Machine Learning Techniques for Nematode Microscopic Image Analysis: A Systematic Review

Abstract

Farmers rely on nematode analysis for critical crop management decisions, yet traditional detection and classification methods remain subjective, labor-intensive, and time-consuming. Advances in Machine Learning (ML) and Deep Learning (DL) offer scalable solutions for automating microscopy-based nematode analyses. This systematic literature review (SLR) analyzed 44 articles published between 2018 and 2024 on ML/DL-based nematode image analysis, selected from 1460 records screened across Web of Science, IEEE Xplore, Agricola, and supplemental Google scholar searches. The quality of reporting was examined by considering dataset documentation and code availability. The results were synthesized narratively, as diversity in datasets, tasks, and metrics precluded a meta-analysis. Performance was primarily reported using accuracy, precision, recall, F1-score, Dice coefficient, Intersection over Union (IoU), and average precision (AP). CNNs were the most commonly used architectures, with models such as YOLO providing the best detection performance. Transformer-based models excelled in dense segmentation and counting. Despite strong performance, challenges include limited training data, occlusion, and inconsistent metric reporting across tasks. Although ML/DL models hold promise for scalable nematode analysis, future research should prioritize real-world validation, diverse nematode datasets, and standardized benchmark datasets to enable fair and reproducible model comparison.

1. Introduction

Plant-parasitic nematodes are microscopic roundworms, typically between 300 and 1000 $\mathsf{\mu }$m long. They are differentiated from non-parasitic nematodes by the presence of a stylet—a hollow needle-like structure used by the nematode to pierce and divert nutrients from the plant. More than 4100 plant-parasitic nematode species have been described, grouped into genera that share similar life cycles, feeding habits, and morphology [1]. Plant-parasitic nematodes have negative impacts in agriculture, causing crop losses and plant diseases [2]. Therefore, being able to detect and treat plants affected by nematodes is critical, especially as concerns regarding food insecurity, environmental protection, and invasive species around the world rise [3]. A crucial step in the detection and diagnosis of plant-parasitic nematodes is evaluation of specimens under the microscope by a nematologist or diagnostician, with the objective of classifying the genera and/or species of the nematodes (based on the observed morphology of the individual nematodes) and quantifying the population count within a sample (to determine whether the population rises above risk thresholds). Morphology-based identification is conducted by observing characteristics unique to a nematode genus, such as overall body length, shape of the head and tail sections, length and shape of the stylet, cuticle adornments such as annulations, and arrangement of internal organs including the esophagus, intestine, and reproductive structures. Conducting these evaluations manually (i.e., by eye) under the microscope requires extensive technical training, and the slow, laborious process limits the total number of samples that can be evaluated by each technician. Furthermore, repetitive laboratory tasks like this can lead to operator fatigue and errors [4]. Taken together, this may limit the quantity and quality of data provided to farmers and plant protection specialists to make nematode management decisions, implement management tactics, and set regulatory policies.

New developments in image processing and computer vision have enabled the automation of various image analysis tasks through the use of artificial intelligence (AI) techniques. Machine learning (ML) and deep learning (DL) models, as specific tools in ML/DL research, have been developed and deployed to automate the process of microscopic nematode image analysis. Prior to the development of ML/DL techniques, nematode image analysis was limited to image processing algorithms that require high tuning and expert-level operation [5]; for example, tuning methods via ImageJ version 1.48 have been employed to automate the nematode assay counting process, albeit with minimal occlusion and limited results [5]. The same was also true for the generation of segmentation masks [6]. There has been a significant increase in the development and use of ML/DL models for automating nematode image analysis over the past decade. The tools and the knowledge necessary to construct these models have become ubiquitous with the increasing popularity of ML—and, in particular, DL—across various scientific applications. Furthermore, modifying a classical computer vision algorithm for image segmentation is often more technically challenging and time-consuming than training and using an ML/DL model, which can be trained with more data or modified for certain applications. One example that illustrates the impact of ML/DL in the field is model performance for species classification, which has improved dramatically from an accuracy of 88.3% for 3 classes in 2020 [7] to 96.9% for 11 classes in 2023 [8] through the use of more sophisticated deep learning techniques, more data, and data augmentation strategies.

ML/DL models have become more accessible for image analysis due to the proliferation of hardware for their acceleration and their democratization through cloud platforms. For example, multiple research articles screened as part of this study used the Google Colab platform [9,10,11,12,13]. The use of these cloud platforms provides a low-cost way to train and evaluate models. Additionally, the availability of various public datasets has enabled easier comparison of models [9,14,15,16,17,18]. However, there is little uniformity among the papers for performance evaluation on public datasets.

In this work, we distinguish between the computational methods used for image analysis tasks and biological inference tasks. Automated microscopic image analysis can be performed using ML/DL models to extract quantitative data from microscopic images. As described by Xing and Yang [19], the main goals are to process large volumes of images and improve consistency by removing the potential for human error and subjectivity. The foundational image analysis tasks covered in this review include image classification (labeling a nematode), object detection (locating a nematode), segmentation (outlining a nematode’s exact shape), and tracking (following a nematode’s movement). Object detection builds on the classification task. In its simplest form, image segmentation can be considered a binary classification problem (e.g., detecting whether a pixel in an image is a nematode or not). However, the most relevant version of this task provides different labels for species as well, particularly in cases of nematode overlap. Finally, tracking is often coupled with detection or segmentation to capture the location (configuration) of a moving object in a video. Segmentation and tracking were observed to be the most common combination in the articles surveyed.

On the other hand, biological inference tasks are those driven by biological targets. As seen in this literary review, researchers have tried to automate main tasks such as species classification, counting specimens, behavior tracking, and lifespan tracking. Figure 1 illustrates the relationships between image analysis and biological inference tasks. Species classification is relevant because plant-parasitic nematodes have a negative impact on agriculture, causing crop damage and plant diseases [1,2]. Counting of specimens is used for ecosystem evaluation, as nematodes serve as good bio-indicators; i.e., they play important roles in the ecosystem [20]. Behavior tracking refers to tracking of the mobility of the nematodes and derived parameters such as speed and aggregation. This is relevant because the nervous and locomotion systems nematode are simple enough to be modeled and tracking behavior informs research on how organisms respond to various stimuli [21]. Finally, lifespan tracking mainly focuses on the detection of live versus dead nematodes; however, researchers have also expanded these tools to track parameters such as speed during development, egg-laying rates, and behavioral decline in aging nematodes. This is relevant for modeling aging and neuro-degenerative diseases and screening drugs, based on their short lifespan and simple nervous system.

In this article, we are interested in addressed the following research questions:

• What challenges are addressed by AI-based methods?
• What are the most effective ML/DL models? What are their strengths and limitations?
• What novel AI-driven techniques have been proposed to improve nematode inference?
• What trends and opportunities exist for improving AI-based nematode inference?

The rest of this article is structured as follows: Section 2 outlines the systematic review methodology, including the literature search, selection criteria, data extraction procedures, and the background on performance metrics and ML models. Section 3 presents the core findings including in the articles reviewed, and performance results organized by biological inference tasks. Section 4 synthesizes and compares model performance across architectures, highlighting trends, limitations, and research gaps based on the results. Finally, Section 5 concludes the review and offers recommendations to guide future research focused on ML-based nematode image analysis.

