Artificial Intelligence for Finding Parasites in Stool

This page looks at one point from a 2026 integrative-oncology presentation: that a deep-learning computer model published in October 2025 finds parasites in stool samples better than the traditional microscope examination. It sets that point beside the published research — the October 2025 study itself, the earlier work it grew from, validations in other laboratories (including one with disappointing results), a before-and-after count from a laboratory that adopted the method, and field studies in places where parasitic worms are common. The research is about detecting parasites; none of it measures cancer.

What the Presentation Said
How Parasites Are Found in Stool Today
What a Deep-Learning Image Model Does
The October 2025 Wet-Mount Study
Earlier Work: Stained Slides and Portable Scanners
Validations in Other Laboratories
After Adoption: A Before-and-After Count
How to Read the Accuracy Figures
Beyond Stool: Malaria Blood Smears and Primary Care
How the Claim Compares With the Research
What Is Not Known
Connections
Key Research Papers
Featured Videos

What the Presentation Said

The presentation came to artificial intelligence at the end of a section on testing. Before it, the speaker described standard stool tests as having low sensitivity, said parasites release eggs and larvae only intermittently so a single test can come back falsely negative, said antibody blood tests cannot tell a past infection from a current one, and described molecular (polymerase chain reaction (PCR)) tests as more specific but costly and available only in specialised laboratories. The speaker also said that less than 10% of parasite cases are detected by standard stool tests — a figure examined on the sibling page Stool Tests for Parasites and Their Limits.

The speaker then described a paper published in October 2025 that, in the speaker's words, states that AI outperforms traditional methods for detecting parasites in stool samples. According to the presentation:

researchers developed a deep-learning model that could transform how intestinal infections are diagnosed;
the method increases the speed, accuracy and accessibility of parasite detection, with potential to transform diagnostics worldwide — a statement the speaker called huge and read out twice;
this is the direction medicine is moving, toward tests that are more sensitive and more specific.

The speaker linked better detection to the presentation's larger theme, that parasitic infection promotes inflammation and contributes to cancer. Answering an audience question later, the speaker returned to the slide, described AI as a likely quantum leap in the ability to diagnose parasites accurately, and expressed the hope that many laboratories would soon have the ability.

The presentation did not name the paper's authors, journal or laboratory. The study that matches the description — a deep convolutional neural network for parasites in stool wet mounts, published online on 21 October 2025 — is described below. It comes from a large US reference laboratory (a laboratory that tests specimens sent in by many hospitals and clinics) working with the company that makes the software.

How Parasites Are Found in Stool Today

The classic stool parasite test is the ova and parasite examination (often shortened to "O&P"; ova means eggs). A trained laboratory scientist looks through a microscope at prepared stool for the eggs and larvae of worms and the cysts and moving forms of single-celled parasites (protozoa). Two preparations are common in clinical laboratories:

Concentrated wet mount — stool is processed to concentrate parasites and a drop is examined unstained or with iodine. Worm eggs and larvae are mainly found this way.
Permanent stained smear — a thin smear is fixed and stained (the trichrome stain is widely used) to show the internal detail of protozoa.

In tropical control programmes the usual method is the Kato-Katz thick smear, a measured amount of stool pressed under cellophane so eggs can be counted per gram.

The authors of the October 2025 study describe stool microscopy as manual, labour-intensive work requiring highly trained staff, performed on millions of specimens a year and largely unchanged in 100 years apart from minor improvements in processing and fixation (Mathison and colleagues, 2025). A 2026 laboratory report adds that in low-prevalence settings such as the United States, the examination is "prone to error" (Anderson and colleagues, 2026). Microscopists also disagree with one another: when two reference microscopists in the Philippines labelled images of worm eggs from Kato-Katz slides for an AI training set, they agreed on 89.24% of images (a statistical agreement score, kappa, of 0.86); the eggs of Schistosoma japonicum were the hardest to agree on, and a third expert was needed to resolve most disagreements (Chua and colleagues, 2025).

These limits are the reason researchers have tried to automate the reading of slides.

What a Deep-Learning Image Model Does

The systems in these studies work in two steps. First, a slide scanner photographs the whole microscope slide at high magnification and stores it as a digital image. Second, a convolutional neural network — a type of computer program that learns to recognise patterns in pictures by being shown many labelled examples — searches the image for objects that look like parasites and labels each one with its most likely identity.

