The process of nematode counting is both labor-intensive and time-consuming, yet it serves as a crucial step in numerous quantitative nematological studies, including the preparation of initial population and final population densities related to nematode management in pot, micro-plot and field trials (Barker and Campbell, 1981). Accurate results heavily depend on the expertise of taxonomy of nematodes of the individuals conducting this task. This study aims to streamline the nematode counting process for Meloidogyne spp. juveniles and eggs by implementing a deep learning algorithm for automation.

Traditionally, nematode counting has been a manual process (Hussey and Barker, 1973; Seinhorst, 1988). Following the extraction of nematodes from plant materials or soil, they are collected in a water suspension and subjected to counting. Prolonged storage in water suspension makes the task of counting and identifying nematodes progressively challenging, even for experienced nematologists, as nematode might die due to depletion of stored food. In some cases, all nematodes in the entire sample are counted, but this can be impractical, especially when dealing with a large number of samples or a high nematode density per sample, due to the considerable time required. Consequently, many laboratories opt to count nematodes by analyzing subsamples from the mother suspension, although this approach can introduce errors and diminish statistical accuracy when few nematodes are counted (Schomaker and Been, 1998). Furthermore, the manual nematode counting process is susceptible to inconsistency when carried out by different individuals or prone to errors that may arise when an individual spends an extended period behind a microscope.

Nematologists have long recognized the challenges associated with nematode counting, prompting a few research initiatives to automate the process. Been et al. (1996) introduced ANECS (Automatic NEmatode Counting System), a software program designed to count juveniles of Globodera spp in water suspension. While this method achieved high detection accuracy, its widespread adoption among nematologists was hindered by the need for specialized and expensive hardware and image analysis systems. In more recent years, Holladay et al. (2016) adapted ImageJ, an open-source image analysis software, to create a standard curve for automated nematode counting based on the black and white pixel sizes in individual images. However, this method is limited to samples containing a single species of similarly sized nematodes and excludes samples with soil and root debris. Given the recent advancements in artificial intelligence, these challenges encountered during automated nematode counting can now be effectively tackled.

Deep learning, a broader realm within machine learning encompassing various architectures like Convolutional Neural Networks (CNNs), has garnered significant research attention among nematologists, as indicated by recent studies (Akintayo et al., 2018; Chen et al., 2020, 2022; Kalwa et al., 2019; Uhlemann et al., 2020). The utilization of neural networks for computer-assisted nematode identification dates as early as 2000 when it was proposed by Diederich et al. (2000). CNNs, the focus of this study, are computer programs designed to replicate how our brains process visual information, proving highly effective in the analysis of biological images, such as microscope pictures. Notably, there are compelling examples of AI-driven approaches for automating nematode counting. In their work, Akintayo et al. (2018) applied a deep learning architecture originally designed for detecting rare objects in cluttered images to the task of identifying eggs of Heterodera glycines. They designed the Convolutional Selective Autoencoder (CSAE) architecture, which facilitated rapid detection, consistency, and accuracy in identifying nematode eggs amidst debris. Likewise, Kalwa et al. (2019) developed a modified version of the U-Net convolutional autoencoder model learning algorithm, specifically customized for detecting H. glycines eggs in purified samples. In contrast to typical approaches using microscopic images, their imaging systems employed a high-resolution scanner and a light-emitting diode (LED) to illuminate the processed sample flowing through a microfluidic flow chip, along with a CMOS image sensor. Both studies necessitated prior staining. In 2020, Chen et al. adapted the standard U-Net architecture to automate the counting of worm-shaped objects. Instead of employing the standard bounding box detection method used in other studies, they utilized the skeleton to address overlapping and curled nematodes. Their initial model was primarily assessed on C. elegans. Subsequently, in 2022, Chen et al. extended their work by developing a segmentation model for cyst detection in soil debris. This was accomplished by leveraging the standard U-Net architecture and ResNet architecture. While not focused on counting, Uhlemann et al. (2020) effectively employed CNNs to differentiate between three entomopathogenic nematode species within the same family. To select the most appropriate architecture for their study, the researchers initially screened 13 CNN architectures available in 2020. They ultimately decided to employ Xception due to its highest accuracy among the options. Additionally, Shabrina et al. (2023) devised a deep-learning model aimed at automatically identifying 11 different genera of plant-parasitic nematodes commonly found in Indonesia. Although their primary focus was classification rather than quantification like Uhlemann et al. (2020), they investigated four distinct architectures, ResNet101V2, CoAtNet-0, EfficientNetV2Bo, and EfficientNetV2M, across various augmentation processes. Their research culminated in the creation of a website capable of analyzing nematode images and providing genus-level identification.

