Franklin.ai - Supporting the Diagnosis and Reporting of Cancer with AI

First, we conducted research to understand how pathologists go about their tasks in their offices. We shadowed eight pathologists, observing them diagnose and report cases. We documented their tasks and workflows with a particular focus on:

1. How they interacted with objects, including their microscopes, desktops and reference materials
2. What information they were using to help diagnose and report, including the notes they wrote or dictated
3. What ’work-arounds’ or ’rules of thumb’ they used to make their work more efficient

This helped the team develop a shared understanding of the opportunities to help pathologists.

By collaborating with our colleagues, we helped define a product scope that would not only help pathologists but also be technologically feasible and financially viable for our business.

Second, conceptual mock-ups went through over ten rounds of iteration. Each of these ”design sprints” involved focussing on a specific product question, designing an experiment and concepts to elicit feedback and having at least five pathologists interact with a prototype.

We first delved into questions and decisions that would have the most far-reaching consequences for our product. These included:

1. When and how should the AI findings be presented to the user?
2. What was the right ’unit of analysis’ - slides, specimens, or something else?
3. How should we summarise findings for reporting?

The feedback helped inform workflow, functionality, information architecture and high-level interaction design decisions.

Once these big, consequential decisions were made, we delved more into more nuanced questions such as:

1. What should the report look like?
2. How should we communicate the probabilistic nature of AI findings?
3. How might our product need to adapt to a variety of labs and jurisdictions?

These helped refine our user interface and information design.

Our product shows the AI’s findings to pathologists. They can then agree or disagree with them. If they agree, they can move through a case faster. If they disagree, they can use our interface to find the cause and edit as they see fit. Our interface purposefully limits information to only what is necessary at a given time. That is, our design answers the most important questions in succession:

1. Have you found cancer?
2. Where have you found it?
3. How severe is it?

This lets pathologists optimize their workflow. They can find the most relevant specimens for diagnosis.

Our aesthetics reflect this focus. We avoided overwhelming the user with data and chose a muted colour palette. We use colours to encode the most important information - the type and location of findings - so they stand out against the background.

We also added design touches that honor pathologists’ real-world practices. An example is rotating a specimen badge 90 degrees once a pathologist has viewed a specimen. This mimics real-world pathologists. They review a slide, then turn it upside down to check their progress on a case.

We also developed a range of checks to prevent illogical reporting. For example, the grading system reflects the percentage of cancer that is high grade versus low grade. If high grade cancer is over 50%, we prevent the user from reporting the specimen as a low grade specimen. Our product can help minimize illogical errors with our built-in checks.

Our software will enable more accurate diagnoses and faster reports. It will make pathologists happier. We designed some novel features with pathologists from Sonic Healthcare. These include:

1. Specimen triage is a quick overview of findings in a case. It helps pathologists find the diagnosis-defining specimen first.
2. Clear, high-level visual and textual overviews of findings and their quantification. They should let pathologists quickly decide if they agree with the AI’s entries.
3. ’Showing our workings’ for cancer grading. This explains how we quantify the ratio of low to high grade cancer, which in turn drives treatment.
4. Showing AI confidence levels. This will help pathologists use the AI better and optimize their workflow. For example, they are more likely to spend time reviewing findings the AI is less confident in.

Our product sets the standard for medical imaging AI software. Since the beginning, our core belief was to augment, not replace, human pathologists. By humanizing AI and making it useful, we can improve patient outcomes. We aim to support human decision-making.

Pathologists will be less likely to miss small bits of cancer.

Less time on reports will help diagnose and treat patients faster.

Referring clinicians will better measure key diagnostic and prognostic indicators, like cancer grading.