Some terms used in the studies:

Training set — the labelled images the model learns from.
Holdout or validation set — specimens kept aside and never shown to the model during training, used to test it.
Presumptive classification — the model's suggested identity, which a human then confirms or rejects. In the clinical studies below, the computer screens and a trained scientist reviews what it flags; the computer does not issue the final result on its own.
Limit of detection — how dilute a sample can be (how few parasites it can contain) before a method stops finding them.

A model can only recognise what it was trained on. A parasite that is absent from its training set, or a preparation it has not seen, is a known weak point, which is why several of the papers below stress validation in each laboratory that adopts it.

The October 2025 Wet-Mount Study

The paper is "Detection of protozoan and helminth parasites in concentrated wet mounts of stool using a deep convolutional neural network", published in the Journal of Clinical Microbiology (online 21 October 2025; issue of 12 November 2025). Its authors were from a large US reference laboratory and its associated university pathology department, and from the company that makes the software (Mathison and colleagues, 2025). The authors describe it as the first comprehensive wet-mount AI model development and validation.

How it was built

The model was trained on 27 different parasites, using 4,049 unique parasite-positive specimens collected in the USA, Europe, Africa and Asia.
Whether a training specimen was positive was determined by traditional microscopy — the method the model was later compared with.
It was validated on a separate holdout set, with a diversity of fixatives and preparation techniques.

The accuracy figures, exactly as published

Before discrepant results were resolved: the AI correctly detected 250 of 265 positive specimens (94.3% agreement) and 94 of 100 negative specimens (94.0%). In other words, it missed 15 specimens that microscopy had called positive and flagged 6 that microscopy had called negative.
The AI also detected 169 additional organisms in the validation specimens that had not been identified before.
These additional detections were then adjudicated by reviewing the scans and by microscopy. After resolution, counting the newly confirmed organisms as true positives and the rejected ones as false positives, positive agreement was 472 of 477 (98.6%).
Negative agreement varied by organism, from 91.8% to 100%.

The comparison with people

In a limit-of-detection study, the authors compared the AI with three laboratory scientists of varying experience, using serial dilutions of specimens containing Entamoeba (an amoeba), Ascaris (the giant roundworm), Trichuris (whipworm) and hookworm. The AI "consistently detected more organisms and at lower dilutions of parasites than humans, regardless of the technologist's experience."

What the authors concluded

The abstract states that AI for wet-mount analysis is highly sensitive, "detects considerably more organisms than traditional microscopy alone", simplifies the parasitology workflow and reduces the reliance on traditional microscopy. It describes the work as bringing parasite screening "into the technological age".

Disclosures and commentary

Several co-authors listed an affiliation with the software company, and the conflict-of-interest statement reports that two authors received speaker compensation from it.
The journal published an accompanying commentary by two pathologists from another US academic medical centre, who described accuracy and analytical sensitivity exceeding that of a cohort of highly trained laboratory scientists, credited the large, globally sourced training set, called the result a robust proof of concept, and discussed how well the technology would transfer to routine laboratories (Kirby and Arnaout, 2026).

The study measured how well a computer reads slides compared with people reading the same kind of slides. It did not test patients' outcomes, and it did not measure or discuss cancer.

Earlier Work: Stained Slides and Portable Scanners

Trichrome-stained protozoa, 2020

The same laboratory group first built a model for permanent trichrome-stained slides, aimed at screening out negative slides and flagging possible parasites for a human to confirm. Clinical validation used 10 unique positive slides per class of protozoa and 125 negative slides, with accuracy measured as slide-level agreement with microscopy. Positive agreement was 98.88% (95% confidence interval 93.76% to 99.98%) and negative agreement 98.11% (93.35% to 99.77%). With serially diluted stool, the model and scanner were 5-fold more sensitive than manual examination by multiple parasitologists (Mathison and colleagues, 2020). The 2025 wet-mount study is described by its commentators as Part II of this work.

Mobile microscopes and schoolchildren in Kenya

A separate line of research comes from Finnish and Swedish groups working with low-cost portable digital microscopes. A 2017 proof-of-concept study digitised 13 iodine-stained stool samples and 4 urine samples with a miniature scanner; a deep-learning algorithm detected Ascaris, whipworm and hookworm eggs with a sensitivity of 83.3% to 100% in a test set of 217 manually labelled eggs (Holmström and colleagues, 2017). That was egg-level testing on a handful of samples, not a diagnosis of patients.

