ETAB: A Benchmark Suite for Visual Representation Learning in Echocardiography

Echocardiography is one of the most commonly used diagnostic imaging modalities 1 in cardiology. Application of deep learning models to echocardiograms can enable 2 automated identification of cardiac structures, estimation of cardiac function, and 3 prediction of clinical outcomes. However, a major hindrance to realizing the full 4 potential of deep learning is the lack of large-scale, fully curated and annotated data 5 sets required for supervised training. High-quality pre-trained representations that 6 can transfer useful visual features of echocardiograms to downstream tasks can help 7 adapt deep learning models to new setups using fewer annotated examples. In this 8 paper, we design a suite of benchmarks that can be used to evaluate echocardio- 9 graphic representations with respect to various clinically-relevant tasks using pub- 10 licly accessible data sets. In addition, we develop a unified evaluation protocol— 11 which we call the echocardiographic task adaptation benchmark (ETAB)—that 12 measures how well a visual representation of echocardiograms generalizes to com- 13 mon downstream tasks of interest. We use our benchmarking framework to evaluate 14 state-of-the-art vision architectures, pre-training and transfer learning algorithms. 15 We envision that our standardized, publicly accessible benchmarks would encour- 16 age future research in high-impact application domains and expedite progress in 17 applying deep learning models to practical problems in cardiovascular medicine.