Neural Scaling Laws Cut Model Parameters 240x for Heart Ultrasound Segmentation

A team led by Clara Rodrigo González has introduced a compute-optimal network design method for myocardial segmentation and perfusion quantification using neural scaling laws, typically used to scale large language models. They applied these laws to predict the best network size for two small imaging datasets: the CAMUS echocardiography dataset and a 25-patient contrast-enhanced ultrasound (CEUS) dataset. By training models on data subsets and extrapolating test loss, they identified two networks that achieved state-of-the-art performance on CAMUS with a staggering 240-fold reduction in parameter count. The scaling law gradient transferred from CAMUS to CEUS with only a bias in predicted losses, demonstrating cross-domain applicability.

Crucially, the automatically segmented masks performed equivalently to a senior cardiologist in myocardial perfusion quantification, proving clinical utility. This approach addresses the chronic challenge of data scarcity in medical imaging by enabling efficient, data-driven model design without massive datasets or GPU clusters. The work establishes neural scaling laws as a practical tool for small imaging datasets, potentially accelerating adoption of AI-assisted contrast echocardiography for bedside, non-ionizing perfusion assessment. It bridges the gap between large-scale NLP efficiency gains and resource-constrained medical imaging applications.