Reducing costs of Giardia spp. cyst enumeration using machine learning-based systems

Giardia duodenalis is one of the main intestinal parasites in humans worldwide and among the most common reasons for diarrheal diseases (Keserue et al. 2012). Each year, an estimated 1.5 million deaths globally result from diarrheal diseases, and most of these deaths occur in low-income settings (Daniels et al. 2018). Drinking water is the major transmission route for infections, thus needing proper water treatment and monitoring to avoid human infection (Feng et al. 2021).

Standard methods for Giardia spp. cyst detection in water samples, such as the U.S. Environmental Protection Agency (EPA) 1623.1 (U.S. EPA 2012), require (1) concentration of samples (filtration) and (2) parasite purification through immunomagnetic separation, followed by (3) staining and (4) visualization using fluorescence microscopy. After preparing water samples through the filtration of large volumes of water (minimum of 10 L), the concentrate undergoes purification using immunomagnetic beads to separate target organisms from process-related impurities. Afterwards, the sample receives fluorescent dyes specific for Giardia spp. cysts (fluorescein isothiocyanate – FITC) following the manufacturer's recommendations. For internal structure visualization, the sample also receives 4′,6-diamidino-2-phenylindole (DAPI). After staining, the prepared cover-slipped microscope slide is kept dry at room temperature overnight. After staining, a well-trained technician places the prepared samples on a fluorescence microscope for cyst detection and enumeration manually. For this task, the technician screens the whole microscope slide for characteristic FITC-fluorescence, as well as the size and shape of the Giardia spp. cysts, confirming their identification through DAPI-stained nuclei and cytoplasm.

Over the last decade, developed countries have widely utilized Method 1623.1, but the effort and associated costs involved in various constituent procedures still make it difficult to deploy in low- and medium-income countries (Fava et al. 2021). When using fluorescence microscopy, the high costs and technical analytical complexity of Giardia-specific fluorescent dyes, laboratory infrastructure, and equipment are evident (Maciel & Sabogal-Paz 2016). The primary cost is the separation process involving immunomagnetic beads (Fava et al. 2021), followed by the two final steps of the analysis procedure involving staining and detection using fluorescence microscopy. From a practical perspective, microbiological methods used to detect parasites should ideally meet the criteria of being simple, easy-to-use, and rapid but reliable and low cost (Shrestha et al. 2020).

There is a wide range of alternative methods used for detecting Giardia spp. cysts in water, including flow cytometry, laser scanning cytometry, and molecular techniques (Efstratiou et al. 2017). However, even if these methods offer more sensitivity (SE) and specificity (SP) than the microscopic technique used in the U.S. EPA protocol, they also have the disadvantages of being time- and resource-consuming (El Mallahi et al. 2013).

Researchers have developed various other detection methods to partially mitigate some of these disadvantages. Mudanyali et al. (2010) investigated the use of a field-portable lens-free holographic microscope to record cell holograms within a custom-made cuvette and also automatically detected and enumerated Giardia spp. cysts at concentrations of ∼190 cysts/mL. This is several orders of magnitude higher than those expected in raw water samples (Efstratiou et al. 2017). Besides, specimen fixation is required. Mobile platforms for Giardia lamblia cyst detection and quantification have also been introduced to replace bulky and costly optical/mechanical components. Koydemir et al. (2015) presented a smartphone-based fluorescence microscope to capture images of fluorescently labelled cysts over a disposable sample processing cassette, achieving a lower detection limit (∼12 cysts/10 mL). Shrestha et al. (2020) developed a smartphone-based brightfield microscope for the simultaneous detection of Cryptosporidium and Giardia spp. (oo)cysts stained with Lugol's iodine. After loading a standard haemocytometer with suspensions, technicians manually analyse magnified images of (oo)cysts using a ball lens directly on the screen of a smartphone to provide estimates based on size, shape, and contrast at a detection limit of ∼1 cyst/g (derived from reported 73 cysts/100 g). This is a subjective, tedious, and time-consuming task. As all these methods involve chemicals affecting the purity of the suspension, which is an expensive input, the analysed cysts cannot be used for other trials or microbiological studies. Furthermore, apart from the proposal by Koydemir et al. (2015), experts only analyse aliquots, meaning representative estimates can only be obtained by sampling, preparing, and analysing numerous samples. More recently, Nakarmi et al. (2023) reported the usage of deep learning-based object detection to automate (oo)cyst detection in images from the smartphone-based microscope developed by Shrestha et al. (2020). Although the ability to detect cysts using a smartphone can be valuable for field testing scenarios, the results showed that because of the curvature effect of the ball lens in the smartphone microscope on the imaged objects, the models could falsely predict Giardia as Cryptosporidium and vice versa. Besides, the field of view using this lens is too small, making scanning an equivalent area to that of the standard microscopy slides a tedious and time-consuming task. Automated microscopy approaches have also combined imaging flow cytometry and deep learning-based models to predict Cryptosporidium and Giardia oocysts in drinking water (Luo et al. 2021). Despite the ability of flow cytometry to rapidly acquire thousands of single-cell images to be analysed at rates of 346 frames per second, the cost of acquiring and maintaining this equipment makes it prohibitive for low-resource settings. Besides, the created machine learning models were not able to generalize the prediction on non-seen data in the training phase.