2. Methodology

The methodological design for this Systematic Literature Review (SLR) was performed according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines [22], ensuring a structured and transparent review process (see Supplementary Materials, Document S3). This comprehensive SLR was conducted based upon predefined research questions, a rigorous literature selection process, and strict eligibility criteria to maintain a clear focus on ML/DL techniques for nematode microscopy image analysis (MIA). The following subsections outline the systematic approach followed in this review, centered around a set of research questions and detailed strategies that ensure the reproducibility and validity of the findings.

2.1. Review Process

Three main phases were executed: Planning, conducting, and reporting the review process [23]. The planning phase defines the research questions, developing a review protocol, and designing a search strategy with relevant keywords. Three databases were selected for the search: Web of Science [24], IEEE Xplore [25], and Agricola [26]. In the conducting stage, a literature search was implemented using structured search queries for each specific database. Each source was last consulted on 16 May 2025. A total of 1460 articles were retrieved. A multiple-layer study selection process was implemented. The final stage encompassed synthesis and analysis of the collected data, centered around answering the research questions. This process integrated both qualitative and quantitative methods. A preliminary search was conducted across each database to identify the most relevant sources for ML-based Nematode MIA. Google Scholar was initially explored to compile a list of highly relevant journals and keywords; however, it was not included as a primary search database due to its limited export capabilities and lack of structured metadata retrieval (e.g., abstracts, authors). The primary database for this review was Web of Science, as it provided the most comprehensive collection of relevant articles. To ensure broader coverage and comparison of our topic in the AI and agricultural space, IEEE Xplore and Agricola were also included. Figure 2 shows the PRISMA flowchart for the overall process.

2.2. Search Strategy

The search strategy for this systematic review focused on retrieving relevant studies from the three major digital scientific databases mentioned in

Table 1. Search Strings and Results from Databases.

Database Name	Search String	Results
Web of Science	TI = (Nematode OR Nematodes OR elegans OR C. elegans) AND TI = (detecting OR detect OR detection OR tracking OR track OR identifying OR identify OR identification OR classifying OR classify OR classification OR segmentation OR segment OR recognizing OR recognize OR recognition OR automating OR automate OR automation OR automated OR sorting OR sort OR counting OR count OR phenotypic OR phenotype OR phenotypes OR deciding OR decide OR decision)	1012
Agricola		178
IEEE Xplore		226

. Initially, exploratory search strings were tested across multiple databases, including Google Scholar, to refine keyword selection. The search process began by focusing on “organism-related” and “computational methods” terms in article titles; for instance, (Nematode OR Worm) AND (“Machine Learning” AND “Artificial Intelligence”). To ensure full coverage of the literature related to nematode inference, the search was expanded to include detection, segmentation, and classification inference tasks: (Nematode OR Worm) AND (“Machine Learning” AND “Artificial Intelligence”) AND (“Detection” AND “Classification” AND “Segmentation”). However, using the expanded search query limited the number of retrieved studies, yielding 189 using Google Scholar, 69 in Web of Science, and 14 in Agricola. The strategy was broadened by including only “organisms” and “inference tasks” terms, allowing for more inclusive search results. The term “worm” was excluded as it frequently returned irrelevant studies on malware worms, email worms, and AI cybersecurity applications. Instead, “C. elegans” was explicitly incorporated, as it is the most studied nematode species in microscopy image analysis research. Additionally, seven studies from Google Scholar were manually included as they were directly relevant to the research questions but did not appear in the primary databases used in this review. The final search keywords were structured using the search query detailed in Table 1.

This review restricted the publication years to 2018–2024 and included only English-language articles. The database search was conducted on 25 December 2024, across all three databases. The Web of Science and Agricola searches in Table 1 were restricted to Title-only (TI) queries to ensure high specificity in the studies. IEEE Xplore was intentionally queried without restrictions (e.g., Title-only), in order to broaden the scope and capture all the relevant ML research related to our topic. After executing the search queries, a total of 1416 articles were returned and screened based on predefined inclusion and exclusion criteria to ensure their relevance to the research objectives.

2.3. Selection Criteria

The selection criteria helped to identify articles that answered the research questions. These criteria were based on the SLR objectives and centered around the research questions. All selected studies were saved in a repository and recorded in a spreadsheet. An article was selected if it included the following criteria: (1) it is related to nematode analysis, (2) it makes use of ML/DL techniques in nematode image analysis research, and (3) it was published within the last 7 years. The exclusion criteria include: (1) the publication is a survey or review paper, (2) the publication is not available in English, and (3) the publication is not focused on nematodes and is not related to detection, classification, segmentation, morphological analysis, and so on using ML/DL techniques.

2.4. Selection Process

The chosen queries produced a total of 1416 articles, and 7 further articles were selected from Google Scholar as they were also considered relevant for inclusion in this study. The study selection process followed the PRISMA framework to ensure a rigorous and transparent methodology. Duplicate records were manually removed and irrelevant studies were excluded, including those focusing on Tree Nematode ML/DL models, Worm Malware, Worm Gear Neural Network, and Convolutional Neural Network (CNN) models trained on Nematode Connectome data. Two reviewers independently screened each article at the title, abstract, results, and/or full-text levels, with disagreements resolved through discussion and consensus or a third reviewer. This process ensured alignment with research objectives. Only studies explicitly focused on ML-driven nematode MIA tasks for detection, segmentation, and classification were included. Metadata were systematically recorded for each of the included studies.

2.5. Data Extraction Process

To systematically address the research questions, a structured data extraction framework was implemented. Each selected study was thoroughly reviewed, with key methodological and performance-related details recorded. A question-based Google Form was designed to standardize data collection, ensuring consistency across all studies. Data extraction was performed independently by the three main authors, with each reviewer responsible for a subset of the 44 included studies (16, 15, and 14 articles, respectively). To ensure accuracy, the data extracted for each study were cross-checked by one of the other two reviewers who had not originally collected data for that article. This framework captured essential metadata, methodological details, inference tasks, model architectures, and performance metrics. Each study was evaluated for dataset and code availability, as well as reported challenges, allowing for the synthesis of trends and research gaps in ML-based nematode microscopy image analysis. No study investigators were contacted to obtain or confirm additional data.

Table 2. Data Collected in Research Papers.

Category	Data Collected
Metadata	Title, Authors, Year, Citations, Publication Type, Location (Country), Institution/University, Journal/Publisher
Methods	ML/DL Model Type, Architecture, Computational and Pre-processing Techniques
Inference Tasks	Counting, Lifespan, Tracking, Morphological Analysis, Segmentation, Classification
Performance Metrics	Accuracy, Precision, Recall, F1-Score, mAP, AP, mAP50, AP@0.5, Avg Mask, IoU, Dice Coefficient
Dataset Information	Dataset Size, Availability, Splits (Train, Validation, Test)
Nematode Species	Species studied in each paper
Code Availability	Publicly available source code
Challenges	Identified gaps, limitations, and proposed improvements

summarizes some of the collected data.

