Introduction
Human parasitic gastrointestinal infections constitute one of the most prevalent neglected global health burdens. Over 3.5 billion people remain at risk, with soil-transmitted helminths (STHs) and intestinal protozoa perpetuating cycles of malnutrition, anaemia and cognitive impairment across low- and middle-income countries (Hotez et al., Reference Hotez, Molyneux, Fenwick, Kumaresan, Sachs, Sachs and Savioli2007; Jourdan et al., Reference Jourdan, Lamberton, Fenwick and Addiss2018). The prevalence of these infections varies significantly across endemic regions, with sub-Saharan Africa and Asia bearing the highest burden – frequently exceeding 40% for STHs and 20% for protozoa (Figure 1) (World Health Organization, 2025). Despite this staggering burden, the reference diagnostic method – optical microscopy of stool samples – has remained essentially unchanged since the early 20th century (Dirani et al., Reference Dirani, Zannoli, Paesini, Farabegoli, Dalmo, Vocale, Liguori, Varani and Sambri2019; Amjad, Reference Amjad2020). Although molecular techniques such as polymerase chain reaction have improved diagnostic sensitivity and specificity, their implementation remains severely restricted by high costs, infrastructure requirements and dependence on highly skilled personnel (Cools et al., Reference Cools, Vlaminck, Albonico, Ame, Ayana, José Antonio, Cringoli, Dana, Keiser, Maurelli, Maya, Matoso, Montresor, Mekonnen, Mirams, Corrêa-oliveira, Pinto, Rinaldi, Sayasone, Thomas, Verweij, Vercruysse and Levecke2019; Miswan et al., Reference Miswan, Singham and Othman2022). Artificial intelligence (AI), particularly deep learning (DL), has emerged as a promising solution to this diagnostic bottleneck. Convolutional neural networks (CNNs) trained on digital microscopy images have repeatedly demonstrated diagnostic accuracies that rival or exceed human performance (Ramos-briceno et al., Reference Ramos-briceno, Flammia-d’aleo, Fernandez-lopez, Carrion-nessi and Forero-pena2025). Yet a notable asymmetry has emerged across parasite classes. For helminth eggs – structures typically exceeding 50 µm in length, with species-specific yet conserved morphologies – AI performance consistently achieves clinical-grade accuracy and generalizes across diverse populations (Dacal et al., Reference Dacal, Bermejo-pelaez, Lin, Alamo, Cuadrado, Martínez, Mousa, Postigo, Soto, Sukosd, Vladimirov, Mwandawiro, Gichuki, Williams, Muñoz, Kepha and Luengo-oroz2021; Cure-bolt et al., Reference Cure-bolt, Perez, Broadfield, Levecke, Hu, Oleynick, Beltrán, Ward and Stuyver2024; Chua et al., Reference Chua, Henderson, de Guzman, Foss, Schub, Bell, Medina, Siao, Mistica, Belleza, Modequillo, Torres and Belizario2025; von Bahr et al., Reference von Bahr, Suutala, Diwan, Mårtensson, Lundin and Linder2026). Here, we define clinical-grade accuracy as a diagnostic performance that meets the minimum acceptable standards for routine clinical use, typically characterized by a sensitivity and specificity of at least 90–95% and a Cohen’s kappa coefficient ≥0.80 when compared to expert manual microscopy, indicating almost perfect agreement (Dacal et al., Reference Dacal, Bermejo-pelaez, Lin, Alamo, Cuadrado, Martínez, Mousa, Postigo, Soto, Sukosd, Vladimirov, Mwandawiro, Gichuki, Williams, Muñoz, Kepha and Luengo-oroz2021). For gastrointestinal protozoa such as Giardia intestinalis, Cryptosporidium hominis and Entamoeba histolytica, the same algorithms exhibit reduced robustness, batch-dependent degradation and poor external validation (Zhang et al., Reference Zhang, Jiang, Jiang, Xi, Chen, Liu, Juhas, Li and Zhang2022). This review argues that this asymmetry is not incidental but structural. It reflects key differences in object complexity, dataset representativeness and the historical inertia of diagnostic priorities. The result is an apparent discrepancy in diagnostic parasitology: a convolution and attention network (CoAtNet) achieved an average accuracy of 93% in classifying 11 different types of parasite eggs under complex imaging conditions (AlDahoul et al., Reference AlDahoul, Karim, Momo, Escobar, Magallanes and Tan2023). However, challenges persist, as automated systems frequently misidentify Giardia cysts when analysing images acquired under suboptimal conditions. For instance, smartphone-based microscopy – often deployed in resource-limited settings – introduces noise and artefacts that increase the likelihood of confusing cysts with air bubbles or lipid droplets (Nakarmi et al., Reference Nakarmi, Pudasaini, Thapaliya, Upretee, Shrestha, Giri, Neupane and Khanal2024). This systematic review therefore synthesizes research on AI-based diagnostics for gastrointestinal parasites, systematically comparing algorithmic approaches for helminths and protozoa. By identifying the methodological and regulatory barriers impeding clinical adoption – especially for protozoan infections – we aim to establish clear priorities for future investigation and accelerate progress towards closing the protozoan gap.
Prevalence of gastrointestinal parasites by endemic region. The chart displays the estimated prevalence ranges (percentages) of 6 key parasitic infections – ascariasis, trichuriasis, soil-transmitted helminths (STHs), giardiasis, amebiasis and cryptosporidiosis – across 4 major endemic regions: South America and the Caribbean, sub-Saharan Africa, Asia and Oceania. Coloured bars illustrate the prevalence range for each parasite within a given region, with the specific low-high percentages shown above each bar. The values in parentheses represent the midpoint of each range, providing an average estimate for comparative purposes. STHs refer to a composite index that includes Ascaris lumbricoides, Trichuris trichiura and hookworms. Data sources: GBD 2021, WHO 2025 and country-level surveillance reports. Prevalence ranges were derived from WHO 2024 Global Estimates and a meta-analysis by Smith et al. (2023). For regions without published data, expert consensus among 3 authors was used.
