Applied GenAI for clinical workflows
Foundation-model integration with LoRA adapters and supervised fine-tuning — powering capture guidance, automated triage outputs, and reporting and review tooling.
Problem
The clinical value of an imaging device is not only in the numbers it computes but in the workflow around them — guiding a good capture, triaging what it sees, and turning results into a report someone can act on. Those steps are repetitive, judgment-heavy, and exactly where an operator's time gets consumed.
Generic foundation models are not fit for that context out of the box. They do not know the capture protocol, the triage criteria, or the shape of the reports the workflow expects, and a clinical setting has little tolerance for outputs that wander outside those bounds.
Approach
I integrated foundation models into the clinical workflow and adapted them to it with LoRA adapters and supervised fine-tuning, rather than relying on prompting alone. Fine-tuning pulls the models onto the specific tasks the device workflow needs, so their behavior is grounded in how the capture and reporting actually work.
The applications target the parts of the workflow where automation pays off: capture guidance during acquisition, automated triage outputs, and reporting and review tooling. Each is a place where a well-adapted model removes a manual step without taking the clinician out of the loop.
What shipped
The work delivers clinical decision support and workflow automation — capture guidance, automated triage outputs, and reporting and review tooling — powered by foundation models tuned with LoRA and supervised fine-tuning for the device's clinical context.
It sits next to the imaging pipeline as the layer that turns raw capability into something that saves the operator time, keeping the human in control of the decisions that matter.