Only a few detection methods that allow for eliminating the labelling steps, which are time-consuming and expensive and simultaneously enable the analysis of larger volumes, have been presented. Göröcs et al. (2020) reported a field-portable holographic imaging flow cytometer to detect and count low levels of Giardia contamination (10 cysts/50 mL) at a volumetric throughput of 100 mL/h. However, a new set of disposable components is required for each measurement. More recently, our research group introduced in situ microscopy for Giardia spp. cyst detection and enumeration in agitated suspensions (Belini et al. 2022). Here, we capture high optical resolution brightfield microscopic images of cysts directly from a 1 mL sampling chamber and process them within 10 min to produce automated estimates at a detection limit of 30 cysts/mL, without requiring either mechanical/optical adjustments or disposable components. Despite providing estimates without involving labelling, the capability of these methods for detecting other waterborne pathogens such as Cryptosporidium spp. oocysts still needs to be demonstrated.

Although improvements have been achieved in several aspects compared to standard detection methods, all the aforementioned methods rely on custom-designed devices with varying degrees of complexity. Therefore, implementing these methods still requires some level of skill, inputs, and costs. From these observations, it seems obvious that the usage of existing laboratory equipment and available personnel would be preferable.

The low contrast of (unstained) Giardia spp. cysts, especially under brightfield microscopy, makes them difficult to visualize without the help of staining procedures (e.g., Lugol's iodine in brightfield) or fluorescence-based dyes specific for this protozoan such as FITC. Recently, Belini et al. (2018) demonstrated the effectiveness of using inexpensive darkfield microscopy setups to acquire higher contrasted images of Giardia spp. cysts of unstained microscopy slides. By using this illumination technique, well-resolved cyst walls, specific intracellular structures, including nuclei, axostyles, and retracted cytoplasm in Giardia spp. cysts could be visualized, all exhibiting characteristic features comparable to those observed in fluorescence-based microscopy (FITC/DAPI) following Method 1623.1.

While sensitive detection methods are crucial, it is equally important to address the high costs associated with standard detection methods to make them viable for water quality assessment in low- and medium-income countries. Despite numerous alternative methods available for detecting this protozoan in water samples, there appears to be a lack of effort in detecting and enumerating Giardia spp. cysts in less complex suspensions, such as those from drinking or treated water samples, using existing standard laboratory settings. The ability to rapidly detect and enumerate Giardia spp. cysts (e.g., post-recovery via EPA 1623.1) in a label-free process would reduce both the detection time and level of human intervention required, as well as contribute to saving resources (McGrath et al. 2017).

The objective of this study was to develop and evaluate an automated method for detecting and enumerating Giardia spp. cysts in darkfield microscopic images of unstained suspensions, aiming to create a technique easily applicable in microbiologists' routine practice. The proposed solution differs from existing methods as it does not require custom-made imaging devices to perform label-free estimates. By using standard microscopes in microbiology laboratories, the proposed method has the potential to make cyst suspension assessments more cost-effective and accessible in regions with limited infrastructure as the method does not require expensive equipment, consumables for labelling, or specialists to conduct and interpret analyses. Specifically, we used standard image processing algorithms for detecting cyst candidates in darkfield images. After extracting characteristic object features, we trained machine learning-based models with Giardia cysts and non-giardia particle images. The purpose of this training was then to determine the number of Giardia cysts found in testing images.