Figure 1 Long description
The heatmap displays prevalence ranges of six parasitic infections across four major endemic regions: South America and Caribbean, Sub-Saharan Africa, Asia and Oceania. The infections listed are Ascariasis, Trichuriasis, Soil-Transmitted Helminths (STHs), Giardiasis, Amebiasis and Cryptosporidiosis. Each cell contains numerical values representing prevalence percentages. For South America and Caribbean, Ascariasis ranges from 10 to 29 percent, Trichuriasis from 7 to 21 percent, STHs from 11 to 49 percent, Giardiasis from 13 to 30 percent, Amebiasis from 5 to 19 percent and Cryptosporidiosis from 4 to 12 percent. Sub-Saharan Africa shows Ascariasis at 10 to 35 percent, Trichuriasis at 10 to 25 percent, STHs at 25 to 55 percent, Giardiasis at 10 to 25 percent, Amebiasis at 5 to 15 percent and Cryptosporidiosis at 3 to 10 percent. In Asia, Ascariasis ranges from 10 to 30 percent, Trichuriasis from 5 to 15 percent, STHs from 20 to 40 percent, Giardiasis from 8 to 22 percent, Amebiasis from 3 to 10 percent and Cryptosporidiosis from 2 to 8 percent. Oceania shows Ascariasis at 5 to 20 percent, Trichuriasis at 2 to 10 percent, STHs at 10 to 30 percent, Giardiasis at 5 to 15 percent, Amebiasis at 1 to 5 percent and Cryptosporidiosis at 0 to 5 percent. The heatmap includes a legend indicating prevalence percentage from 0 to 40 percent.
Methods
Search strategy and selection criteria
We conducted systematic searches in MEDLINE (via PubMed), Scopus and IEEE Xplore electronic databases from database inception to 01 February 2026. We derived search criteria from preliminary scoping searches and structured them to combine terms related to 2 core concepts: (1) AI or machine learning – including but not limited to CNNs, DL or traditional machine learning – and (2) target parasites of interest – human gastrointestinal parasites. We adapted the search strategy for each database while maintaining the same core conceptual structure and search terms, with syntax modifications to comply with each platform’s requirements. Two independent reviewers screened all titles and abstracts, followed by full-text review. A third reviewer resolved any disagreements through discussion until the team reached consensus. We extracted data using a standardized form in Covidence, with 2 reviewers independently extracting data from each included study. We registered this systematic review with PROSPERO (CRD420261328256) and followed PRISMA 2020 reporting guidelines.
Inclusion criteria: Articles were eligible if they (1) evaluated any AI or machine learning algorithm for the detection, classification or quantification of gastrointestinal parasites that naturally infect humans as definitive or intermediate hosts; (2) used microscopy images of human clinical samples; (3) reported quantitative performance metrics (e.g. sensitivity, specificity, accuracy or area under the curve) and (4) were original peer-reviewed research articles published in English.
Exclusion criteria: We excluded articles if they (1) reported only qualitative findings without quantitative performance metrics; (2) were not primary literature (e.g. reviews, editorials, conference abstracts, protocols without data); (3) used AI algorithms for purposes other than direct patient-sample analysis (e.g. image enhancement only, without diagnostic classification); (4) focused exclusively on parasites that do not naturally infect humans (e.g. plant parasites, parasites of veterinary importance only or experimental animal models without human relevance); (5) focused on haemoparasites (Plasmodium, Babesia, Trypanosoma, Leishmania); (6) focused on tissue parasites (Echinococcus, Schistosoma, Fasciola, Toxoplasma); (7) used non-microscopy imaging modalities (ultrasound, computed tomography, Raman spectroscopy without microscopy) or (8) were environmental or water quality studies without direct human sample analysis.
Data extraction and analysis
From each included study, 2 independent reviewers extracted: (1) study characteristics (first author, publication year, corresponding author country, journal, target parasite species and reference standard used as ground truth); (2) dataset characteristics (total number of unique patients or samples, total number of images, number of positive and negative samples and class balance); (3) technology and imaging specifications (microscope type, magnification, staining method and image resolution); (4) AI model details (algorithm architecture, DL framework and primary diagnostic task – detection, classification, quantification or segmentation); (5) validation strategy (internal validation defined as random split or cross-validation of single-institution data; external validation defined as testing on data from a different laboratory, scanner, geographic region or time period) and (6) quantitative performance metrics (sensitivity, specificity, accuracy, positive predictive value, negative predictive value and area under the receiver operating characteristic curve (AUROC), with 95% confidence intervals where available). When studies reported multiple algorithms, we extracted results from the best-performing algorithm as reported by the authors as reflecting the state-of-the-art per study. This approach may overestimate performance; results should be interpreted cautiously. When studies reported results across multiple parasite species, we treated each species as a separate data entry to enable species-specific synthesis.
Risk-of-bias assessment
Two reviewers independently assessed risk of bias using the QUADAS-2 tool, adapted for AI-based diagnostic studies. The tool evaluates 4 domains: patient selection, index test, reference standard and flow and timing. We rated each domain as low, high or unclear risk of bias and assessed applicability concerns for the first 3 domains. A third reviewer resolved disagreements.
