Intracellular protozoan parasites are responsible for dramatic infectious diseases such as malaria, leishmaniasis, Chagas disease, toxoplasmosis, etc., which threaten the lives of billions of people around the world. Among these diseases, leishmaniasis has become the second most common cause of human death after malaria and is currently considered by the World Health Organization as an emerging and neglected disease. Leishmaniasis is caused by the intracellular amastigote form of Leishmania, which resides within the macrophages of the infected host.

The quantification of intracellular parasites within the host cell is a fundamental step in the validation of a new treatment, which involves evaluating the activity of new compounds, drug susceptibility, assessment of clinical strains and target validation using genetically modified parasites. This task is routinely performed by manually counting the intracellular parasites within the in vitro infected macrophages on Giemsa-stained slides or fluorescence microscopy images (Schindelin et al., 2012), which is a very time-consuming and laborious process that may generate errors due to inaccurate counts as well as observer bias (Górriz et al., 2018, Moraes and Alcântara, 2019). These difficulties lead to manual cell counting of microscopy images that may present variations between different expert observers. Therefore, the development of automatic counting methods derives from the need to overcome the subjectivity of the counting process, rendering it reproducible and efficient (Alahmari et al., 2019, Xing et al., 2021, Zhou et al., 2020).

Recently, automatic counting methods emerged as an alternative to manual quantification (Evangeline et al., 2020, Falk et al., 2019, He et al., 2021, Riccio et al., 2019, Zhang et al., 2020); these methods aim to count only one cell type in the field of view. However, counting intracellular parasites is a more complex task. At least three cell categories can be defined in an image, i.e., parasites, infected cells and uninfected cells. This makes the task of counting intracellular parasites more difficult than just counting a single category of cells.

Automatic counting of intracellular parasites has mainly been based on traditional image-processing methods, in which single-cell segmentation and subsequent classification are performed (Yazdanparast et al., 2014, Schmid et al., 2015, Gomes-Alves et al., 2018). The INsPECT method (Yazdanparast et al., 2014), for example, is a software developed to automatically measure, using phase constrat and fluorescence microscopy images, the infection level of Leishmania. It shows the number of intracellular parasites detected, allowing the determination of the parasite index. However, this index is significantly higher using INsPECT than a manual count of Giemsa-stained samples obtained under the same conditions, and INsPECT is not currently available online. Another method for automatic intracellular Leishmania amastigote counting using cell segmentation was developed with the IN Cell Investigator Developer Toolbox (GE Healthcare) and made available for open-source CellProfiler software (Gomes-Alves et al., 2018, McQuin et al., 2018) with good results. However, using ad hoc generated images, we did not obtain a good correlation between the infection index calculated using CellProfiler software compared with that from manual counting.

The aforementioned works also reported the limitations of conventional image processing. Using classic techniques, the parasite burden may be over or underestimated due to several factors: visualization or acquisition of images in a single focal plane, where not all macrophages and/or parasites are properly focused; and with overlapping, high confluence or fuzzy borders of the cells in the monolayer (He et al., 2021). When one or more of these conditions are present, image-processing methods can generate erroneous results, making traditional methodologies difficult to generalize and transfer to other laboratories with different equipment. Also, the methods from classic techniques for estimating intracellular parasite numbers are not easily transferred between different datasets, thus requiring manual adjustment of different parameters. This is not the case with deep learning, whose approach is to infer general rules from a training set using neural networks (He et al., 2017, Stringer et al., 2021).

Automated counting methods based on deep neural networks using object detection (Redmon et al., 2016), segmentation (Stringer et al., 2021, Yang et al., 2022, Hu et al., 2022) or density maps (He et al., 2021) have been proposed. YOLO is an algorithm based on deep learning developed for object detection and classification (Redmon et al., 2016). This method learns very general representations of objects to detect them in real time, fast and with satisfactory results. Nevertheless, being a method based on object detection, it is necessary to define a bounding box for each object. The bounding box labeling turns out to be a tedious and time-consuming task when hundreds of small objects have to be annotated (Zaji et al., 2022), e.g., intracellular parasites in microscopy images. In addition, bounding box annotation requires more information compared with the point annotation used in counting objects by density maps. A bounding box is defined by four values; in contrast, a point is defined by two values, i.e., the coordinates of the centroid of the object in the image. Several studies have highlighted the limitations of detection methods for counting hundreds of objects in an image, mainly due to their increased susceptibility to errors in the presence of object overlap or extremely small-sized objects (Ranjan et al., 2021, Ma et al., 2015). This can become a disadvantage for intracellular parasite counts when there is high infective capacity. In parasite burden studies, there are many parasite lines with high virulence, which is reflected in the great capacity of replication that the parasite has inside the host cell (Orrego et al., 2019). In these cases, counts above 100 amastigotes in a single image can be obtained, requesting the use of reliable tools with a count high accuracy.

Cellpose (Stringer et al., 2021) is another automatic method that also uses deep learning for the segmentation of morphologically diverse cells, with good performance. In general, this method segment one cell category at a time or cells that fall within a size range. Therefore, the use of this method in intracellular parasite counting involves the adjustment of a size value for the host cell and another for the intracellular parasite, as well as a classification to differentiate infected from uninfected cells. It is important to mention that neither YOLO nor Cellpose have been used for intracellular parasite segmentation, as of the current writing.

Deep learning methods have begun to be explored for the quantification of protozoan parasites (Hu et al., 2022). Within the deep learning methods reported so far for Leishmania quantification, one trains a U-net model that segments Leishmania parasites and classifies them into promastigotes, amastigotes and adhered parasites in Giemsa staining (Górriz et al., 2018). This technique requires a series of steps: segmentation, feature extraction and classification. The performance of the count depends on each of these stages, which limits the method’s application to another set of images, sacrificing its performance. The resulting precisions were cytoplasm identification 0.882, nucleus 0.938, promastigote 0.512, adhered promastigote 0.677 and amastigote 0.757, a correlation too low to be used in biological studies. Another work used the Viola–Jones algorithm to develop a Leishmania parasite detection system; however, only a 65% recall and 50% precision were obtained in the detection of macrophages infected with the Leishmania parasite (Zare et al., 2022). Therefore, as far as we know, no high-performance algorithms have been developed that use neural networks for the automatic quantification of intracellular parasites and, therefore, parasite burden.

In this work, we proposed the FiCRoN tool based on Fully Convolutional Regression Networks (FCRNs), a promising new tool to estimate parasite burden from fluorescence microscopy images. The advantage of an automatic deep learning tool is that it relies on a training phase, based on data generated by experts, wherein it learns specific features of the data that allow it to adapt to the task autonomously. This method generates density maps, whose integral is related to the number of each of the three cell categories in the image. Density maps are created from the geometric center of each cell, so the segmentation problems that occur with traditional methods do not occur here. A count estimation from density maps generated by neural networks allows counting in fewer steps, reducing the propagation of errors and time for dataset creation as data labeling for training only needs to mark the centroid of the cells instead of a mask that completely delimits the cell.

We, therefore, focused on the multiple advantages and virtues of FCRN deep learning to develop a tool for application in the quantification of intracellular parasites like Leishmania with potential application in other intracellular trypanosomatids like Trypanosoma cruzi, responsible for Chagas disease. The FiCRoN method was implemented in Python 3 using open-source packages (Van Rossum and Drake, 2009, Abadi et al., 2015, Van der Walt et al., 2014), and was generated with an easy-to-install and easy-to-use free GUI (Graphical User Interface) that can be installed locally and is freely available (see Code available in Supplementary Material).