NTH

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

AuthorsMoein Heidari, Alvin Kimbowa, Mahdi Tabatabaei, Afshin Bozorgpour, Sina Ghorbani Kolahi, River Jiang, Wenjin Chen, David J. Foran, Jasmine Grewal, Dorit Merhof, Ilker Hacihaliloglu

AffiliationsSchool of Biomedical Engineering, The University of British Columbia, Vancouver, BC V6T 1Z4, Canada

November 13, 2026 2 min read
Watch on YouTube
The one-line take

P2Echo uses view- and structure-aware text prompts to guide one model in segmenting cardiac anatomy across multiple echocardiography views.

Key results

96.94%
2CH Dice

Internal test-set Dice score for the 2CH view.

2.33%
Median Dice gain

Median improvement over the strongest baseline across evaluated datasets.

82.22%
HMC_QU Dice

Held-out external-dataset Dice score.

92.95%
CardiacNet Dice

Held-out external-dataset Dice score.

6.90
Compute

Inference cost in GFLOPs.

64.68
Throughput

Frames per second measured on an NVIDIA H100.

What the paper found

Prompt2Echo, or P2Echo, replaces separate echocardiography segmentation specialists with one model controlled by structured prompts naming the view and cardiac structure. Its core innovation, Dynamic Image-Text Alignment, or DyITA, uses straight-through hard routing, adaptive feature-sharpening kernels, and a cross-modal differential operator to inject prompt guidance into decoder features at multiple resolutions. The model combines a PVTv2-B2 visual encoder with frozen Qwen3-Embedding-0.6B text embeddings, and handles 2CH, 3CH, 4CH, and PSAX views. Across five development datasets, P2Echo reports a 2.33% median Dice gain over the strongest baseline; on 2CH, Dice reaches 96.94%. On three held-out datasets, Dice is 82.22% on HMC_QU, 87.79% on EchoNet-Dynamic, and 92.95% on CardiacNet. It requires 6.90 GFLOPs and runs at 64.68 FPS, measured on an NVIDIA H100. Compared with alternatives including SAM3 and EchoONE, the results show strong accuracy with substantially lower computational cost. One caveat: performance depends on the canonical structured prompt format; free-form clinical paraphrases reduce accuracy.

Original abstract

Automated segmentation in multi-view echocardiography is essential for comprehensive cardiac assessment, yet remains challenging due to substantial anatomical and appearance variability across standard acoustic windows. In this paper, we present P2Echo, a unified framework capable of segmenting diverse cardiac structures across multiple standard echocardiographic views using a single model controlled by structured view-aware text prompts. Unlike recent foundation model adaptations that rely on implicit latent priors, P2Echo explicitly conditions the segmentation process via a novel Dynamic Image-Text Alignment (DyITA) module. DyITA integrates a differentiable routing mechanism and token differential operators to dynamically modulate decoder features according to the requested view and target structure. This allows the network to adaptively switch its processing pathways for different views without architectural changes. Extensive experiments on a large-scale internal dataset (seen during training) and held-out evaluation on three external datasets demonstrate that P2Echo matches or exceeds both view-specific specialists and state-of-the-art universal models at a fraction of their computational cost, offering an efficient, controllable solution for multi-plane cardiac analysis. Code is available at: https://github.com/moeinheidari7829/P2Echo.

Read the original paper

More in Computer Vision

Browse all 82 papers →