Data synthesis
Due to substantial heterogeneity such as clinical (parasite species, sample preparation), methodological (CNN architectures, training protocols, data split strategies) and statistical (pooling not feasible, narrative), we performed a narrative synthesis rather than a meta-analysis, consistent with our prespecified synthesis plan. For this review, we extracted and synthesized studies reporting 3 key performance metrics: the Cohen’s kappa coefficient (κ), which measures inter-rater agreement between model predictions and human experts while accounting for chance; the mean average precision (mAP), a standard metric in object detection that summarizes the precision-recall trade-off across different classes and intersection-over-union thresholds and the AUROC, which provides a threshold-independent measure of a model’s ability to discriminate between positive and negative cases. We compared helminth and protozoan studies across multiple dimensions: sample size, validation strategy, algorithm architecture and diagnostic performance. For continuous variables, we used Mann–Whitney U tests; for categorical variables, we used Fisher’s exact tests. We considered P values < 0.05 to be statistically significant. We applied the GRADE approach to assess the certainty of evidence for each parasite class.
Results
Study selection
The systematic search yielded 4323 records across all databases (PubMed: 560; Scopus: 3132; IEEE Xplore: 631). After removing duplicates, 3082 unique records underwent title and abstract screening, of which we excluded 2305 as clearly irrelevant to the research question. We assessed 777 full-text articles for eligibility, excluding 700. Primary reasons for exclusion included: focus on haemoparasites (n = 142), tissue parasites (n = 156), veterinary or environmental applications without human samples (n = 184), non-microscopy imaging modalities (n = 98), review articles or protocols without original data (n = 76) and non-English publications (n = 44). Ultimately, 77 studies met all inclusion criteria and were incorporated into this systematic review. Figure 2 presents the complete study selection process in a PRISMA 2020 flow diagram.
PRISMA 2020 flow diagram of systematic literature search. The diagram illustrates the systematic process of study selection following PRISMA guidelines. A total of 4323 records were identified through database searching. After removing 1241 duplicates, 3082 records were screened, of which 2305 were excluded based on title and abstract. Full-text assessment was conducted for 777 articles; 700 were excluded with reasons (haemoparasites: 142; tissue parasites: 156; veterinary/environmental: 184; non-microscopy imaging: 98; reviews/protocols: 76; non-English: 44). Finally, 77 studies met all inclusion criteria and were included in the review.
Figure 2 Long description
The flowchart outlines the systematic literature search process across four stages: identification, screening, eligibility and inclusion. In the identification stage, 4,323 total records were identified through database searching, including MEDLINE, PubMed, Embase, Scopus, Web of Science and Cochrane. After removing 1,241 duplicates, 3,082 records were retained. During the screening stage, 3,082 records were screened based on title and abstract, with 2,305 records excluded. In the eligibility stage, 777 full-text articles were assessed and 700 were excluded for various reasons: 142 for haemoparasites (Plasmodium, Babesia, Trypanosoma, Leishmania), 156 for tissue parasites (Echinococcus, Schistosoma, Fasciola, Toxoplasma), 184 for veterinary/environmental applications (no human samples), 98 for non-microscopy imaging modalities, 76 for reviews or protocols without original data and 44 for non-English publications. Finally, in the inclusion stage, 77 studies met all inclusion criteria and were included in the systematic review, comprising 42 on helminths, 28 on protozoa and 7 on both.
Study characteristics
Table 1 summarizes the characteristics of all 77 included studies. Of these, 42 studies (54.5%) focused on helminth detection, including Ascaris lumbricoides, Trichuris trichiura, hookworm species (Necator americanus and Ancylostoma duodenale) and Enterobius vermicularis. Twenty-eight studies (36.4%) focused on gastrointestinal protozoa, including G. intestinalis, Cryptosporidium spp. (C. hominis and C. parvum) and Entamoeba histolytica. Seven studies (9.1%) evaluated algorithms applicable to both parasite classes. Studies originated from 23 countries across 5 continents, with the majority from China (n = 16), India (n = 12), Brazil (n = 10) and the United States (n = 9). Sample sizes varied widely, ranging from 42 to 4283 images per study (median: 347 images; interquartile range [IQR]: 128–956). Helminth studies demonstrated significantly larger sample sizes (median: 623 images; IQR: 287–1342) compared to protozoan studies (median: 189 images; IQR: 78–412; Mann–Whitney U test, P = 0.008). CNNs dominated the field, appearing in 51 studies (66.2%). You Only Look Once (YOLO) variants represented the most common object detection architecture, employed in 21 studies (27.3%). Traditional machine learning approaches – including support vector machines, random forests (RFs) and k-nearest neighbours – appeared in 15 studies (19.5%). Specialized architectures, including deep cycle transfer learning, fully convolutional regression networks and vision transformers, appeared in 9 studies (11.7%). Note that these categories are not mutually exclusive; YOLO variants are a subset of CNNs, and some studies employed multiple architectures for comparison (Figure 3). Regarding imaging modalities, brightfield microscopy predominated (53 studies, 68.8%), followed by fluorescence microscopy (13 studies, 16.9%), holographic imaging (6 studies, 7.8%) and smartphone-based platforms (5 studies, 6.5%). Staining methods varied considerably across studies. For helminth detection, Kato–Katz represented the most common method (28 studies, 66.7% of helminth studies). For protozoan detection, iodine (12 studies, 42.9% of protozoan studies), trichrome (10 studies, 35.7%) and auramine (6 studies, 21.4%) predominated.
Distribution of included studies by parasite group and AI architecture. The Sankey diagram illustrates the distribution of the 77 included studies across 4 dimensions: parasite group (helminths: 42 studies; protozoa: 28 studies; both: 7 studies), target species (helminths: Ascaris lumbricoides, Trichuris trichiura, hookworm, Enterobius vermicularis; protozoa: Giardia intestinalis, Cryptosporidium, Entamoeba histolytica), imaging modality (brightfield, fluorescence, holographic, smartphone) and AI architecture (CNN/YOLO, traditional ML, specialized). Bar width is proportional to study count.
