Fully automated Quantitative Lung Ultrasound spectroscopy for the differential diagnosis of lung diseases: The first multicenter in-vivo clinical study

Quantitative Lung Ultrasound (LUS) analyses are demonstrating significant improvements in the diagnoses of lung diseases. However, recent approaches are limited to the manual segmentation of artifactual patterns (i.e., vertical artifacts), whose detection is subjective, dependent on imaging settings, and time-consuming. These limitations can be partially addressed by segmenting anatomical patterns [e.g., the pleural line (PL)], which are less affected by variations in imaging settings. Additionally, an automated segmentation algorithm would reduce segmentation time and subjectivity. In this study, we proposed an Automatic Segmentation Algorithm of the Pleural line (ASAP). ASAP was first optimized on one dataset [composed of 52 patients affected by Chronic Obstructive Pulmonary Disease (COPD), Pulmonary Fibrosis (PF), and Pulmonary Fibrosis Exacerbated (PFE)] and then tested on an external dataset [composed of 34 patients affected by Pneumonia (PNE) and Cardiogenic Pulmonary Edema (CPE)]. In total, 12761 multifrequency radiofrequency images from 86 patients were segmented into three regions: pre-pleura, pleura, and post-pleura. These regions were quantified with five IT parameters (where R = 1, 2, 3, 4, and 5), each characterized with three features: Center Frequency, Bandwidth, and maxIT. The discrimination potential of these features was tested through different binary classifiers. Results show how ASAP segments in quasi real-time with 10 images/s and achieves a segmentation accuracy of 97.70%. The differential diagnosis of lung diseases shows accuracies up to 87.48, outperforming the state of the art.