Study uses artificial intelligence to accelerate leprosy diagnosis

A technological innovation may contribute to the quest for the leprosy elimination, one of the humanity oldest diseases. An international team of scientists, led by the Oswaldo Cruz Institute (IOC/Fiocruz), in partnership with Microsoft AI for Health and the Novartis Foundation, has developed a diagnostic assistant, based on artificial intelligence, which can help identify suspected leprosy lesions. The technology has been called AI4Leprosy.

Photographer records skin lesion on project volunteer (Photo: Fareed Mirza/AI4Leprosy)

An article published in the scientific journal The Lancet Regional Health - Americas shows that, based on photos of the lesions on patient's skin and clinical data observed by physicians, the diagnostic assistant indicates the disease likelihood, hitting more than 90% of the cases. According to the scientists, the publication is a proof of concept of the method, which should serve as the basis for the mobile application creation for use by health professionals.

"Our study proves that it is possible to reach the suspicion of leprosy diagnosis with an artificial intelligence algorithm. This tool can support the physician's decision to start the investigation, accelerating the confirmation of the diagnosis and the start of treatment, which is fundamental to interrupt the transmission of the disease and prevent sequelae," says Milton Ozório Moraes, head of the Leprosy Laboratory at the IOC.

"The further validation and global rollout can help cover the last few miles of leprosy elimination, using the newest technology available to end one of the oldest scourges known to man," says Novartis Foundation President, Ann Aerts.

The delay in diagnosis is one of the biggest challenges to leprosy elimination. Since it was introduced in the 1980s, multidrug therapy (MDT) - treatment based on a combination of antibiotics - has cured about 18 million people, reducing the prevalence of the disease by 95%. MDT is donated by Novartis for all leprosy patients worldwide through the World Health Organization (WHO). In addition to promoting cure, the treatment blocks the transmission of Mycobacterium leprae, which causes the infection.

However, because of the delay in identifying the disease, the bacteria continues to spread and many people still develop visible deformities, loss of movement of the feet or hands, and vision problems. In 2019, more than 200,000 new cases were reported worldwide, with approximately 10,000 with advanced lesions. In Brazil, the second most affected country, there were 27,000 new cases detected, including 2,300 with advanced damage.

With the impact of the COVID-19 pandemic, which disrupted the health services, leprosy diagnoses dropped. According to the WHO, the global drop in new case detection was 37% by 2020. In Brazil, a survey by the Brazilian Society of Dermatology (SBD) indicated a 35% reduction in records in 2020 and 45% in 2021, compared to 2019. Considering the first results, scientists believe that artificial intelligence can contribute to achieve the goals set by the WHO, such as reducing new cases of infection by 70% by 2030 and, in the long term, interrupting The disease transmission.

Machine Learning

The starting point for the development of the virtual assistant for leprosy diagnosis was a type of image recognition algorithm that has been applied, for example, in supporting the diagnosis of melanoma, a form of skin cancer. One of the study's authors, Paulo Thiago Souza Santos, PhD in oncology with an emphasis in bioinformatics and postdoctoral fellow at the IOC's Leprosy Laboratory, explains that the technology is based on the computer's ability to distinguish subtle variations in images.

"Artificial intelligence can see more than the human eye can see. For the computer, each point of the image is a bit, translated into a number. An untrained person may not perceive the difference between two colors that are very close, but when the computer transforms these colors into numbers, it 'sees' a clear difference. It is on this basis that we can train the machine to try to make a differential diagnosis," says the researcher.

However, the scientists needed to adapt the methodology to face one of the great challenges of leprosy: the diversity of forms of the disease. Considering only the manifestations on the skin, the infection can manifest with one or many lesions, small or large, flat or raised, whitish or reddish, concentrated in one region or spread throughout the body.

"The number of similar diseases is large. Since the beginning of the research we have been looking for alternatives to increase the algorithm's accuracy", points out the dermatologist and researcher at IOC's Souza Araújo Outpatient Clinic, Raquel Barbieri, who shares the first authorship of the article with Microsoft's senior scientist of applied research, Yixi Xu.

One of the project's strengths was to rely on a large image bank of lesions to train the system to differentiate leprosy from other skin diseases. In all, 1,229 photographs were registered for 585 lesions, including both confirmed cases of leprosy and diseases with similar presentations. The images were obtained with the collaboration of 222 patients seen at the Souza Araújo Outpatient Clinic, maintained by the Leprosy Laboratory of the IOC.

The scientists also developed an artificial intelligence model capable of combining image recognition and clinical data analysis. The tests carried out pointed out ten main characteristics to establish the probability of the disease. For example, loss of heat sensitivity in the lesion and sensitivity changes in the feet were associated with a high probability of leprosy, while itching, which is more present in other dermatological diseases, was associated with a lower chance of infection.

Analyzing the images alone, the system achieved a 70% hit rate. Combining this analysis with the processing of the clinical data, the rate exceeded 90%. In the next phase of the research, the scientists will train the algorithm by collecting images and data through a mobille app, improving the system to operate with lower resolution images and in situations similar to day-to-day life in health services.

Milton Moraes estimates that a beta version of the app may be available in two years. The technology validation stage should also expand the geographical area of the research, including urban and rural regions in Asia and Africa, in order to evaluate the system in different contexts.

In addition to the IOC, Microsoft AI for Health and the Novartis Foundation, the Universities of Basel, Switzerland, and Aberdeen, Scotland, participated in the study. The research was funded by the Novartis Foundation in partnership with Microsoft. The image and data banks from the study are available for use by other scientists in open access.

Serial Advancements

National reference center with the Ministry of Health, the IOC Leprosy Laboratory is leading pioneering researches aiming to contribute to the Unified Health System (SUS). Last year, after 15 years of studies, the scientists obtained the registration from the National Health Surveillance Agency (Anvisa) for the first molecular diagnostic test for leprosy developed in Brazil.

The NAT Leprosy Kit was developed by the IOC in partnership with the Carlos Chagas Institute (Fiocruz-PR) and the Institute of Molecular Biology of Paraná (IBMP), linked to Fiocruz and the Paraná state government. In November, the test received a favorable opinion from the National Commission for the Incorporation of Technologies into SUS (Conitec) and should be part of the new Leprosy Clinical Protocol and Therapeutic Guidelines, currently being prepared by the Ministry of Health.