An automated workflow for fungal cell counting, cell size and data analysis: enhancing throughput and accuracy with the cell analysis and counting tool using ilastik software (caactus)

Abstract

Background

The development of filamentous fungal cells from spore to mature mycelium is influenced by a myriad of environmental conditions. Traditional methods for quantifying various fungal cell morphologies in liquid media via microscopy are labor-intensive and prone to user-dependent variability.

Methods

This study addresses these challenges by introducing a streamlined workflow for fungal cell analysis, utilizing ilastik-based image segmentation and Python for fast data preparation and analysis. The goal is to offer a user-friendly method for counting and analyzing different fungal cell morphologies and their sizes that is free of charge and accessible to users with minimal programming experience.

Results

The presented workflow significantly reduces the time required to determine cell developmental stages from microscopy images. It enables rapid results, as compared to manual counting. It also minimizes inter-user variability, enhancing consistency across analyses. It provides a high-throughput, automated solution for fungal cell analysis allowing users to process hundreds of images (n = 48-438) for analysis in a short time period (9-56 min).

Conclusion

This approach shows great potential for accelerating research and improving data reliability in biological studies. Furthermore, it provides laboratories in resource limited settings with a free to use solution.

Introduction

The fungal kingdom comprises a large group of organisms that are both beneficial and detrimental to humanity and our ecosystems. Fungi are used extensively for food production and biotechnology, provide key ecosystem services, and are also a major group of human and plant pathogens that harm our health and livelihoods [1–3]. Compared to other microbes, this kingdom has been neglected for a long time. Studying their morphology and developmental stages is of key importance for advancing mycology in all its sub-fields. Comparison of inter- and intra-species differences in early fungal development and the growth of genetically modified versus wild type strains are important questions in mycological research. Furthermore, studying the development of filamentous fungal cells under varying conditions, such as different media, concentrations of essential micronutrients, or exposure to antifungal agents in vitro, provides valuable insight into fungal physiology and their responses to various stressors.

As our research focus lies on the field of human pathogenic molds, the tools presented in this study were developed for studying Mucorales and Aspergilli. In 2022, the World Health Organization (WHO) has published the first ever priority list of fungal pathogens to highlight and coordinate the need for global action against fungal diseases [4]. The goal of the project is to raise awareness of the correct diagnosis of causative pathogenic agents and guide research efforts to find new treatment options for this neglected group of infectious diseases. Here, one of the causative agents of aspergillosis, Aspergillus fumigatus, ranks on third place of relevant human fungal pathogens in the critical group and globally causing the highest number of invasive fungal infections and chronic pulmonary syndromes [2]. Besides A. fumigatus, multiple other species belonging to this genus are known to cause infections in humans due to their naturally ubiquitous occurrence of their spores across the globe [5]. Furthermore, the rise of antifungal-resistant Aspergillus strains due to the dual-use of antimicrobial compounds in the environmental and clinical settings has brought more attention to these groups of fungi and the need for studying them [6].

Similarly, Mucorales were placed in the High Priority Group of the WHO’s multi-criteria decision analysis [4]. In recent years, Mucorales, the agents causing Mucormycosis, have witnessed more interest due to the increasing number of clinical cases, the severity of these infections and their high resistance to antifungal treatment [7, 8]. The causative agents belong to many genera, but in the clinical setting Rhizopus spp., Mucor spp., Lichtheimia spp. are the most frequently encountered [8].

For many mycological research questions, determining the abundance of different fungal developmental stages under different conditions is of interest to answer questions concerning the influence of effectors (drugs, temperature, nutrients, light, oxidative stress) or mutations of genes involved in germination or hyphal development [9–13]. When evaluating fungal growth and development in static liquid cultures different methods are widely used which all have their benefits and drawbacks. Traditionally, cell distribution across different developmental stages has been assessed using bright-field microscopy, hemocytometers, and manual counting using mechanical palm-click counters. This is highly labor-intensive and susceptible to operator bias [14]. Moreover, monitoring cell development across a wide range of conditions simultaneously can be challenging, as manually counting cells in each well of a 96-well plate is time-consuming and could thus lead to inaccuracy. Still, manual counting remains a cheap and widely used method.

In standardized antifungal susceptibility assays employed in microbiology diagnostic laboratories, results for filamentous fungi are typically read from microtiter plates by eye [15]. Alternatively, optical density (OD) measurements can be taken from the wells, although inhomogeneous growth pattern or onset of sporulation on the surface may lead to either overestimation or underestimation of fungal growth and are more suitable for later stages of fungal development and do not allow the study of the early onset of fungal growth [16].

In contrast, advanced methods such as flow cytometry provide a less subjective and high-throughput alternative, although they require additional staining steps, specialized equipment, and well-trained operators [14]. Furthermore, optical detection systems such as oCelloScope^TM have been developed to assess morphologies, size changes, and cell development over time. These systems enable automated examination of fungal growth at early time points and real-time monitoring over extended periods for organisms belonging to the genera Aspergillus, Beauveria, Candida, and Penicillium [16, 17].

Wurster et al. [18] demonstrated the repurposing of a module originally designed for tracking neural cell growth to analyze live mycelial growth dynamics under different conditions. Automated detection of hyphae and spores from fluorescence images for pathological diagnostics has also been reported [19]. Similarly, for the morphological analysis of yeast mutants from fluorescence microscopy, the Java-based program CalMorph was developed, enabling quantitative assessment of cell shape, actin organization, and nuclear DNA using multiple fluorescent input channels [20].

In addition, hyphal development under antifungal stress has been investigated using microfluidic CellASIC plates combined with three-dimensional confocal microscopy, revealing how different morphotypes respond to varying drug concentrations in vitro [21].

However, the disadvantages of these approaches are their reliance on specialized equipment and commercial software, which may not be affordable for all laboratories. Barry et al. [22] developed a classical image processing pipeline in ImageJ for quantifying morphological development of Aspergillus oryzae on solid substrate, measuring spore size, hyphal length, tip number, and growth unit over a 24-hour time course. Spores are immobilized on cellulose nitrate membranes grown on agar, then fixed, dried, and stained before imaging. While accurately capturing spore swelling and hyphal growth dynamics at the population level, the multi-step preparation is laborious, kills the culture, and is incompatible with high-throughput applications.

Li et al. [23] demonstrated the development of a tool for automated counting of yeast cells and spores of filamentous fungi. For their approach, they employed free and open-source software, such as ilastik [24] - an interactive and supervised machine learning software for image analysis - and ImageJ [25]. Their tool could significantly decrease time and inter-operator variability. Similarly, with a special focus on Saccharomyces cerevisiae, Dietler et al. [26] developed YeaZ, a U-Net-based convolutional neural network (CNN) for the segmentation of phase contrast microscopy images of budding yeast, including stressed cells, mutants, and time-course experiments. A key feature is a cell-cell boundary test that avoids oversegmentation without requiring fluorescent markers. The tool is accessible via a Python-based Graphical User Interface (GUI) and web application and can be used to study the influence of environmental or genetic perturbations on cell morphology at large scale. Building upon different tools, among other YeaZ, Padovani et al. [27] developed Cell-ACDC (Cell Analysis of the Cell Division Cycle), an open-source, GUI-based Python framework for live-cell imaging analysis with a special focus on yeast cell cycle analysis. The tool provides a complete pipeline from raw microscopy file conversion to segmentation, cell tracking, semi-automated error correction, and cell cycle annotation. It embeds four state-of-the-art deep learning segmentation models selectable by the user — YeaZ and YeastMate for yeast, and the generalist models Cellpose and StarDist for other organisms. Cell-ACDC enables quantification of genetic contributions to cell cycle progression and growth, with cell volume estimated from 2D segmentation masks as a key readout. While primarily developed for budding yeast, the tool is generalizable to other cell types like hematopoietic stem cells.

Using classical image segmentation approaches, Hardy et al. [28] and Sten et al. [29] designed workflows for studying mycelial networks and hyphal morphologies. Hardy et al. [28] developed a fully automated classical image analysis pipeline for 3D morphological characterization of Trichoderma reesei in bioreactor conditions. For analyzing hyphal networks from microscopy images, Sten et al. [29] proposed a six-step workflow in MATLAB that prepares data by detecting scale bars from 2D microscopic images and estimating pixel size from it. The pipline then analyzes mycelium connectivity, branching, hyphal length and density from microscopic images. Praetorius et al. [30] developed an automated image analysis pipeline in the open-source visual programming tool JIPipe for label-free quantification of Candida albicans filamentation and hyphal length from live-cell microscopy images.

However, microscopic images of filamentous fungi grown in liquid media are often challenging to analyze due to variable image characteristics, touching cells and cells in various developmental stages present at the same time. To not only separate cells from its background, but also from each other and then classifying different morphologies as well as cut-off objects, cell debris and media impurities, we established an ilastik-based workflow to fill this gap. To enhance usability, it includes data pre-processing, automated data summary and analysis accessible in a GUI which allows for rapid processing of hundreds of images. This study presents a novel workflow designed to accelerate research on fungal cell morphological development at early stages, facilitating high-throughput analysis of extensive microscopic image datasets and ultimately enhancing the pace of data acquisition, analysis and comparison in this field also for users with minimal to no programming and scripting experience. The fungal kingdom is vast, with many filamentous species that have their unique morphological characteristics. The workflow described here allows researchers to tailor the analysis quickly to the species they are studying and their morphological particularities. Therefore, caactus aims in supporting the investigation of fungi of emerging interest as well as popular model organisms alike.

The remainder of the paper is organized as follows: Fungal strains and culturing to Defining fungal morphologies sections illustrate the method, Validation section provides information on the validation process, and Predicting mycelium formation from early morphologies for antifungal susceptibility testing (AFST) section introduces a clinical use case. Results section will discuss the presented results.

Methods

Fungal strains and culturing

For acquiring images for growth analysis with the cell analysis and counting tool using ilastik software (caactus) workflow, the fungi in question were cultivated in flat-bottom 96-well-transparent-polystyrol plates (Greiner, Germany), shown in Fig. 1 step (a). Depending on the growth rate of the fungal species in question, images were taken at time points at which cells in different developmental stages are still distinguishable, meaning mycelial growth has not yet been established.

Fig. 1.

