Background

When growing microalgae, being it at a laboratory or industrial scale, cell counting methods are essential for monitoring population growth. There are several alternative methods of estimating population size and growth, such as the optical density of the culture or in vivo fluorescence of chlorophyll a, but in many instances knowing the number of cells is necessary so that physiological processes and products may be referred in a per-cell basis (Guillard and Sieracki 2005).

The most common and simple counting method for microalgae cultures is by light microscopy using specific counting devices (counting chambers) selected based on the size of the organism and the cell concentration (Guillard 1973; Karlson et al 2010; Reguera et al 2016). This method has the advantage that the cells can be observed by the researcher, and the physiological status of the culture maybe assessed. However, it is also very time-consuming and tedious, particularly when a high number of cultures need to be examined, a common practice in laboratory experiments and industrial production units.

Already in the second half of the twentieth century, electronic devices aiming at automated counting systems started to be developed to help overcome these constraints (Parsons 1973). In the last decades, automated counting methods for cell cultures have been increasingly implemented and developed and there are now in the market several bench-top solutions available (e.g. CellDrop, DeNovix Inc., USA; Countess, Thermo Fisher Scientific, USA; EVE, NanoEntek, South Korea) characterized by having high efficiency and precision, and requiring low application times and low expertise. These methods are now starting to be used in microalgae research laboratories (e.g. Salbitani et al 2022). The main drawback of these automated methods is their initial high cost (above 15 000€) rendering them inaccessible to many laboratories, particularly in underdeveloped countries.

Notwithstanding, other high technological approaches and equipment are nowadays easily accessible. Cell phones are increasingly accessible and feature high-quality cameras, with images useful in many computer vision tasks such as object detection and recognition. In the last decades, the detection of objects using deep learning techniques has been intensively exploited, mainly due to its broad application (Cao et al. 2018; Viso.AI n. d.; Wang et al. 2018) and good results namely in almost all fields of biology (Webb 2018; Ching et al 2018), microscopy imagery (Sorzano et al 2009; Ito et al 2018; Xing et al 2018; Devan et al 2019) and microorganisms in general (Zhang et al. 2021). What motivated our interest in deep learning algorithms was the limited success of conventional image processing techniques applied to images where context is a decisive factor and the operator reasoning must be incorporated. In addition, inherent in deep learning is the possibility of data augmentation. Data augmentation is the introduction of small randomized changes to train data, in the form of radiometric and geometric alterations in dimensions, rotation, translation, and shearing. This is particularly interesting because whatever the number of train images we use, we cannot guarantee that all the possible positions, sizes, and aspects of the cells are contemplated in the training set.

The main objective of this work was to develop a low-cost cell counting method for microalgae cultures that were simultaneously easy, fast, and accurate. To achieve this objective, we used a stereo microscope and images acquired with a common cell phone, which were pre-processed in Matlab environment, with basic image processing operations available in any open-source image processing software (e.g. ImageJ), and a deep learning algorithm also open-source, running on a regular laptop.

Methods

Concerning microalgae strains and culture conditions in this study, two different strains of the marine dinoflagellate species Protoceratium reticulatum (Claparède & Lachmann) Bütschli 1885 (IO116-01 and IO116-02) were used both in the training and validation stage of the counting experiments. The cultures were obtained from the algae culture collection of the University of Lisbon (ALISU) and were grown under controlled laboratory conditions (Fitoclima 600PL, Aralab, Portugal), in L1 medium (Andersen et al 2005), at a salinity of 33, at 19 ± 1 °C under a 12:12 h light: dark cycle, 100–110 μmol photons m−2 s−1.

In what concerns sample preparation for cell counts, 3 ml culture samples were fixed with approximately 0.15 ml of Lugol’s solution (Karlson et al 2010). Immediately before filling the counting chamber samples were homogenized by gently rotating the flask 25 times. A sub-sample was then used to fill the chamber of a Palmer–Maloney counting slide (100 µl) (LeGresley and McDermott 2010). No dilution steps were used. The sample was allowed to settle for 5 min, and the chamber was placed under a stereo microscope at 10 × magnification (Zeiss Stemi 305, Germany). The images were then acquired through the eyepiece with a cell phone (Samsung M21, South Korea) equipped with a 48.0 MP camera (Samsung S5KGM1—f/2.0, 26 mm (wide), 1/2.0″, 0.8 µm). The image covered the whole area of the counting chamber, the equivalent of a100 µl culture sample.

