AI malaria diagnosis uses machine learning to find parasites in digitised blood smears, and similar tools now read stool and skin images for neglected tropical diseases. Field studies show good accuracy at moderate and high parasite levels. At low parasite densities, and without careful slide preparation, a trained microscopist, or a microscopist checking the AI, still does better.
Malaria microscopy has barely changed in a century. A drop of blood on a glass slide, a stain, and a trained person counting parasites through a lens for as long as it takes. In a busy district laboratory in the rainy season, that person may read dozens of slides a day, and by late afternoon their eyes are not what they were at eight in the morning.
It is an obvious target for automation. And yet the most interesting finding from a decade of AI microscopy research is not that machines can replace the microscopist. It is that the best-performing setup is usually a machine that finds candidates and a person who confirms them, and that the value of that pairing depends less on the algorithm than on the slide, the stain and the programme around it.
Why is malaria diagnosis still a problem?
The scale alone justifies attention. WHO’s World Malaria Report 2025 estimates 282 million malaria cases and 610,000 deaths in 2024, both slightly higher than the year before. The African Region carried about 94% of cases and 95% of deaths, and three quarters of deaths in the region were among children under five.
Most malaria is now diagnosed with rapid diagnostic tests, which are cheap, quick and need little training. But RDTs have limits. They cannot reliably count parasites, which matters for judging severity and monitoring response to treatment. Some cannot distinguish species. And in parts of the Horn of Africa and elsewhere, parasites with deletions in the gene that the most common RDTs detect can produce false negatives.
Microscopy fills those gaps, but it depends on skilled people. Good microscopists are scarce, their skills fade when they see few positive slides, and quality varies widely between laboratories. That is the problem AI is being asked to solve.
How accurate is AI malaria microscopy in the field?
The most useful evidence comes from field studies rather than laboratory benchmarks. A multicentre evaluation of EasyScan GO, a digital microscope that uses machine learning to analyse thick and thin blood films, was published in Malaria Journal in 2022 after testing across 11 sites in 2018 and 2019.
The results were respectable and revealing. Overall sensitivity was about 91% and specificity about 76%. With good-quality slides, specificity rose to 85%. Species were correctly identified in 93% of Plasmodium falciparum samples and 92% of P. vivax. Measured against WHO’s microscopy competence scheme, the device reached level 2 for detection and species identification and level 4 for parasite counting and false positives.
The detail that matters most is buried in the breakdown. Sensitivity was only 57% when parasite density was below 200 parasites per microlitre, rising to 90% and above at higher densities. And the authors found every performance measure was significantly affected by slide quality.
That combination defines where AI helps and where it does not. At moderate and high parasite counts, in a patient who is clearly sick, the machine reads as well as a competent microscopist and never gets tired. At very low counts, which is where elimination programmes live, it misses a lot. And a badly prepared slide defeats the algorithm just as it defeats the person.
Does AI help with neglected tropical diseases?
Neglected tropical diseases, the group of mostly parasitic and bacterial infections that WHO tracks under its 2021 to 2030 road map, are in some ways a better fit for AI than malaria. Many NTD programmes rely on mass screening of stool, urine or skin, with huge volumes of samples, low per-sample budgets and not enough trained readers.
A study published in Scientific Reports in 2025 gives a useful picture. Researchers in Kwale County, Kenya, digitised Kato-Katz stool smears from schoolchildren with a portable scanner and compared three approaches for soil-transmitted helminths: manual microscopy, fully autonomous AI, and AI whose detections were verified by an expert. Almost all infections were light-intensity, the hard case.
For whipworm, manual microscopy found about 31% of infections. Autonomous AI found 84%. Expert-verified AI found 94%. For hookworm, the figures were roughly 78%, 88% and 92%. All three methods kept specificity above 97%. For roundworm, autonomous AI did no better than manual reading, and expert-verified AI caught everything.
Two lessons stand out. First, in light infections, the unaided human missed a lot, largely because some eggs, notably hookworm, fade from Kato-Katz slides within an hour or two and whipworm eggs are easily overlooked. Second, the best configuration was neither human nor machine alone. The software found candidate objects; a person decided what they were.
This matters because NTD programmes increasingly need to prove that transmission has stopped. That means finding the last few light infections in a population, which is exactly where manual reading is weakest.
Where does the microscope and a trained person still win?
It is worth being specific, because vendors are not always.
Very low parasite densities. In malaria elimination settings, the goal is to find infections below the range where current AI devices are reliable. A skilled microscopist who spends longer on a suspicious slide still outperforms devices whose sensitivity drops sharply below a few hundred parasites per microlitre.
Unusual findings. Mixed-species infections, rarer species such as P. knowlesi or P. ovale, and artefacts that mimic parasites all trip up models trained mostly on common cases. A person can recognise that something is odd and send it for a second opinion.
Bad slides. Every field study says the same thing: slide preparation and staining drive performance. An algorithm cannot compensate for a smear that is too thick, badly stained or scratched. A trained person can often still extract an answer, or at least know that the slide should be redone.
