One model segments four echocardiography views: P2Echo swaps per-view specialist networks for typed text prompts
Synopsis
P2Echo folds multi-view echocardiography segmentation into a single network that takes a structured text prompt naming the view and target structure and uses it to modulate decoder features. Previously each standard acoustic window typically needed its own specialist model, or relied on implicit priors that were hard to control explicitly. On internal data, P2Echo improves over EchoONE, the other multi-view model, by 5.3, 1.1 and 2.8 Dice on 2CH, 3CH and 4CH, is comparable on PSAX, and holds competitive results on three external datasets at 6.90 GFLOPs.
Interpretation
Multi-view echocardiography segmentation no longer needs a separate network per acoustic window; one model switches its processing path according to the prompt. Earlier approaches either trained a specialist per view or used multi-branch architectures in which a new view required a new parallel branch, while classic architectures trained on pooled multi-plane data performed markedly worse than plane-specific training. On the internal test split, P2Echo improves over EchoONE, the only prior work targeting multi-view echo segmentation, by 5.3, 1.1 and 2.8 Dice on 2CH, 3CH and 4CH and is comparable on PSAX, using one set of parameters rather than per-plane weights or learnable dense prompts.
The control signal is text an operator can write directly, not an implicit prototype learned from data. SAM-based universal models derive dense prompts from clustering-based structural prototypes, so the guidance depends on how well those priors represent the input anatomy and is difficult to specify explicitly. Prompts follow a fixed template of View, Target and Description and are encoded by a frozen Qwen3-Embedding-0.6B; on 450 images and 950 annotated structure instances from the internal test split, the canonical template reaches 93.53% Dice.
Text guidance continuously modulates visual features at every decoder resolution rather than only localizing targets before mask decoding. Earlier text-guided segmentation methods mainly used language to guide target localization or visual-prompt generation before mask decoding, without continuously modulating decoder features at multiple spatial resolutions. Ablations on HMC_QU show that replacing dynamic projection with a static linear layer, removing the dynamic measure kernel, and removing the cross-modal differential operator cost 2.18, 1.70 and 2.19 mDice and add 0.39, 0.43 and 0.92 HD95.
Perspective
The experiments cover four standard echocardiographic views (2CH, 3CH, 4CH and PSAX) and assume the view identity is available when constructing the prompt; five datasets are used for internal training and evaluation, with three external datasets held out for evaluation. The authors propose extending P2Echo to additional views and investigating whether improved segmentation translates into more accurate and reproducible cardiac measurements and functional assessment.
The authors state that the present results do not establish generalization to previously unseen views; 3CH and PSAX supervision each come from only one development dataset; the model is sensitive to prompt wording, with clinical paraphrases and concise prompts reaching 80.48% and 35.83% Dice respectively and concise prompts showing high variance; and evaluation covers segmentation accuracy only, not downstream clinical measurements.