While the aforementioned studies have demonstrated the effectiveness of deep learning for automating nematode counting, they have primarily been limited to single-class object identification within individual images. To make this technology more practical, there is a need for an algorithm capable of simultaneously detecting and classifying multiple objects, even when they cohabit with other objects that are confusing shapes and sizes, such as root and soil debris or non-target nematodes within individual samples. This study aims to address this challenge by developing a deep learning algorithm that can simultaneously identify and classify nematode eggs and juveniles of Meloidogyne spp. while distinguishing them from free-living nematodes and other clutter in the sample. The resulting algorithm will be shared as open-source software on GitHub for public use.

Our research showcased the effectiveness of the trained models, which attained a detection accuracy of over 90% for both life stages of Meloidogyne spp. eggs and juveniles, a genus known for its economic importance in various crops (Jones et al., 2013). To make these models easily accessible, we have created an open-source application called SEGNEMA, which can be found on the author’s GitHub account (https://github.com/bresilla/nematode_counting). SEGNEMA enables the simultaneous detection and classification of nematodes and eggs. In the future, additional models can be independently created for different genera, including cyst nematodes like Globodera spp. and Heterodera spp., as well as Pratylenchus spp. These models can then be seamlessly incorporated into the same GUI through the use of transfer learning techniques.

While our models have exhibited high accuracy in detecting and classifying objects, they have encountered challenges that are commonly reported in similar studies, as noted by Akintayo et al. (2018), Been et al. (1996), and Chen et al. (2022). Like the models developed by Akintayo et al. (2018) and Been et al. (1996), our models faced difficulties in distinguishing overlapped objects as distinct entities and in discerning nematodes from organic debris within samples. Furthermore, our models exhibited reduced detection accuracy when it came to FLN compared to the other class objects in this study. This can be attributed to both the limited number of instances for this particular class and the diverse range of lengths and shapes that FLN can assume. Nevertheless, these challenges can be mitigated by incorporating additional images that specifically address these conditions. This study has introduced a framework that can serve as a universal automation solution for nematode detection amidst multiple objects. However, it’s crucial to recognize that no model is flawless, and ongoing improvements in object detection and classification are essential. Therefore, developers and users alike should remain vigilant in identifying scenarios where models may fall short.

In addressing some of the challenges encountered during the development of our model for this study, we experimented with various techniques. Initially, we employed bounding boxes for annotating nematodes, a method commonly used in object detection (Huang et al., 2016), as seen in the work by Akintayo et al. (2018). However, we found that this approach did not yield the desired level of detection accuracy, especially when dealing with objects like nematodes that exhibit diverse shapes, overlap with each other, or are surrounded by debris in the sample. Subsequently, after the unsuccessful attempt with bounding boxes, we explored alternative annotation methods and ultimately concluded that segmentation offered the highest level of detection accuracy. We also investigated the utilization of the skeleton, referred to as key-point detection, as suggested in the study by Chen et al. (2020), which demonstrated successful detection of intertwined and overlapping worm-shaped objects. However, for the small preliminary dataset we used, segmentation proved to be more effective in distinguishing RKN J2s from FLN. Additionally, it’s worth mentioning that even with 10 key-points required to annotate an object, the process was quite labor-intensive and frequently necessitated additional key-points, particularly when annotating curled nematodes. With Darwin V7, the annotation system we employed, which allows for auto-segmentation of objects, facilitated the generation of the large annotated dataset in our study. The utilization of segmentation has also been applied successfully in other studies, such as the detection and phenotyping of cysts in samples with debris, as demonstrated by Chen et al. (2022). Segmentation proved to be a valuable technique for annotating nematodes with diverse shapes, leading to noticeable enhancements in the model’s ability to correctly identify curled nematodes as nematodes, rather than misclassifying them as eggs. However, it’s important to note that further improvement in this regard would benefit from an increased dataset comprising more images depicting such cases.

