Align
We began by deeply understanding ATUSA’s context: the product, the technology, and the people who rely on it.
Our team interviewed radiologists, breast-cancer surgeons, and imaging specialists to capture their expectations, frustrations, and level of trust in AI. We also reviewed industry studies showing how ABUS-based AI can reduce unnecessary recalls compared to standard mammography — framing AI not as a novelty, but as a potential game-changer in breast cancer detection.
In parallel, we ran a full Design Sprint approach supported by interviews and benchmarking to understand both iSono’s internal challenges and the realities faced by their end users.
Prioritize
From this discovery work, we identified two high-impact workstreams:
UX/UI workstream
Focus: streamline radiologist workflows, reduce friction and clicks, and surface the most relevant images first to reduce review time and cognitive load.
AI/ML workstream
Focus: reduce false positives, improve confidence scores, and ensure that model performance improvements translated into real downstream clinical value.
We conducted qualitative and quantitative research, stakeholder interviews, usability testing, and industry benchmarking to refine the interface across three key stages of the radiologist journey. In AI/ML, we audited training and validation data, reviewed and tested the model’s training code, and quantified false positives across cases.
90%
Model accuracy
increasing diagnostic reliability
28.7%
False detections
reduction in false positives
Prove
With priorities clear, we moved to prove impact in both tracks:
UX/UI track
We streamlined the workflow by reducing clicks and removing friction, enabling doctors to access critical 3D images and AI-marked insights faster. This helped them more quickly distinguish between easy and complex cases, significantly reducing time spent per review.
AI/ML track
We introduced targeted modifications to the learning process — including hyperparameter changes within the focal loss function and experiments to assess their contribution.
As a result, we:
Boosted model accuracy from 70% to 90% (a 28.6% relative improvement)
Achieved a 28.7% reduction in false detections
This led to faster, more trustworthy scan reviews and more reliable confidence scores.
We validated our UX and AI improvements through 4 prototype iterations and 2 rounds of user testing, plus a deep dive into radiologist workflows to ensure changes mapped to real practice.
4 prototype iterations
2 rounds of user testing
A deep dive into the radiologist workflow
Integrate
To make these improvements usable in the real world, we focused on integration:
In UX/UI, we redesigned the review interface to surface only the most relevant 3D images (those flagged by AI), with an option to expand into a multi-plane comparison mode. The redesigned Scrubber allows radiologists to navigate slices quickly and correlate findings across planes, improving interpretation speed and clarity.
In AI/ML, we built internal evaluation tools using Jupyter notebooks and prototyped lightweight interventions in the existing codebase. This allowed iSono to see how different training decisions influenced false positive rates and recall, and to incorporate these learnings into their model lifecycle.
We also delivered a clear Improvement Roadmap so the team could integrate changes step by step into ATUSA’s product and AI pipeline.
Reduced time and cognitive load
reduction in false positives
Scale
Although the project ran over just 8 weeks, we didn’t stop at a “better model now” — we aimed for a path forward.
We established a clear path for further enhancement, recommending additional hyperparameters and training strategies that could be explored to keep improving the model beyond this engagement.
By analyzing false positives from both the original iSono model and our best-performing approach, we identified artifacts in images as a key source of noise — and proposed R&D ideas such as a model that simultaneously predicts artifacts and potential lesions to better discriminate between them.
These insights gave iSono not just higher accuracy in the short term, but a structured foundation for ongoing AI/ML evolution.