In accordance with the PRISMA 2020 guidelines, the primary outcomes of interest were the performance metrics. These metrics were used to evaluate ML/DL models for nematode microscopy image analysis (e.g., accuracy, precision, recall, F1-score, Intersection over Union (IoU), Dice coefficient, and average precision). When multiple metrics were reported, all were included in this study. Accuracy and recall were the most considered performance metrics, as they were frequently reported. To provide context for these outcomes, additional variables such as metadata, model type and architecture, inference tasks, data information, and code availability were extracted, as summarized in Table 2. Categorizing the inference tasks in each paper helped to map the range of MIA objectives targeted across studies. Model architectures were categorized to identify top-performing frameworks, offering insight into architectural preferences among researchers. Dataset information and code availability were collected to assess dataset reproducibility and accessibility for replication. It was also recorded whether authors reported challenges, limitations, or future research directions. When information was missing or unclear (e.g., dataset split, code availability), it was recorded as not reported rather than inferred.

2.6. Other Methods

The reviewed studies were not subjected to a standardized risk of bias tool (e.g., QUADAS-2 or RoB-2); instead, three reviewers qualitatively evaluated potential biases in dataset design, code/data availability, and reporting clarity, resolving internal disagreements through discussion. Randomness in model training and the lack of cross-study replication further complicated efforts to apply formal bias or certainty assessment between studies. The primary effect measures extracted were model performance metrics that are commonly used in the context of image analysis (see Section 2.7 Performance Metrics). These measures were synthesized narratively rather than pooled meta-analytically, as differences in datasets, task definitions, and reported metrics precluded quantitative aggregation. This narrative method involves the textual summary and interpretation of findings from the included articles, which were organized thematically to identify the key patterns, data trends, and research gaps.

Studies were included in each synthesis if they reported at least one performance metric relevant to nematode detection, segmentation, tracking, or classification. To prepare the data for synthesis, missing or unclear details (e.g., dataset splits, code availability) were recorded as “not reported” without restraint. The results were summarized through structured tables and figures (e.g., task distributions, model architectures, geographic trends), in order to facilitate comparison across studies. No meta-analysis, subgroup analysis, or sensitivity analysis was conducted, as variations across models and outcomes were substantial. Heterogeneity was noted by comparing results across ML and DL model families, dataset types, and reported inference tasks.

Reporting bias was assessed indirectly by noting the absence of dataset availability or selective reporting of performance metrics, which limited reproducibility in several studies. Certainty in the body of evidence was therefore judged qualitatively: while CNN- and YOLO-based models consistently showed strong results across multiple tasks, the overall confidence in generalizability remains moderate due to dataset limitations and inconsistent reporting standards.

2.7. Performance Metrics

The reviewed studies employed a variety of performance metrics, depending on the image analysis task. To compare ML models, it is necessary to understand their respective evaluation metrics. For nematode detection, researchers applied well-known performance metrics such as accuracy, precision, recall, and F1-score. For segmentation, researchers commonly used Intersection over Union (IoU), Jaccard Index, Dice Coefficient, average precision (AP), and F1-Score as well. In this section, some of the most common metrics reports are briefly described.

2.7.1. Standard Classification Metrics

The metrics presented in this section are derived from the true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN). The corresponding formulas can be found below:

$$\begin{array}{cc}\mathrm{Accuracy}={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}},& \mathrm{Precision}={\displaystyle \frac{TP}{TP+FP}},\\ \mathrm{Recall}={\displaystyle \frac{TP}{TP+FN}},& \mathrm{F}1-\mathrm{Score}={\displaystyle \frac{2·\mathrm{Precision}·\mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}}.\end{array}$$

(1)

All of these values range between 0 and 1, and a higher value indicates better performance. Accuracy measures how often a model correctly classifies samples. However, it can be a misleading measure when class imbalance is present in a dataset [28]. Precision quantifies the proportion of true positives (TP) among all predicted positives, making it especially valuable in imbalanced datasets where false positives (FP) must be minimized. Recall measures a model’s ability to correctly identify all positive classes, making it essential for evaluating binary segmentation and classification performances [29]. F1-Score is one of the most commonly used performance metrics in machine learning for binary and multi-class segmentation and classification. It is best-known as the harmonic mean of precision and recall, balancing both metrics [30].

2.7.2. Additional Metrics

Intersection over Union (IoU), also known as the Jaccard Index, is widely used in object detection and segmentation to measure the overlap between predicted and ground truth regions. It is defined as the ratio of the intersection to the union of the predicted and actual segments, making it scale-invariant and well-suited for microscopic inference tasks such as nematode detection [31]. Its formula is given by:

$$\mathrm{IoU}={\displaystyle \frac{|A\cap B|}{|A\cup B|}}={\displaystyle \frac{TP}{TP+FP+FN}}$$

(2)

where A represents the prediction region, B the ground truth, and $|·|$ represents the area of the region.

The Dice Coefficient (DC) is a commonly used evaluation metric in image segmentation to quantify the overlap between the predicted shape and ground truth regions. It is defined as the ratio of twice the intersection over the total number of pixels in both sets:

$$\mathrm{DC}={\displaystyle \frac{2·|A\cap B|}{\left|A\right|+\left|B\right|}}{\displaystyle \frac{2·TP}{2·TP+FP+FN}}.$$

(3)

A dice score of 1 indicates perfect overlap, while 0 indicates no overlap. It is especially useful for assessing quality of pixel-level segmentation in nematode inference tasks [32].

Average precision (AP) measures overall model performance by computing the area under the precision–recall curve [33]. It captures the trade-off between precision and recall across varying confidence thresholds [34,35]. AP is typically approximated by the discrete summation:

$$\mathrm{AP}=\sum _n(R_n-R_{n-1})P_n,$$

(4)

where $R_n$ and $P_n$ are the recall and precision at the n-th threshold [34]. The PASCAL VOC benchmark introduced AP@0.5 using interpolated precision, which smooths the curve by selecting the highest precision at or above each recall level [35]. In contrast, the COCO benchmark extended this by averaging APs over IoU thresholds AP@[0.5–0.95] [36]. Both metrics are commonly used in the context of nematode image analysis.

For all these metrics, the values range between 0 and 1, with higher values indicating better performance.

2.8. Machine Learning Models

As mentioned in the Introduction, the recent decades have seen exponential growth in the use of deep learning models for image analysis tasks. Before these developments, classical machine learning methods would make use of hand-crafted features extracted from images, defined manually by experts, as inputs to a model that would turn vectors into a classification output. Classification models such as Support Vector Machine (SVM), Random Forest (RF), and artificial neural networks (ANNs) [37] fall in this category. However, thanks in part by the progress in computer hardware and the creation of large datasets enabled by the Internet, deep neural networks (DNNs) that can operate directly on the image domain became popular. These DNNs were the starting point of the deep learning revolution, and are capable of learning features directly from a dataset without the need for hand-crafted features. At present, there exist a large number of pre-trained models with different architectures that can be fine-tuned for many applications. Some of the standard architectures include Convolutional Neural Networks (CNNs), AutoEncoders (AEs), and Transformers [29]. Going over the details of all these architectures is beyond the scope of this article, but we provide some highlights on common architectures found in our survey in Appendix A.