Figure 3 Long description
The Sankey diagram shows the distribution of studies across four dimensions: parasite group, species target, imaging modality and AI architecture. The parasite group is divided into GI Protozoa with 28 studies (36.4 percent), Both with 7 studies (9.1 percent) and Helminths with 42 studies (54.5 percent). Species targets include Mixed (protozoa), Cryptosporidium spp., Giardia intestinalis, Mixed/both, Entamoeba histolytica, Hookworm spp., Trichuris trichiura and Ascaris lumbricoides. Imaging modalities are Smartphone, Holographic, Fluorescence and Brightfield. AI architectures are categorized as Other, Specialized with 9 studies (11.7 percent), Traditional ML with 15 studies (19.5 percent) and CNN/YOLO with 51 studies (66.2 percent). The diagram visually represents the flow and connection between these categories, with bar widths proportional to study count.
Summary characteristics of included studies
Table 1 Long description
Characteristics are reported for Helminth Studies, Protozoa Studies, Both Classes, and Total, covering geographic distribution, sample size in images, algorithm architecture, imaging modality, and staining method. Geographic distribution totals list China 16, India 12, Brazil 10, United States 9, Thailand 5, Turkey 4, Belgium 3, and Other countries 18. Sample size in images shows median and interquartile range of 623 with 287 to 1,342 for helminths, 189 with 78 to 412 for protozoa, 347 with 156 to 876 for both classes, and 347 with 128 to 956 for total; ranges are 42 to 4,283 for helminths, 42 to 1,200 for protozoa, 98 to 1,847 for both classes, and 42 to 4,283 for total. Algorithm architecture counts and percentages are CNN general 28 at 66.7 percent for helminths, 18 at 64.3 percent for protozoa, 5 at 71.4 percent for both classes, and 51 at 66.2 percent total; YOLO variants 15 at 35.7 percent, 4 at 14.3 percent, 2 at 28.6 percent, and 21 at 27.3 percent total; Traditional ML 8 at 19.0 percent, 6 at 21.4 percent, 1 at 14.3 percent, and 15 at 19.5 percent total; Specialized architectures 5 at 11.9 percent, 3 at 10.7 percent, 1 at 14.3 percent, and 9 at 11.7 percent total. Imaging modality counts and percentages are brightfield microscopy 32 at 76.2 percent for helminths, 16 at 57.1 percent for protozoa, 5 at 71.4 percent for both classes, and 53 at 68.8 percent total; fluorescence microscopy 4 at 9.5 percent, 8 at 28.6 percent, 1 at 14.3 percent, and 13 at 16.9 percent total; holographic imaging 2 at 4.8 percent, 3 at 10.7 percent, 1 at 14.3 percent, and 6 at 7.8 percent total; smartphone based 4 at 9.5 percent, 1 at 3.6 percent, 0 at 0 percent, and 5 at 6.5 percent total. Staining method counts and percentages are Kato Katz 28 at 66.7 percent for helminths, 0 at 0 percent for protozoa, 2 at 28.6 percent for both classes, and 30 at 39.0 percent total; iodine 8 at 19.0 percent, 12 at 42.9 percent, 2 at 28.6 percent, and 22 at 28.6 percent total; trichrome 2 at 4.8 percent, 10 at 35.7 percent, 1 at 14.3 percent, and 13 at 16.9 percent total; auramine 0 at 0 percent, 6 at 21.4 percent, 0 at 0 percent, and 6 at 7.8 percent total; other or unspecified 4 at 9.5 percent, 0 at 0 percent, 2 at 28.6 percent, and 6 at 7.8 percent total.
a Categories are not mutually exclusive; percentages may sum to >100%. YOLO variants are a subset of CNNs, and specialized architectures may also employ CNN backbones.
Risk of bias in included studies
Table 2 presents the risk-of-bias assessments using the adapted QUADAS-2 tool. Overall, 51 studies (66.2%) demonstrated high or unclear risk of bias in at least 1 domain. The patient selection domain showed the highest concern, with 31 studies (40.3%) lacking consecutive or random sampling and instead using convenience samples or case-control designs. The index test domain revealed high or unclear risk in 48 studies (62.3%), primarily due to the absence of prespecified thresholds (35 studies, 45.5%) and lack of external validation (55 studies, 71.4%). Only 22 studies (28.6%) performed any form of external validation using independent test sets from different laboratories, scanners or geographic regions. The reference standard domain demonstrated the lowest risk of bias, with 52 studies (67.5%) using appropriate microscopy reference standards. However, only 24 studies (31.2%) reported explicit blinding of microscopists to AI results, raising concerns about incorporation bias. The flow and timing domain showed high or unclear risk in 39 studies (50.6%), primarily due to exclusion of samples with poor image quality or low parasite burden after data collection. Applicability concerns remained generally low across studies, as most used clinically relevant samples and standard diagnostic workflows. Protozoan studies demonstrated significantly higher risk of bias compared to helminth studies: 78.6% of protozoan studies (22/28) showed high or unclear risk in the index test domain vs 47.6% of helminth studies (20/42) (Fisher’s exact test, P = 0.001).
Risk of bias and applicability concerns summary
Table 2 Long description
Risk of bias is reported as low, high, or unclear for four domains, and applicability concerns are reported as low, high, or unclear for three domains. For patient selection, risk of bias was low in 32 studies, 41.6 percent, high in 31, 40.3 percent, and unclear in 14, 18.2 percent; applicability concerns were low in 62, 80.5 percent, high in 8, 10.4 percent, and unclear in 7, 9.1 percent. For index test, risk of bias was low in 29 studies, 37.7 percent, high in 28, 36.4 percent, and unclear in 20, 26.0 percent; applicability concerns were low in 66, 85.7 percent, high in 5, 6.5 percent, and unclear in 6, 7.8 percent. For reference standard, risk of bias was low in 52 studies, 67.5 percent, high in 12, 15.6 percent, and unclear in 13, 16.9 percent; applicability concerns were low in 64, 83.1 percent, high in 6, 7.8 percent, and unclear in 7, 9.1 percent. For flow and timing, risk of bias was low in 38 studies, 49.4 percent, high in 22, 28.6 percent, and unclear in 17, 22.1 percent, and applicability concerns are not reported for this domain.
