User research, Product design
Product Manager, Product Designer, 2 Full-stack Developers
AI Scope, a non-profit organisation, teamed up with ThoughtWorks to help build their app. The vision: artificial intelligence that can diagnose parasitic diseases such as malaria and tuberculosis.
In order to get to this stage, a large database is needed with images of previously detected parasites. The project was therefore divided into two phases. The first phase involved the development of an app that microbiologists and trained analysts can integrate with their workflow when doing blood analysis. With their phone mounted to the microscope, they capture positive blood samples, mark the disease in the photo, and save it to a database.
The second phase will then combine open-source image recognition technology with a portable, resistant and ultra cheap microscope to help early diagnosis of curable diseases, globally.
As a small ThoughtWorks team, we kicked off with the first phase: building an app that analysts would use while diagnosing malaria, to help build an image database. The team faced many challenges. Due to the pro bono nature of the project, we were especially short on resources, as most of us were contributing in our free time. This forced us to think carefully about the bare minimum requirements, and get feedback as quickly as possible.
Working asynchronously and lacking understanding about the context in which the app would be used, presented additional problems. Collecting user insights was vital but had to be done by instructed people on the ground. We had to continuously experiment with phone and image quality as well as internet connectivity in remote areas.
I started off by interviewing users, to gather insight into their usual workflow and all the complexities around diagnosing malaria. Talking to subject matter experts both in developed and underdeveloped countries helped us refine the app’s purpose, as it turned out there were significant differences to their working conditions and processes. When we started, we considered universities as our users too, but soon decided to focus our MVP efforts on health facilities in areas where malaria is common.
To get the app off the ground as quickly as possible, we used existing patterns and libraries. As we were building an Android app, Material Design was taken as a starting point for the interface. A dark skin was chosen for the app, to avoid eye strain for users who are looking through a microscope for the larger part of their day. Similarly, due to the red and pink tones in the microscopic images, green was chosen to stand out as accent color.
For microbiologists to use the app and collect the necessary images for machine learning, we were depending on their goodwill. Rather than us helping them solve a problem, they were helping us; they just believed in and were willing to contribute to the cause. Therefore, the most important principle was that the app was seamlessly integrated with their way of working, and that engagement with it was as short as possible.
With the help of a health facility in Zambia, the team was able to get realtime user feedback. To our surprise, as it turned out, the app was actually very useful to them. This was because analysing the cells through a phone screen was much more convenient than staring through a microscope lens. A win-win, encouraging the team to deliver fast.
There were a few things that needed to work more smoothly. Zooming and panning while being in capturing mode turned out to be essential and didn’t work well initially. Another important finding was that the app was not providing enough context for the analysts to understand the intended workflow. More flexibility was also needed between the capturing and masking of images. Here, we intentionally decided to compromise speed of the workflow in exchange for clarity.
While working on the app, we were discussing the next phase with other people from ThoughtWorks that had experience in building AI, which informed our decisions regarding masking images and storing data. At the same time, we consulted legal expertise and had many ethics discussions in relation to storing privacy-sensitive information. We needed to make absolutely sure that there was no way to trace the stored images back to individuals, while at the same time collecting information relevant to the infection: a delicate balance.
The project is currently ongoing and is open source. Anyone looking to contribute can find the project in this GitHub repository.