3. Results

The database searches yielded a total of 1423 unique records, of which 1416 were screened based on their titles and abstracts. Following full-text assessment for eligibility, 44 studies were included in the final review (see the PRISMA flow diagram in Figure 2). Some papers that initially appeared to meet the inclusion criteria were excluded because one was a systematic review, another a Master’s thesis, and the rest either did not use appropriate imaging techniques or were not related to AI-based techniques [5,6,38,39,40,41,42,43].

Substantial heterogeneity was observed across the included studies in terms of study designs, dataset composition, and reported outcome metrics. Consequently, quantitative meta-analyses were not performed, and pooled summary estimates, precision intervals, or measures of statistical heterogeneity could not be generated. Similarly, subgroup analyses, formal investigations of heterogeneity among study results, and sensitivity analyses were not conducted due to incomplete and non-standardized reporting across studies. No formal assessment of reporting bias was made, as comparable effect sizes were rarely available. Nevertheless, the prevalence of high-performance results, coupled with limited dataset accessibility and small sample sizes, suggests that selective reporting cannot be excluded. A formal certainty-of-evidence framework, such as GRADE, was also not applied. Overall, confidence in the synthesized evidence was considered to be low to moderate, reflecting heterogeneous methods, small sample sizes, and incomplete reporting across studies.

This section covers the various tasks performed in the selected papers and provides a summarized view of recent research performed for the analysis and processing of nematode assays. Section 3.1 highlights the trend in the publications, Section 3.2 compares performances across the different image analysis tasks, and Section 3.3 provides an overview of trends on the biological inference tasks.

3.1. Publication Trends

This SLR adopted a narrative synthesis approach instead of a meta-analysis due to the substantial differences in model architectures, dataset structure, performance metrics, which made it difficult to mathematically combine the performance results. Instead, however, this study employs structured tables and visualizations to display trends in model usage, research objectives, publication patterns, and performance outcomes. The number of publications has increased significantly since 2021 (see Figure 3 (left)), reflecting a growing interest in ML/DL models for nematode microscopy analysis. A geographic breakdown showed that China, Germany, Spain, the United States, Australia, and Indonesia were the most active contributors by country (see Figure 3 (right)), pointing to regional hubs of innovation and investment in agriculture ML applications. China and Germany lead the field in publication output.

As can be observed from Figure 4 (left), publication venues ranged from AI-focused to general science venues, with a total of 13 conference publications and 31 journal articles. Publication patterns revealed a wide array of venues, yet recurring appearances of Scientific Reports, Sensors, and IEEE-affiliated conferences highlight a multidisciplinary interest in the application of AI-based biological image analysis for nematodes. The citation distribution (Figure 4 (right)) shows that most studies received fewer than 10 citations; however, 9 studies stood out with 20 or more citation, and one particularly influential study reached 124 citations [44], indicating varying degrees of influence within the field over the past 7 years.

The trends in image analysis tasks are shown in Figure 5. The left plot illustrates the distribution of tasks across the different publications, from which it can be observed that detection was the most common task. The right plot illustrates the set of ML models used across all these tasks, indicating that the YOLO architecture for object detection was the most common.

3.2. Image Analysis Trends

This section summarizes the general trends found regarding the main three image analysis tasks studied; i.e., classification, detection, and segmentation. The papers associated with tracking are discussed in the biological inference task section, as there are fewer papers reporting metrics that can be compared across the board and they are more specific to the biological task at hand.

Table 3. Classification Results. (S) under Class indicates that the task was species-related, while (L) indicates lifespan-related.

Ref					Classification
and Year	Model Used	Dataset: Size	Train/Test	Class	Acc	Prec	Rec	F1
[48] 2018	Cust. Siamese	Cust.: 50 k	-	5 (S)	0.93			0.94
	Neural Net
[45] 2019	Cust. Fully Conv. Nets (FCN)	Open Worm Mov. [49] (Single): 10,476 (V)	-	5 (S)	0.81			0.53
		Open Worm Mov. [50] (Multiple): 308 (V)	95 / 5	12 (S)	0.99			0.98
[51] 2021	Custom CNN Model	Cust.: 390	330/60	2	0.83
[7] 2020	Xception	Cust.: 188	-	3 (S)	0.88
		Cust.: 234	-	3 (S)	0.69
[46] 2022	Cust. CNN-WormNet	BioImage Archive (EMBL) [52]: 734	-	2 (L)	0.72
[47] 2022	ResNet101	Cust.: 257	190 / 67	19 (S)	0.94
[11] 2022	EfficientNet	Cust.: 957	862 / 95	11 (S)	0.97	0.98	0.97	0.97
[15] 2022	Cust. CNN-NemaNet	Cust.: 3063	-	5 (S)	0.98	0.98	0.98	0.98
[53] 2022	Vision Transformer	Cust.: 2769	2215 / 554	19 (S)	0.83	0.82	0.81	0.81
[54] 2023	Inception V3	I-Nema Dataset [16]: 277	227 / 50	2 (S)	0.9
[8] 2023	EfficientNet	Cust.: 957	-	11 (S)	0.96	0.97	0.96	0.96

, Table 4 and Table 5 summarize the performances and datasets associated with these tasks. The “Year” column in the tables illustrates the chronological evolution of the models, helping readers to understand how model usage has progressed over time. Table 3 highlights the main classification results. Most of these results are associated with species classification from images, while one paper focused on videos [45] and one paper predicted lifespan [46]. It can be observed that the papers did not use consistent dataset sizes, with the smallest dataset including around 200 samples, largest dataset including around 50k samples, and a median of less than 1000 samples. The number of classes ranged from 2 to 19. Some papers specified the training and testing split used for evaluation, while most did not. The most common metrics are included in the table, with accuracy and F1 score having the most overlap across all the papers. Accuracy ranged from 0.69 to 0.99, and the paper including 19 classes [47] reported an accuracy of 0.94. These are relatively high values; however, due to the lack of accessibility to the datasets, it was not possible to reproduce the results for most of the models. The F1-score values ranged from 0.53 to 0.98.

Table 4 highlights the main detection results. In this case, all articles used their own custom dataset with sizes ranging from around 200 to 20 k (excluding the one obtained with data augmentation), with the only exception being [55], which used the public dataset BBBC010/11 [56]. The studies used accuracy, precision, recall, and F1-score more consistently as metrics for evaluation. The accuracy ranged from 0.57 to 0.99, precision from 0.59 to 1.00, recall from 0.68 to 0.99, and F1-score from 0.47 to 0.99. As in the previous case, relatively high scores reported were reported in several cases; however, it was not possible to reproduce or directly compare them due to the lack of accessible datasets. Over time, there is a clear trend showing that YOLO-based architectures have become increasingly dominant in recent years, reflecting the shift toward faster and more efficient object detection models.

Table 4. Detection Results.