Values represent the number of studies (N = 77) with percentages in parentheses. Applicability concerns are not assessed for the flow and timing domain.
Diagnostic performance: helminth detection
Helminth detection studies consistently demonstrated high diagnostic performance, with sensitivities ranging from 85% to 98% and specificities from 88% to 99% across externally validated studies (Table 3). Mohammed et al. (Reference Mohammed, Anzaku, Ward, Levecke, Krishnamoorthy, De Neve and Van Hoecke2025) reported mAP of 0.92 for simultaneous detection of 4 helminth species using YOLOv4 trained on 2847 Kato–Katz images (Mohammed et al., Reference Mohammed, Anzaku, Ward, Levecke, Krishnamoorthy, De Neve and Van Hoecke2025). Mathison et al. (Reference Mathison, Knight, Potts, Black, Walker, Markow, Wood, Bess, Dixon, Cahoon, Hymas and Couturier2025) provided perhaps the most compelling evidence for AI superiority. Analysing over 4000 unique clinical specimens, their CNN detected 50% more positive samples than experienced microscopists examining the same slides. This advantage proved most evident in low-burden infections (fewer than 5 eggs per slide), where AI achieved 96% sensitivity compared to 64% for human experts (P < 0.001) (Mathison et al., Reference Mathison, Knight, Potts, Black, Walker, Markow, Wood, Bess, Dixon, Cahoon, Hymas and Couturier2025). AI may complement microscopy in specific scenarios, but prospective clinical validation is required before any challenge to current gold standards can be substantiated. Corpuz et al. (Reference Corpuz, Kusolsuk, Wongphan, Chonsawat, Naing, Boonsang, Kittichai, Fan, Chuwongin and Watthanakulpanich2025) extended these findings to out-of-distribution evaluation, demonstrating that DINOv2-large achieved 98.93% accuracy for intestinal parasite identification, with strong agreement compared to medical technologists (κ > 0.90) (Corpuz et al., Reference Corpuz, Kusolsuk, Wongphan, Chonsawat, Naing, Boonsang, Kittichai, Fan, Chuwongin and Watthanakulpanich2025). Kumar et al. (Reference Kumar, Arif, Ahamad, Chaudhary, Khan and Ali2023) reported mAP of approximately 97% using YOLOv5 algorithms with detection times of only 8.5 ms per sample – a critical metric for clinical workflow integration (Kumar et al., Reference Kumar, Arif, Ahamad, Chaudhary, Khan and Ali2023). Multiple architectural approaches proved effective. Butploy et al. (Reference Butploy, Kanarkard, Intapan and Magalhães2021) achieved 93.33% classification accuracy for A. lumbricoides eggs using standard DL architectures (Butploy et al., Reference Butploy, Kanarkard, Intapan and Magalhães2021). AlDahoul et al. (Reference AlDahoul, Karim, Momo, Escobar, Magallanes and Tan2023) applied CoAtNet to the Chula-ParasiteEgg dataset, achieving 93% average accuracy (AlDahoul et al., Reference AlDahoul, Karim, Momo, Escobar, Magallanes and Tan2023). Chaibutr et al. (Reference Chaibutr, Pongpanitanont, Laymanivong, Thanchomnang and Janwan2024) developed a CNN for Enterobius vermicularis egg detection, achieving 90.0% accuracy after data augmentation, with the Xception architecture reaching 99.0% accuracy (Chaibutr et al., Reference Chaibutr, Pongpanitanont, Laymanivong, Thanchomnang and Janwan2024). Hassan et al. (Reference Hassan, Alqahtani, Elbedwehy and Talaat2025) applied YOLO with convolutional block attention modules, further improving detection of small or overlapping eggs (Hassan et al., Reference Hassan, Alqahtani, Elbedwehy and Talaat2025). Mirzaei et al. (Reference Mirzaei, Ilhan, Guler, Suer and Sekeroglu2025) compared multiple state-of-the-art architectures for Ascaris and Taenia classification, achieving F1-scores of 98.6% (ConvNeXt Tiny), 97.5% (EfficientNet V2 S) and 98.2% (MobileNet V3 S) (Mirzaei et al., Reference Mirzaei, Ilhan, Guler, Suer and Sekeroglu2025).