The caactus workflow. a The fungi are cultured in 96-well plates; b Images are acquired using software of a scientific microscope; c The images are prepared for further processing (renaming and conversion to HDF5-format); d Pixel classification is performed in ilastik; e Boundary-based Segmentation is performed in ilastik; f The background is removed from the segmentation layer; g Object classification is performed in ilastik; h Data post-processing for batch-mode: (1) The individual tables are merged into one global table. (2) Summary statistics are generated. (3) PLN Modelling with PCA-plot. Created with www.biorender.com

For the initial setup and proof of concept for caactus, five strains belonging to five different Mucoralean species were used (Table 1). For antifungal susceptibility testing (AFST) as a use case, one Aspergillus strain was used (Table 1).

Table 1.

Species and their respective reference strains used in the development of caactus

Group	Species	Strain ID	Lab ID
Mucoralean fungi used for validation of the workflow	Rhizopus (R.) microsporus	ATCC 11559	ATCC 11559
	Rhizopus (R.) oryzae	CBS 126971	CD2
	Lichtheimia (L.) corymbifera	CBS 109940	CD4
	Mucor (M.) circinelloides	CBS 394.68	CD6
	Mucor (M.) lusitanicus	R7B	R7B
Aspergillus spp. used for AFST	Aspergillus (A.) fumigatus	ATCC 204305	R7

To obtain spore solutions to be used in the growth assays, Murcorales strains were cultivated on Yeast Peptone Glucose agar (YPG, 3 g/L yeast extract (Thermo scientific, Germany), 10 g/L peptone from soy (Carl Roth, Germany), 20 g/L glucose (Millipore, Germany), 15 g/L agar (VWR, Austria)). For A. fumigatus Sabouraud (SAB) agar (30 g/L Sabouraud Dextrose Broth (Liofilchem, Germany), 15 g/L agar (VWR)) was used [31, 32]. Spores were harvested and washed as described previously [31]. The Mucoralean strains were either used directly after harvest (fresh) or snap frozen in liquid nitrogen and stored for later use at (frozen). A. fumigatus was always used fresh. Fungi were then cultivated in 96-well culture plates (flat bottom, Greiner, Germany) in the respective media. For this, spores were adjusted to a final concentration of cells/ml according to the European Committee on Antimicrobial Susceptibility Testing (EUCAST) guidelines for AFST [33].

For testing Mucorales, YPG liquid medium was used. For A. fumigatus, RPMI1640 (Sigma-Aldrich, Germany) was used, following EUCAST guidelines [33]. Prior to imaging, plates were incubated according to the species’ growth requirements at 30 or 37C, respectively, without shaking.

For testing the five Mucoralean strains experiments were repeated twice with 8 technical replicates. For EUCAST AFST with A. fumigatus strain R7 experiments were repeated three times with two to four technical replicates.

Microscopy and image acquisition

This step corresponds to step (b) of the workflow (see Fig. 1). Images were taken using a Zeiss Axio Observer 7 microscope equipped with a Zeiss Axiocam 503 color camera (2.83 Megapixel, pixel size: 4.54 m x 4.54 m, Carl Zeiss AG, Jena, Germany) at 4, 7, 10 and 24 h post incubation for Mucorales and for Aspergilli alike. Time points were based on preliminary experiments and experience from previous studies with similar time points [34, 35]. The microscopy software ZEN (blue edition, version 2.3.) (Carl Zeiss AG, Jena, Germany) was used in LIVE mode for controlling the microscope and maneuvering to the desired image section. Images were taken with the following settings: depth of focus: 17.61 m, lens: LD A-Plan 10x, (numerical aperture 0.25, serial number 421241-9910), final magnification 100x, exposure time 4.8 to 6 ms. The light source used was TL LED with an intensity between 6 and 8.61 V. Final effective image pixel size was 0.454 x 0.454 m. Images were acquired in 14-bit per channel RGB format. To achieve comparable images and conditions for each well, it was aimed to photograph the center of each respective well. Also, this resulted in the best illuminated images with no disturbing shadows. Bright-field images were acquired using identical illumination and camera settings. No post-acquisition adjustments to white balance or RGB channel intensities were performed. Images reflect the raw camera output. Following acquisition, images were saved uncompressed in ZEN’s native.czi-file format. For image analysis with caactus, images were exported to tagged image file format (TIF) with linear conversion to 8-bit and LZW compression, without modification of color balance or contrast. For the workflow and file conversion steps, caactus currently supports grayscale (1-channel) and RGB (3-channel) images in TIF format.

caactus – workflow

The main goal of caactus was to provide a Python [36] package to prepare data for working with ilastik [24] and provide a solution for summarizing and analyzing the resulting large data sets. Source code for caactus is openly available at https://github.com/mr2raccoon/caactus and is easily installable from https://pypi.org/project/caactus/. An overview of the workflow is given in Fig. 1. All ilastik training steps were performed in ilastik version 1.4.0 on Microsoft Windows 11 Pro Operating System [24]. We used an HP Z2 Tower G9 Workstation with an Intel64 Family 6 Model 183 Processor, 32 GB Physical Memory and an Nvidia t1000 8GB Graphics Processing unit for training the three ilastik models. The minimum requirements for running ilastik can be found on the developer’s website https://www.ilastik.org/documentation/basics/installation. With the computational resources described above, a trained user required between two to three working days to complete all three training steps in ilastik, depending on the number of training images (see Table 3).

Table 3.

Species and image counts used for training and testing of caactus, as well as time needed for a user of caactus to complete the workflow steps batch-renaming to step csv_summary as shown in Fig. 1c) to h)

Species	Lab ID	Training	Testing	Analysis
		images()	images()	time (min)
Rhizopus microsporus	ATCC 11559	48	96	8.77
Rhizopus oryzae	CD2	11	48	11.72
Lichtheimia corymbifera	CD4	29	144	19.43
Mucor circinelloides	CD6	19	96	14.62
Mucor lusitanicus	R7B	11	48	9.05
Aspergillus fumigatus	R7	20	438	55.27

caactus – batch-renaming

After the user has successfully acquired and exported the images of interest to TIF file format, the first step in caactus serves the purpose of renaming all files according to different conditions and metadata. In the example shown in Fig. 2, the individual wells of the microtiter plate correspond to one individual image. However, any logic given by the user is welcome (the user could also record multiple images per well or a different order on the microtiter plate). The only limitation is that the order of the images saved in a folder and the rows of the table need to be in line with each other. A CSV table serves as the template for the renaming. This format was chosen to provide a low hurdle for users. Spreadsheet editing software like free and open source OpenOffice Calc or standard Windows application Microsoft Excel are widely used and usually preinstalled on most systems. Different conditions and meta-data are listed as the columns (e.g. date of experiment, medium, strain etc.) The information for each well (image) is stored in one row of the table. Figure 2 panel (a) shows a screenshot of CSV-file used for the batch-renaming. The order of the table’s rows follows the order in which images were acquired from the plate (Fig. 2 panel (b)). The column names can be adjusted according to the user’s needs.

Fig. 2.

a Screenshot of an exemplary csv-table as template for batch-renaming, and b scheme of the corresponding 96-well plate with the direction of images being acquired. Compare to Fig. 1 step (c)

caactus – batch-conversion from TIF to HDF5

Following the recommendations of the ilastik documentation for performance enhancement, the second step for data preparation of the workflow converts the images from TIF to hierarchical data format version 5 (HDF5). During the annotation process in ilastik, HDF5-format allows for only partial loading of the image and thus more economical use of computational resources [37]. The step uses parts of a Python script published by the ilastik developers on their website for this purpose.

Machine learning-based image analysis with ilastik

For identifying cells in the images, segmentation and object classification was realized by chaining three ilastik workflows: pixel classification for semantic segmentation of cell boundaries, instance segmentation based on the boundary images in multicut, and object classification [24]. ilastik is a free of charge, open-source software for interactive machine learning under the supervision of rater annotators. After sufficient training of the classifier(s) by the user, new data can be automatically processed by the models in “batch-mode” not requiring user supervision [24].

ilastik – pixel classification

For the first step, the pixel classification workflow was used, which semantically segments images. Here, a random forest classifier is used to classify pixels according to image filters. The basis for decisions are pixel properties like pixel color, intensity and texture. Through learning from user annotations, ilastik estimates the probability that other pixels fall into one of the user defined semantic classes. Three classes were defined: “background”, “boundary” (of cells), and “interior” (of cells). For feature selection in the pixel classification step, all features were selected (for further information, consult ilastik documentation [37]). The resulting probability maps are then used to define cell boundaries in the next step. This allows to separate touching cells as well [24]. Figure 3 panel (a) and (b) shows how pixel in an exemplary image are classified.

Fig. 3.

a Screenshot of original image for pixel classification. b Pixel classification in ilastik: yellow = cell interior, orange = cell boundary, light blue = image background. c Boundary-based segmentation with multicut in ilastik: green = edges to drop, red = edges to preserve, yellow = unannotated edges. d Different categories of cell types for object classification, 1 = resting, 2 = swollen, 3 = germling, 4 = hyphae, 5 = not usable (see Table 2). e Object classification, original image. f Object classification in ilastik: orange = resting spore annotated, yellow = resting spore predicted, dark blue = swollen spore annotated, light blue = swollen spore predicted, dark purple = germling annotated, light purple = germling predicted, magenta = hyphae annotated, pink = hyphae predicted (category not usable not shown). g Raw image with dense mycelial growth. h Image during annotation process: dark green = mycelium annotated, light green = mycelium predicted, dark blue = not usable predicted

ilastik – boundary-based segmentation with multicut

The multicut workflow aims at separating objects by boundaries defined in the pixel classification step. It allows for separating touching objects likely to occur in complex biological images, in our case spores/fungal cells that are sticking to each other. In this step, the user annotates edges between superpixels. Thus, the user defines boundaries belonging to objects as “true” or “false” when they are true edges of an object or belong to the background or are inside of an object. With these user labels, a random forest classifier is fed. For distance transform (DT) watershed settings were put as follows: Threshold 0.30, Min. Boundary size = 2, Smooth = 1.0, Alpha 0.9. For training and multicut, the threshold was set at 0.50. Figure 3 panel (c) shows how edges in an exemplary image are annotated.

caactus – background processing

After the multicut workflow, the background still exists as a segmented object. In order to exclude the background as an object from the object classification, an intermediate step is included. ilastik assigns an ID to each object as well as pixel counts as a result of the multicut process. Here, the largest object is filtered out and set to zero. Thus, the background is excluded from analysis in the object classification step.

ilastik – object classification

The object classification step is designed to classify objects based on object-level traits, such as shape and size, into categories defined by the user [24]. For object feature selection, all features of the selection were included (for further information, consult ilastik documentation [37]). For the caactus workflow, the categories were defined based on fungal morphologies in different developmental stages (compare to Fig. 3 panel (d)):

1. resting spore;

2. swollen spore;

3. germling;

4. hyphae;

5. not usable;

6. mycelium (For the antifungal susceptibility testing (AFST) characterization of A. fumigatus, this sixth category was defined. Shown in Fig. 3 panel (g) and (h)).

