Improving malaria diagnosis with AI

Why did you choose to work on malaria diagnosis?

Renaud Piarroux: 600,000 people a year worldwide still die of malaria. Improving the way we diagnose the disease is already half of the job in achieving better treatment. This is why we have launched a project, supported by the Institute of Global Health at Sorbonne University, to improve the quality of diagnosis in Africa, where the tools available are not ideal. The idea is simple: anyone who sees a doctor with suspected malaria should receive a reliable diagnosis, regardless of the hospital or clinic.

How does this system actually work?

R.P.: Malaria diagnosis relies primarily on microscopy. Other methods exist, such as molecular biology (PCR), but they are expensive, and rapid diagnostic tests using test strips are unreliable. Microscopes remain the most relevant tool, provided that the person using them is well trained and supervised, as the forms of the malaria parasite vary according to species and the stages of its cycle, making it difficult to identify.

Our idea is to use smartphones to photograph the slide observed by a local microscopist and share these images via an application to experts who can confirm or rule out the presence of the parasite. Given the millions of diagnoses made each year, it is unrealistic to constantly call on a small group of experts. The project therefore relies on deep learning models capable of filtering images and identifying those which are likely to contain a parasite, optimising human verification. 
 

Which partners are you working with?

R.P.: The project involves the Sorbonne Centre for Artificial Intelligence (SCAI), the Pitié-Salpêtrière Hospital, the Montpellier University Hospital Centre and a network of French hospitals. In Africa, the project includes three partners: the National Biomedical Research Institute (INRB) in Kinshasa, with funding from the French Development Agency; the University of Lomé, as part of a partnership with Sorbonne University, and CIGASS, an institute that trains microscopists to diagnose malaria in French-speaking countries in Africa.
Two start-ups are also involved in developing the application. The first is working on improving data using adapted microscopes which analyse images using AI and reconstruct more accurate images for better quality learning. The second is authorised to manage health data. It is developing a secure application to which users will send their photos. These will be processed by our models and then sent back to identified experts, once they have been pre-sorted by the AI. At a glance, they will know which elements to study in order to confirm or rule out the presence of the parasite. 

This project is not limited to Africa. In France, in hospitals far from major cities or during on-call shifts, diagnoses are not always made by experienced parasitologists. The project therefore aims to offer a remote support solution to improve diagnosis wherever and whenever expertise is lacking.
 

What are the main technical difficulties?

R.P.: Correctly labelled, good quality images are rare. We first need to obtain a large number of them to train computer vision models. Unlike traditional image recognition – there are countless photos of cats and dogs on the Internet – there are almost no databases of correctly annotated Plasmodium images, nor are there any collections of images of artefacts to learn how to distinguish them from the parasite.

The first job is to collect and label these images correctly in collaboration with our network of French hospitals and partners in Kinshasa, Lomé and soon Dakar, before moving on to the machine learning stage.

What are the next steps for the project? 

R.P.: Developing an application and setting up all of the necessary partnerships takes time. We have been working on malaria detection for several years now. We have obtained some excellent results and started to publish on the quality of AI-based diagnosis.

In the coming months, our goal is to possess models which are efficient enough to implement a remote supervision system; AI will not provide the diagnosis but it will be able to follow the work carried out by microscopists and identify training needs. Our approach aligns with national plans for the elimination of malaria, supported by the World Health Organisation (WHO) and the World Bank, which promote improved treatment in numerous countries around the world. 

In a second phase, we will improve the models, particularly by enhancing staining techniques and imaging conditions. Then, once the tool has been sufficiently tested and validated, it will be possible to make an immediate diagnosis without consulting experts. At this point, microscopists will be able to rely entirely on the results produced by the algorithm.

But we do not want to rush things: statistical reliability must be proven. For example, a person with malaria may have only one parasitised red blood cell in 10,000. If the model produces a 0.1% error rate, this means that it will produce ten times more false positives than real cases, which is unacceptable in practice. We therefore need to determine precise thresholds and adapt the models to this type of data.
 

Besides malaria, are there any other international projects that you are currently pursuing?

R.P.: Yes, I am working on another project which looks at the identification of the fungal agents responsible for infections. There are more than 700 different species that infect humans, each with their own characteristics. Diagnosis, which is often carried out using a microscope, is extremely complex. An alternative is to use mass spectrometry, a technique in which a laser interacts with a protein extract from the fungus to produce a kind of molecular fingerprint. Analysing these fingerprints requires high-precision digitisation, but major diagnostic companies have shown little interest in mycology, which is much less profitable than bacteriology, and which accounts for 95% of their business. They have limited themselves to the most common species, leaving out many fungi which are essential for specialised laboratories.

To fill this gap, we have developed an online application capable of analysing mass spectra and comparing them with an extensive database. We built this database in collaboration with a European collection in Brussels, thanks to an agreement which allowed us to use their strains to create a reference library. This application, which is free and accessible to all, is now used worldwide. Its use is growing every year; in 2024, we expect to exceed 800,000 spectra sent for identification.

For the 2025 honorary doctorate ceremony at Sorbonne University, you supported and introduced Jean-Jacques Muyembe, a Congolese virologist who co-discovered the Ebola virus. What does this relationship mean to you?

R. P.: For many years, I worked on epidemics, particularly cholera epidemics in the Democratic Republic of Congo, and Jean-Jacques Muyembe was head of the National Biomedical Research Institute, our national partner on this issue.

We have worked together for over twenty years. I even coordinated a mission on the plague in eastern Congo at his request. Sorbonne University wished to honour the career of a researcher who, aside from Ebola, has played a leading role in monitoring epidemics in his country. He is a model researcher and a formidable example of scientific commitment. We should remember that he was the first Congolese doctor to obtain a PhD in science. Presenting him with this distinction on behalf of Sorbonne University was a very moving moment.
 

What lessons have you learned from major health crises in Africa? How have they changed the way we think about global health today?

R. P.: By improving my understanding of other issues such as malnutrition, obesity and global warming, the Institute for Global Health has helped me to take a step back and look at infectious diseases from a different perspective. I have come to understand that all epidemics are part of a whole, governed by the same determinants as other health issues: the way we live on Earth and our lifestyles. Transport, for example, contributes both to sedentary lifestyles and to the spread of viruses. Urbanisation, access to water and essential services influence quality of life, as well as the risk of waterborne diseases such as cholera and the proliferation of malaria vectors. The way cities are organised therefore plays a role in obesity and infectious diseases.

Underlying these health issues, there is also an economic rationale. Priority is given to what sells, i.e. vaccines, rather than fundamental work on access to water or sanitation, which are not profitable. I saw this in Mayotte, where I was during the 2024 cholera epidemic. The first instinct was to vaccinate and spread messages promoting hygiene, before extending measures to ensure minimum access to drinking water. Acting almost exclusively in an emergency does not seem to me to be a satisfactory approach, and I find it hard to believe that we can protect people by telling them to wash their hands without providing them with the clean water they need to do so.
 

Interview by Justine Mathieu