Diagnostic performance of selected studies
Table 3 Long description
Rows list studies by year with parasite class, target species, algorithm, sample size, validation type, and reported sensitivity, specificity, accuracy, and AUC ROC or other metrics. He et al. 2024 used YOLOv4 on 2,847 images with external validation and reported mean average precision 0.92. Kumar et al. 2023 used YOLOv5 on 5,393 images with internal validation and reported mean average precision about 97%. Corpuz et al. 2025 used DINOv2 large with external validation and reported sensitivity 78.0%, specificity 99.6%, accuracy 98.93%, and AUROC 0.97, with sample size not given. Al Khlifeh et al. 2025 used random forest on clinical data with internal validation and reported AUC 0.98, with sensitivity, specificity, and accuracy not reported. AlDahoul et al. 2023 used CoAtNet on 11,000 images with internal validation and reported accuracy 93% and F1 93%. Mirzaei et al. 2025 used ConvNeXt Tiny with internal validation and reported F1 98.6%, with sample size and other metrics not reported. External validation rows include Boonyong et al. 2024 with FA280 AI on 1,000 samples and Cohens kappa 0.367; Wang et al. 2024 with KU F40 on 1,030 samples with sensitivity 71.2%, specificity 94.7%, and Cohens kappa 0.633; Aggarwal et al. 2026 with MISPA F30 on 275 samples with sensitivity 66.7%, specificity 86.2%, and agreement 83.3%; Yang et al. 2025 with an auto analyser with sensitivity 84.3% and specificity 98.7% for AI plus user audit 94.1% and 99.7%, with sample size not given; Aupaix et al. 2026 with KU F40 on 267 samples with sensitivity 86% and specificity 45%. Additional internal validation rows include Yarahmadi et al. 2025 on Giardia with EfficientNet B0 using 1,610 images with sensitivity 96.2%, specificity 98.2%, accuracy 96.3%, and precision 0.96; Salguero et al. 2023 on Cryptosporidium with YOLOv5x using 520 plus 1,200 images with the other metric listed as, Deteccion automatica; Butploy et al. 2021 on Ascaris lumbricoides with deep learning reporting accuracy 93.33% with sample size not given; Koydemir et al. 2015 on Giardia with DenseNet using 1,200 images reporting accuracy 94%; Gorocs et al. 2020 on Giardia and Cryptosporidium with holography plus deep learning reporting accuracy 94% with sample size not given; Mathison et al. 2025 on mixed helminths with CNN using more than 4,000 specimens with external validation reporting sensitivity 96%.
mAP, mean average precision; AUC, area under the curve; AUROC, area under the receiver operating characteristic curve; κ, Cohen’s kappa.
a Direct cross-study comparability is limited by metric heterogeneity.
Diagnostic performance: gastrointestinal protozoa
Protozoan detection studies painted a more complex picture. Studies relying exclusively on internal validation frequently reported accuracies exceeding 90%. Koydemir et al. (Reference Koydemir, Gorocs, Tseng, Cortazar, Feng, Chan, Burbano, McLeod and Ozcan2015) achieved 94% accuracy for Giardia cyst detection using a smartphone-based fluorescent microscope with a DenseNet classifier on 1200 images (Koydemir et al., Reference Koydemir, Gorocs, Tseng, Cortazar, Feng, Chan, Burbano, McLeod and Ozcan2015). Gorocs et al. (Reference Gorocs, Baum, Song, de Haan, Ceylan Koydemir, Qiu, Cai, Skandakumar, Peterman, Tamamitsu and Ozcan2020) eliminated the need for fluorescent labelling entirely, using lens-free holographic imaging to identify unstained Giardia and Cryptosporidium cysts with 94% accuracy on internal validation (Gorocs et al., Reference Gorocs, Baum, Song, de Haan, Ceylan Koydemir, Qiu, Cai, Skandakumar, Peterman, Tamamitsu and Ozcan2020). Yarahmadi et al. (Reference Yarahmadi, Ahmadpour, Moradi and Samadzadehaghdam2025) applied EfficientNet-B0 to classify G. lamblia cysts and trophozoites, achieving precision of 0.96 and accuracy of 0.96 (Yarahmadi et al., Reference Yarahmadi, Ahmadpour, Moradi and Samadzadehaghdam2025). Salguero et al. (Reference Salguero, Rendón, Valencia, Gil, Galvis, Londoño, Calderón, Osorio and Soto2023) applied YOLOv5x for automatic detection of Cryptosporidium in optical microscopy images (Salguero et al., Reference Salguero, Rendón, Valencia, Gil, Galvis, Londoño, Calderón, Osorio and Soto2023). However, studies attempting external validation or using independent test sets revealed substantial performance degradation. Aggarwal et al. (Reference Aggarwal, Choudhary, Pallavi, Saxena and Dey2026) evaluated the MISPA-F30 fecal analyser in a tertiary care hospital setting, reporting 66.7% sensitivity and 86.2% specificity against a composite reference standard, with 83.3% agreement with microscopy (Aggarwal et al., Reference Aggarwal, Choudhary, Pallavi, Saxena and Dey2026). Aupaix et al. (Reference Aupaix, Filippin, Jaumot, Cannoot, Chemais, Martiny, Miendje Deyi, Deffontaine, Deckers, Verbelen, Potters, Drieghe, Mzougui, Soleimani, Philippart and Brauner2026) assessed the KU-F40 automated microscope across 6 Belgian laboratories, finding overall sensitivity of 86% and specificity of 45% for intestinal parasite detection – though sensitivity increased to 95% when considering only clinically relevant parasites (Aupaix et al., Reference Aupaix, Filippin, Jaumot, Cannoot, Chemais, Martiny, Miendje Deyi, Deffontaine, Deckers, Verbelen, Potters, Drieghe, Mzougui, Soleimani, Philippart and Brauner2026). Boonyong et al. (Reference Boonyong, Hunnangkul, Vijit, Wattano, Tantayapirak, Loymek and Wongkamchai2024) compared the FA280 automatic fecal analyser with formalin–ethyl acetate concentration technique. They demonstrated perfect agreement for protozoa identification with user audit (κ = 1.00) and strong agreement for helminths (κ = 0.857), but AI-alone reports showed only fair agreement (κ = 0.367) (Boonyong et al., Reference Boonyong, Hunnangkul, Vijit, Wattano, Tantayapirak, Loymek and Wongkamchai2024). Wang et al. (Reference Wang, Liao, Huang, Tang and Lin2024) evaluated the KU-F40 in routine clinical settings, reporting 71.2% sensitivity and 94.7% specificity with a kappa value of 0.633 (Wang et al., Reference Wang, Liao, Huang, Tang and Lin2024). Yang et al. (Reference Yang, Deng, Wang, Wang, Gong, Jia, ZQ, RX, GH and YH2025) compared automatic fecal analyser performance with direct wet smear microscopy, demonstrating sensitivity of 84.3% for AI-alone reports and 94.1% for user-audited results, with specificity of 98.7% and 99.7%, respectively (Yang et al., Reference Yang, Deng, Wang, Wang, Gong, Jia, ZQ, RX, GH and YH2025). For Entamoeba histolytica, Al-khlifeh et al. (Reference Al-khlifeh, Tarawneh, Almohammadi, Alrashidi, Hassanat and Hassanat2025) circumvented image-based limitations by shifting to clinical feature mining. Using electronic medical records and routine laboratory data – including neutrophil percentages, fecal leukocytes and mucus presence – an optimized RF model achieved an area under the curve of 0.98 in distinguishing amebiasis from other gastroenteritis causes. Although this approach falls outside direct microscopy analysis, it demonstrates the potential of multimodal diagnostic integration (Al-khlifeh et al., Reference Al-khlifeh, Tarawneh, Almohammadi, Alrashidi, Hassanat and Hassanat2025).
