EchoONE: Segmenting Multiple echocardiography Planes in One Model

TL;DR

EchoONE presents a uniform segmentation framework for multi-plane echocardiography by integrating a SAM-based backbone with a prior-semantic dense-prompt module (PC-Mask) and a CNN-driven local feature adaption (LFFA). PC-Mask generates plane-aware prompts without plane labels by learning prototypes in a latent space and composing tailored prompts for SAM, while LFFA fuses local CNN features into SAM's decoder to enhance segmentation and convergence. The approach achieves state-of-the-art results across six echocardiography planes on internal and external datasets, demonstrating strong cross-plane generalization and robustness to domain shifts. This work delivers the first unified solution for multi-plane echocardiography segmentation and suggests a path toward broader MPS applications in medical imaging.

Abstract

In clinical practice of echocardiography examinations, multiple planes containing the heart structures of different view are usually required in screening, diagnosis and treatment of cardiac disease. AI models for echocardiography have to be tailored for each specific plane due to the dramatic structure differences, thus resulting in repetition development and extra complexity. Effective solution for such a multi-plane segmentation (MPS) problem is highly demanded for medical images, yet has not been well investigated. In this paper, we propose a novel solution, EchoONE, for this problem with a SAM-based segmentation architecture, a prior-composable mask learning (PC-Mask) module for semantic-aware dense prompt generation, and a learnable CNN-branch with a simple yet effective local feature fusion and adaption (LFFA) module for SAM adapting. We extensively evaluated our method on multiple internal and external echocardiography datasets, and achieved consistently state-of-the-art performance for multi-source datasets with different heart planes. This is the first time that the MPS problem is solved in one model for echocardiography data. The code will be available at https://github.com/a2502503/EchoONE.

Figures (8)

• Figure 1: Comparison of existing solutions for the Multi-plane segmentation problem in echocardiographic images. The performance of classic U-Net significantly decreases when trained on multiple planes together (b), compared to training on each plane individually (a). The direct application of SAM on different echocardiographic planes yields poor results (c). EchoONE demonstrates excellent segmentation performance across all planes (d).
• Figure 2: Overview of our EchoONE, which contains a SAM-based segmentation architecture (middle), a prior-composable mask learning (PC-Mask) module (top) for dense prompt generation, and a learnable CNN-based local feature branch (bottom) for SAM encoder tuning and decoder adapting.
• Figure 3: Details of our prior-composable mask learning (PC-Mask) module, which leverages prior structural information to automatically generate high-quality semantic-aware dense prompts for images without knowing the plane information.
• Figure 4: Local feature fusion and adaption (LFFA). The output of each CNN layer (local feature) and image embeddings from a previous transformer is first concatenated and then mixed with a convolution layer to obtain the fused image embedding.
• Figure 5: (a) Remapping annotations to unified mask representation. (b) Preprocessing to generate LV cavity from the annotation of MYO in multiple planes. Landmarks are required to be detected first for planes of 2CH, 3CH and 4CH.