For the image train process, a total of 6 images of 2250 × 4000 pixels at 24 bits were acquired from cultures in different phases of the growth curve to cover a variety of particle properties (e.g. range of cell sizes, cell debris, and thecal plates).

The pre-processing of the images was achieved in Matlab R2021a environment and consisted of four steps: first, a modified homomorphic filter was applied with a sigma of 11 to compensate for irregular illumination of the background; second, all images were histogram matched to one reference image, chosen for its ideal radiometric range; the third step consisted in producing a binary mask for each image with a global threshold, followed by morphological operations to consolidate the area of interest and eliminate surrounding structures included in the field of view (FOV); finally, this mask was applied to the processed image (Fig. 1).

Fig. 1
figure 1

a Original image acquired with a cell phone through the stereo microscope eyepiece; b the same image after being pre-processed and masked

Full size image

The algorithm used is one of the latest developments in one-stage algorithms based on convolutional neural networks (CNNs), the 5th version since the introduction of the concept You Only Look Once (YOLO) (Redmon et al 2016). YOLO v5 was made publicly available in a GitHub repository in 2020 (GitHub Ultralytics n.d.). The algorithm has been retrained for the task using a transfer learning technique: since many basic features are common to all detection problems (edges, contrasts, forms, etc.), an already heavily trained network can be used to implement a new problem. The new discriminators will define the last layers of the CNN, tuning the detector according to the details of the specific problem. After the download and successful installation of YOLOv5, the algorithm was trained on our data set as described below. The set of images acquired was segmented into tiles of 800 × 800 pixels to accelerate training procedure, as it was our objective to use the most complete model of YOLOv5, the model x that uses a CNN with 476 layers.

The train demands a set of images with all the objects of interest identified with bounding boxes, and the respective list; the data set annotated in this way, called ground truth, is then split into train and validation subsets, between which the algorithm will converge to the best possible achievement in terms of precision (percentage of true positives correctly classified) and recall (percentage of true positives detected).

A metric usually considered in object detection applications is the mean average precision (mAP) that quantifies the stability and consistency of the model, within a confidence threshold related to the intersection-over-union (IoU) areas between the anchor boxes estimated from the train data and the bounding boxes predicted by the model in the annotated data. With a threshold of X, the box is assigned to an object of interest if the IoU quotient is above X and considered background in the opposite case. Non-maxima suppression ensures most multiple detections are avoided, by considering only the box with maximum probability in each set of overlapping boxes.

The annotation of a subset of images for train and validation purposes can be made with online tools, such as (Makesense.AI n.d.) used in the present work, with a user-friendly graphic interface. The images to be annotated (usually 30% of the images available, further split in 20% for train, 10% for validation) are uploaded to the site and the graphic tools available in the interface are used to draw boxes around all the objects of interest in each image, using zoom, correction, delete and pan functionalities. At the end, a text file for each image is exported in a user-defined format, with all the annotations (image coordinates for the boxes) made in that image. The images and corresponding text files are then distributed between validation and train, because YOLOv5 requires a fixed directory tree, with names that it will recognize to know where to find image and label files during the train stage. Once trained, the algorithm was applied to a test data set of 43 images of P. reticulatum cultures as described above. Results were assessed by manually verifying the false negatives and false positives in each image, and the performance of the model was evaluated based on precision and recall.

Results

The 6 images acquired were cropped in 72 initial tiles of 800 × 800 pixels, and the algorithm was trained with 21 annotated images, using 14 images for train and 7 for validation. We verify later that several images only contain the pieces surrounding the area of interest, so we tested on 43 remaining images containing 3659 microalgae cells. With a confidence threshold of 0.25 for inference, YOLOv5 detected 3681 objects, of which 100 were false positives, leaving out 78 false negatives, giving an overall precision of 97.4% and a recall of 97.9%. Other parameterizations for the confidence threshold (0.20, 0.30 and 0.35) were tried with less success, as the improvement in accuracy led to a decrease in recall performance. An IoU threshold of 0.25 had the best results, avoiding most multiple detections. The mAP@0.5 was 0.862.