Comparative analysis: helminths vs protozoa
Statistical comparisons revealed systematic differences between helminth and protozoan studies. Helminth studies demonstrated significantly larger sample sizes (median 623 vs 189 images; Mann–Whitney U test, P = 0.008). They also showed lower risk of bias in the index test domain: 52.4% of helminth studies (22/42) were rated low risk vs 21.4% of protozoan studies (6/28) (Fisher’s exact test, P = 0.001). Protozoan studies were significantly more likely to use single-institution datasets without external validation: 82.1% (23/28) compared to 61.9% (26/42) for helminth studies (Fisher’s exact test, P = 0.012). Across all studies and validation types, the pooled range of diagnostic accuracy (sensitivity/specificity pairs) was 85–98% for helminths vs 70–94% for protozoa. When restricting to studies with external validation, the range narrowed to 85–95% for helminths but widened substantially to 55–85% for protozoa, confirming the hypothesized performance gap. For studies reporting AUROC, helminth studies achieved a median of 0.97 (IQR: 0.94–0.99) compared to 0.86 (IQR: 0.79–0.93) for protozoan studies. Algorithm architecture alone did not explain performance differences; YOLO variants achieved comparable performance to other CNN families when sample sizes were adequate. Rather, sample size (Spearman’s ρ = 0.62, P = 0.001) and dataset diversity (number of distinct sources/institutions; ρ = 0.71, P < 0.001) emerged as the strongest predictors of generalizable performance. Geographic validation proved particularly critical: studies validating only within the same institution showed 12–28 percentage point overestimates of accuracy compared to those with multicentre or multi-country validation. Subgroup analyses revealed that studies using standardized Kato–Katz thick smears for helminths showed significantly less performance variability (coefficient of variation for sensitivity: 6.2%) than those using diverse staining protocols for protozoa (coefficient of variation: 18.7%).
Certainty of evidence
Applying the GRADE approach for diagnostic studies, we graded the certainty of evidence for AI-based helminth detection as MODERATE. We downgraded 1 level due to inconsistency in outcome definitions and threshold selection across studies, but the consistent direction of effect and generally adequate sample sizes supported moderate confidence. For gastrointestinal protozoa, we graded the certainty of evidence as LOW, downgrading 2 levels due to serious risk of bias (predominantly lack of external validation and non-representative patient selection) and imprecision (small sample sizes and wide confidence intervals around accuracy estimates). For specific protozoan species, we graded the evidence as VERY LOW for Entamoeba histolytica differentiation from Entamoeba dispar, reflecting the near-absence of validated image-based AI studies for this critical diagnostic challenge. Consequently, we do not anticipate that reliable differentiation can be achieved using current AI-based detection methods without significant advances in data availability, annotation quality and standardization of laboratory procedures.
Discussion
Our systematic review confirms a performance gap between AI-based detection of helminths and gastrointestinal protozoa, a disparity that does not appear to stem primarily from algorithmic architecture but rather from 3 interconnected factors: object size and structural complexity, data poverty and representational bias, and imaging heterogeneity. Helminth eggs possess features that align well with CNN inductive biases – they are large (45–75 µm), morphologically conserved and well-contrasted – whereas protozoa such as Cryptosporidium (4–6 µm) and Giardia (8–12 µm) approach the diffraction limit of light microscopy, occupy few pixels and are easily confused with background elements such as yeasts, lipid droplets and stain precipitates (Centers for Disease Control and Prevention, 2019). Beyond these morphological challenges, practical diagnostic difficulties deserve greater attention than they have received in the AI literature. Intermittent shedding of cysts and oocysts in stool means that parasite burden varies considerably over time, and a single sample may not capture infection status – a challenge that AI algorithms cannot address without repeated sampling or integration with clinical metadata. Preservation artefacts – including shrinkage, distortion and loss of diagnostic features – are more pronounced for protozoa than for helminth eggs, particularly when samples are transported from field settings to reference laboratories, yet very few studies have evaluated AI performance on suboptimally preserved specimens. Variable parasite burden complicates algorithm training, as models must perform reliably across a wide dynamic range, from low-burden infections (few organisms per slide) to high-burden samples where overlapping and clumping occur. Finally, microscopy variability – differences in illumination, focus, slide preparation and observer interpretation – affects protozoan identification more severely than helminth detection, given the smaller size and subtler morphological features of protozoa. These practical challenges, which are underdeveloped in the current AI literature, represent critical barriers that cannot be resolved through algorithmic advances alone and highlight a significant imbalance in the literature: while the AI and machine learning aspects of the included studies have been discussed extensively, the parasitological foundation – including the biological, clinical and operational complexities of protozoan diagnosis – has received comparatively less attention. Public repositories of protozoan images are orders of magnitude smaller than those for helminths, and the absence of large, multicentric, annotated datasets perpetuates a feedback loop in which poor generalization discourages further research and development (Ibrahim et al., Reference Ibrahim, Liu, Zariffa, Morris and Denniston2021; Zhang et al., Reference Zhang, Jiang, Jiang, Xi, Chen, Liu, Juhas, Li and Zhang2022). Moreover, unlike the globally standardized Kato–Katz protocol for STHs, no equivalent standard exists for protozoan stool examination; variations in concentration methods, staining techniques and magnification produce domain shifts that current adaptation methods cannot fully reconcile (Utzinger et al., Reference Utzinger, Botero-kleiven, Castelli, Chiodini, Edwards, Köhler, Gulletta, Lebbad, Manser, Matthys, N’goran, Tannich, Vounatsou and Marti2010). Helminth detection has shown promising results in research settings – with YOLO-based architectures achieving high accuracy on internal validation – but even these findings rest on limited external validation and single-institution datasets, and the claim that AI detects more positive samples than expert microscopists (Mathison et al., Reference Mathison, Knight, Potts, Black, Walker, Markow, Wood, Bess, Dixon, Cahoon, Hymas and Couturier2025) requires prospective confirmation before any definitive conclusions can be drawn (Kumar et al., Reference Kumar, Arif, Ahamad, Chaudhary, Khan and Ali2023).
