Accessing AI-driven image search and generation for medical education, training, and practice

Introduction

Radiologists read images in dim rooms, often under time pressure, and a tool meant for their training has to respect the habits they already have while introducing something they have not used before.

Medical education depends on seeing many examples of a condition, and rare or early findings are, by definition, hard to collect. IndaigoMed approaches this by generating CT scans from a written description, so that a trainee can see what a described finding looks like and compare it with a real study. The generator is useful only if the surrounding viewer lets people read, compare and question the images it produces.

The work moved through four stages, each beginning with an artifact from the stage before it. A later phase widened the question from how the viewer should behave to what a resident’s working day in the tool should look like.

Auditing the existing build

The build the participants used was already capable: it could list studies, display PET and CT series, and generate a new scan from a prompt in a panel beside the image.

Its interface, however, had grown from an open radiology viewer, and the generative features had been added at the edges of a layout that was not designed for them. I worked through each screen and left a written note wherever the design made a task harder, then grouped the notes by kind. The six groups below became the brief for the redesign.

Smaller decisions in the viewer

Once the layout was settled, the remaining notes from the audit could be addressed one by one.

For each decision I record the problem it addressed, what I chose and, where there was one, what the choice cost.

Questioning the word “generate”

The prompt screen in the existing build looked like the AI image generators people already knew, with a single input and a button labelled Generate, and that resemblance was partly deliberate.

Matching a familiar pattern helps people know what to do, but the label carried implications that did not fit the setting. In medicine, a generated image can sound like a fabricated one. I proposed looking for a call to action that matched the tool’s purpose more closely, and argued that the brand itself should present the model as a trusted instrument for trainees and medical professionals, since concerns about data privacy and unfamiliarity with prompting were already making people sceptical.

The home screen I designed keeps a single question at the centre and offers three example prompts beneath it, so that a newcomer can begin from a well-formed description instead of a blank field. The call to action changed from a bare “Generate” to the more declarative “Generate a CT scan”, which tells the reader exactly what the tool will produce before they commit to it.

Two layouts competed for this screen. The first set the headline and prompt on the left with three brain scans on the right, and the second centred the logo, a one-line description and the prompt, with example prompts drawn from different parts of the body. The centred layout was chosen for its prominence, because it gives the prompt the centre of the screen with nothing competing beside it.

A simulation environment for residents in training

The later part of the project asked what the viewer would need around it to serve as a full training environment, in which a resident manages a list of studies the way a working radiologist does.

I knew little about how radiologists organise their day, so before designing I used Perplexity’s research model to find published work on the measures radiology departments track, such as report turnaround, critical results and workload, and on dashboards that had been built for them. I read seven of those sources closely, and they are listed at the end of this study.

The resulting dashboard opens on the resident’s priority tasks and deadlines, with a study list below in which each patient expands to show the series available and the modes in which they can be opened: the basic viewer, segmentation and the generative AI mode. Segmentation compares two series side by side and keeps tracked measurements in a panel that can be exported. I drew two options for its side panels. Option A gave each study a large thumbnail with tabs for primary, recent and all studies, and split the measurements panel into segmentation and measurement tabs with grouped tracked and non-tracked lists. Option B showed studies in a two-column grid under a single dropdown, which fits more series on screen, and simplified the measurements into one list with counts. Option B became the final design.