With this trade-off in parameterization, among the 3681 particles identified 2.7% were wrongly detected (false positives) and 2.1% of the 3659 particles present were not detected (false negatives), which seems like reasonable numbers compared to human curation.

If we look at the results considering each image as a unity, 72% reach a precision of 95%, 74% attain 95% of recall and 67% attain both recall and precision better than 95% simultaneously. The result of the inference for the image that exemplifies the pre-processing is shown in Fig. 2.

Fig. 2
figure 2

Points of interest (red squares) detected by YOLOv5 in the image shown in Fig. 1

Full size image

Inference over an 800 × 800 pixel image is computed in 1800 ms in the laptop described below, giving two or three outputs: a numerical output in the command window with the total count of detections and a file containing a copy of the image with all occurrences flagged (Fig. 2), as well as an optional text file featuring the coordinates of all the windows around the occurrences in image coordinates, allowing to extract every single particle for further processing if desired.

Discussion

The estimation of cell numbers when culturing microalgae is essential because the physiological processes under study or the target products to be produced usually need to be referred in a per-cell basis. Manual counting of microalgae cells for estimation of cell concentrations with acceptable accuracy levels consumes considerable human resources and is a tedious procedure sometimes with results affected by operator subjectivity and fatigue. This limits the number of samples that can be processed which may partly constraint the outputs of the work. Nevertheless, it is still used today in routine estimation of cell culture concentrations (e.g. Pereira et al 2016; Sheward et al 2017; Rocha et al. 2022) probably because of the lack of affordable, accurate and precise automatic alternatives in the market.

In the present work we provide the tools for the implementation of an automated cell counting method with no costs or only residual costs to the user. Compared to other automated cell counting methods routinely used in many laboratories, the methodology described is accessible in financial terms in facilities where a stereo microscope is available, as it only requires a regular laptop and a cell phone, and software available online without any associated costs. The learning curve is fast for anyone used to software with parameterizable interfaces.

The same methodology can be easily extended to other particles, requiring however a separate training stage for each kind of organism to be counted. The high number of epochs or iterations in the training stage is time-consuming, but it is done just once. As a reference, in our case, 700 iterations of model x, with images of 800 × 800 pixels took 21.2 h in a HP-Pavilion 17-cd1006np laptop with an Intel Core i7 processor, 16 GB SDRAM and a graphic unity NVIDIA GeForce RTX 2060 Max-Q 6 GB. In microalgae production units or when developing a particular laboratory experiment, the number of species that are cultured is relatively small, and this methodology could easily be implemented for all species of interest.

The annotation of the subset of images for train and validation purposes should be done by someone knowing well the characteristics of the organism to be counted. This will allow high accuracy output once the system is implemented. From the technical operator perspective, the system does not require an expensive training program since the only needed expertise is image acquisition using a standard cell phone and the basic informatic skills of any computer user to run a new image.

In addition, the method here proposed allows the counting of a higher number of cells in a significantly shorter time. This increases the confidence limits of the concentration estimates while allowing for a larger number of samples to be counted each day.

Conclusions

The present work describes a processing chain that allows the implementation of a particle counting procedure with application in the field of microalgae research and production. It allows monitoring population growth in microalgae cultures using a deep learning method with accurate results, and much less cost in both human labour, laboratory technical equipment and high-end image acquisition equipment. As with other automated cell counting methods, it is faster than using the light microscope and allows for the enumeration of high particle numbers, increasing statistical robustness (Guillard & Sieracki 2005). The human uncertainty involved in the tedious labour of counting a large number of cultures is eliminated, as well as the errors due to interruptions or fatigue.

Future work will include an evaluation of the performance of the same train in cultures of other microorganisms of similar gross morphology, which could lead to a portfolio of ready-to-run train files appropriated for the set of organisms most usually processed in each facility.