Annalise.ai

Showing what the AI saw, and where it was unsure

Radiologists repeatedly told us that an AI which flags too much becomes another source of work. They needed to see the evidence behind a finding to judge whether it was credible. So each finding showed the model's probability and a pixel-level heatmap marking where on the image it sat.

The heatmap came from how the model already worked. It calculated a probability for every pixel as it assessed a scan, but the map was not being stored anywhere. Once we could see it existed, engineering had a clear case to keep it and bring it into the product. Before the application was built, I had real patient cases run through the early model and made a prototype from the findings and heatmaps it returned.

A radiologist could move from a finding straight to the relevant part of the image, see why the AI had raised it and decide whether to confirm or dismiss it. I also worked to cut what I called eye miles, the distance a radiologist's eyes travel across three or four monitors, and gave radiologists control over how the app behaved. This design for confidence and uncertainty was later granted an Australian patent.

Deciding what earns a radiologist's attention

The initial engineering approach ranked findings by model confidence, with the highest probability at the top. Our first reader study showed the problem. A rib fracture detected at 95% could sit above a suspected pneumothorax detected at 40%, and missing the pneumothorax is far more dangerous.

With the clinical team, I helped move the ranking to clinical consequence, so potentially life-threatening findings came first even when the model was less confident. We split the interface into Priority and Other findings. Separately from the app, we built integrations into the radiologist's worklist that moved urgent cases forward on the same logic. In testing, radiologists identified critical conditions earlier, and the teams had a shared rule for later prioritisation calls.

The version with no interface

Some hospitals used legacy PACS workstations that could only receive static DICOM secondary captures. On those, finding text had to be burned directly into the image, and every pixel used to explain a finding took space from the diagnostic image.

The same output had to work for a radiologist reading their hundredth case of the day and for a non-specialist in an emergency department seeing it for the first time. I defined the typography, contrast and layout across English, non-Latin scripts and different monitor resolutions. We also put structured finding data into custom DICOM tags, so the information could travel with the record and leave the image itself unchanged.

93% of clinicians still using it after two months

Annalise went on to process more than 2.1 million cases across 12,000 facilities worldwide. In post-rollout research, 93% of clinicians were still using the product after two months and 90% said it positively affected their reporting. Those figures measure the product. The model's diagnostic accuracy was the work of the clinical and data science teams.

I built the design system that carried the same core interaction across four diagnostic modalities. The work won an IxDA Interaction Award (Best in Category, shortlisted alongside Google and Philips), Good Design Australia Best in Class and Good Design Korea.

What moved engineering, the founders and clinicians on costly or contested decisions was research and testing evidence showing precisely what the clinical user needed.