for automated microscopy, because of the need for high-throughput parasite quantitation in laboratory settings: The labor-intensive process requires quantitating blood films drawn every few hours from treated patients, in order to plot parasite clearance curves (White, 2011). Since several film quantitations are combined to calculate a clearance curve, the exact performance specification for individual blood film quantitations is not well defined and partly depends on the calculation method. An informal guideline recommends that most quantitations be accurate to within +/- 25% (excepting very low parasitemia samples) (Dhorda M, WWARN. Personal communication). WHO’s microscopist evaluation protocol looks at whether P. falciparum samples with parasitemia between 200 and 2000 p/µL have a quantitation error within 25% (WHO, 2016). In this study, EasyScan GO’s accuracy was within this margin for 33% of such samples. This may be too low to usefully calculate accurate clearance curves, though it might be close to sufficient. An experiment on a set of time-series blood films as used in clearance studies, comparing clearance curve log slopes as calculated from quantitations done by light microscopy and by the EasyScan GO, would help clarify this point.

An important detail to note is that light microscopy parasite quantitations in this study were based on accurate total WBC counts which varied widely by individual (mean 5430, std dev 2070), while EasyScan GO assumed a fixed value of 8000 WBCs/µL. When correcting parasitemia estimates provided by EasyScan GO using accurate individual WBC counts, the percentage of samples having quantitation error within the 25% margin increases to 50%. This indicates that quantitation error was affected by how WBC counts per µL were defined for the samples examined.

The EasyScan GO system, and very similar algorithms deployed on different hardware, have been evaluated in four other field trials (Torres et al., 2018; Vongpromek et al., 2019; Horning et al., 2021; Das et al., 2022) and two internal tests (Mehanian et al., 2017; Delahunt et al., 2019) allowing a broad perspective on system reliability in the face of diverse slide presentations.

The perspective afforded by five separate field trials is, to our knowledge, unique for an automated malaria diagnosis system. This perspective is highly valuable because a system’s performance can vary in different settings. Since machine learning-based systems can be brittle in the face of new data sources, the high variability of slide preparation at different clinics is a serious challenge for automated malaria diagnosis systems. In this context, therefore, multiple data points on system performance are especially important to understanding a system’s suitability for deployment. Collectively, these trials demonstrate strong performance of a fully automated system for assessing a diversity of Giemsa-stained blood films.