Interpretable Saab Subspace Network for COVID-19 Lung Ultrasound Screening

Abstract

In addition to bedside Point-of-care diagnosis, lung ultrasound imaging classifier has been used to triage of COVID-19 symptomatic patients after emergency room admission. There is a more urgent need for a simple and low cost portable ultrasound device for each elderly resident at risk in retirement communities and independent living facilities to monitor their lung conditions, to find out if their lung condition is getting worse with COVID-19 during self-quarantine phase or if it's getting better during their recovery phase. Various complicated convolutional neural networks have been developed with very high accuracy but health professionals find it hard to understand and trust such complex models due to the lack of intuition and explanation of their predictions. In this work, we proposed to use an interpre table Subspace Approximation with Adjusted Bias (Saab) multilayer network to screen the lung ultrasound images. Such subspace representations learned from a successive subspace network will provide more invariance to intra-class variability and thus give better discrimination for a task such as classification. We demonstrated the advantage of using Saab Subspace Network to design a low-complexity, low-cost, low-power-consumption solution for interpreting and visualizing features of the lung ultrasound images to confirm the classifier recommendation. Since both training and inference can be done on-device, make it a potential solution to deliver personalized healthcare to underserved senior communities without any internet connection.