Battling river blindness disease with the help of AI

The potential for AI and deep learning to accelerate analysis and deliver superior results with fewer errors has excited medical researchers for several years. For example, one of the first successful applications of these technologies to the medical research field occurred in 2017, when researchers at Stanford University in California used a deep convolutional neural network and image analysis to successfully classify skin lesions. The system performed as well as a panel of 21 dermatologists and results such as this made deep learning a promising candidate for medical applications to automate visual inspections.

But many early AI projects failed to develop into full solutions. In fact, it wasn’t until September of 2021 that the US Food and Drug Administration approved the first use of an AI-powered system for diagnostic use. It is a clinical-grade AI-based solution from Paige AI in New York that assists pathologists in detecting prostate cancers. The company invested more than 10 years in its development.

A condition affecting more than 20 million people

Still, the effort and time invested in such projects has tremendous potential to improve health outcomes around the world. As an example, it could help with the elimination of onchocerciasis – more commonly known as “river blindness.” It belongs to a group of conditions known as Neglected Tropical Diseases and has currently infected more than 20 million people and caused permanent blindness in more than one million.

Common to sub-Saharan Africa, Central America, and South America, river blindness is an infection spread by black flies that typically live near fast flowing rivers. When the flies bite humans, worm larvae invade the body. These grow into worms which reside in nodules under the skin where they can produce millions of baby worms (microfilariae), which then travel through the body. If these microfilariae reach the eyes they can cause irreversible blindness. Typical symptoms include extreme irritation, inflammation, and itching. To combat this disease, the World Health Organization’s Onchocerciasis Technical Advisory Subgroup is spearheading a global effort to eliminate transmission in 10 countries by 2030.

The road to elimination

While drugs currently exist to kill the microfilariae, no medication has yet been developed to target the adult worms. Since adult worms can live and reproduce for up to 15 years, it is difficult to stop transmission through current drug regimens and patients often must receive treatment for decades. Hence the need for new drugs that also target the adult worms.

Evaluating the efficacy of different treatment regimens involves a histological examination of tissue samples. This is a time-consuming process for several reasons. There are only a handful of experts worldwide trained to do the work. Proper examinations can take up to three months for 150 nodules and more than a year for 2,000 nodules. As a consequence, evaluating and registering new treatments can take years.

Using deep learning to speed up clinical trials

At Capgemini’s AI Center of Excellence, we believe AI can be used to address these challenges and help bring new treatments to market. To that end, Capgemini and the Institute for Medical Microbiology, Immunology and Parasitology (IMMIP) at the University Hospital Bonn in Germany have joined forces to apply the power of deep learning to automate the evaluation of tissue samples.

Capgemini’s role in the project is multifaceted. In addition to developing the algorithms used to evaluate tissue samples, we are also creating suitable training data and designing and implementing the solution. Other project partners include the German Center for Infection Research (DZIF), the Drugs for Neglected Diseases Initiative (DNDI), the Kumasi Centre for Collaborative Research in Tropical Medicine (KCCR) in Ghana, and Washington University in St. Louis.

Our project – AI for Onchocerciasis – plans to use AI to standardize testing and accelerate the examination of tissue samples. We expect these two improvements to shave several months off the time required to evaluate clinical studies.

How it works

Our approach starts with digitizing and labeling microscope slides of tissue samples subject to four different staining’s. We use these to train and test deep-learning models to detect and classify worm sections in a histological image. To handle the massive images involved (up to 35GB) we have developed a two-phase approach. First, we apply object detection on low-res images to identify regions of interest. The underlying models are based on Faster Region- Based Convolutional Neural Networks (R-CNN) and fine-tuned via transfer learning on labeled images.

In the second phase, we crop high-res section images that contain the necessary level of detail to make a judgment and classify them using Convolutional Neural Networks (CNN). This results in a separate model for each attribute we seek to identify when inspecting a new tissue sample. The initial results are very promising, and we will do a first test on ongoing clinical trials in 2022.

Crowd-sourcing AI development with the Global Data Science Challenge

While our results are already positive, the field continues to develop rapidly and there is room for improvement. To help with this, we are engaging with the 5^th edition of the Capgemini Global Data Science Challenge (GDSC). We will launch this challenge, called Code for a Cure, in 2022, and task participants with creating an AI-based solution to automate the current manual evaluation process.

The GDSC is an opportunity to accelerate our progress, and we’re looking forward to the outcome of this competition. At the same time, using the GDSC to tackle river blindness represents an accessible, hands-on introduction to this exciting field. It will help teach AI to the next generation of data scientists, and help non-experts identify opportunities to add AI to their projects.

Looking ahead

The partners in this project are already working on next steps. We want to verify our results in ongoing clinical trials and develop a user interface to allow researchers who lack programming knowledge to use our solution. And because this project represents a significant opportunity to improve health outcomes for millions of people, we will share our results through open-source venues with the goal of enabling other researchers, in keeping with the vision of AI for good. This is fitting, as our solution builds on the work of others and would not have been possible without sharing via the open-source movement. AI modeling techniques are ready to automate visual tasks and improve their results. They can shave precious time off drug discovery and approval processes and contribute to a massive transformation of the health industry.

The biggest obstacle to this is access to high-quality data for training the AI systems. To fully realize the benefits of AI in the medical sector, data sharing is essential. The good news here comes from an unlikely source: the global pandemic. COVID-19 has demonstrated the value of working together to create medical solutions to pressing problems. Already, we’re seeing a big push towards data sharing from both the research community and the health sector.

Innovation Takeaways

Let’s have a look at you

In medical practice, diagnosis often relies on visual examination. AI-powered systems can support this, and even automate the process.

Garbage in, Garbage out

The algorithms are ready to enable AI-powered visual examinations. Getting the right data set is the biggest issue. Data sharing is essential.

Crowd-Sourcing AI with GDSC

Capgemini’s Global Data Science Challenge combines AI for good with hands-on opportunities to teach AI to data scientists and introduce it to those in other fields of study.

Interesting read?

Data-powered Innovation Review | Wave 3 features 15 such articles crafted by leading Capgemini experts in data, sharing their life-long experience and vision in innovation. In addition, several articles are in collaboration with key technology partners such as Google, Snowflake, Informatica, Altair, A21 Labs, and Zelros to reimagine what’s possible. Download your copy here!