Prompt2Echo (P2Echo): Text-Prompted Multi-View Echocardiography Segmentation with One Model

Prompt2Echo (P2Echo). Text-Prompted Multi-View Echocardiography Segmentation with One Model. By Moein Heidari and ten others, from the University of British Columbia, EPFL, and five other institutions. Published on the tenth of November.

A full cardiac assessment looks at the heart through several standard acoustic windows, and automated segmentation across those views is essential to it. The difficulty is that the views do not resemble each other. Anatomy and appearance vary substantially from one standard window to the next, so a structure that is easy to outline in one plane presents quite differently in another. The usual answer is a specialist per view, which multiplies the models you have to train and maintain.

P2Echo is a unified framework that segments diverse cardiac structures across multiple standard echocardiographic views with a single model. What steers it is a structured text prompt that is view-aware: the request names the view and the target structure. Recent foundation model adaptations lean on implicit latent priors to handle this variation. P2Echo instead conditions the segmentation process explicitly, on what was asked for.

That conditioning lives in a module the authors call Dynamic Image-Text Alignment, or DyITA. It combines a differentiable routing mechanism with token differential operators, and uses them to modulate the decoder's features dynamically according to the requested view and structure. The consequence is that the network can switch its processing pathways for different views without any change to the architecture itself. One set of weights, different routes through it.

Experiments run on a large-scale internal dataset, seen during training, plus held-out evaluation on three external datasets. Across these, P2Echo matches or exceeds view-specific specialists and also state-of-the-art universal models, and it does so at a fraction of their computational cost.

Multi-plane cardiac analysis usually trades breadth against expense. Here one controllable model covers the windows that a specialist each would have covered, and the text prompt is the control surface: say which view and which structure, and the routing follows.