For EUCAST AFST the category mycelium was recorded as a binary outcome. Since the object classification step failed to determine the whole image as mycelium, but would also misdetect other object categories and not annotate part of the image (Fig. 3 panel (h)), we chose a workaround. If any large enough mycelial structure was detected (green), the percentage of Predicted Class, was set to 100% for those images in the summary_statistics_eucast.py-script. Results of this 24 h readout were always cross-validated by standard EUCAST macroscopic readout via mirror [38].

Figure 3 panel (e) and (f) shows how objects in an exemplary image are classified during the training step of object classification.

Next to their use for classifying the objects, the characteristics of the objects can be exported along their predicted class. This allows for deeper analysis of the morphologies. For showcasing, we only used the parameter size in this paper (e.g. Fig. 4 panel (a) and (d)).

Fig. 4.

Exemplary output of the summary statistics step. The dataset originated from analyzed images of all five Mucoralean strains (4 h post incubation in YPG medium). Spores were directly incubated after spore harvest (fresh) or snap frozen in liquid nitrogen and stored at (frozen). Data shown are objects classified from images from two independent experiments with eight technical replicates each (one image per well). a screenshot CSV-file with summary statistics and b stacked bar chart of the relative frequency of different cell types grouped by experimental variables (here strain and storage). c Principal Component Analysis (PCA) plot created with pyPLNmodels and caactus for dimension reduction. Each dot corresponds to one image. Graph shows influence of storage condition (fresh vs. frozen). Output (df_clean.csv) used for subsequent analysis. d Correlation circle on the transformed variables of dimension reduction with PCA. The plot shows how much each variable (here storage condition and strain) contributes to the variance explained (difference in counts of cell types in each image). e Violin plots of cell size per cell type and strain. Vertical lines in violins represent 25th, median and 75th percentile. For some samples (CD2 (germling, frozen) R7B (hyphae, frozen), CD4 (germling, fresh) and CD6 (hyphae, fresh)) only a single cell was classified across the 16 images for each category. Hence, there is only a singular vertical line

caactus – data post-processing

The output of object classification from ilastik generates one individual spreadsheet file (csv) per image analyzed with all the information on the objects stored in it. To facilitate subsequent data analysis, the individual files per image are:

1. summarized into one global table, which is then used to;

2. calculate summary statistics and prepares the data for;

3. statistical modeling.

The user may define experimental conditions of interest (e.g., strain, media, time point etc.) to group the data accordingly. The hard coded variables biorep and techrep for biological replicate and technical replicate are used to calculate averages across experiments. The counts of different cell morphologies per condition are thus conveniently summarized for user interpretation.

Figure 4 panels (a) and (b) show exemplary output of the summary statistics step: a CSV-file with descriptive summary statistics and stacked bar chart of the relative frequency of different cell types grouped by experimental variables.

Furthermore, the global output table (df_clean.csv) generated by the script csv_summary.py can also be used for analysis with different statistics software by the choice of the user (GraphPad, R, Python etc.). As an exemplary use case (see Fig. 4 panel (e), we used it for displaying cell size differences of the five Mucoralean strains after 4 h incubation in YPG medium when incubated directly after harvest (fresh).

caactus – Poisson lognormal (PLN) modeling

In the final post-processing step (Fig. 1 step (h)), the pyPLNmodels package is used for analyzing multivariate count data [39]. It can be utilized for performing a multitude of analysis like multivariate sample comparison, clustering of sites or samples and dimension reduction (ordination) for visualization purposes [40]. The script prepares the data frame for the structure demanded by pyPLNmodels and fits a model for analyzing the multivariate counts (of different cell morphologies) depending on the experimental conditions specified. Figure 4 panels (c) and (d) show how objects in an exemplary image are classified. By default, Zero inflated Poisson Log normal Model is applied. This model is particularly suitable for growth dynamics data, because many zero counts are expected, especially at early time points when specific morphotypes may not yet be present. The results of this output are meant for visual interpretation, for rapid diagnostics. Tests like PERMANOVA etc. for separating groups based on metadata need to be performed by users independently.

Defining fungal morphologies

For consistent algorithm training in object classification within ilastik, precise definitions of filamentous fungal morphologies were derived from the relevant literature [41–43]. Where definitions were elusive, such as for hyphae, empirical definitions were adopted based on observational data. Figure 3 panel (d) shows an exemplary image with different object categories. The categories are defined in more detail below and given as an overview in Table 2.

Table 2.

Overview of cell morphology categories used in the caactus workflow

Cell morphology	Definition	Source
Resting spore	Lack of hydration	[42]
Swollen spore	Isotropic expansion upon hydration, initiation of germination	[42, 43]
Germ tube	Disruption of the polarity axis in a swollen sporea	[41, 43]
Hyphae	Length of germ tube 4 diameter of originating cell, branching in maturationb	[41, 43]
Mycelium	Network of branched hyphaec	[41]
Not usable	category for objects not of interestd	–

^aThe disruption of the polarity axis in a swollen spore marks the transition to a germ tube, indicating the onset of directional growth [41, 43]. However, it has been shown for Penicillium roqueforti that swelling is not a necessary prerequisite for the formation of a germtube

^bHyphae result from the continuous apical expansion of polar growth. The transition from germ tube to hyphae remains ambiguously defined in the literature. In septate fungi, the initiation of septa formation signifies the beginning of hyphal maturation [41, 43]. For non-septate Mucorales, a working definition considers a structure as hyphae when its total length, from its origin, equals or exceeds four times the diameter of its originating cells. Hyphal maturity is further indicated by the occurrence of branching [42]

^cFinally, this fifth category was defined for classifying images for the AFST workflow in which dense fungal growth with branching hyphae was evident. One of the features for the formation of a mycelium is branching to form secondary and tertiary hyphae [42]. Thus, a mycelium is a “network of branched hyphae” [41]. For operational purposes, during the object classification step, patches of dense hyphal growth that showed branching and thus hyphal growth were defined as mycelium. Since the object classification for hyphae resulted in patchy and incomplete segmentation masks, we decided to record mycelium as binary outcome. The occurrence of mycelium was additionally confirmed macroscopically according to the EUCAST protocol with mirror reading [38]

^dThis category served as a “residual” category. Objects that were partially visible (‘cut-off’), artifacts of the media, or clusters of cells that were not separable by the algorithm were classified as ‘not usable’ and excluded from further analysis

Validation

After completing the training of the three ilastik workflow steps (steps (d), (e) and (g) in Fig. 1) for each strain, images were analyzed in batch mode by a user. The output generated with the caactus workflow for the five Mucorlean strains by this user was then compared with the manual count of three raters trained in mycology. To achieve this, ten images per strain were randomly selected showing cells in different developmental stages. The raters used printed images to count manually and record the number of objects per category (resting, swollen, germling, hyphae, not usable) by entering them into spreadsheet file.

The number of training and testing images is recorded in Table 3. This table also includes the time needed from workflow steps batch-renaming to step csv_summary as shown in Fig. 1c) to h). This time was used as comparison against the human raters. The time needed for the AFST analysis was recorded, but not used in comparing to human raters. The number of images for validation against human raters for the Mucoralean species was always .

Modelling

To assess the agreement between the raters and caactus, the count data were modeled using the Poisson lognormal (PLN) approach [39], which accounts for overdispersion and correlations between counts. First, all images across all five species were modeled globally, followed by separate analyses per species. The counts of the five morphological categories (resting, swollen, germling, hyphae, not usable) were treated as a multivariate response. Specifically, the response is a matrix with m rows (number of images) and n columns (number of cell morphologies, here ), i.e., The predictors included the rater (i.e., rater1 to rater3, with the caactus method as the reference group) and species (see Table 1 for details on the species). Initially, two models—the Null Model (species only) and the Full Model (species + rater)—were compared. In matrix notation, the PLN model can be written as:

1

where is the multivariate response (counts of each cell morphology), is the covariate vector (species for the Null Model; species + rater for the Full Model), and is the variance-covariance matrix. Here, denotes the Poisson distribution and the multivariate normal distribution. The latent vector represents the log-rate of each morphology for image i, capturing the correlation structure between the different cell types.

Similarly, the model specification in Eq. 1 was fitted for each species individually. For each species, the Null Model included only an intercept, and the Full Model included rater as a covariate. For strain CD4, the hyphae column, which contained only zeros, was excluded from the count matrix to ensure identification, as no hyphae were present at the time points used and hence no rater counted this category.

Model fit was evaluated using the log-likelihood (logLik), Akaike Information Criterion (AIC; [44]), and Bayesian Information Criterion (BIC), which are standard statistics for comparing statistical models. The log-likelihood quantifies how well the model explains the observed data; higher values indicate better fit. In contrast, AIC and BIC balance model fit and model complexity by penalizing the number of estimated parameters. Lower AIC and BIC values therefore indicate a more parsimonious model with a better trade-off between goodness-of-fit and complexity. These criteria were used for relative comparison between competing models rather than as absolute measures of fit. Because the Poisson-lognormal model uses latent variables, the likelihood is approximated using a variational method rather than calculated exactly. In the pyPLNmodels package, AIC and BIC are computed on the same scale as this approximates likelihood, which differs from the classical formulas by a factor of 2. This ensures the reported criteria are consistent with the model’s fit while still penalizing for model complexity [39, 40].

The relevance of rater identity was assessed using a likelihood ratio test (LRT), which compares the goodness-of-fit of a full model (including rater identity) to that of a nested null model (excluding rater identity). A statistically significant LRT indicates that inclusion of rater identity significantly improves model fit.

For both global (all species combined) and species-level analyses, rater-specific regression coefficients were exponentiated to obtain rater/caactus ratios for each cell type, with 95% confidence intervals (CIs) derived from the fitted model. Exponentiation allows interpretation on the ratio scale. Ratios > 1 indicate higher counts relative to caactus (overestimation), whereas ratios < 1 indicate lower counts (underestimation). Statistical significance was determined based on whether the 95% CI excluded 1. Specifically, if the entire CI lies above 1, this indicates significant systematic overestimation; if the entire CI lies below 1, this indicates significant underestimation.

