Improving echocardiogram view classification using diffusion models

Abstract

In the field of medical science datasets are often highly imbalanced, where rare datapoints are of high importance. This study aims to explore the usage of synthetic datasets to improve the classification of echocardiogram views. We join together different echocardiogram datasets (EchoNet-LVH, EchoNet-Dynamic, TMED-2) to form a custom imbalanced dataset (N=38,915) including Apical two-chamber (A2C), Apical four-chamber (A4C), Parasternal long axis (PLAX), and Parasternal short axis (PSAX) views. We study the results of training a diffusion model on differently sized subsets of this dataset. For each size of real subset available, we train two echocardiogram view classifiers on (i) the real data subset and (ii) on the synthetic subset, generated from training on the real subset.

Our results show that synthetic data can be used to improve echocardiogram view classification performance. Specifically we prove that the classification performance of minority classes is significantly improved when the imbalanced dataset is limited. The percentage of images that get correctly classified, for the minority classes A2C and PSAX, increases from 0 to 0.83 and from 0 to 0.77 respectively, when the real subset is limited to 1,000 examples.