A dual modal collaborative learning network for echocardiography video segmentation

Abstract

As a core parameter for heart failure classification and therapeutic decision-making, the accuracy of ejection fraction measurement directly influences patient management and prognostic evaluation. Accurate left ventricular segmentation is fundamental to the reliable calculation of ejection fraction.Although existing artificial intelligence based medical image segmentation methods have achieved certain progress, their effectiveness remains limited under conditions of small sample sizes and insufficient annotations. To address this issue, we propose a dual-modal collaborative learning-based echocardiography video segmentation framework. Specifically, we first introduce a teacher-student network based dual modal collaborative learning framework. In this framework, the main branch acts as the teacher, taking high-quality blood flow mode ultrasound images as input, extracting deep features through the DINOv3 pretrained model, and learning valuable knowledge from the limited annotated samples. The auxiliary branch serves as the student network, taking low-quality B-mode ultrasound images as input and performing supervised learning based on the teacher network's output. Inspired by the concept of mutual growth through teaching, we design a spatial consistency injection block, which the student network optimizes the teacher network in a reverse manner, achieving a bidirectional mutual enhancement effect. Additionally, the main branch network introduces a motion aligned cross attention, which uses context reference frames to provide extra temporal information for the current frame, thereby further improving segmentation accuracy. Extensive experiments on a collected mouse echocardiography dataset, we validate the effectiveness of the proposed method, achieving state-of-the-art segmentation performance and accurate ejection fraction calculation.MethodResult