Lung Ultrasound Segmentation and Adaptation between COVID-19 and Community-Acquired Pneumonia

Lung ultrasound imaging has been shown effective in detecting typical\npatterns for interstitial pneumonia, as a point-of-care tool for both patients\nwith COVID-19 and other community-acquired pneumonia (CAP). In this work, we\nfocus on the hyperechoic B-line segmentation task. Using deep neural networks,\nwe automatically outline the regions that are indicative of pathology-sensitive\nartifacts and their associated sonographic patterns. With a real-world\ndata-scarce scenario, we investigate approaches to utilize both COVID-19 and\nCAP lung ultrasound data to train the networks; comparing fine-tuning and\nunsupervised domain adaptation. Segmenting either type of lung condition at\ninference may support a range of clinical applications during evolving epidemic\nstages, but also demonstrates value in resource-constrained clinical scenarios.\nAdapting real clinical data acquired from COVID-19 patients to those from CAP\npatients significantly improved Dice scores from 0.60 to 0.87 (p < 0.001) and\nfrom 0.43 to 0.71 (p < 0.001), on independent COVID-19 and CAP test cases,\nrespectively. It is of practical value that the improvement was demonstrated\nwith only a small amount of data in both training and adaptation data sets, a\ncommon constraint for deploying machine learning models in clinical practice.\nInterestingly, we also report that the inverse adaptation, from labelled CAP\ndata to unlabeled COVID-19 data, did not demonstrate an improvement when tested\non either condition. Furthermore, we offer a possible explanation that\ncorrelates the segmentation performance to label consistency and data domain\ndiversity in this point-of-care lung ultrasound application.\n