Cost at low volume. A digital microscope is a capital purchase with maintenance, power and software needs. In a health centre that sees a few malaria slides a week, an RDT and a well-trained microscopist are cheaper and simpler.
The clearest role for AI is the reverse of these: high-volume, routine reading where fatigue degrades human accuracy; quality assurance, where software re-reads a sample of slides to catch drift in a laboratory’s performance; and standardisation of parasite counts across sites in clinical trials and drug-resistance surveillance.
A worked example: an NTD survey team
This is a hypothetical example, using round numbers, to illustrate the trade-offs.
A national programme wants to confirm that soil-transmitted helminth prevalence has fallen below the threshold for stopping mass deworming in a set of districts. It plans to test 5,000 schoolchildren. Its microscopists can each read around 40 Kato-Katz slides a day with care, and there are six of them for a two-month survey.
Under manual reading, the survey is feasible, but on this study’s evidence many light whipworm infections would be missed. The programme might wrongly conclude prevalence is low enough to stop treatment, and transmission could quietly resume.
With portable scanners and AI-assisted reading, slides are digitised in the field, the software proposes candidate eggs, and microscopists verify them on screen, including remotely from the capital. Each reader can check more slides per day, and more light infections are found. Because the images are stored, an independent team can audit a sample later.
The cost is real: scanners, power, image storage and training. And the programme has to decide how to handle the harder decision the better data now forces on it. Finding more infections may mean mass deworming continues for longer than planned, which is the right public health answer and an awkward budget conversation.
The strongest counterargument: is AI even needed for malaria?
A thoughtful critic would argue that malaria diagnosis is not really a technology problem. RDTs are already cheap and widely used. The biggest failures in malaria control are funding, drug and insecticide resistance, supply chains and access to care. Spending scarce money on digital microscopes, the critic would say, is solving the problem that donors find interesting rather than the one that kills children.
There is a lot of truth in this. For the typical febrile child in a high-burden area, a quality-assured RDT and prompt treatment matter far more than any AI microscope. Our piece on how the biggest killers in emerging markets get the smallest share of aid makes the wider point about where money goes versus where the burden sits.
But the critic’s case weakens in three places: counting parasites, where RDTs cannot help; surveillance for RDT-evading parasites and drug resistance, where standardised microscopy data is valuable; and elimination settings, where finding the last low-density infections decides whether a programme can stop. In each of those, AI-assisted microscopy is a narrow tool for a specific job, which is exactly the kind of AI that tends to work, as we argued in AI diagnostics: the evidence is strongest where the question is narrowest.
What this means for programmes, donors and developers
For national malaria and NTD programmes, the practical advice is to buy AI microscopy for specific tasks, such as quality assurance, surveys and reference laboratories, rather than as a general replacement for microscopists. Invest in slide preparation and staining first, because every study shows that is where performance is won or lost. Keep training microscopists, because verification is where the best results come from.
For donors, the evidence argues for funding hybrid workflows and the people in them, not just devices. A scanner that allows one expert in the capital to verify slides from twenty districts is a way of stretching scarce expertise, not replacing it.
For developers, the message from the field is to publish performance at low parasite densities and on poor-quality slides, not just overall sensitivity. Programmes need to know exactly where your device fails, because in elimination, the failure zone is the whole game.
Frequently asked questions
Can AI replace malaria microscopy?
Not yet, and probably not entirely. Field evaluations show AI microscopes perform well at moderate and high parasite densities, but sensitivity drops sharply at low densities and depends heavily on slide quality. The most promising approach pairs software, which finds candidate parasites and counts them consistently, with a trained person who verifies them. That combination has outperformed either alone in recent studies.
Is AI better than rapid diagnostic tests for malaria?
They do different jobs. Rapid diagnostic tests are cheap, fast and suited to diagnosing individual patients at the point of care. AI-assisted microscopy can count parasites, identify species and create a record that can be audited, which matters for severe malaria, drug trials and surveillance. For most routine diagnosis in high-burden settings, RDTs remain the practical first choice.
What are neglected tropical diseases?
Neglected tropical diseases are a group of around 20 conditions, including soil-transmitted helminths, schistosomiasis, lymphatic filariasis, trachoma and several skin diseases, that mainly affect poor communities in tropical regions. WHO’s road map for 2021 to 2030 sets targets for controlling, eliminating or eradicating them. Many programmes rely on mass screening, which makes them a natural fit for AI-assisted image reading.
Why does slide quality matter so much for AI microscopy?
AI models learn from images, and a thick, poorly stained or damaged smear produces an image unlike the ones the model learned from. Field studies of AI microscopy consistently find that performance falls with poor slide quality. Unlike a person, the software may not recognise that a slide is unreadable. Training staff in preparation and staining is often the cheapest way to improve AI results.
The takeaway
AI in malaria and NTD programmes earns its place in the unglamorous corners: counting, checking, surveying and verifying at volume. It does not earn a place as a replacement for the person at the microscope, and the best field evidence says it works best when that person is still there. Fund the slide, the stain and the reader first, and the software will make all three go further.
In your programme, would you rather have one more trained microscopist or one AI microscope, and why?
