Can artificial intelligence assess image quality in point-of-care ultrasound?

Over the past decade, a growing number and range of physicians have implemented point-of-care ultrasound (POCUS) into clinical practice. Since diagnostic ultrasound is one of the most operator-dependent imaging modalities, there is a need for standardization of POCUS use among novice users to reduce the risk of misdiagnoses and incorrect treatments. Our goal is to help novice operators to obtain and use high quality images for interpretation by providing real-time artificial intelligence (AI) based feedback on image quality. This study evaluates the accuracy of the AI system for assessing image quality, as compared to subjective expert analysis (ASE Level III echocardiographer), in both cart-based transthoracic echocardiogram and POCUS. For a prospective study of 53 patients, an experienced sonographer performed both POCUS (Lumify, Philips) and cart-based ultrasound (iE33, Philips; Vivid E9, General Electric) to acquire 9 standard image views: parasternal long axis, parasternal short axis (aortic, mitral, papillary muscle, and apical levels), apical 2 chamber, apical 4 chamber, subcostal 4 chamber, and subcostal inferior vena cava. The images were randomized and scored by a clinical expert (ASE Level III echocardiographer) based on a subjective rating system. This previously validated system classifies image quality into 4 categories, based on percentage of endocardial border visibility (0-25%, 25-50%, 50-75%, 75-100%). Images were also assigned image quality scores by the AI system, which uses a combination of deep convolutional layers to extract relevant features and Long Short-Term Memory layers to extract temporal dependencies. The system has been trained on 17,400 cart-based cines of 14 different cardiac views. Clinical expert and AI scores were compared. On a database of 1,267 images, we achieved an overall accuracy of 84% with respect to subjective expert scores. The AI system in cart-based ultrasound had comparable accuracy to POCUS (Table 1). In a clinical setting, the AI system can assess image quality with modest accuracy. While trained on cart-based images, the AI system can function equally well in compact, less sophisticated POCUS devices.