Designing lightweight deep learning models for echocardiography view classification

Abstract

Transthoracic echocardiography (echo) is the most common imaging modality for diagnosis of cardiac conditions. Echo is acquired from a multitude of views, each of which distinctly highlights specific regions of the heart anatomy. In this paper, we present an approach based on knowledge distillation to obtain a highly accurate lightweight deep learning model for classification of 12 standard echocardiography views. The knowledge of several deep learning architectures based on the three common state-of-the-art architectures, VGG-16, DenseNet, and Resnet, are distilled to train a set of lightweight models. Networks were developed and evaluated using a dataset of 16,612 echo cines obtained from 3,151 unique patients across several ultrasound imaging machines. The best accuracy of 89.0% is achieved by an ensemble of the three very deep models while we show an ensemble of lightweight models has a comparable accuracy of 88.1%. The lightweight models have approximately 1% of the very deep model parameters and are six times faster in run-time. Such lightweight view classification models could be used to build fast mobile applications for real-time point-of-care ultrasound diagnosis.