Method Overview
Overview of CHD-FM for first-trimester congenital heart disease analysis.
Effective first-trimester screening for congenital heart disease (CHD) remains an unmet clinical need, and automated analysis is critical for clinical decision-making. Despite recent advances of foundation models in clinical workflows, the ability to analyze CHD remains limited, primarily due to the lack of CHD-oriented reasoning annotation. This paper aims to bridge this gap. Specifically, we collect a large cohort of first-trimester cardiac screenings across multiple regions in China, comprising 10,821 image-pathology pairs with clinical imaging phenotypes. Based on this cohort, we develop CHD-FM, a versatile foundation model tailored for first-trimester CHD analysis. Through a multi-stage training paradigm, including visual domain adaptation and clinical phenotypes reinforcement, CHD-FM systematically acquires disease diagnosis knowledge and reasoning interpretive capabilities for the CHD scope. In rigorous evaluations, CHD-FM consistently outperformed state-of-the-art commercial and open-source foundation models, and matched or surpassed experienced clinicians across retrospective, prospective, and external cohorts.
Overview of CHD-FM for first-trimester congenital heart disease analysis.
Performance comparison with closed commercial models and open-source foundation models across retrospective and out-of-domain cohorts.
Diagnostic reasoning examples, confusion matrices, and clinical physician assessment of CHD-FM and comparison models.
The public repository is code-only and does not include fetal ultrasound images, patient tables, annotations, trained checkpoints, generated predictions, commercial API outputs, logs, or doctor-scoring results.