The group's 2024 study was much larger. Stool from 1,335 schoolchildren in Kwale County, Kenya, was prepared by the Kato-Katz method at a local primary-care laboratory, scanned with a portable whole-slide scanner and uploaded over mobile networks; 1,180 samples were of adequate quality, split into training (388) and test (792) sets (Lundin and colleagues, 2024). With expert manual microscopy as the reference:

sensitivity for Ascaris, whipworm and hookworm was 80%, 92% and 76%;
specificity was 98%, 90% and 95%;
more than 90% of positive cases were light-intensity infections;
in 79 samples (10%) that manual microscopy had called negative for a species, the system found eggs that were confirmed by visual inspection of the digital slide.

Two of the authors declared that they founded and co-own an image-analysis company, and patents for the mobile microscope were reported as pending.

Spiked slides in China

A small Chinese study added known numbers of Schistosoma japonicum eggs to stool from uninfected residents to make 80 Kato-Katz slides. An AI imaging system was correct on 97.50% of slides for presence or absence, against 81.25% for manual microscopy, and counted eggs accurately on 91.67% of infected slides, against 31.67%. It missed 3.33% of infected slides (manual: 13.33%) and made no false detections on the 20 negative slides (manual: 35.00%). It was slightly slower on average — 16.70 against 15.78 minutes per slide (Chen and colleagues, 2025). Laboratory-spiked slides are easier to score than real patient samples, because the true answer is known.

Validations in Other Laboratories

A result from the laboratory that built a system is the first step; how it performs elsewhere is the test that matters for routine use. Published validations from other laboratories give a mixed picture.

A Swiss university laboratory, 2025

The Swiss group tested the same company's wet-mount algorithm with a compact slide scanner in routine clinical work (Büschlen and colleagues, 2025).

On 135 reference samples, positive slide-level agreement with light microscopy was 97.6% (95% confidence interval (CI) 94.4% to 100%), reached after the confidence threshold for Schistosoma mansoni was adjusted, and negative agreement was 96.0% (86.6% to 98.9%).
In dilution series the digital system had slightly lower analytical sensitivity than microscopy at higher dilutions — the opposite direction from the developer laboratory's finding.
On 208 routine clinical samples tested prospectively, overall agreement was 98.1% (95% confidence interval 95.2% to 99.2%), kappa 0.915; minor disagreements involved Blastocystis, where the digital system was slightly more sensitive.
The authors described it as a "low-throughput screening solution" that assists technicians by pre-classifying structures for expert review, and wrote that each laboratory needs "extensive site-specific validation" and organism-by-organism tuning of thresholds. They declared no conflict of interest.

A Spanish hospital laboratory, 2026

A university hospital in Barcelona retrospectively tested an iodine wet-mount version of the same company's software on 178 stool specimens: 116 with one parasite, 31 with mixed infections and 31 negatives (Rubio Maturana and colleagues, 2026).

With a human operator reviewing the AI's findings, positive percent agreement was 96.62% and negative percent agreement 93.33% (kappa 0.865).
The system detected additional parasites in 74 of 178 specimens (41.57%).
The authors listed limitations including the small volume of specimen analysed and the detection of artefacts (debris that resembles parasites). The study was funded by a company, which the authors disclosed as a potential conflict of interest.

A different automated analyser in Belgium, 2026

Not every system has performed well. Six Belgian hospital laboratories, including two national reference centres, evaluated a commercial automated feces analyser that combines automated microscopy with deep learning, on 267 clinical stool samples with light microscopy as the reference (Aupaix and colleagues, 2026).

With standard settings, overall sensitivity was 86% and specificity 45% — meaning it raised many false alarms.
For clinically relevant parasites only, sensitivity was 95%.
Taking more pictures per sample improved detection, but it stayed under 90% for several targets.
The authors concluded that its performance "does not allow it to be used as a screening tool with automatic validation of negative results", and listed missing features, including a 10x objective lens and parasites absent from its database.

An Indian teaching hospital, 2026

In New Delhi, wet-mount microscopy found parasites in 204 of 7,267 stool samples (2.8%) over three years. In 275 samples run on both microscopy and another AI-assisted feces analyser, the analyser agreed with microscopy in 83.3% and with an expert audit in 84.7%; the audit found 35 misidentified cases and 6 missed detections, while the analyser found 10 cases that microscopy had overlooked. Against a composite reference standard, its sensitivity was 66.7% and specificity 86.2%. The authors wrote that with expert review its performance was excellent and that cost-effectiveness still has to be weighed (Aggarwal and colleagues, 2026).

