Point of care lung ultrasound (POCUS-L) and the interpretation of lung artifacts has recently emerged as key assessment for patients with acute respiratory illnesses. In the era of COVID-19, POCUS-L can be used to risk stratify, and diagnose COVID-19 pneumonia preceding a formal diagnosis by polymerase chain reaction (PCR). There has been recent development of artificial intelligence (AI) software to automate analysis of ultrasound images for clinicians without formal training. Given the magnitude of COVID-19, AI-assisted POCUS-L has the potential to change the diagnostic landscape empowering frontline clinicians from tertiary care centres to rural communities. While there have been published deep learning models for the interpretation of POCUS-L, images can be challenging for AI networks for 2 reasons: 1. large datasets are not currently available and 2. the presence of lung artifacts can be sparse within the numerous frames of a video. To overcome these limitations, we propose a knowledge transfer approach for the former, and an attention-based model for the latter. Using an attention-based model allows the network to narrow focus on relevant frames improving diagnostic precision and accuracy.
METHODS AND RESULTS
In our design we modify the current open-source state-of-the-art model (POCOVID-NET). A convolutional neural network (CNN) extracts spatially encoded features from POCUS-L images, which are fed to a novel attention-based transformer encoder to capture temporal information across frames, which then narrows focus to key frames. We guide the network learn relevant features (A and B lines) by training it on a pulmonary biomarker detection task from our own private dataset of lung images (Fig 2, step 1). We then transferred the knowledge learned and apply this novel attention-based model to a publicly available POCUS-L image dataset set consisting of patients with 1. healthy lungs 2. bacterial lobar pneumonia, and 3. COVID-19 viral pneumonia. We performed a performance study comparing diagnostic precision and recall compared to the POCOVID-Net model. Our novel attention-based model achieves 85% precision and 90% recall for COVID-19, an improvement of 4 % and 10% respectively over the previous model (Table 1).
In this study, our novel attention-based machine learning model outperformed the state-of-the-art model for COVID-19 diagnosis by POCUS-L. By integrating an attention mechanism across video frames, our models leverages both spatial and temporal information highlighting key frames used to make the prediction. To our knowledge, this is the highest performing model on this POCUS-L dataset diagnosing COVID-19.