To account for multiple comparisons across raters and cell types, p-values were adjusted using the Bonferroni correction. This method controls the family-wise error rate by dividing the chosen significance level (e.g., = 0.05) by the number of tests performed, thereby reducing the probability of false-positive findings due to multiple testing. Forest plots (Fig. 5) display point estimates together with their 95% CIs.

Fig. 5.

Forest plots for rater/caactus count ratio across cell morphologies. Each point represents the ratio of rater to caactus counts for a given cell type, with horizontal lines indicating 95 % confidence intervals. The ratio is given as its numerical value after the rater pair that is compared. Ratios above 1 indicate higher counts by raters relative to caactus, ratios below 1 indicate under-counting, and confidence intervals overlapping 1 suggest no significant difference. Reported p-values are Bonferroni-corrected for multiple testing. Asterisks denote statistical significance of the deviation from parity for . For validation, ten images per species were counted by human raters and caactus. Images were chosen from larger datasets with two biological and eight technical replicates per species. For species ID refer to Table 1.

Total number of objects

As an additional validation check to compare the caactus workflow with the raters, the total number of objects per image was calculated for each rater by summing all cell categories. The total counts were modeled using Negative Binomial generalized linear models (NB-GLMs) with a log-link function to account for overdispersion. To evaluate potential rater effects, a null model without rater and a full model including rater effect were fitted both for the global dataset (all species combined) and separately for each individual species.

• Global models

  • Global Null model (no rater effect): The total counts were modeled as a function of species only:

    
  • Global Full model: The total counts were modeled as a function of species and rater:

    
• Species-level models

  • Species-level Null model (no rater effect): The total counts were modeled with an intercept only:

    
  • Species-level Full model: The total counts were modeled as a function of rater:

    

Likelihood ratio tests (LRTs) were used to compare the null and full models globally and per species. As no significant rater effects were detected, no post hoc pairwise comparisons were performed.

Time

For comparing analysis time among the raters and caactus, each rater stopped the time needed for counting all objects per image and filling the count data into a spreadsheet. The time was compared across all five species. It was then compared with the time a pre-trained user needed to perform image analysis with the caactus workflow from the step batch-renaming to step csv_summary as shown in Fig. 1c) to h) for all test images per species (compare to Table 3). Here, the average time was calculated for each species by dividing the total time by the number of images processed.

Predicting mycelium formation from early morphologies for antifungal susceptibility testing (AFST)

Data preprocessing

Raw counts of each morphotype were aggregated at the image level to obtain total object counts per image. For each image, the relative abundance (%) of each morphotype class was computed as the fraction of the total number of objects of that image. This normalization ensures that each image contributes equally, independent of object density of individual images.

For the 10 h dataset, the resulting variables and represent the percentage of each morphotype within a given well–replicate combination.

After preprocessing, datasets from 10 h and 24 h time points were merged by matching concentration, well, biological replicate, and technical replicate identifiers to ensure correspondence between early and late morphological states. Missing or incomplete combinations were excluded from downstream analyses.

To determine whether early morphotypes could predict subsequent mycelium formation, logistic regression models were fitted, linking the proportion of hyphae or germlings observed at 10 h to the binary outcome of mycelium at 24 h. In standard EUCAST AFST testing, MIC-values are read by human raters at 24 h hours for determining growth per well in a binary fashion with a mirror [33].

The dependent variable, , was defined as:

Predictors () were the measured proportions (in %) of hyphae or germlings at 10 h within the same well-replicate combination.

Model 1 – hyphae predictor

The first model assessed whether the proportion of hyphae at 10 h predicted mycelium presence at 24 h. The logistic regression model was expressed as:

where:

• is the conditional probability that mycelium is present at 24 h of observation , given the predictor (here, the percentage of hyphae at 10 h),

• is the intercept term, and

• represents the change in log-odds of mycelium presence per 1 % increase in hyphae at 10 h.

The corresponding odds’ ratio (OR) was calculated as:

which represents the multiplicative change in odds of mycelium presence for each 1 % increase in hyphae.

Model 2 – germling predictor

A second logistic regression model was fitted analogously using the proportion of germlings at 10 h as the predictor:

where parameters have the same interpretation as above, with denoting the percentage of germlings observed at 10 h.

Model estimation and performance

The models were estimated using maximum likelihood estimation (MLE) implemented in the scikit-learn [45] and statsmodels [46] Python packages.

For the binary logistic regression models (mycelium presence vs. absence), model discrimination was quantified using the area under the receiver operating characteristic curve (AUC), which measures the ability of the model to correctly distinguish between presence and absence across all possible classification thresholds.

For both the logistic and the (multivariate) count data models, model fit was assessed using McFadden’s pseudo- and likelihood ratio (LR) tests. McFadden’s pseudo- is defined as

where denotes the log-likelihood of the fitted model and that of the intercept-only model. As a likelihood-based measure, it quantifies the proportional improvement in log-likelihood relative to the null specification and is appropriate for generalized linear models estimated by MLE, including logistic and Poisson-type count models. Although bounded above by 1, values of McFadden’s pseudo- are typically substantially smaller than classical values in linear models. In practice, values in the range of approximately 0.2–0.4 are generally considered indicative of a good fit for discrete outcome models.

Reporting both AUC and pseudo- provides complementary information: while the AUC evaluates discriminatory performance in the binary classification setting (i.e., the ability to correctly rank presences versus absences), pseudo- and LR tests assess improvement in overall log-likelihood relative to a null model across likelihood-based specifications.

Parameter estimates () were reported with standard errors, Wald -statistics, -values, and 95% confidence intervals. For the logistic models, exponentiated coefficients () were interpreted as odds ratios (ORs), indicating the multiplicative change in the odds of mycelium presence associated with a one-unit (here: 1%) increase in each morphotype, holding all other variables constant. For the count models with log link, exponentiated coefficients were interpreted as incidence rate ratios (IRRs), representing multiplicative changes in the expected fungal counts for a one-unit increase in the predictor, holding other variables constant.

Results

caactus – experimental workflow

Using ilastik and Python scripting [36], we developed a workflow capable of rapidly detecting, separating, and classifying different cell types from bright-field microscopic images. To enable users with minimal or no programming experience to apply this workflow reliably, we created a series of Python scripts that automate data preparation and result summarization. To enhance usability, the scripts are wrapped into a GUI. We present the resulting cell analysis and counting tool based on the ilastik software (caactus) for the analysis of large sets of microscopic images.

caactus provides consistency in counting across cell types and species

To assess annotation consistency, Poisson log-normal (PLN) models were fitted and compared using likelihood ratio tests (Table 4). For the global dataset including all species, adding rater identity significantly improved model fit compared to a model including only species ((15) = 94.72, 0.001). The 15 degrees of freedom correspond to the additional regression parameters introduced by including rater effects across the multivariate morphotype counts.

The full model (log-Likelihood (logLik) = –3562.3, AIC = 3617.3, BIC = 3708.0) provided a significantly better fit than the null model without rater effects (logLik = –3609.7, AIC = 3649.7, BIC = 3715.7), as indicated by a likelihood ratio test: (15) = 94.7 ( 0.001). This global result indicates that differences between human raters and caactus exist across the full dataset. However, species-specific analyses revealed that these differences were not uniform. For ATCC11559 ((15) = 13.25, ), CD2 ((15) = 24.54, ), and CD4 ((12) = 14.66, ), inclusion of rater identity did not significantly improve model fit, indicating no evidence for systematic differences between human raters and caactus for these strains. The lower degrees of freedom for CD4 reflect the fact that the category hyphae was not present in this dataset.

In contrast, significant rater effects were detected for CD6 ((15) = 65.29, 0.001) and R7B ((15) = 80.90, 0.001), suggesting species-specific discrepancies in morphotype counts. Thus, although overall statistical differences between raters were detectable, these differences were primarily driven by two species rather than reflecting a systematic deviation across all strains.

The results for each model and the individual cell types are presented in detail below.

Global analysis – across all species

Resting

When examining the specific effects of raters in reference to caactus at the global level (Table 5, Fig. 5 panel (a)), no significant differences in cell counts between raters and caactus were observed, although ratios were smaller for rater1 and rater2 (all Bonferroni-adjusted 0.05).

Swollen

All human raters produced ratios close to 1 (0.80–0.89, all Bonferroni-adjusted 0.05), with wide confidence intervals overlapping 1, indicating no systematic difference relative to caactus on a global scale (Table 5, Fig. 5 panel (a)).

Germling

For the category swollen, no significant differences were observed for the cell type germling between all raters and caactus on the global scale (all Bonferroni-adjusted ; Table 5, Fig. 5 panel (a)).

Hyphae

When comparing across all species (Table 5, Fig. 5 panel (a)), rater3 significantly overcounted the cell category hyphae (ratio = 3.57, CI = 1.72–7.40, ).

Not usable

On the global scale, rater2 under-estimated counts (ratio = 0.456, CI 0.28–0.74.28.74, ), whereas rater1 and rater3 did not significantly differ from caactus (Table 5, Fig. 5 panel (a)).

Species level analysis

Likewise, models were fitted for each species. In general, caactus agreed with human raters, except for minor deviations described below. As listed in detail and shown in Fig. 5, human raters agreed with caactus for all cell types (resting, swollen, germling, not usable) except hyphae, for which raters either over- or undercounted. However, no systematic pattern of human versus machine differences was apparent. Results, described in detail below, are summarized in Supplementary Table 4 and Tables 6–10.

ATCC 11559

For ATCC 11559 no influence of rater could be observed when fitting the model (Table 4). This indicates that no differences between raters and caactus could be observed for ATCC 11559 for any morphotype (Table 6 and Fig. 5 panel (b)).

CD2

For CD2 no influence of rater could be observed when fitting the model (Table 4). Although not statistically significant, the result showed a trend toward a rater influence (). This can be accredited to rater3 who overcounted hyphae for CD2 (ratio = 3.07, CI 1.97–6.87.97.87, ). For the other morphotypes (resting, swollen, germling and not usable), no differences could be observed for CD2 (Table 7 and Fig. 5 panel (c)).

CD4

For CD4 no influence of rater could be observed when fitting the model (Table 4). Like for ATCC 11559, this indicates that no differences could be observed for CD4 for any morphotype (Table 8 and Fig. 5 panel (d)).