Ref				Detection
and Year	Model Used	Dataset: Size	Train/Test	Acc	Prec	Rec	F1
[44] 2018	Conv. Selective AE	Cust: 3044	2435/609	0.95	0.94	0.94	0.94
[57] 2018	CNN	Cust: 11,820	10,047/1773	0.93
[58] 2019	HoG+SVM	Cust: 1200	996/204				0.87
[45] 2019	FCN			0.99
[59] 2021	CNN+LSTM	Cust: 23,584	17,923/5661	0.91
[60] 2022	YOLOv5s	Cust.: 1901	1520/380	0.93
	Faster R-CNN			0.94
[61] 2022	YOLOv4-1 Class				0.91	0.79	0.85
	YOLOv4-Multiclass	Cust.: 20,000 (augmented images)	19,000/1000		0.93	0.93	0.92
[62] 2022	Image processing	N/A			0.98	0.98	0.98
[55] 2023	YOLOv5s	BBBC010/11 [56]: 17,213	14,385/2828		0.88	0.76
[63] 2023	YOLOv7-640	Cust: 415	332/83	0.99	0.99	0.99	0.99
[64] 2023	YOLOv5	Cust:13k (augmented)	10,400/2600		1.00	0.99	0.99
[65] 2023	Faster R-CNN	Cust.: 182	163/19	0.86	0.94	0.83	0.88
	YOLOv4-CSP			0.77	0.91	0.80	0.85
[66] 2023	YOLOv7	Cust.: 3931	3537/394	0.99	0.97	0.96	0.97
	Faster R-CNN			0.98	0.74	0.99	0.85
[67] 2023	YOLOv5x	Cust: 3503	-		0.85	0.75
[68] 2023	Faster R-CNN	Cust: 1625	1413/212	0.87	0.61	0.68
[69] 2024	CNN and contour methods	Cust: 10,468		0.99
[70] 2024	YOLOv8	Cust: 19,929	15,677/4252	0.57	0.59	0.94	0.47
[71] 2024	YOLOv5	Cust: 894	788/106		0.78	0.78
[72] 2024	YOLOv5	Cust: 6000			0.97	0.94	0.95
[73] 2024	HoG+SVM	Cust: 20,647		0.91	0.92	0.97	0.95

Table 4. Detection Results.

Ref				Detection
and Year	Model Used	Dataset: Size	Train/Test	Acc	Prec	Rec	F1
[44] 2018	Conv. Selective AE	Cust: 3044	2435/609	0.95	0.94	0.94	0.94
[57] 2018	CNN	Cust: 11,820	10,047/1773	0.93
[58] 2019	HoG+SVM	Cust: 1200	996/204				0.87
[45] 2019	FCN			0.99
[59] 2021	CNN+LSTM	Cust: 23,584	17,923/5661	0.91
[60] 2022	YOLOv5s	Cust.: 1901	1520/380	0.93
	Faster R-CNN			0.94
[61] 2022	YOLOv4-1 Class				0.91	0.79	0.85
	YOLOv4-Multiclass	Cust.: 20,000 (augmented images)	19,000/1000		0.93	0.93	0.92
[62] 2022	Image processing	N/A			0.98	0.98	0.98
[55] 2023	YOLOv5s	BBBC010/11 [56]: 17,213	14,385/2828		0.88	0.76
[63] 2023	YOLOv7-640	Cust: 415	332/83	0.99	0.99	0.99	0.99
[64] 2023	YOLOv5	Cust:13k (augmented)	10,400/2600		1.00	0.99	0.99
[65] 2023	Faster R-CNN	Cust.: 182	163/19	0.86	0.94	0.83	0.88
	YOLOv4-CSP			0.77	0.91	0.80	0.85
[66] 2023	YOLOv7	Cust.: 3931	3537/394	0.99	0.97	0.96	0.97
	Faster R-CNN			0.98	0.74	0.99	0.85
[67] 2023	YOLOv5x	Cust: 3503	-		0.85	0.75
[68] 2023	Faster R-CNN	Cust: 1625	1413/212	0.87	0.61	0.68
[69] 2024	CNN and contour methods	Cust: 10,468		0.99
[70] 2024	YOLOv8	Cust: 19,929	15,677/4252	0.57	0.59	0.94	0.47
[71] 2024	YOLOv5	Cust: 894	788/106		0.78	0.78
[72] 2024	YOLOv5	Cust: 6000			0.97	0.94	0.95
[73] 2024	HoG+SVM	Cust: 20,647		0.91	0.92	0.97	0.95

Table 5 highlights the main segmentation results. In contrast to the detection studies, most segmentation works relied on publicly available datasets (Table 6), with several articles utilizing common datasets (D1–D5), thus enabling more consistent benchmarking across studies. The most frequently used models included U-Net, Mask R-CNN, and transformer-based architectures, with the Dice coefficient (DC) and AP50 being the most commonly reported evaluation metrics. Some studies also reported Precision and Recall values. Custom datasets (C1–C5) generally resulted in lower performance, particularly with conventional architectures such as U-Net. The highest results were consistently achieved by transformer-based models, reaching up to 0.99 on CSB-1 and 0.98 on the Mating Dataset. Despite these strong performances, not all studies provided clear details about the number of dataset images or train/test splits, which limits reproducibility and direct comparisons. Two segmentation papers [46,74] were not included in Table 5, as their reported evaluation metrics were not comparable with those adopted in our segmentation results table.

Table 5. Segmentation Results.

	Ref		Segmentation
Data	and Year	Model Used	DC	AP50	Prec/Rec @ 0.5
C4	[75] 2019	Image proc.	0.76
C6	[13] 2020	U-Net			0.90/0.87
D1	[76] 2021	Mask R-CNN		0.9
C1	[77] 2022	U-Net	0.46
C2			0.47
C3	[78] 2022	Mask R-CNN		0.89
D2	[18] 2023	Transformer-WormSwin		0.98
D3	[33] 2024	Mask2Former		0.62
D4	[13] 2020	U-Net			0.96/0.97
	[18] 2023	Transformer-WormSwin		0.95
	[33] 2024	Mask2Former		0.92
	[79] 2024	Mask2Former		0.71
D5	[18] 2023	Transformer-WormSwin		0.99
	[79] 2024	Mask2former		0.87
D3	[33] 2024	Multiple Instance Learning		0.65
D4				0.92

Table 5. Segmentation Results.

	Ref		Segmentation
Data	and Year	Model Used	DC	AP50	Prec/Rec @ 0.5
C4	[75] 2019	Image proc.	0.76
C6	[13] 2020	U-Net			0.90/0.87
D1	[76] 2021	Mask R-CNN		0.9
C1	[77] 2022	U-Net	0.46
C2			0.47
C3	[78] 2022	Mask R-CNN		0.89
D2	[18] 2023	Transformer-WormSwin		0.98
D3	[33] 2024	Mask2Former		0.62
D4	[13] 2020	U-Net			0.96/0.97
	[18] 2023	Transformer-WormSwin		0.95
	[33] 2024	Mask2Former		0.92
	[79] 2024	Mask2Former		0.71
D5	[18] 2023	Transformer-WormSwin		0.99
	[79] 2024	Mask2former		0.87
D3	[33] 2024	Multiple Instance Learning		0.65
D4				0.92

Table 6. Segmentation Datasets.