The novelty of this study is the usage of darkfield images in conjunction with machine learning-based methods to enumerate Giardia cysts in microscopic images of unstained suspensions. Access to a more economical and accurate detection technique, as described here, can potentially reduce people's vulnerability to these parasites in developing countries, where infection is likely underdiagnosed and underreported due to limited detection resources (Shrestha et al. 2020).

To obtain the suspensions for the darkfield experiments, commercial suspensions of G. duodenalis cysts (Waterborne™, Inc., New Orleans, LA, USA) containing ∼5 × 10⁶ cysts in 8 mL were homogenized for 2 min using a vortex and vigorously pipetted 20 times. Our research group adopted these suspensions in various experiments using them as stock suspensions in the present study. Afterwards, we serially diluted the suspensions with distilled water. For each resulting suspension, we prepared two aliquots of 50 μL with a Merifluor^® (Meridian Bioscience, Inc.) kit according to the manufacturer's recommendations and used them to count (in two wells of a microscopy slide) the number of cysts on a fluorescence microscopy (BX51, Olympus^®, λ_excitation = 475 nm, λ_emission = 520 nm). From this process, we obtained average concentrations of Giardia spp. cysts at 100, 250, 500, and 1,000 cysts/mL. We also prepared a negative control using cyst-free distilled water. For each established concentration of organisms, we loaded 25 μL replicates onto a standard glass 1.0 mm thickness microscope slide and covered the mounting with a coverslip (0.17 mm thickness), sealing it with enamel for darkfield imaging.

Darkfield microscopy is an inexpensive imaging technique that enables the typical oval-elliptical shape of Giardia cysts to be seen as a much brighter structure than the background of the image, as confirmed by the inspection of images published by our research group (Belini et al. 2018). This seemingly trivial observation motivated us to choose darkfield imaging because this facilitates cyst detection using standard edge detection techniques (Gonzalez & Woods 2008). As also shown, specific intracellular structures such as nuclei, axostyles, and (retracted) cytoplasm, as assessed using fluorescence microscopy, are much easier to be observed in darkfield than in brightfield images (Belini et al. 2018). These structures also help discriminate cysts from morphologically similar non-cyst particles in the images.

Darkfield microscopy was used to acquire images of distilled water that had been contaminated with Giardia spp. cysts. The images were acquired using an Olympus BX60 microscope, which was equipped with an Olympus DP74 digital camera (1/1.2″ colour CMOS imaging sensor, 5.86 × 5.86 μm² pixel size, 1,600 × 1,200 pixels, RGB, tiff, 12-bit). The images were acquired in darkfield by using a condenser for darkfield illumination and an Olympus UPlanFL objective lens (40× magnification and 0.75 numerical aperture). The field of view resulting from the imaging system corresponds to a slide region of 0.27 × 0.20 mm².

The prepared microscope slides were placed onto the microscope stage for darkfield imaging. The microscope light intensity was set for proper imaging (∼50%), as recommended by the microscope's manufacturer. The aperture of the diaphragm of the illumination source was set close to its minimum so that the image background is dark, and potential artefacts such as dust onto the slide are minimized. The darkfield condenser was positioned just beneath the mounted slide.

The exposure time of the camera was adjusted to 60 ms, so that the internal content of the cysts was not overexposed. Just as for the light intensity, the exposure time was kept constant during the whole experiment. We adjusted the focus manually, so that the cyst wall appeared well defined and brighter than the internal content of the cyst. When needed, we readjusted the focus within different fields of view of the same microscope slides.

The whole microscope slide was screened for Giardia spp. cysts. Darkfield images of unlabelled slides were acquired after they had been thoroughly scanned in a systematic scanning pattern following the haemocytometer counting principle (Mundt & Shanahan 2020). The aim is to avoid missing Giardia spp. cysts or counted cysts more than once. In this process, the cysts were visually detected based on their typical oval shape, size (8–18 μm long by 5–15 μm wide, U.S. EPA 2012), well-defined wall (under darkfield illumination), and either by the presence of retracted or non-retracted cytoplasm (Belini et al. 2018).

