Search papers, labs, and topics across Lattice.
The paper introduces Diff-MOCS, a diffusion model-based framework for multi-object segmentation of fetal four-chamber ultrasound images, addressing the need for accurate and objective prenatal evaluation of the fetal heart. Diff-MOCS leverages the iterative denoising process of DDPMs to capture fine morphological details and incorporates a MedCLIP-based relative position encoding module to model spatial relationships between cardiac structures. Experiments on a real fetal heart ultrasound dataset demonstrate that Diff-MOCS outperforms existing methods in segmenting multiple structures, including cardiac chambers and the thoracic cavity.
Diffusion models can now accurately segment multiple structures in noisy fetal cardiac ultrasound images, outperforming existing methods by explicitly modeling spatial relationships.
Fetal echocardiography is an essential component of prenatal screening; rapid and accurate analysis of cardiac anatomy can significantly improve the detection rate of congenital heart disease (CHD). However, given the complexity inherent in fetal cardiac ultrasound assessment and the variability among sonographers' diagnostic experience, achieving objective and precise prenatal evaluation of the fetal heart has become an urgent clinical need. Among the scanning standard views used in fetal echocardiography, the thoracic four-chamber view is the most valuable for CHD detection. We therefore propose Diff-MOCS, a novel framework built on the Denoising Diffusion Probabilistic Model (DDPM), for multi-object segmentation of the fetal four-chamber plane. To the best of our knowledge, this is the first attempt to leverage diffusion models for multistructure segmentation in fetal cardiac ultrasound. Diff-MOCS leverages the iterative denoising mechanism of DDPMs to effectively capture fine morphological details and precise boundary information of fetal cardiac structures. To further enhance the model's spatial awareness, we incorporate a relative position encoding module based on MedCLIP, which improves segmentation accuracy by modeling the spatial relationships between different cardiac structures. We validated Diff-MOCS on a real fetal heart ultrasound dataset that included multiple structures, including the fetal thoracic cavity (including and excluding the subcutaneous soft tissue), the cardiac chambers, the left and right atria, and the left and right ventricles. Experimental results demonstrate that Diff-MOCS outperforms competing methods in segmentation performance, proving its effectiveness in tackling the challenges of low-quality medical imaging and providing a new avenue for robust medical image analysis systems.