ID	Name and Ref	No. of Images	Train/Test Split	Avail.
D1	C. elegans [80]	1908	1711/197	Yes
D2	Mating Dataset [18]	450 patch images	0/450	Yes
D3	Haemonchus Contortus [17]	95	76/19	Yes
D4	BBBC010 [56]	100	N/A	Yes
D5	CSB-1 [18]	4631 (10 videos)	N/A	Yes
C1	Custom Nematode Dataset [77]	3967	3367/667	No
C2	Custom Cyst Dataset [77]	487	387/100	No
C3	Custom Nematode Dataset [78]	16k	14,880/1120	No
C4	Custom Cyst Dataset [75]	132	N/A	No
C5	Potato Cyst Nematode [13]	1973	1606/367	No

Table 6. Segmentation Datasets.

ID	Name and Ref	No. of Images	Train/Test Split	Avail.
D1	C. elegans [80]	1908	1711/197	Yes
D2	Mating Dataset [18]	450 patch images	0/450	Yes
D3	Haemonchus Contortus [17]	95	76/19	Yes
D4	BBBC010 [56]	100	N/A	Yes
D5	CSB-1 [18]	4631 (10 videos)	N/A	Yes
C1	Custom Nematode Dataset [77]	3967	3367/667	No
C2	Custom Cyst Dataset [77]	487	387/100	No
C3	Custom Nematode Dataset [78]	16k	14,880/1120	No
C4	Custom Cyst Dataset [75]	132	N/A	No
C5	Potato Cyst Nematode [13]	1973	1606/367	No

Some papers addressed multiple tasks explicitly and, therefore, appear in more than one category. For example, ref. [46] used separate models for classification and segmentation. Thus, while it is shown under segmentation in Figure 5, it is also listed in Table 3. Likewise, ref. [45] is categorized under detection in Figure 5 but is also included in Table 3. Similarly, ref. [65] is grouped under tracking in Figure 5, yet it is also listed in Table 4.

CNNs were the most frequently used architectures across the 44 selected studies, appearing in 14 implementations and showing high adaptability across tasks such as classification, detection, and segmentation. YOLO-based models followed in adoption (12 studies), with YOLOv7 achieving the best performance for real-time species classification (99.56% mAP@0.5), with other studies also performing tasks in tracking and counting. Transformer-based models (4 studies) demonstrated exceptional accuracy for instance segmentation and counting when data were synthetically augmented. For instance, WormSwin reached 99% AP@0.5. R-CNN variants (5 studies) showed strong spatial localization but incurred higher computational cost. U-Net and encoder–decoder architectures (4 studies) performed moderately in segmentation tasks, while the Convolutional Selective Autoencoder (CSAE) achieved over 94% F1-Score for nematode egg counting. Finally, classical ML models (4 studies), including HOG+SVM pipelines, proved highly effective for nematode counting tasks (up to 95% F1-Score) but underperformed in segmentation applications.

Across the models, performance generally correlated with task complexity. YOLOv7 delivered the highest results, achieving 99.56% mAP@0.5 and 52.5 FPS, outperforming Faster R-CNN in both speed and accuracy for species classification. Transformer-based models such as WormSwin excelled in dense segmentation and counting tasks using synthetic data, while classical ML pipelines such as the HOG+SVM ML model remained effective for simple detection tasks. However, U-Net models showed limited segmentation accuracy and R-CNN variants incurred slower inference times, exceeding 43 ms per prediction.

Given the considerable heterogeneity in study design and inconsistent reporting, a formal risk-of-bias tool was not utilized. Many studies lacked essential details regarding dataset composition, annotation protocols, and train/test splits, which hindered a structured evaluation. Thus, the risk of bias across studies is deemed unclear to high, and the results should be interpreted with caution.

Datasets

The nematode public dataset shown in Table 3 focuses primarily on behavioral tracking, genetic diversity, lifespan analysis, and species recognition. The behavior dataset presented by Javer et al. [49] is among the most extensive resources, comprising nearly 15,000 single-worm tracking experiments across hundreds of genotypes, including over 2700 videos from the L4 larval stage to death for aging studies. The CeNDR dataset [50], in contrast, is not an image repository but a critical community resource that provides genetically diverse wild C. elegans strains, allowing for population-level genetics and strain-specific imaging experiments. A complementary microscopy dataset available on the BioImages Archive [52] offers wide-field time-lapse images of C. elegans annotated for lifespan, movement, and body part segmentation. Finally, the I-Nema dataset [16] extends beyond C. elegans, providing an open access dataset benchmark for species-level recognition with more than 2700 images representing 19 different species collected in both natural and laboratory environments.

Complementing the previously described datasets, the public datasets of nematodes in Table 6 differ considerably in species focus, imaging conditions, annotation strategies, and experimental design, which directly affects model performance and reproducibility. For example, the DIY microscope dataset by Fudickar et al. [80] provides high-resolution (3280 × 2464 px) Petri dish images of C. elegans acquired with a Raspberry Pi-based system, enabling affordable large-scale data collection but with variability in illumination. In contrast, the dataset used in WormSwim [18] includes three complementary sets: the CSB-1 dataset, which contains video-derived frames of C. elegans with pixel-level binary masks; a Synthetic dataset generated by compositing segmented worms onto background images to expand training diversity; and the Mating Dataset (MD), consisting of 450 dense, overlapping grayscale patches designed to challenge instances in segmentation under crowding conditions. Beyond C. elegans, Zofka et al. [17] introduced a dataset of Haemonchus Contortus (an animal-parasitic nematode) motility assays, pairing manually annotated microscopic images with motility videos across live–dead ratios. Lastly, the BBBC010 benchmark set [56] provides curated microscopy images of C. elegans subjected to pathogen and antibiotic treatments, annotated with foreground–background segmentation masks and biological phenotype (dead/alive) labels. Together, these datasets highlight key differences: controlled versus low-cost imaging setups, real versus synthetic data generation, single versus multiple nematode species, and annotations ranging from binary masks to biological treatment outcomes. The BBBC10 benchmark was also used in Table 4, together with the BBBC11 benchmark found in [56]. The difference in datasets here underscores the importance of standardized benchmarks such as CSB-1 and BBBC010 for fair model evaluation and reproducibility.

3.3. Biological Inference Trends

Table 7. Proportion of Papers Addressing Biological Inference Tasks.

Biological Inference Task	Proportion of Papers
Species Classification	$14/44=0.32$
Counting Specimens	$16/44=0.36$
Behavior Monitoring	$11/44=0.25$
Lifespan Tracking	$4/44=0.09$

shows the distribution of the biological inference tasks across all the publications. The number of papers do not add up to 44, as some papers addressed multiple biological tasks. It can be observed that the majority of the articles focused on species classification and counting tasks, while lifespan tracking was present in a relatively smaller percentage of articles. The rest of this section highlights some of the trends observed regarding the biological inference tasks.

3.3.1. Species Classification

Species classification is one of the most researched tasks for nematode microscopic image inference, as determined by this systematic review. The first major approach to species classification using DL was presented by Uhlemann et al. [7], who used CNNs to identify nematodes based on the variations present in the dataset of images to train where nematodes are classified.