The first step is to detect objects from the darkfield images to create portraits containing cyst candidates. Before detecting objects in the acquired images, we performed the following image preprocessing steps: first, a noise reduction step was carried out using a 7 × 7 median filter (Gonzalez & Woods 2008) to reduce high-frequency noises.

Complex visual data analysis over the entire original (large) image is a time-consuming task. To save computational time, the algorithm initially generates portraits containing cyst candidates for further analysis. This involves computing the centroid of the detected objects in the final binary image and then cropping a corresponding portrait image (sized 80 × 80 pixels) from the original darkfield RGB image, which is subsequently converted to a grayscale image. The corresponding lengths of the portrait are 12 μm × 12 μm by counting the pixels forming the sides of the portrait. Each pixel corresponds to a real length given by the pixel length over optical magnification. Before detecting objects within the portrait, we performed a high-frequency noise reduction step using a 5 × 5 median filter (Gonzalez & Woods 2008).

Afterwards, the Canny operator (threshold values equal to 0.07 and 0.17, respectively) prepared the portraits for segmentation because the portraits may still contain more than one object. To generate portraits containing only one centralized object inside the portrait, a morphological closing operation (disk-shaped, 5 pixels) is performed, followed by hole filling and portrait border cleaning. After the morphological opening operation (disk-shaped, 7 pixels), only the largest object within the portrait is kept. Afterwards, the centroid of the segmented object is computed, and this information is used to centralize the selected object within the portrait.

The solidity factor (Soltys et al. 2005) of each object was computed as the scalar that specifies the proportion of pixels in the convex hull that are also inside the detected object.

The length and the width of the cyst candidates were determined as the maximum and minimum Feret's diameters (MathWorks^™ 2023), respectively.

The final value for each texture feature was determined by averaging the four co-occurrence matrices, each one computed for each direction of adjacency θ.

We used example images of Giardia spp. cysts and non-cysts to train and validate the machine learning-based classifiers. Developing algorithms for automatic detection and segmentation facilitated the process of generating training sets of annotated images from the entire darkfield image. From this process, object portraits (80 × 80 pixels²) were extracted from original RGB darkfield images. These images were manually annotated as either a Giardia spp. cyst or non-cyst.

Aiming at providing a more diverse training image dataset, accounting for, e.g., the variability of the cyst's optical appearance and different imaging conditions, we acquired hundreds of microscopic images of cysts from different microscope slides containing different cyst concentrations.

The final step is classification, whereby Giardia spp. cysts are identified and counted within a target darkfield image using a trained classifier. To achieve this, we developed a supervised machine learning algorithm to distinguish Giardia spp. cysts from non-cysts. The learning technique is based on bagging (an acronym from bootstrap aggregating), as introduced by Breiman (1996). This learning system manipulates the training data by randomly replacing the original training data with replicated training sets. These sets are sampled with replacement from the original training data in order to train different base classifiers. In this process, some features describing the cyst candidates may not appear in the training sets, while others appear more than once. Finally, every base-classifier's output is then combined by the majority vote to form the final prediction.

To evaluate the ability of the model to generalize to similar but unseen data in the training/validation stage, 10-fold cross-validation was used for tuning the learning system parameters and also for assessing the overall prediction performance of the (embedded) classifier. In the validation method, the training dataset was randomly partitioned into 10 equal-sized subsets. Each subset, in turn, was then used as test data for the classifier from the remaining nine validation subsets. This procedure was repeated 10 times, whereby a different subset was used as the test set, and the performance was computed as the average of the 10 repetitions.

We visually inspected the classified objects within a target darkfield image to verify the correct classification by the trained machine learning algorithm. In this context, we manually classified the classifier outcome as true positive (TP), false positive (FP), or false negative (FN). The classification is based on the following criteria: complete cyst bodies, the largest fragment of over-segmented cysts (i.e., cysts divided into many parts), and one cyst for each case of under-segmentation (more than one cyst detected as one cyst) were classified as TP. FP was non-cyst particles predicted as cysts. FNs were missed cysts in cyst clusters.