Taken together: systems that pair a scanner with a well-trained network and human review have matched or exceeded microscopy in several laboratories, while a different analyser on the market performed well below that standard. "AI" is not one test; each system, scanner and laboratory set-up has its own figures.

After Adoption: A Before-and-After Count

In 2026 the US reference laboratory that published the October 2025 study reported what happened in routine work. It compared two six-month periods, before and after it began using AI to pre-screen concentrated wet mounts (Anderson and colleagues, 2026):

The number of patients with a helminth (worm) detected rose from 29 to 104.
Positivity measured at the level of the laboratory's clients rose from 4.5% to 12.3% (P < 0.0001).
Pinworm (Enterobius vermicularis) positivity rose about fivefold, from 0.03% to 0.15%.
Median turnaround time fell from 5.2 to 4.0 days.
Most infections were found in the first or only specimen submitted, and more than half of positive specimens contained 5 or fewer eggs or larvae — light infections that are easy to miss.

The authors describe the data as observational. A before-and-after comparison cannot fully separate the effect of the new method from changes in who was tested in each period, and the paper reports detections, not whether patients were treated differently or did better. One author disclosed being a collaborator of the software company.

How to Read the Accuracy Figures

Agreement is not the same as sensitivity

Most of these studies report positive and negative percent agreement — how often the AI's answer matched a reference method, usually human microscopy. If the reference itself misses parasites, a test that finds them will look like it is producing "false positives". Lundin and colleagues (2024) and the October 2025 study both found organisms that the human reading had missed, which means microscopy is an imperfect yardstick.

Discrepant resolution

The October 2025 study reports two sets of figures: 94.3% and 94.0% before discrepant results were resolved, and 98.6% positive agreement after. The second figure comes from re-examining only the specimens where the AI and the original reading disagreed. This method, called discrepant analysis, has been criticised in the diagnostic-statistics literature: a 1996 analysis in The Lancet by a U.S. Centers for Disease Control and Prevention (CDC) statistician showed that sensitivity and specificity estimated this way are biased upward, even when a perfect test is used to resolve the disagreements (Hadgu, 1996). This does not mean the newly found organisms are not real — the study reports they were confirmed by scan review and microscopy — but the pre-resolution figures are the more conservative ones.

How the reference standard was built

In the October 2025 study, a specimen counted as positive for training and validation according to traditional microscopy, and additional finds were adjudicated by scan review and microscopy.
In the 2020 trichrome study, accuracy was slide-level agreement with microscopy.
In Kenya, the reference was expert manual microscopy, plus visual review of the digital slides when the two disagreed.
In New Delhi, a composite reference standard combined microscopy and expert audit.
None of the stool studies above used a molecular (polymerase chain reaction) test as an independent reference.

Other things that change the numbers

Which parasites were included — negative agreement in the October 2025 study ranged from 91.8% to 100% depending on the organism.
Specimen mix — studies that pick known positives and negatives (Barcelona, the Swiss reference set) are not the same as unselected routine samples.
Threshold settings — the Swiss figure for Schistosoma mansoni was reached after the threshold was adjusted.
Funding and affiliation — the developer laboratory's papers include company co-authors or collaborators; the Swiss and Belgian validations declared no conflicts.

Beyond Stool: Malaria Blood Smears and Primary Care

The same kind of image analysis has been studied for malaria, where parasites are seen inside red blood cells on a stained blood smear. One research group built a deep-learning method that runs on a smartphone to find malaria parasites in thick blood smears, and released a dataset of 1,819 images from 150 patients. In patient-level five-fold cross-evaluation (repeatedly training on part of the data and testing on the rest), its accuracy at telling parasite from background image patches was 93.46%, with sensitivity 92.59% and specificity 94.33% (Yang and colleagues, 2020). These are figures for image patches in a research dataset, not for diagnosing patients in a clinic.

