The main findings of our study were as follows: (1) AI-POCUS had excellent performance at detecting patients with COVID-19 pneumonia (sensitivity 92.3% and specificity 100%), and (2) the zone-level sensitivity was moderate (37.3%), although specificity was very high (97.8%). Even with moderate sensitivity in each zone, pneumonia in COVID-19 usually spreads to multiple lung zones, leading to a high sensitivity for each patient. Notably, these excellent results were achieved with AI-POCUS by a novice observer who had minimal experience in lung-POCUS.
With the recent miniaturization of ultrasound devices and advancements in image quality, lung-POCUS is becoming a popular examination, especially in intensive care [8, 16]. Although the lung itself, which is filled with air, cannot usually be observed by ultrasound beams, the noise and artifacts generated by ultrasound beams provide useful clinical information in lung-POCUS. Ultrasound B-lines are one of the most useful artifacts that present with pulmonary congestion, either by pneumonia or cardiogenic pulmonary edema. These reverberation artifacts originate at the pleura, reflecting an air-fluid mixture, which occurs when the subpleural interlobular septa surrounded by subpleural air-filled alveoli become edematous. Previous studies have reported that three or more B-lines visible in a single ultrasound plane are fully sensitive and specific to demonstrate subpleural thickened interlobular septa and/or ground-glass areas with a CT scan as a reference .
Diagnosis of pneumonia using lung-POCUS techniques, including the B-line, is expected to be an effective tool in the ongoing COVID-19 pandemic. Reports have already shown that B-lines are not only sensitive, but also associated with disease severity, future deterioration, and treatment effects in COVID-19 [15, 18–22]. However, chest radiography and CT remain the leading examinations used in the management of COVID-19 worldwide, and lung-POCUS has not been sufficiently used. Technical difficulties and problems in interpreting images are major concerns when using lung-POCUS; the number of experts is not large enough to teach and supervise the use of lung-POCUS by clinicians, including general practitioners, although this technique is relatively new and may be difficult for novice observers.
In the present study, we demonstrated that with AI technology, lung-POCUS can be effective to diagnose pneumonia, with excellent accuracy, even by a novice observer. Recent advancements in AI technology, more specifically machine learning, have enabled automated image recognition with similar or even higher accuracy compared to expert clinicians, and the application of such technologies to POCUS has been enthusiastically studied [10, 11, 23, 24]. The automated B-line counting application that we used in this study was also developed using machine learning technology, and showed excellent agreement with expert readings [25, 26].
The moderate zone-level accuracy observed herein may be due to the limited ability to detect slight pneumonia in the central area of the lung far from the body surface. Such slight pneumonia in a deep area of the lung may not affect the pleura, and would therefore not generate a B line. However, pneumonia caused by COVID-19 usually spreads to multiple regions of the lungs, appearing in the form of acute respiratory distress syndrome . Therefore, diffuse pneumonia caused by COVID-19 is a suitable target for AI-POCUS.
COVID-19 is expected to persist for some time. AI-POCUS allows non-specialist general practitioners to diagnose pneumonia, a problem in viral infections, in an access-free manner, which may be useful in the treatment of pneumonia during a pandemic, and is expected to become more widespread.
This study has some limitations. First, this was a single-center study conducted in a university hospital, and included a small number of cases. Larger studies involving various settings are necessary before the current results can be widely implemented. Second, we did not check the accuracy of AI-POCUS compared with traditional lung POCUS by an expert. However, a previous study showed that AI-POCUS has over 90% sensitivity and specificity for detecting expert-annotated B-lines . Next, we studied only hospitalized patients who were considered sicker than those who were cared for at home or at other facilities. The performance of the present application should be tested in less intensively sick patients to uncover it’s applicability for screening purposes.