At present, AI should be viewed as a research tool with potential future applications, but prospective clinical validation is required before its role in routine diagnosis can be determined. Several structural factors perpetuate the protozoan gap: funding for AI diagnostics has disproportionately favoured malaria, tuberculosis and high-profile elimination campaigns, while giardiasis and cryptosporidiosis – despite substantial morbidity – receive comparatively limited resources (Collaborators, 2025). Real-world adoption faces substantial unresolved implementation barriers, including the need for GPU (Graphics Processing Unit)-equipped workstations and reliable power in endemic regions, and only 3.9% of included studies provided economic data, making cost-effectiveness impossible to assess. AI systems that perform well under controlled research conditions often encounter difficulties when applied to routine laboratory workflows due to user interface design, laboratory information system integration and trust calibration; for protozoan diagnostics, the consequences of false positives vs false negatives differ substantially, making these human factors particularly consequential.
Ethical and regulatory considerations are also underexplored: none of the 77 included studies reported explicit anonymization procedures compliant with HIPAA (Health Insurance Portability and Accountability Act) or GDPR (General Data Protection Regulation), and federated learning – a promising privacy-preserving alternative – has not been applied to gastrointestinal parasitology (European Parliament and Council, 2016; Kaissis et al., Reference Kaissis, Makowski, Rückert and Braren2020; Rieke et al., Reference Rieke, Hancox, Li, Milletarì, Roth, Albarqouni, Bakas, Galtier, Landman, Maier-hein, Ourselin, Sheller, Summers, Trask, Xu, Baust and Cardoso2020).
Algorithmic bias is a further concern, as most image repositories originate from high-income countries and may underrepresent parasite strains and microscopy conditions prevalent in low- and middle-income regions (Ibrahim et al., Reference Ibrahim, Liu, Zariffa, Morris and Denniston2021); we recommend that AI-based systems undergo equity audits prior to any consideration of deployment (Rajkomar et al., Reference Rajkomar, Hardt, Howell, Corrado and Chin2018; Wiens et al., Reference Wiens, Saria, Sendak, Ghassemi and Liu2019). Regulatory pathways – including FDA (Food and Drug Administration) SaMD (Software as a Medical Device) classification and EU (European Union) IVDR (In Vitro Diagnostic Medical Devices Regulation) conformity assessment – remain unaddressed in the primary literature, and no included study reported any form of regulatory approval, underscoring the substantial gap between research and any potential clinical use (European Parliament and Council, 2017; FDA, 2019).
Addressing the protozoan gap will require sustained efforts across multiple fronts, including the development of standardized imaging protocols, the creation of open and demographically diverse image repositories such as the ParasitoBank (Muñoz Galindez et al., Reference Muñoz Galindez, Vásquez Arteaga and Vargas Cañas2025), the adaptation of advanced machine learning techniques, early engagement with regulatory science and a design philosophy that aims to augment rather than replace microscopists. However, until these foundational resources and validation studies are in place, the clinical readiness of AI for gastrointestinal parasite diagnosis – particularly for protozoa – remains low, and considerable work lies ahead before any transition from research to practice can be realistically contemplated.
Finally, this systematic review has several limitations. First, despite comprehensive searches across 4 databases, we may have missed relevant studies published in non-indexed journals or languages other than English. Second, the quality of included studies varied considerably, and our conclusions are necessarily constrained by the methodological weaknesses of the primary literature. Third, the rapid evolution of AI methodologies means that studies published during the review process may not be captured; we mitigated this by including preprints and early online publications where available. Fourth, we were unable to perform meta-analysis due to substantial heterogeneity in outcome definitions, reference standards and reporting formats – a limitation we prespecified but one that limits quantitative synthesis.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0031182026102522.
Acknowledgements
The authors acknowledge the Universidad de Antofagasta and Universidad de Concepción for institutional support. The authors used OpenAI ChatGPT (GPT-4-turbo, model gpt-4-0613; accessed via https://chat.openai.com) on 1–11 February 2026 solely for language polishing, grammar correction, and reference formatting. No proprietary or unpublished data were uploaded, and no modifications were made to the tool. All AI-generated text was reviewed and approved by the authors, who take full responsibility for the content.
Author contributions
Conceptualization: J.G. and J.L.V. S.Z., K.M., J.S.F., I.A., C.G., B.G., J.G. and J.L.V. wrote the article.
Financial support
This work was partially funded by ANID-FONDECYT 1261166 (to J.L.V.) and by the Faculty of Biological Sciences of the Universidad de Concepción (to J.L.V.).
Competing interests
The authors declare that there are no conflicts of interest.
Ethical standards
Not applicable.
Open access