A 2026 scoping review looked for studies of AI-supported digital microscopy performed in primary health-care laboratories and compared with a standard diagnostic method. Of 3,403 papers screened, 22 qualified: 12 on blood (blood counts and malaria), 4 on urine, 3 on cervical or oral cells, 1 on sputum, and only 2 on stool parasites. The review reported accuracy comparable to the reference standard across these uses, and higher sensitivity than manual microscopy in 6 of the 7 studies whose design allowed that comparison. It named scalability and cost-effectiveness as unresolved questions (von Bahr and colleagues, 2026). One of its authors, also an author of the Kenyan stool studies, declared fees from and co-ownership of an image-analysis company.

How the Claim Compares With the Research

Supported

The paper exists and was published in October 2025. A deep convolutional neural network for stool wet mounts appeared online on 21 October 2025 in the Journal of Clinical Microbiology (Mathison and colleagues, 2025).
It found more parasites than people did. Published studies report that the AI detected 169 additional organisms in the validation specimens and detected parasites at lower dilutions than three laboratory scientists. Separate studies in Kenya, Spain and the developer laboratory's own routine work also report parasites found that human readers had missed.
Speed. After adoption, the developer laboratory reported median turnaround falling from 5.2 to 4.0 days (Anderson and colleagues, 2026).

Supported only in part

"Outperforms traditional methods". The abstract's own wording is that AI "detects considerably more organisms than traditional microscopy alone". Before discrepant resolution, its agreement with microscopy was 94.3% for positives and 94.0% for negatives, so it also missed specimens that microscopy found and flagged some that microscopy called negative. In the clinical studies the AI screens and a human confirms; it works with traditional microscopy rather than replacing the expert.
"More sensitive and more specific". Several systems show high sensitivity. Specificity varies a great deal: 91.8% to 100% by organism in the October 2025 study, 45% for one commercial analyser in Belgium, and 86.2% for another in India.
Accessibility. Field studies in Kenya point toward use in primary-care settings, but the scoping review found only two stool studies done in primary health-care laboratories, and the Swiss and Belgian groups stressed site-by-site validation.

Not found in the research

"Potential to transform diagnostics worldwide". This wording was not found in the paper's abstract, which ends by describing the work as bringing parasite screening into the technological age. The source of the quoted sentence on the slide could not be confirmed.
A link to cancer. None of the studies on this page measured cancer, studied cancer patients, or tested whether detecting more parasites changes any cancer-related outcome. Whether particular parasites contribute to particular cancers is a separate question, covered on Parasites Classified as Cancer-Causing Agents.
Support for the "less than 10%" figure. The AI studies show that human microscopy misses some parasites, especially in light infections, but none of them estimates that standard tests detect fewer than 10% of cases.

Decisions about parasite testing, and about cancer treatment, belong with a clinician.

What Is Not Known

Patient benefit. No study found here tested whether AI-assisted stool testing leads to better treatment decisions or better health, rather than more detections.
Clinical meaning of extra finds. Many additional detections were light infections or organisms such as Blastocystis whose role in disease is debated; how many of the extra finds would change care is not reported.
Independent replication of the headline figures. The largest wet-mount figures come from the developer laboratory with company co-authors. Independent laboratories report similar agreement for the same software, but with smaller samples and, in one case, lower sensitivity than microscopy at high dilution.
Comparison with molecular tests. No study found here compared AI-read microscopy head-to-head with multiplex polymerase chain reaction stool panels on the same patients as an independent reference.
Parasites outside the training set. A model recognises only what it was trained on; rare or unusual parasites remain a question, and the Belgian evaluators listed parasites missing from one system's database.
Cost. Scanner, software and validation costs, and cost-effectiveness in low-income settings, are named as open questions by the New Delhi authors and the scoping review.
Infections the stool test cannot see. Parasites that live in tissue (for example, the larval cysts in the brain described on Parasites in the Brain) do not appear in stool however good the reading is; AI improves the reading of the sample, not the limits of the sample.

