Purpose or Case Report: Point of Care Lung ultrasound has proven in multiple studies to be superior to CXR to diagnose pneumonia in children especially in limited resource settings. This non-radiating, portable and adaptable technique, brings an opportunity to detect pneumonia with higher accuracy than CXR. Ultrasound imaging interpretation is challenging. To deal with this complexity, we created a "brightness profiles" data reduction technique to identify specific anatomical structures identified by lung ultrasound using artificial intelligence. We use this technique to demonstrate how data reduction can help identify common anatomical landmarks and abnormal findings, and aid in the interpretation of ultrasound diagnosed pediatric pneumonia.
Methods & Materials: Brightness Profiles compress the high dimensional data of an ultrasound sweep. This approach distills an ultrasound sweep into frames and further portions the frames into columns of a set width (20 pixels). Each column is termed a brightness profile – where the average brightness of the column width is captured as a function of depth (or length of the column). Each profile is used to build a graphical representation of different structures within the lung (i.e. pleural line, ribs, lung parenchyma). From a cohort of 50 pediatrics patients ages 0-59 months, we used lung ultrasound cine images and applied image processing to create Brightness Profiles. We then applied machine learning techniques to automate the detection of the pleural line, and the classification of pneumonia consolidation vs. non-pneumonia consolidation in these images.
Results: The Brightness profiles when used to build a consistent graphical representation of different structures within the lung successfully and consistently identified the pleural line as well as other normal anatomical structures from the ultrasound images selected. It also recognized accurately pleural line disruptions and areas of lung consolidation detected by lung ultrasound images. This machine learning technique performs well to automate the detection of normal and abnormal lung structures initially detected by the ultrasound images.
Conclusions: Using brightness profiles to incorporate normal and abnormal anatomic findings from the ultrasound images into a predictable and accurate graphical representation is a useful artificial intelligence technique that can be integrated into a model for interpretation of ultrasound images of pediatric lung pneumonia based on specific predetermined patterns of lung ultrasound findings.