CD6

For CD6 an influence of rater could be observed when fitting the model (Table 4). For CD6 (Fig. 5 panel (e), rater1 undercounted resting spores (ratio = 0.181, CI 0.12–0.27.12.27, 0.001), whereas rater2 and rater3 showed no significant difference in resting spores across the ten sample images.

Similarly for hyphae, rater1 undercounted for CD6 (ratio = 0.011, CI 0.003–0.04.003.04, 0.001). In contrast, rater3 overcounted hyphae for CD6 (ratio = 12.9, CI 8.96–18.63.96.63, 0.001). Rater2 showed no differences to caactus in counting hyphae. For categories swollen, germling and not usable no significant differences between the raters and caactus were apparent (Table 9 and Fig. 5 panel (e)).

R7B

For R7B, an influence of rater could be observed when fitting the model (Table 4). In the category resting, rater1 (ratio = 0.125, CI 0.09–0.17.09.17, 0.001), and rater2 (ratio = 0.101, CI0.06-0.17.17.17, 0.001), both undercounted while with rater3 no difference could be observed (Fig. 5 panel (f). Similarly for hyphae, rater1 undercounted for R7B (ratio = 0.376, CI 0.26–0.55.26.55, 0.001), whereas rater3 over counted hyphae for R7B (ratio = 3.12, CI 1.91–5.10.91.10, 0.001). For hyphae, rater2 showed no differences to caactus. Similar to CD6, no significant differences between the raters and caactus were apparent for categories swollen, germling and not usable (Table 10 and Fig. 5 panel (f)).

In summary, differences between human raters and caactus were observed for certain cell types, but these deviations were minor and not systematic. Rater 1 tended to undercount, whereas Rater 3 tended to overcount, yet no consistent bias relative to Caactus was apparent. Overall, caactus is a reliable tool that enables consistent and reproducible quantification of cell morphology abundances from bright-field microscopic images.

The total number of objects does not differ between caactus and human raters

Likelihood ratio tests (see Table 11) confirmed no significant rater effect globally ( = 0.33, ). The 3 degrees of freedom reflect the three additional regression coefficients introduced when including rater identity (three human raters compared to caactus as reference).

Similarly, for the individual species no rater effect could be observed (ATCC11559: = 0.04, ; CD2: = 0.01, ; CD4: = 0.04, ; CD6: = 0.11, ; R7B: = 0.99, ).

These results indicate that including rater identity did not improve model fit, showing that expected total object counts were statistically indistinguishable between caactus and human raters. Descriptive statistics are provided in Table 12 and visualized in Fig. 6.

Fig. 6.

Total number of objects per image analyzed between raters and caactus. The top line represents the 75th percentile, the middle horizontal bar the median and the bottom line the 25th percentile. Error bars indicate last points within 1.5x interquartile range. Open circles represent outliers are outside 25th percentile - 1.5 interquartile range or 75th percentile +1.5interquartile range, respectively. Outliers were not excluded from analysis. For validation, ten images per species were counted by human raters and caactus. Images ( per species) were chosen for cross-validation by human raters from larger datasets with two biological and eight technical replicates per species each. The underlying data is summarized in Table 12. For species ID refer to Table 1.

The absence of significant rater effects in total object counts indicates that caactus detects overall object numbers as consistently as human raters. Building on this, the agreement observed across most morphotype-specific comparisons suggests that caactus not only detects a consistent number of objects overall, but also classifies them consistently into morphotypes with comparable accuracy to human raters.

Overall, these results show that the total object counts in the testing images are consistently detected between caactus and its human counterparts. Taken together with the analysis per morphology type, these results underline that the caactus workflow reliably detects all relevant objects from images and can effectively discriminate them based on morphological characteristics.

caactus significantly reduced the time needed for analysis compared to human raters

The time for analysis was compared between human raters and caactus from workflow step batch_renaming to step csv_summary.

As expected, image analysis with caactus significantly decreased the time needed per image ( per species) when comparing each rater’s time to the time caactus needed (Conover’s post hoc test, ). Interestingly, highly significant differences were observed in between the group of raters, revealing that rater2 was much faster in counting than rater1 and rater3. The results of Conover’s post hoc test for the comparisons of the human raters vs. caactus and among each other are provided in Table 13. The human raters needed three and more minutes on average per image (rater1 = 3:55 min, rater2 = 3:00, rater3 = 3:34 min) whereas caactus only nine seconds. The average evaluation time per image was decreased 24.5x when comparing the average time of all three raters with the average time of the caactus workflow needed for each species’ model for the subset of analysis steps (see Fig. 7), clearly showing the biggest advantage of using caactus. In summary, this time saving potential allows for rapid and reliable analysis of larger image datasets for statistically robust results with multiple biological and technical replicates.

Fig. 7.

Time per image for counting. Boxes represent average time [min] for all sample images per rater or caactus. The top line represents the 75th percentile, the middle horizontal bar the median and the bottom line the 25th percentile. Error bars indicate last points within interquartile range. Open circles represent outliers are outside 25th percentile - interquartile range or 75th percentile interquartile range, respectively. Outliers were not excluded from analysis. The average time per validation image () was calculated across all five Mucoralean species. It was then compared with the time a fourth user needed to perform the analysis with the caactus workflow from the step batch-renaming to step csv_summary as shown in for all test images per species (). Here, the average time was calculated for each species by dividing the total time by the number of images processed

caactus can discriminate growth dynamics of different fungal species at an early time point

As shown in Fig. 4 panels (a) and (b), the five species exhibit distinct proportions of resting spores, swollen spores, germlings, and hyphae, indicating that germination dynamics are species-specific and are additionally influenced by post-harvest storage conditions (frozen at or used directly). We chose 4 h as an early time point for cross-species comparison, since for rapidly germinating species such as R. arrhizus strain CD2 and M. lusitanicus R7B, analysis with caactus was not possible at later time points; by 6 h, extensive hyphal growth and mycelium formation occurred within the wells. In contrast, the slower-growing species could be analyzed at 6 h or 8 h post-inoculation (data not shown).

Similarities in growth dynamics are illustrated in Fig. 4 panel (c). The PCA plot shows that the faster-growing species (CD2 and R7B) cluster together when incubated directly after harvest (fresh), consistent with similar proportions of cells in a more advanced development stage (germlings and hyphae) (Fig. 4 panel (b) fresh). In contrast, the remaining strains and storage conditions form individual clusters, generally suggesting species specific reactions to spore storage condition. PCA1, accounting for 37% of the variance, primarily separated samples by storage condition, whereas PCA2 accounted for 31% of the variance based on strain identity (Fig. 4 panel (d)). These differences in growth dynamics (proportion of different morphotypes per image) were further reflected by cell size differences at the analyzed time point (Fig. 4 panel (e)). Average morphotype counts and cell sizes for all strains are summarized in Table 14. Consistent with their faster growth, these species develop larger germlings and hyphae at 4 h. Taken together, this example demonstrates how caactus enables cross-species comparisons of fungal growth behavior under different experimental conditions, here spore storage prior to incubation) at defined time points (4 h in this case).

caactus can predict minimum inhibitory concentrations at early time points

Finally, we analyzed standard EUCAST AFST assay in microtiter plates with the widely employed antifungal drug Amphotericin B (AmB) against A. fumigatus R7. We wanted to investigate whether the percentage of hyphae and germlings at earlier time points than 24 h (here 10 h), can serve as a predictor of drug responsiveness (Fig. 8 panel (a) and Table 2). In our data, complete mycelial growth at 24 h was observed only at Amphotericin B concentrations between 0–0.125 g/ml, which corresponds to a MIC of 0.25 g/mL. The MIC value correspond to official EUCAST MIC ranges of 0.25–1.25 g/mL for A. fumigatus ATCC 204305 (Lab ID R7) [38]. The average percentage for each morphotype over time are displayed in Fig. 9. To statistically support the morphology analysis at early time points with caactus, we fitted logistic regression models predicting mycelial presence at 24 h from percentage of hyphae or germlings at 10 h, respectively. The models are described below.

Fig. 8.

a Average percentage of hyphae and germlings at 10 h relative total number of objects per image (averages were computed from images of three independent experiments with two to four technical replicates per time and concentration). Presence of mycelium at 24 h recorded as binary outcome 1 = mycelium growth, 0 = no mycelium was multiplied by 100 solely to rescale the value to the same 0–100% axis used for the other variables in the figure, allowing all measures to be displayed together. Since the object classification for hyphae resulted in patchy and incomplete segmentation masks, we decided to record mycelium as binary outcome. The occurrence of mycelium was confirmed macroscopically according to the EUCAST protocol with mirror reading (data not shown) [38]. b and c Logistic regression model fits predicting mycelium presence at 24 h based on (b) hyphae and (c) germling proportions at 10 h. Model performance assessed by the area under the curve (AUC) yielded values of 0.83 and 0.92, respectively, indicating strong discriminatory ability

Hyphal model

The logistic regression model using the proportion of hyphae at 10 h as a predictor of mycelium presence at 24 h showed good discrimination between positive and negative wells (AUC = 0.83, Fig. 8 panel (b)). The exact percentage of hyphae at each concentration can be viewed in Table 15. Model fit was significant (); likelihood-ratio test , indicating that early hyphal abundance explained nearly half of the variance in mycelium detection (Table 16). Classification accuracy was 86%, with perfect recall for wells lacking mycelium (recall = 1.00) and moderate recall for wells containing mycelium (recall = 0.56).

The model coefficients were significant , yielding an odds’ ratio (OR) of . The intercept corresponds to baseline odds of mycelium formation of 0.15 in wells without hyphae. A note of partial quasi-separation (16% of observations perfectly classified) suggests a threshold-like effect, where high hyphal percentages almost invariably led to mycelial development.

The proportion of hyphae at 10 h was a strong, statistically significant positive predictor of mycelium formation at 24 h. Each 1% increase in hyphal proportion increased the odds of subsequent mycelium detection by a factor of approximately 258 . This strong relationship implies a nonlinear, threshold-driven transition from early hyphal growth to established mature mycelium, consistent with hyphae being a direct morphological precursor to mature mycelium.

This proofs, that hyphal development, even at its starting point, reliably predicts subsequent mycelia growth. This indicates at which concentrations an antifungal drug is not effective. Our data reinforce the potential of early analysis of fungal morphology to differentiate between wildtype (=susceptible) and non-wildtype (resistant) populations faster than in the usual clinical setting, with reading by eye at 24 h and later [38].