All algorithms developed in this study were run on a standard Desktop (Intel PC, Quad-CPU, 2.66 GHz, 4.0 GB RAM). All parameters involved in both image analysis procedures, and machine learning algorithms were kept fixed for the whole experiment.

We used the average Giardia spp. cyst counts obtained from both manual and proposed methods for statistical analyses. To assess the linearity between both methods, a correlation analysis was performed over a range of cyst concentrations. Agreement of the manual and the automated methods regarding cyst concentrations were also assessed using Bland–Altman plots (Bland & Altman 1999). Statistical analyses were carried out using Excel software (Microsoft, USA).

Microscopic images acquired under darkfield illumination show not only the target microorganism but also the whole background complexity, which appears as much brighter structures than the background itself (Figure 2). As seen in Figure 2(a), most objects in the image are non-cyst particles. Artefacts, such as dust particles on the cover glass (Figure 2(b)), might be highly refractive in darkfield microscopy and, since they are not usually in the focal plane of the target, they appear in the form of bright spotted halos (Ciurte et al. 2018). Giardia spp. cysts from different suspensions show different optical appearances and retracted cytoplasm (Belini et al. 2018). The cyst-free image in Figure 2(d) shows higher numbers of non-cyst particles.

Different densities of non-cyst particles may change the overall brightness of the image, as can be observed by comparing the darker background in Figure 2(c) to that in Figure 2(a) and 2(b). The slight interimage brightness variances and intraimage brightness inhomogeneities (Figure 2(b)) observed in the acquired images can influence the predictions made by the proposed system, since all chosen descriptors are intensity level-based. This could be avoided or reduced by using an enhanced image acquisition system and slide cleaning protocols but were omitted to present the proposed system to some degree of real-world interference factors.

As shown in Figure 3, the input RGB darkfield image is firstly converted into a grayscale image for further processing. After applying a smoothing median filter to reduce high-frequency noise, a Canny edge detector (Gonzalez & Woods 2008) is performed to highlight sharp intensity changes in foreground and background pixels (Figure 3(c)). Appropriate Canny threshold values (empirically determined as 0.07 and 0.18 by inspection in hundreds of images) yield a binary image having detected objects as white structures on a black background (Figure 3(d)). In this process, most pixels belonging to optical artefacts (as shown in Figures 2(b) and 3(a)) are not detected because their pixels have much lower intensities than those belonging to objects imaged in the focal plane.

The portraits show images of Giardia spp. cysts exhibiting cells of different sizes, shapes, and distinct internal content. In common with most visually inspected portraits, there was the presence of oval-shaped cells and a well-defined bright contour corresponding to the cyst wall (Figure 4(a)). From the point of view of image analysis, the cyst border helps to detect them, while the internal content, characterized by texture features, is useful in helping to distinguish cysts from non-cyst particles that are morphologically similar to the target (Figure 4(b)).

After extracting the features for each cyst candidate within portraits, we used this information to train the machine learning algorithm to distinguish Giardia spp. cysts from non-cyst particles. From 820 acquired darkfield images, a total number of 681 and 528 Giardia spp. cyst and non-cyst portraits, respectively, were generated by the automatic segmentation algorithm. The number of training examples used was balanced to avoid the model from becoming biased towards one class (i.e., Giardia spp. cyst or non-cyst). For each object within the portraits, 11 different features were computed and then used for training machine learning-based classifiers. Figure 4 shows examples of both classes of objects. The classifier-type model was based on the decision tree and the classification training used 10-fold cross-validation, achieving an accuracy rate of 88%. This prediction performance means that the classifier can properly identify most of the Giardia spp. cysts in the validation step.

Once we trained and validated the machine learning algorithm, we actively applied it to independent testing slides containing different cyst concentrations and compared its predictions to manual counts carried out in the same darkfield images.

Besides images of Giardia spp. cysts, we also acquired an average of 25 images from different fields of view within each microscopy slide (for 0, 100, 250, 500, and 1,000 cysts/mL), all containing a variety of non-cyst particles. The acquisition of cyst-free images aimed to capture the complexity of the background and also helped assess the robustness of the classification method.