Connections

Key Research Papers

Mathison BA, Knight K, Potts J, Black B, Walker JF, Markow F, Wood A, Bess D, Dixon K, Cahoon B, Hymas W, Couturier MR. Detection of protozoan and helminth parasites in concentrated wet mounts of stool using a deep convolutional neural network. J Clin Microbiol. 2025;63(11):e0106225. PubMed PMID: 41117595
Kirby JE, Arnaout R. What's not to learn? AI meets parasitology. J Clin Microbiol. 2026;64(1):e0145125. PubMed PMID: 41358779
Mathison BA, Kohan JL, Walker JF, Smith RB, Ardon O, Couturier MR. Detection of Intestinal Protozoa in Trichrome-Stained Stool Specimens by Use of a Deep Convolutional Neural Network. J Clin Microbiol. 2020;58(6):e02053-19. PubMed PMID: 32295888
Anderson CP, Couturier MR, DeMaranville M, Jensen R, Sant M, Barker A, Mathison BA. Diagnosing helminth infections in a large reference laboratory in the United States: a 6-month pre- and post-implementation analysis of AI-augmented screening of concentrated fecal wet mounts. J Clin Microbiol. 2026;64(7):e0016526. PubMed PMID: 42283511
Büschlen C, Rotzer D, Sidler N, Nguyen HTT, Oberli A. Validation of Digital Slide Scanning and a Convolutional Neural Network for the Detection of Intestinal Parasites in Human Stool Samples. Diagnostics (Basel). 2025;15(23):2974. PubMed PMID: 41374355
Rubio Maturana C, Sulleiro E, Zarzuela F, Martínez-Vallejo P, Mediavilla A, Silgado A, Turró C, Balladares M, Gracia A, Paz C, García-Vega D, José-Villar SMS, Blanco-Grau A, Moreno F, Larrosa N, Goterris L. Performance of an [product name omitted] AI-based image analysis system for coproparasitological diagnosis in clinical stool specimens: a retrospective evaluation study. J Clin Microbiol. 2026;64(8):e0043126. PubMed PMID: 42324608
Aupaix A, Filippin L, Jaumot J, Cannoot S, Chemais M, Martiny D, Miendje Deyi VY, Deffontaine M, Deckers C, Verbelen V, Potters I, Drieghe C, Mzougui S, Soleimani R, Philippart P, Brauner J. Evaluation of [product name omitted] automated microscope for parasitology: when artificial intelligence meets old school microscopy. J Clin Microbiol. 2026;64(3):e0074325. PubMed PMID: 41665376
Aggarwal P, Choudhary J, Pallavi, Saxena S, Dey S. Prevalence of intestinal parasites in a tertiary care hospital and utility of AI-assisted fecal analyzer for their detection. Exp Parasitol. 2026;282:109110. PubMed PMID: 41628712
Lundin J, Suutala A, Holmström O, Henriksson S, Valkamo S, Kaingu H, Kinyua F, Muinde M, Lundin M, Diwan V, Mårtensson A, Linder N. Diagnosis of soil-transmitted helminth infections with digital mobile microscopy and artificial intelligence in a resource-limited setting. PLoS Negl Trop Dis. 2024;18(4):e0012041. PubMed PMID: 38602896
Holmström O, Linder N, Ngasala B, Mårtensson A, Linder E, Lundin M, Moilanen H, Suutala A, Diwan V, Lundin J. Point-of-care mobile digital microscopy and deep learning for the detection of soil-transmitted helminths and Schistosoma haematobium. Glob Health Action. 2017;10(sup3):1337325. PubMed PMID: 28838305
Chen L, Luo H, Wu C, Wan J, Wu J, Liu H, Li B, Yang S. [Effectiveness of an artificial intelligence-enabled microscopic imaging recognition system for detection of Schistosoma japonicum eggs]. Zhongguo Xue Xi Chong Bing Fang Zhi Za Zhi. 2025;37(4):415-419. PubMed PMID: 41167950
Chua RSCS, Henderson KA, de Guzman LMC, Foss V, Schub N, Bell C, Medina JRC, Siao TG, Mistica MS, Belleza MLB, Modequillo MCR, Torres NJC, Belizario VY Jr. Variability of interobserver interpretation of selected helminth ova in the development of a training image set. Int Health. 2025;17(5):836-842. PubMed PMID: 39782878
Hadgu A. The discrepancy in discrepant analysis. Lancet. 1996;348(9027):592-3. PubMed PMID: 8774575
Yang F, Poostchi M, Yu H, Zhou Z, Silamut K, Yu J, Maude RJ, Jaeger S, Antani S. Deep Learning for Smartphone-Based Malaria Parasite Detection in Thick Blood Smears. IEEE J Biomed Health Inform. 2020;24(5):1427-1438. PubMed PMID: 31545747
von Bahr J, Suutala A, Diwan V, Mårtensson A, Lundin J, Linder N. AI-Supported Digital Microscopy Diagnostics in Primary Health Care Laboratories: Scoping Review. J Med Internet Res. 2026;28:e78500. PubMed PMID: 41490467