Germling model

A second logistic regression model assessed whether the proportion of germlings at 10 h predicted mycelium presence at 24 h. The exact percentage of germlings at each concentration can be viewed in Table 15. This model exhibited excellent predictive performance (AUC = 0.92, Fig. 8 panel (c)) and strong model fit (pseudo-; LLR ) (Table 17). Overall accuracy reached 89.7 %, with precision = 0.88 and recall = 0.99 for wells lacking mycelium, and precision = 0.96 with recall = 0.69 for wells with mycelium.

The fitted coefficient for germling percentage was highly significant (, ), corresponding to an odds’ ratio of (95 % CI: 1.16–1.58). The intercept () reflects low baseline odds of mycelium formation in wells devoid of germlings. A small subset of cases (13 %) were perfectly classified, again suggesting biological rather than numerical separation in morphological outcomes.

The proportion of germlings at 10 h was also a significant positive predictor of mycelium formation at 24 h. Each 1 % increase in germling proportion increased the odds of detecting mycelium by approximately 35 % (). Unlike the hyphae model’s abrupt transition, the germling–mycelium relationship was smoother and more proportional, indicating that at this time point many of these germlings will still form into hyphae and subsequently contribute to the development of mature mycelium.

Discussion

The workflow created proved to be a fast and reliable method for counting different fungal morphology types from bright-field microscopic images. In contrast to human raters, the method is clearly superior in terms of time saving. Although differences in counted objects per cell morphology existed for some species, these differences were not uniform among the three human raters when compared to caactus, indicating that variability was more likely due to subjective interpretation of human raters than to systematic bias of the automated workflow.

Although the global model indicated significant rater effects, these differences were species-specific and not observed for the majority of strains. For three of five species, inclusion of rater identity did not improve model fit. Where significant effects were detected (CD6 and R7B), discrepancies were restricted to specific morphotypes and did not consistently affect total object counts. Thus, while statistical differences were present, their practical impact appears limited.

Across all species, there was no significant difference in counts between caactus and the human raters for the categories swollen, germling, and not usable. This suggests that these categories are most easily distinguishable not only for the machine learning classifiers implemented in ilastik, but also for the human eye. In general, for species ATCC 11559 and CD4, no differences emerged across any category between caactus and the human raters. Similarly, for CD2, only one rater overcounted the cell category hyphae.

To our surprise, the related category germling did not show corresponding undercounting by this rater (rater 3). A similar pattern was observed for strain CD6, where rater 1 undercounted resting spores and hyphae, but this did not result in overcounting of related categories such as swollen spores or germlings. For R7B, rater 1 and rater 2 undercounted resting spores without overcounting swollen spores. However, this contributed to a lower total object count in R7B, although total counts did not differ significantly overall (see Fig. 6 panel (f)). Taken together, this suggests that a) in general all cell categories are readily distinguished by human raters and the ilastik models and that b) human raters do not count consistently. These results indicate that individual perception and different ability to concentrate while counting lead to deviating results by human raters. Rater specific perception is supported by the fact that rater1 only undercounted in four incidences (2x resting and 2x hyphae) and rater3 only overcounted (1x resting and 3x hyphae), whereas rater2 was always in agreement with caactus except for 1 undercounting of resting.

Thus, using a tool such as caactus ensures consistency across experiments, independent of the operator. When comparing total object numbers, no significant differences emerged between the raters. Nevertheless, it should be kept in mind that, when adapting the workflow to other fungal species or cell types, performance also depends on the quality of the initial annotation steps performed in ilastik. Time saving and consistency remain the main advantages of such workflows [23].

Regarding image acquisition, we observed that consistent lighting conditions are crucial both for training images and for subsequent analysis. In this respect, human raters may have an advantage in post hoc analysis, as they can visually compensate for variations in lighting between images, whereas automated approaches rely on standardized acquisition settings. Image pre-processing should be considered if acquisition settings cannot be optimized sufficiently.

In addition to object size, a wide range of geometric and color features (>30 parameters, especially when additional staining is applied) can be extracted to further distinguish and classify objects of interest. In microscopic imaging, white balance must be properly adjusted, as variations in lighting and color calibration can influence color-based measurements and therefore represent a limitation of the current approach. At the same time, this feature extraction offers the possibility of more in-depth analyses of morphological differences in fungi grown under varying conditions. The computable features are described in detail on the developer’s website. This represents another major advantage compared to human raters. Accurately assessing geometric parameters of numerous objects and distinguishing subtle size differences—such as swollen versus resting spores or germlings versus hyphae—is nearly impossible to perform consistently by manual scoring. By analyzing the geometry of fungal cells throughout their growth, biologically relevant traits can be quantified in a reproducible manner.

Commercial solutions such as oCelloScope^TM undoubtedly provide more advanced analytical tools, user-friendly interfaces, and faster processing of entire 96-well plates (approximately 2 min 36 s per complete scan, according to the manufacturer). In contrast, acquisition of images for a full 96-well plate within the caactus workflow required approximately 20 minutes for an experienced user. Furthermore, oCelloScope^TM offers the advantage of automated image acquisition at predefined time points directly within an incubator. Users of caactus, in contrast, must remove plates from incubation, which limits experiments under strictly controlled environmental conditions such as hypoxia or assays requiring tightly regulated temperatures.

Nevertheless, caactus and ilastik [24] are free of charge, open source, and independent of the user’s operating system, and they can be readily adapted to different organisms. Reusing ilastik models trained on one operating system on another was possible if relative file path were used when importing the raw images, prediction masks and segmentation images into the project file (e.g. sharing of ilastik models trained in ilastik 1.4.0 on Windows 11 Pro worked in the same ilastik version on Linux Mint 21.3 Cinammon). This allows using trained models on different machines and fostering sharing between groups. Technically, images from any microscope—or even images captured using a smartphone with an ocular adapter—can be analyzed. Imaging hardware settings such as camera resolution or magnification are independent of caactus, and available equipment therefore constitutes the main limitation. In contrast, oCelloScope™ is restricted to a fixed optical magnification (4) with defined optical resolution and pixel size [17], whereas standard research microscopes allow flexible adjustment of magnification according to experimental requirements.

Beyond hardware and cost considerations, the choice of image analysis methodology itself fundamentally shapes what biological questions can be addressed and how readily a workflow can be adapted to new experimental contexts. Workflows based on classical image processing, as presented by Barry et al. [22] and Hardy et al. [28], require no training data or machine learning expertise, are computationally minimal and fully transparent, but rely entirely on hand-crafted operations such as thresholding, Gaussian filtering, and skeletonization. Performance may therefore degrade significantly with variable illumination, complex morphologies, or touching and overlapping objects. Random forest classifier-based approaches, such as ilastik, require relatively little annotation time and few training images, allowing rapid adaptation to new experimental systems. Furthermore, the direct user feedback in the ilastik GUI during training, allows for accurate expert supervised results. In their study of C. albicans hyphal formation, Praetorius et al. [30] found their classical MVP pipeline outperformed an ilastik-based approach for hyphal length measurement. However, ilastik training was restricted to only two samples, and accuracy would likely improve with a larger training set. CNN-based approaches, as presented by Padovani et al. [27] and Dietler et al. [26], generally provide the highest segmentation accuracy but require large annotated training datasets. Notably, Praetorius et al. [30] demonstrated that their classical pipeline achieved an F1-score of 0.82 for crossing hyphal detection — a task of considerable complexity — suggesting that for specific, well-defined morphological measurements, carefully tuned classical methods can remain competitive with more data-intensive approaches. For research groups requiring a practical, flexible, and accessible solution without the overhead of large training datasets or deep learning infrastructure, random forest-based workflows, such as ilastik, therefore remain a well-justified choice that allows segmentation and classification of multiple cell categories.

We believe that caactus represents a valuable tool for fungal cell counting, size determination, and morphological analysis. We demonstrated its reliability in comparison with human raters and showed how it can effectively describe distributions of morphotypes and cell sizes across different filamentous fungal species grown under varying conditions. Therefore, caactus is not only suitable for comparisons across species, but may also be applied to mutant libraries within a species. Since users can easily and quickly train models themselves in ilastik, the workflow is highly flexible and can be adapted quickly to specific research questions or organisms under investigation with sufficient computational resources.

Importantly, using caactus does not require additional financial investment if a microscope with an attached camera system is already available. Even a standard light microscope equipped with a smartphone adapter could be used, although this would limit precision in pixel-based size measurements. Our aim was to provide a tool with a low barrier to entry, enabling laboratories worldwide to study less frequently investigated fungal organisms, including those responsible for neglected tropical diseases [47].

Finally, we presented a clinical use case. Early assessment of antifungal susceptibility at an early stage of fungal development is essential for timely initiation of effective therapy in humans and animals, and could also support drug screening studies. In the example shown for A. fumigatus strain R7 treated with Amphotericin B, the workflow demonstrated clear applicability.

There are two distinct models of how Amphotericin B acts on the fungal cell membrane. The traditional model, named “barrel-stave ion channel model”, proposed in the 70 s, suggests, that Amphotericin B acts by binding to ergosterol in the fungal cell membrane, which results in the formation of ion channels, resulting in ion leakage and thus causing depolarization of membrane and subsequent death of cells. A newer model, the “sterol sponge model” assumes, that Amphotericin forms aggregates with ergosterol and these aggregates accumulate laterally on the fungal cell membrane surface [48]. In standardized assays to determine antifungal susceptibility, like the one performed here and which is also performed in diagnostic laboratories, the drug of interest is directly added to resting spores. If the concentration is high enough, the cells cannot develop from spores to hyphae and subsequently mycelium. In these microdilution tests, the cutoff is in relation to the internal growth control (0 g/ml) AmB. MICs therefore help to distinguish wild-type (=susceptible) and resistant isolates. Thus, it helps to predict if a patient responds to treatment or not. Results were always cross-checked with standard readout by eye after 24 h. Where mycelium was detected in images, it is also visible with analog readout.

Such assays could be performed in diagnostic laboratories equipped with a microscope and camera system. According to EUCAST protocols, MIC testing is typically performed with one row per isolate ( per concentration) [33, 38]. Within the caactus workflow, image acquisition of one or two rows of a 96-well plate requires only a few minutes for an experienced user, and subsequent analysis delivers results within minutes. This could potentially allow detection of growth and adjustment of therapy substantially earlier than with standard EUCAST AFST, providing valuable time for appropriate treatment initiation. Whether this approach is practical in routine diagnostic settings requires further investigation, clinical validation and quality assurance procedures. Taken together, it may complement existing AFST methodologies [49] and holds promise for large-scale susceptibility testing [50], evaluation of novel drugs, or mutant library screening studies [51].