Researchers have made use of high-magnification and high-resolution images to focus on specific areas of interest on the nematode’s body, such as the head, tail, and genitalia [8,11,15,47]. This approach focuses on the fact that nematode species tend to express most species-level differences around these areas, especially in adults. The use of close-up images led most researchers to focus on using “off-the-shelf” CNN architectures for training models, while sidestepping challenges associated with nematode detection in complex visual conditions such as low-lighting, high debris, occlusion, and overlapping worms.

The current state-of-the-art (SOTA) species classification model, developed by Qayyum et al. [53], uses close-up images and train a majority voting Vision Transformer with test time augmentation—a method that augments the test data during evaluation to get average predictions of the variations of the test data—to perform species classification across 19 classes with overall accuracy of 83%, precision of 82.%, and recall of 81%. This method still requires further validation with more complex datasets; however, it is a promising approach based on the improved performance of vision transformers over CNNs [81,82,83]. Abade et al. [15] also developed a SOTA model using NemaNet—a custom CNN based on a combination of DenseNet121 and InceptionV3—and close-up images obtained using traditional imaging techniques. NemaNet was trained to classify 5 different plant-parasitic nematodes with an accuracy of 98.82% and precision of 98% [15]. A similar approach was evaluated by Verma et al., who used InceptionV3 to classify two nematode species: Acrobeles and Acrobeloides [54].

A variety of modifications have been made to the problem to reduce the complexity of inference, such as including pose, nematode length, and focal length. Javer et al. [45] used the pose to classify nematode species by postural dynamics alone using a model inspired by a Dilated ResNet architecture, and made their dataset public [49]. They found that using the worm angle is the most effective feature for identifying species based on postural dynamics. Another approach is to focus on nematodes in the infective juvenile (IJ) stage, as they can be more easily classified by their body length [7,63,70,71]. Finally, some researchers have combined custom computer vision algorithms with DL models to produce effective results. One such example is that of Thevenoux et al. [51], who introduced an algorithm that isolates the nematode head from the image background using contour detection, mathematical morphology, and a landmark-based EB-Net architecture CNN model to automate nematode classification. These kinds of algorithms may be difficult to scale to additional species due to the parameters and design of the custom computer vision algorithms, which may require redesign. Lui et al. [48] captured images of the same nematode at multiple focal lengths. Passing the image stack through a CNN, they then performed high-dimensional data analysis to classify species; however, the used techniques impose a heavy data collection burden, reducing their possibility for use in automated approaches to nematode species classification.

3.3.2. Counting

Nematode counting is a common task in handling nematode assays and samples, and having a high-performance model ultimately enables many of the more complex tasks that follow to be accurate. The models most commonly employed for nematode counting are YOLO models; however, other models have been implemented as well (see Table 4). The main challenges in nematode counting are noise handling, occlusion, and overlapping nematodes. The limitations of using non-DL techniques is that they require specific conditions for the nematodes to be detected, such as the entire nematode body being visible [64]. Many researchers have tried to create robust models to handle these challenges using YOLO models, Faster R-CNN, and Histogram of Gradients (HOG) with Support Vector Machine (SVM). Rico-Guardiola et al. [60] showed that both YOLOv5 and Faster R-CNN models performed similarly in terms of accuracy; however, the smaller version of the YOLO model, YOLOv5s, obtained better results when considering computational costs. This indicates that YOLO models may be more suited for high-speed nematode inference applications when considering the one-stage detector architecture, which is faster than the two-stage detector architecture used in other CNN models. The most accurate model was given by Pun et al. [64], who trained various YOLO models and found that the YOLOv5s model performed almost perfectly on the validation set with a precision of 100% and mAP of 99.5% under conditions such as debris and overlapped nematodes. They implemented a mosaic augmentation technique to improve the model robustness and found that it slightly improves the model performance as it models the use of varied lenses in nematode counting. Similarly, Rico-Guardiola et al. [60] obtained an accuracy of 93.2%. Mori et al. [61] addressed the problem of counting overlapping worms by making an object classification for two, three, or four overlapping worms with a comparable performance of AP@.5 at 90%, following the approach cited by Pun et al. [64]. Another technique implemented by researchers to improve model robustness when dealing with noise, occlusion, and overlapping worms is to use dataset augmentation techniques such as Style transfer, Pix2Pix, and Supervised self-labeling [55]. The Pix2Pix method allows researchers to generate synthetic images of a dataset which are capable of matching a given real appearance.

Finally, there is a small, growing body of nematode detection literature focused on counting eggs to estimate the number of nematodes—specifically, root-knot nematodes (RKNs)—within a collected sample. The studies below highlight the use of CNN-based models to estimate egg counts within microscopic images of a sample. Akintayo et al. [44] have been cited frequently in the literature, as they used a Convolutional Sparse Auto-Encoder (CSAE) model which has served as a baseline in other works; however, other approaches are emerging. The best performance reported to date is with the YOLO model, which provides an accuracy of 99.1% with images that contain debris and occlusions at an inference speed of up to 1.3 FPS for eggs and 4 FPS for nematodes [63]. Saikai et al. [70] used two YOLOv8 models to detect RKN eggs and second-stage juvenile (J2) nematodes, respectively. This paper differs in that it considers the inclusion of free-living nematodes (FLN) within the microscopic image, thereby introducing complexities to the inference task. Another method by Fraher, Watson et al. [69] is a hybrid approach employing a contour-based method and a CNN to count RKN eggs. Methods employing ML techniques yield higher performance than techniques utilizing traditional image processing techniques, such as the novel extreme point method proposed by Pun et al. [62], which achieved a detection coefficient of determination ($R^2$) of 90.5%. The benefit of methods employing traditional image processing techniques is that they are not black-box models, such that the algorithms can be modified to account for new factors.

3.3.3. Behavioral Tracking

Another area of research has developed, which involves detecting and tracking nematodes in order to measure and evaluate their responses to various controlled stimuli. Research focused on behavioral tracking of nematodes is quite new, with all research having been published starting from 2023. A major consideration in behavior tracking research is the inference speed, which should be performant due to the requirements of real-time tracking. These algorithms are further derived for applications such as lifespan tracking, where behavior is often an important parameter [9,59].

One of the most relevant results comes from Deep-worm [10], using YOLOv5 and Strong SORT tracking backbones to detect nematodes and track their movement and predicted trajectories in real-time. They use the Strong SORT algorithm to make the model more robust to overlapping and occluded worms, and they have made the dataset publicly available for use [10]. The Deep-worm nematode tracking pipeline obtained an impressive precision of 98%, recall of 94%, and mAP.5 of 98% with a 9 min training time [10]. Ribeiro et al. [65] used a different approach, involving Faster R-CNN and ResNet18 for nematode tracking by implementing tracking algorithms (e.g., IoU Tracker, SORT, Centroid-Kalman) directly onto the raw holographic patterns. Additionally, Ribeiro et al. extracted biomechanical parameters such as nematode speed and movement patterns [65].