For each acquired image, we visually inspected the objects classified as Giardia spp. cyst by the algorithm to evaluate the accuracy, SE, and SP of the proposed method. Table 1 shows the results in detail.

Table 1 shows that the proposed system can properly identify and count an average of 87% Giardia spp. cysts over a wide range of cyst concentrations. The percentages of correct identifications, despite the varying numbers of non-cyst particles exhibiting a diversity of shapes and sizes, are an indication of the robustness of the classification method. The presence of higher numbers of non-cysts resembling the target in the highest cyst concentrations showed no significant effect on the performance of the classification algorithm, where most Giardia spp. cysts were recognized correctly by the algorithm.

The two missed single cysts at the 500 cysts/mL concentration were due to impurities stuck to the detected cysts, making the classifier fail to identify them because this category of object was not considered in the proposal. At the highest cyst concentration, we also acquired images of two cysts touching each other and a four-cyst agglomerate. In the former case, the algorithm was able to detect and identify two cysts but counted only two cysts in the agglomerate. This result was expected as the detection algorithm was not designed to cope with either cyst-cluster or cyst-agglomerates: the two identified cysts in the agglomerate were segmented correctly by the algorithm because they were just touching each other, thus facilitating their segmentation as two single cysts. On the other hand, the two missed cysts were partially overlapping each other and were detected as a much larger structure than that expected for single cysts and therefore were not classified by the algorithm as the target.

Table 2 shows the relative error observed between automatic and manual counting for the five cyst concentrations analysed. The average value of the error observed in the highest concentrations of cysts is within that acceptable by conventional methods (U.S. EPA 2012) and can be explained by the inability of the detection algorithm to segment clusters of cysts, thus increasing the number of lost cysts (not detected). It should be noted that Method 1623.1 considers recovery rates of Giardia spp. cysts from 8 to 100% as an acceptability criterion for modified detection methods, i.e., losses of 92% are considered acceptable.

The strong correlation (R² = 0.99) between cysts counted by both automated and manual methods demonstrates the efficacy of the proposed method as a tool to estimate numbers of unlabelled Giardia spp. cysts directly from darkfield images in the range from 0 to 1,000 cysts/mL.

The proposed system has some distinct advantages over approved methods using immunofluorescence microscopy that make it an attractive tool for automated detection and enumeration of (unlabelled) Giardia spp. cysts in resource-limited settings: (i) darkfield imaging is a low-cost illumination technique that can be easily implemented on standard laboratory microscopes, as demonstrated by Belini et al. (2018); (ii) no preparation of samples using either stains or fluorescent dyes is required, thus saving money and time.

The significant costs associated with fluorescence microscopy-based detection methods (e.g., U.S. EPA 2012) make them difficult to be performed in less-developed countries. Table 3 shows comparative costs of implementing both the proposed system and the 1623.1 method (U.S. EPA 2012). The cost associated with the purchase of a computer to register either fluorescence or darkfield images was not included as standard desktop computers can be used for this purpose, as well as to run image analysis and machine learning algorithms.

The estimated values do not include costs associated with electricity, maintenance, and also the basic staff qualification to prepare samples and operate the equipment. As a matter of fact, by eliminating the need for specific components for fluorescence microscopy such as optical filters, UV lamp, and its power supply, the operating complexity and costs associated with the maintenance of darkfield microscopes might even be reduced. Darkfield imaging-related costs can be further reduced by converting a standard laboratory microscope into a darkfield imaging system by using low-cost custom-made darkfield setups, as described in detail by Belini et al. (2018).

In addition to markedly reducing costs (Table 3), the proposed method quantifies suspensions without modification, as it does not require any chemicals. This enables the quantitative transfer of counted organisms to test volumes of water, as described in Ongerth & Saaed (2017).

Compared with the standard method, the proposed system is less labour-intensive; it only requires preparing the microscopy slide with the suspension to be analysed. Once the prepared slide is placed onto the microscope stage, manual focusing, slide scanning, and image acquisition were performed by one of the authors, taking an average of 1.5 h per slide. Afterwards, a computer automatically analysed an average of 40 images within ∼30 s to produce an estimate of the number of cysts for a single sample. In contrast, ∼4 h are required for labelling the Giardia spp. cysts with fluorescent dyes (U.S. EPA 2012) followed by 30–45 min, so that an expert on fluorescence microscope can complete the counts (Keserue et al. 2012). However, at the current state of the proposed method, cysts cannot be distinguished from another genus of protozoan such as Cryptosporidium parvum. This would certainly be interesting to be explored in future studies. When needing this information, one still requires standard methods that involve antibody-specific fluorescent labelling.