Conclusion

With caactus, we provide a free and reliable workflow for the analysis of bright-field microscopy images of filamentous fungal growth. Our goal was to establish a tool that is accessible to a wide range of research laboratories and potentially also diagnostic laboratories, including those operating in resource-limited settings, and that can be easily adapted to virtually any fungal species. The workflow can be used to answer a number of questions regarding fungal growth dynamics at early time points under different environmental growth conditions, as well as evaluating the impact of gene deletions and other induced mutations on the growth behavior of fungi of interest.

Availability and requirements

• Project name: caactus.

• Project home page: https://github.com/mr2raccoon/caactus/.

• Operating system(s): Platform independent.

• Programming language: Python.

• Other requirements: ilastik version 1.4.0, 64-bit machine, minimum 8 GB RAM.

• License: MIT License.

Abbreviations

AFST

Antifungal susceptibility testing

AIC

Akaike information criterion

AmB

Amphotericin B

BIC

Bayesian information criterion

CI

Confidence interval

CNN

Convolutional neural network

DT

Distance transform

EUCAST

European Committee on Antimicrobial Susceptibility Testing

GLM

Generalized linear model

GUI

Graphical User Interface

HDF5

Hierarchical data Format version 5

MIC

Minimum inhibitory concentration

OD

Optical density

OR

Odds’ ratio

PCA

Principal Component Analysis

PLN

Poisson log-normal models

SAB

Sabouraud

SE

Standard error

TIF

Tagged Image File Format

wt

Wild type

WHO

World Health Organization

YPG

Yeast Peptone Glucose

Appendix A supplementary tables and figures

Table 4.

Results of model fitting for the global comparison (full = counts 1 + rater + species vs. null = counts 1 + species) and each individual species (full = counts 1 + rater vs. null = counts 1) using Poisson log-normal (PLN) models. AIC = Akaike information criterion, BIC = Bayesian information criterion, logLik = Log-likelihood function, LRT_chi2= Likelihood Ratio Test chi-square statistic, LRT_df = Degrees of freedom for the Likelihood Ratio Test, LRT_p = p-value associated with the Likelihood Ratio Test. Ten images per species and rater ( human raters and caactus) were analyzed

Species	logLik_full	AIC_full	BIC_full	logLik_null	AIC_null	BIC_null	LRT_chi2	LRT_df	LRT_p
all species	-3562.33	3617.33	3708.03	-3609.69	3649.69	3715.65	94.72	15
ATCC11559	-695.19	730.19	759.74	-701.81	721.81	738.70	13.25	15	0.583
CD2	-668.70	703.70	733.26	-680.97	700.97	717.86	24.54	15	0.057
CD4	-588.36	614.36	636.31	-595.69	609.69	621.51	14.66	12	0.261
CD6	-692.75	727.75	757.31	-725.39	745.39	762.28	65.29	15
R7B	-681.54	716.54	746.10	-721.99	741.99	758.88	80.90	15

Table 4: Likelihood ratio test (LRT) results for PLN models. Significant p-values indicate that including rater information improved model fit

Table 5.

Rater/caactus ratios of counted objects across cell morphologies. Shown are log coefficients from the PLN model, back-transformed ratios (rater/caactus), 95% confidence intervals (lo95–hi95), and p-values testing deviation from parity (ratio = 1). Ratios >1 indicate higher counts by raters; ratios <1 indicate lower counts compared to caactus

Rater	Cell type	Log coef	Ratio	95% CI lower	95% CI upper	p-value	Adj.p
rater1	resting	-0.887	0.412	0.196	0.863	0.020	0.299
rater2	resting	-0.494	0.610	0.376	0.989	0.047	0.702
rater3	resting	0.215	1.240	0.776	1.980	0.369	1.000
rater1	swollen	-0.226	0.798	0.376	1.696	0.558	1.000
rater2	swollen	-0.119	0.888	0.559	1.410	0.615	1.000
rater3	swollen	-0.198	0.820	0.364	1.846	0.632	1.000
rater1	germling	0.080	1.084	0.514	2.286	0.833	1.000
rater2	germling	-0.037	0.963	0.505	1.838	0.910	1.000
rater3	germling	0.037	1.037	0.508	2.118	0.920	1.000
rater1	hyphae	-0.643	0.526	0.306	0.904	0.021	0.321
rater2	hyphae	-0.133	0.876	0.435	1.762	0.710	1.000
rater3	hyphae	1.272	3.569	1.720	7.404	0.001	0.012
rater1	notusable	-0.174	0.840	0.519	1.359	0.479	1.000
rater2	notusable	-0.785	0.456	0.280	0.743	0.002	0.029
rater3	notusable	-0.379	0.685	0.341	1.376	0.289	1.000

Table 6.

Rater/caactus ratios of counted objects across cell morphologies for ATCC 11559. Shown are log coefficients from the PLN model, back-transformed ratios (rater/caactus), 95% confidence intervals (lo95–hi95), and p-values testing deviation from parity (ratio = 1). Ratios >1 indicate higher counts by raters; ratios <1 indicate lower counts compared to caactus

Rater	Cell type	Log coef	Ratio	95% CI lower	95% CI upper	p-value	Adj. p
rater1	resting	-0.005	0.995	0.267	3.714	0.994	1.000
rater2	resting	-0.398	0.671	0.103	4.387	0.682	1.000
rater3	resting	0.145	1.156	0.107	12.467	0.906	1.000
rater1	swollen	-0.328	0.720	0.200	2.596	0.621	1.000
rater2	swollen	0.090	1.094	0.174	6.865	0.925	1.000
rater3	swollen	-0.551	0.576	0.381	0.873	0.017	0.255
rater1	germling	-0.083	0.920	0.251	3.378	0.902	1.000
rater2	germling	-0.116	0.890	0.164	4.827	0.894	1.000
rater3	germling	-0.049	0.952	0.447	2.028	0.900	1.000
rater1	hyphae	-0.337	0.714	0.185	2.749	0.629	1.000
rater2	hyphae	-0.327	0.721	0.055	9.373	0.805	1.000
rater3	hyphae	0.725	2.064	1.133	3.762	0.028	0.422
rater1	notusable	0.582	1.790	0.279	11.481	0.546	1.000
rater2	notusable	-0.469	0.626	0.051	7.710	0.718	1.000
rater3	notusable	0.088	1.092	0.578	2.063	0.790	1.000

Table 7.

Rater/caactus ratios of counted objects across cell morphologies for CD2. Shown are log coefficients from the PLN model, back-transformed ratios (rater/caactus), 95% confidence intervals (lo95–hi95), and p-values testing deviation from parity (ratio = 1). Ratios >1 indicate higher counts by raters; ratios <1 indicate lower counts compared to caactus

Rater	Cell type	Log coef	Ratio	95% CI lower	95% CI upper	p-value	Adj. p
rater1	resting	-0.044	0.957	0.195	4.680	0.957	1.000
rater2	resting	0.112	1.118	0.153	8.150	0.913	1.000
rater3	resting	-0.406	0.666	0.057	7.776	0.749	1.000
rater1	swollen	-0.408	0.665	0.149	2.962	0.598	1.000
rater2	swollen	-0.171	0.843	0.125	5.668	0.862	1.000
rater3	swollen	-0.367	0.693	0.451	1.065	0.110	1.000
rater1	germling	0.295	1.343	0.259	6.959	0.729	1.000
rater2	germling	0.036	1.036	0.157	6.826	0.971	1.000
rater3	germling	0.060	1.062	0.611	1.847	0.832	1.000
rater1	hyphae	0.116	1.123	0.278	4.541	0.872	1.000
rater2	hyphae	-1.775	0.170	0.016	1.789	0.155	1.000
rater3	hyphae	1.309	3.701	1.989	6.886	0.001	0.008
rater1	notusable	0.172	1.187	0.204	6.925	0.851	1.000
rater2	notusable	-0.561	0.571	0.036	8.963	0.694	1.000
rater3	notusable	-0.096	0.908	0.514	1.604	0.744	1.000

Table 8.

Rater/caactus ratios of counted objects across cell morphologies for CD4. Shown are log coefficients from the PLN model, back-transformed ratios (rater/caactus), 95% confidence intervals (lo95–hi95), and p-values testing deviation from parity (ratio = 1). Ratios >1 indicate higher counts by raters; ratios <1 indicate lower counts compared to caactus

Rater	Cell type	Log coef	Ratio	95% CI lower	95% CI upper	p-value	Adj. p
rater1	resting	-0.494	0.610	0.187	1.992	0.421	1.000
rater2	resting	0.141	1.152	0.283	4.687	0.845	1.000
rater3	resting	0.172	1.188	0.784	1.800	0.424	1.000
rater1	swollen	-0.035	0.966	0.164	5.681	0.970	1.000
rater2	swollen	-0.088	0.916	0.106	7.902	0.937	1.000
rater3	swollen	0.019	1.019	0.390	2.664	0.970	1.000
rater1	germling	0.382	1.465	0.256	8.368	0.672	1.000
rater2	germling	0.225	1.252	0.168	9.315	0.828	1.000
rater3	germling	0.269	1.309	0.603	2.842	0.503	1.000
rater1	notusable	-0.185	0.831	0.165	4.178	0.824	1.000
rater2	notusable	-1.024	0.359	0.053	2.431	0.304	1.000
rater3	notusable	-0.890	0.411	0.191	0.883	0.032	0.381

Table 9.

Rater/caactus ratios of counted objects across cell morphologies for CD6. Shown are log coefficients from the PLN model, back-transformed ratios (rater/caactus), 95% confidence intervals (lo95–hi95), and p-values testing deviation from parity (ratio = 1). Ratios >1 indicate higher counts by raters; ratios <1 indicate lower counts compared to caactus

Rater	Cell type	Log coef	Ratio	95% CI lower	95% CI upper	p-value	Adj. p
rater1	resting	-1.710	0.181	0.122	0.268	0.000	0.000
rater2	resting	-0.159	0.853	0.112	6.504	0.880	1.000
rater3	resting	0.819	2.268	0.462	11.128	0.325	1.000
rater1	swollen	-0.252	0.778	0.486	1.244	0.307	1.000
rater2	swollen	-0.267	0.765	0.089	6.549	0.810	1.000
rater3	swollen	-0.216	0.806	0.601	1.080	0.164	1.000
rater1	germling	-0.211	0.810	0.558	1.174	0.279	1.000
rater2	germling	-0.142	0.867	0.239	3.143	0.831	1.000
rater3	germling	-0.202	0.817	0.523	1.278	0.387	1.000
rater1	hyphae	-4.547	0.011	0.003	0.038	0.000	0.000
rater2	hyphae	1.338	3.813	0.904	16.088	0.083	1.000
rater3	hyphae	2.559	12.917	8.956	18.630	0.000	0.000
rater1	notusable	-0.880	0.415	0.048	3.548	0.431	1.000
rater2	notusable	-0.731	0.481	0.102	2.272	0.367	1.000
rater3	notusable	-0.500	0.606	0.418	0.879	0.016	0.236

Table 10.