Some other uses of behavioral tracking include pose prediction and analysis. Castro et al. [84] developed a novel method for predicting C. elegans poses in low-resolution images of multiple worms in full 55 mm Petri dishes. The predicted poses are images of worm skeletons, and are the result of a neural network trained to predict aggregation behaviors between worms. Zhang and Chen [85] simplified the problem by implementing microfluidic chips; in particular, they proposed an algorithm for automatic counting and analysis of the body-bending behavior of nematodes for microfluidic chips.

3.3.4. Lifespan Tracking

Nematode lifespan tracking and prediction is a newer topic of research, with most of the papers published after 2021 and the majority being in 2024. Nematodes have become model organisms for understanding neuro-degenerative diseases and aging, based on their short lifespan and simple nervous system.

Dead or alive nematode classification has served as a model problem for researchers. For example, Garvi et al. [59] presented an automated system for classifying dead or live C. elegans using a ResNet18 with a unidirectional long short-term memory (LSTM) model. Additionally, they evaluated the system using an original, simulated, and mixed dataset, and showed that the mixed dataset improved the hit ratio by 3.20%, resulting in an accuracy of 83.66% for dead C. elegans and 98.56% for live C. elegans. Galimov and Yakimovich [46] used WormNet and HydraNet to show that the posterior part of the C. Elegans is most useful for classifying long-lived C. elegans, which was achieved by predicting the lifespan class of C. elegans nematodes based on young adult images (1-3-day-old adults), using HydraNet to classify age and WormNet to segment the worms into anterior, mid-body, and posterior parts. Radvansky et al. [73] are the only researchers to use ML techniques, who proposed using a HoG and SVM and tracking the nematodes over time to determine whether they are dead or alive. Finally, Bates et al. [9] demonstrated the variety of inferences possible through the use of a Faster R-CNN to detect and identify nematodes; another one to detect the stage of life they are in; and another to determine whether something is an egg, nematode, or neither. They utilized 3 different Faster R-CNN models to track nematode speeds during development, measure egg-laying rates, map spatial distribution in reproductive adults, and show behavioral decline in aging nematode populations [9].

Researchers have used custom and off-the-shelf CNN architectures for lifespan tracking, with YOLO being the most commonly used architecture in the last year. For example, Phuyued et al. [71] used a YOLOv5 model with contrast-limited adaptive Histogram Equalization (CLAHE) to improve image clarity and detect nematodes in the infective juvenile stage.

4. Discussion

4.1. Advantages

This review highlights the increasing application of deep learning methods in nematode image analysis tasks such as classification, detection, segmentation, and tracking, with a focus on the first three tasks. Biological inference tasks included species classification, counting, behavior monitoring, and lifespan tracking. We observed that YOLO-based architectures consistently performed well in detection and counting tasks, offering fast inference and high accuracy, even in visually complex environments with overlapping worms or background noise. For classification, CNN-based models and transformer architectures achieved strong results, particularly when applied to high-resolution or close-up images of key morphological regions. Transformer-based models also demonstrated superior performance in segmentation tasks, especially when applied to well-annotated public datasets such as BBBC010 and CSB-1. These findings indicate that data-driven approaches are well positioned to improve diagnostic accuracy, reproducibility, and efficiency compared with manual microscopy.

4.2. Challenges and Proposed Solutions

Despite these encouraging outcomes, there are notable limitations in the current body of research. Many studies do not provide detailed information about the dataset composition, train/test splits, or annotation protocols, which undermines their reproducibility and limits the ability to perform fair comparisons. The utilized evaluation metrics also vary across studies, with some using accuracy and F1-score, while others use Dice coefficient, AP, or mAP, further complicating benchmarking efforts. In addition, three major challenges identified in this review are detailed below.

• Limited training data—Custom datasets are often small or inconsistently annotated. Another barrier identified in this review is the lack of standardized benchmark (SB) datasets for nematodes. In the context of ML, an SB is a common, objective framework that is used to evaluate and compare the performance of different models and algorithms. It refers to a curated and publicly available dataset with clearly defined tasks, consistent annotation standards, and well-established evaluation metrics. Solution: Wider use of data augmentation (e.g., rotation, scaling, and synthetic image generation), transfer learning, and collaborative dataset sharing can improve the generalizability of models. Many studies relied on private datasets and inconsistent metrics, making direct evaluation difficult. A prime example of the SB approach is the I-Nema dataset presented by Lu et al. [16], which provides images of 19 nematode species with a rigorous, repeatable experimental protocol. Adopting such benchmarks would directly address the challenges of reproducibility and inconsistent reporting.
• Occlusion and overlapping objects—Worms and eggs often overlap, complicating detection and counting. Solution: Advanced models such as transformer-based detectors and instance segmentation approaches can help to resolve occluded objects more reliably.
• Inconsistent metric reporting—Variation in reported evaluation metrics makes cross-study comparisons difficult. Solution: Adoption of standardized evaluation frameworks and transparent reporting of datasets and splits will facilitate reproducibility and fair benchmarking.

A practical consideration for applying the models reviewed in this study is their handling of large-scale, high-resolution images. CNN- and YOLO-based models are trained at fixed input resolutions, meaning that high-resolution images must be either rescaled—potentially losing critical details—or divided into smaller tiles for analysis. In practice, patch-based processing with overlapping tiles and stitched outputs is the standard approach for dealing with these large images. This allows high-resolution images to be analyzed accurately, though at the cost of additional computational effort.

This review itself is also subject to certain limitations. It focused primarily on three major databases and English-language articles, which may have led to the omission of relevant studies published elsewhere. In addition, no formal risk of bias assessment or meta-analysis was conducted; therefore, the synthesis presented here is qualitative in nature and may be influenced by reporting variations across studies.

4.3. Future Perspectives

These and future advancements in deep learning methods for nematode image analysis stand to improve plant-parasitic nematode detection and diagnostics, thus allowing nematologists and diagnosticians to improve the speed and ease with which samples are processed, while maintaining highly accurate specimen classification and counting. This, in turn, will directly support farmers and plant protection specialists in making timely nematode management decisions and avoiding yield losses. Additionally, future cloud-based imaging processing may support farmers in under-resourced regions of the world, where a nematologist or diagnostician may not be physically present to conduct manual microscopic evaluations. Expanding and sharing plant-parasitic nematode datasets and the application of image analysis programs in plant inspection facilities can assist in preventing the introduction of invasive nematode species to new areas and support regulatory policies.

5. Conclusions

Based on this Systematic Literature Review (SLR), machine learning (ML) and deep learning (DL) models offer a scalable solution for automating microscopic nematode image analysis, which has traditionally been labor-intensive and slow. A total of 44 articles were reviewed in detail. It was observed that Convolutional Neural Networks (CNNs) were the most frequently used models, with specific architectures such as YOLO excelling in detection, while Transformer-based models were more effective for dense segmentation and counting. Despite these advancements and strong performance metrics (with accuracy values reported as high as 0.99 for classification, detection, and segmentation), challenges persist—including a limited number of diverse nematode datasets for training, issues relating to occlusion, and inconsistent reporting of performance metrics across different studies. It is recommended that future work should focus on addressing these gaps.