Deep Learning Video Classification of Lung Ultrasound Features Associated with Pneumonia

Charles B. Delahunt, Daniel E. Shea, Sourabh Kulhare

Abstract:

Ultrasound (US) imaging holds promise as a low-cost versatile,
non-invasive point-of-care diagnostic modality in lowand
middle-income countries (LMICs). Still, lung US can be
challenging to interpret because air bronchograms are anechoic
and the US images mostly contain artifacts rather than
lung anatomy. To help overcome these barriers, advances in
computer vision and machine learning (ML) provide tools to
automatically recognize abnormal US lung features, offering
valuable information to healthcare workers for point-of-care
diagnosis. This paper describes deep learning algorithms
that target three key US features associated with lung pathology:
pleural effusion, lung consolidation, and B-lines. The
algorithms were developed and validated using a large and
varied dataset of 22,400 US lung scans (videos) from 762
patients of all ages (newborn to adult) in Nigeria and China.
The architectures include effective methods for leveraging
frame-level and video-level annotations, are light enough to
deploy on mobile or embedded devices and have high accuracy
(e.g., AUCs ≈0.9). Coupled with portable US devices,
we demonstrate that they can provide expert-level clinical
assistance for diagnosis of pneumonia, which is the leading
cause of both childhood mortality and adult hospitalization
in LMICs. We also discuss some of the challenges associated
with determining ground truth for pneumonia, which impact
the question of how to leverage ML models for lung US to
support clinical diagnosis of pneumonia.