The proposed method does have some limitations: if cysts physically overlap each other, our method could not separate them properly, producing underestimates. Although the presence of clustered cysts is not common in environmental samples, purified suspensions often contain agglomeration of cysts. In addition, although the method performed well on images exhibiting different complexities, such as optical artefacts and varying densities of non-cyst particles, we did not robustly validate it in more complex water matrices because the number and diversity of non-cyst particles are limited in the present conditions. As for the former limitation, a promising avenue is to extend the segmentation algorithm to include, for instance, methods based on the morphological watershed (Meyer & Beucher 1990) to separate touching cells in microscopy images as published elsewhere (e.g., Chiang et al. 2018; Belini et al. 2021). We prepared the suspensions to eliminate potential interferences from, e.g., organic matter, metals, and other microorganisms that naturally exist in environmental water matrices, aiming to understand the principles, potential, and limitations of the proposed system. The full potential of the method in assessing water samples from natural matrices will be explored in further studies by incorporating images of particulates found in water samples into the training phase of the machine learning models.

By using the standard method 1623.1 (U.S. EPA 2012), the estimated cost per sample due to fluorescent dyes was ∼$120 (Brazilian quote in June 2023), whereas no cost was associated with the proposed method as no chemicals were required. This cost refers to the Merifluor^® (Meridian Bioscience, Inc.) kit and Fluoroshield^™ with the DAPI (Sigma-Aldrich^®) solution, computed per three 50 μL aliquots (U.S. EPA 2012).

We believe that when using microscopes featured with image processing-based autofocusing systems (e.g., Zhang & Yao 2021) integrated into the process of motorized microscope stages, the analysis of a large routine batch of suspensions could benefit from a fully automated, cost-effective, fast, reliable, label-free, and expert-independent Giardia spp. cyst counting system.

In this study, we developed a machine learning-based approach and evaluated it to identify and count unlabelled Giardia spp. cysts in darkfield microscopic images. Approved standard methods for detecting Giardia spp. cysts are expensive, laborious, and time-consuming procedures based on sample preparation that require specific fluorescent dyes and trained technicians to conduct and interpret analyses using fluorescence microscopy. The developed system demonstrated the general feasibility of label-free detection and enumeration for Giardia spp. cysts based on darkfield imaging, shape and textural features, and machine learning techniques. The results showed satisfactory performance exhibiting an average identification ratio of 88%, thus proving the applicability of the developed system to perform automatic counting. By using the proposed system, we simultaneously overcame the cost limitation and lack of trained personnel to analyse Giardia spp. cyst suspensions. We believe that the Giardia spp. cyst detection system proposed herein has the potential to be a promising alternative method for fluorescence microscopy in approved standard methods, thus allowing the routine analysis of suspensions in low-income countries as the proposed system is expert-independent and free of any kind of consumables. By eliminating the subjectivity inherent to manual counting using standard methods, the application of the proposed method could enable more accurate estimates in water treatment-related research involving, for instance, the assessment of filters efficiency for pathogen removal.

We acknowledge the Biological Processes Laboratory, Department of Hydraulics and Sanitation, São Carlos School of Engineering, University of São Paulo, for their support with the darkfield imaging presented herein.

V.L.B. conceptualized the study, developed the methodology, conducted the software investigation, performed validation, and contributed to the original draft and editing. N.M.N.F. contributed to the conceptualization, methodology, and reviewed and edited the writing. L.P.S.-P. obtained funding, provided resources, validated the study, and reviewed and edited the writing. All authors have read and approved the final manuscript.

This work was partially supported by The Royal Society (ICA/R1/201373 – International Collaboration Awards 2020) and by the National Council for Scientific and Technological Development (CNPq-Brazil, processes n° 308070/2021-6 and 442074/2023-9).

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.