Moreover, we observed a notable enhancement in our models’ object detection performance across different images and a notable increase in the mAP50 after integrating annotated images obtained from an additional microscope system. Specifically, the mAP50 for the egg model improved from 0.79 to 0.86, while for the juvenile model, it increased from 0.82 to 0.87. Prior to the inclusion of annotated images from the LEITZ DM IRB microscope, our models exhibited high accuracy in object detection primarily on images taken by the Stellaris 5 Confocal LSM system only. However, this performance improved considerably, as illustrated in Table 1 for the final model. Despite the fact that we utilized only two imaging systems in this study, this expansion in the variety of images has a profound impact on the models’ ability to achieve accurate detection. As machine learning models benefit from exposure to a wider array of images, this development translates to enhanced detection accuracy across different imaging systems, beyond those used in this particular study.

An issue previously raised in the context of automated counting systems was their reliance on specialized and costly hardware and image analysis systems, as highlighted by Been et al. (1996). In our study, we addressed this concern by creating a user-friendly GUI called SEGNEMA, which is open source and freely accessible (https://github.com/bresilla/nematode_counting). Once the optimal detection threshold is determined for eggs and juveniles, users should be able to maintain it consistently for the same imaging system, streamlining the usability of SEGNEMA for their specific needs. While SEGNEMA does necessitate a computer environment equipped with substantial graphics processing units (GPUs), this requirement remains reasonably accessible to a wide range of users. The processing time per image using SEGNEMA was typically around 4 to 5 seconds for our images, though this duration may vary depending on the image size. This provides a significantly faster alternative to manual labor for nematode counting, contributing to increased efficiency and productivity in the analysis process. For example, at a magnification of x10 and a speed of 600 µm/s with the Stellaris 5 Confocal LSM, capturing 10 stitched images required 58 minutes. Adjusting the speed can enhance image resolution, but this comes at the cost of a longer image capture time. It’s essential for readers to recognize the trade-off between efficiency (speed) and reliability (resolution). Even when considering the processing time using SEGNEMA, the total time is still less than half of what is typically needed for the conventional nematode counting method involving aliquot subsampling. Moreover, because the imaging system’s stitching process is automated, the waiting period during image acquisition can be utilized for other tasks. Crucially, SEGNEMA ensures consistency in nematode counting, uninfluenced by variations in conditions, unlike manual counting methods.

Our research has effectively demonstrated the promising practical applications of these models, particularly through the user-friendly GUI that enables simultaneous detection and classification of nematodes. Presently, this GUI is tailored for nematode juveniles and eggs, but it possesses significant potential for broader applications across various nematode genera and in different media beyond aqueous solutions. The potential of acquiring fluorescent images simultaneous with the bright field images obtained from the confocal microscope is large and opens doors to automated multiplex (high content) acquisition with no time loss. Without any incubation or staining the autofluorescent images or spectra of nematodes can e.g. provide information on their viability (Forge and MacGuidwin, 1989), and probes can be used to add further specificity to the nematode discrimination.

It’s also crucial to address the challenges identified in our study and similar research endeavors. However, tackling these challenges cannot be the sole responsibility of a single laboratory or research institution. Instead, it requires a collaborative effort, bringing together expertise from both nematologists and AI researchers across multiple organizations and research groups. For our GUI to be thorough in nematode detection and diagnostics, it’s imperative to acquire a wider array of images captured through diverse imaging systems and develop models for different genera. Achieving this goal necessitates enhanced collaboration with fellow nematologists to access their image collections and tap into their expertise in nematode diagnostics. Given the growing interest among nematologists in harnessing the potential of machine learning, we are confident that the development of a universal automated nematode counting system accessible to everyone is within reach. The authors of this paper hope that it serves as a framework and catalyst for initiating global collaboration toward this important goal.

The subsequent authors following the first author are arranged in alphabetical order.

The raw data used to develop models presented in this article will be made available by the authors, without reservation.

KS: Writing – original draft, Writing – review & editing. TB: Writing – original draft, Writing – review & editing. JK: Writing – review & editing. NdR: Writing – review & editing. CvS: Writing – review & editing. MT: Writing – review & editing.