Rater/caactus ratios of counted objects across cell morphologies for R7B. Shown are log coefficients from the PLN model, back-transformed ratios (rater/caactus), 95% confidence intervals (lo95–hi95), and p-values testing deviation from parity (ratio = 1). Adjusted p-values are based on Bonferroni correction

Rater	Cell type	Log coef	Ratio	95% CI lower	95% CI upper	p-value	Adj. p
rater1	resting	-2.082	0.125	0.094	0.166	0.000	0.000
rater2	resting	-2.297	0.101	0.059	0.170	0.000	0.000
rater3	resting	0.009	1.009	0.198	5.137	0.991	1.000
rater1	swollen	-0.218	0.804	0.610	1.059	0.136	1.000
rater2	swollen	-0.156	0.855	0.520	1.407	0.545	1.000
rater3	swollen	-0.122	0.886	0.578	1.356	0.583	1.000
rater1	germling	0.028	1.028	0.763	1.385	0.856	1.000
rater2	germling	-0.022	0.979	0.275	3.482	0.974	1.000
rater3	germling	-0.072	0.930	0.587	1.476	0.763	1.000
rater1	hyphae	-0.978	0.376	0.256	0.552	0.000	0.001
rater2	hyphae	0.225	1.252	0.212	7.378	0.807	1.000
rater3	hyphae	1.137	3.117	1.905	5.102	0.000	0.003
rater1	notusable	-0.594	0.552	0.333	0.916	0.032	0.486
rater2	notusable	-1.268	0.281	0.050	1.575	0.164	1.000
rater3	notusable	-0.667	0.513	0.310	0.849	0.017	0.259

Table 11.

Likelihood ratio tests comparing null vs. full Negative Binomial generalized linear models (GLMs) for total object counts across raters. The null model includes only an intercept, while the full model additionally includes the rater effect. Degrees of freedom for the null and full models are df_null and df_full, and their difference (df_diff) corresponds to the number of parameters added. The likelihood ratio test (LRT) statistic, , is compared to a chi-squared distribution with df_diff degrees of freedom. p-values indicate a statistically significant improvement by including the rater

Species	df	df	df	logLik	logLik	LRT stat	p-value
All species	4	7	3	-1211.90	-1211.73	0.33	0.95
ATCC11559	0	3	3	-240.96	-240.94	0.04	0.10
CD2	0	3	3	-236.40	-236.39	0.01	1.00
CD4	0	3	3	-240.43	-240.40	0.04	0.10
CD6	0	3	3	-254.52	-254.47	0.11	0.99
R7B	0	3	3	-239.59	-239.10	0.99	0.81

Table 12.

Summary statistics of total counts globally and per species across the images ( images per species) used for validation. sd = standard deviation

species	rater	mean	sd	min	25%	median	75%	max
global	caactus	166.1	57.3	57	124.5	158.0	197.3	315
	expert 1	156.5	44.6	56	128.3	149.0	178.8	286
	expert 2	149.8	48.2	57	118.0	139.5	176.0	294
	expert 3	164.0	57.5	47	129.5	156.5	192.5	373
ATCC 11559	caactus	147.2	36.6	95	130.3	141.0	154.5	229
	expert 1	159.1	29.0	126	146.8	150.5	172.5	225
	expert 2	147.6	27.5	114	129.8	142.5	158.5	193
	expert 3	152.2	29.2	118	134.8	146.0	154.5	211
CD2	caactus	134.0	41.5	57	113.3	143.0	159.0	198
	expert 1	137.2	33.1	56	130.3	144.5	148.3	187
	expert 2	131.8	31.9	57	121.3	135.0	152.5	173
	expert 3	137.5	34.7	56	122.8	146.5	158.5	179
CD4	caactus	147.5	38.5	105	120.0	137.5	166.5	223
	expert 1	156.7	42.9	110	120.3	142.5	190.3	220
	expert 2	143.8	38.2	103	115.8	133.5	172.3	217
	expert 3	150.0	62.9	47	117.3	137.0	180.0	269
CD6	caactus	222.1	69.4	93	178.8	215.0	273.5	315
	expert 1	198.1	57.5	89	163.8	197.0	239.8	286
	expert 2	206.2	59.9	92	168.0	210.5	248.8	294
	expert 3	225.1	73.7	105	178.8	220.5	254.0	373
R7B	caactus	179.9	53.4	90	169.0	192.5	203.5	270
	expert 1	131.3	26.0	88	115.3	131.0	144.8	179
	expert 2	119.4	29.7	80	98.3	119.5	131.3	177
	expert 3	155.0	34.3	86	136.0	164.0	168.5	199

Table 13.

Results of Conover’s post hoc test for time needed for counting per image. Only adjusted p-value shown (Bonferroni correction)

Comparison	Adj. p
rater1 vs. caactus
rater2 vs. caactus
rater3 vs. caactus
rater1 vs. rater2
rater1 vs. rater3	0.0584
rater2 vs. rater3	0.0388

Table 14.

Absolute count, average count, cell type (%) and mean object size (m) across Mucoralean strains (see Table 1) incubated in YPG medium at 4 h post inoculation, either directly (fresh) or after storage at C (frozen)

Strain	Storage	Predicted Class	Absolute Count	Avg. count	%	Size (m)
ATCC11559		Resting	331	20.69	11.56	23.33
	fresh	Swollen	2361	147.56	82.46	53.74
		Germling	24	1.50	0.84	121.08
	frozen	Resting	1562	97.62	96.19	14.85
		Swollen	2	0.12	0.12	38.75
CD2		Resting	52	3.25	2.11	23.11
	fresh	Swollen	540	33.75	21.94	118.77
		Germling	1484	92.75	60.30	269.73
		Hyphae	32	2.00	1.30	643.20
	frozen	Germling	2	0.12	0.10	141.50
		Resting	1830	114.38	93.42	23.49
		Swollen	17	1.06	0.87	81.74
CD4		Resting	1369	85.56	53.17	18.03
	fresh	Swollen	1092	68.25	42.41	27.60
		Germling	2	0.12	0.07	83.68
	frozen	Resting	2033	127.06	95.13	12.11
		Swollen	36	2.25	1.68	28.42
CD6		Resting	72	4.50	2.25	10.32
	fresh	Swollen	2869	179.31	89.63	73.83
		Germling	18	1.12	0.56	306.12
		Hyphae	1	0.06	0.03	1150.13
		Resting	286	17.88	10.00	13.87
	frozen	Swollen	2306	144.12	80.60	63.82
		Germling	3	0.19	0.11	160.01
R7B		Resting	276	17.25	8.24	14.81
	fresh	Swollen	823	51.44	24.57	116.75
		Germling	1137	71.06	33.95	390.48
		Hyphae	155	9.69	4.63	1264.71
		Resting	809	50.56	31.43	22.51
	frozen	Swollen	1213	75.81	47.13	110.43
		Germling	295	18.44	11.46	317.68
		Hyphae	1	0.06	0.04	1124.16

Table 15.

Summary of hyphal and germling percentages at 10 h and mycelium presence (1 = present, 0 = not present) at 24 h across tested concentrations of AmphotericinB. Images from three independent experiments with two to four technical replicates per time and concentration

Concentration (g/mL)	% Hyphae at 10h	% Germlings at 10h	Mycelium at 24h
0	31.53	51.312	1
0.03	6.74	41.48	1
0.0625	1.50	21.78	1
0.125	0.15	5.23	1
0.25	0	2.61	0
0.5	0	2.45	0
1	0	1.78	0
2	0	1.06	0
4	0	1.54	0
8	0	1.35	0
16	0	1.23	0

Fig. 9.

EUCAST antifungal susceptibility testing for Aspergillus fumigatus strain R7 with Amphotericin B ranging from 0-16 g/mL in a 96-well plate according to [33]. Images from three independent experiments with two to four technical replicates per time and concentration

Table 16.

Parameter estimates from the logistic regression model predicting mycelium presence at 24 h based on the proportion of hyphae at 10 h (Model 1). SE = standard error, OR = Odds’ ratio, CI = confidence interval

Predictor	Estimate ()	SE	z	p	OR ()	95% CI (OR)
Intercept ()	-1.8688	0.305	-6.120		0.154	[0.085, 0.281]
Hyphae (%) ()	5.5538	2.164	2.566	0.010	258.21	[3.71, 17948.50]
Model fit: Pseudo- = 0.4633, Log-Likelihood = -38.76, AUC = 0.830

Table 17.

Parameter estimates from the logistic regression model predicting mycelium presence at 24 h based on the proportion of germlings at 10 h (Model 2). SE = standard error, OR = Odds’ ratio, CI = confidence interval

Predictor	Estimate ()	SE	z	p	OR ()	95% CI (OR)
Intercept ()	-2.9201	0.495	-5.894		0.054	[0.020, 0.142]
Germling (%) ()	0.3028	0.080	3.803		1.354	[1.158, 1.582]
Model fit: Pseudo- = 0.5396, Log-Likelihood = -33.25, AUC = 0.923

Authors' contributions

JaS and UB designed the approach. JaS built the workflow and performed all annotations in ilastik. HS, AR and UB counted by hand has human raters. JaS, HS and HH cultivated the fungi. JaS completed image acquisition. DK, JoS and VB provided coding assistance and reviewed the manuscript. JaS and JoS performed the statistical analysis for validation. SP designed the GUI. JaS and UB wrote the manuscript. CLF provided strains and computational resources. All authors read and approved the final manuscript.

Funding

Not applicable.

Data availability

• Data availability: images, trained ilastik models and results from manual counting accessible on zenodo https://doi.org/10.5281/zenodo.19468159. • Materials availability: A sample dataset for trying the functionality can be accessed on zenodo: https://doi.org/10.5281/zenodo.19728045. • Code availability: https://github.com/mr2raccoon/caactus/.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.