Real-time and automatic lung diagnostics by quantitative lung ultrasound spectroscopy, a multicenter study

Lung Ultrasound (LUS) is a real-time, non-invasive imaging modality used to assess the lung surface. However, LUS currently relies on the subjective evaluation of imaging patterns. Quantitative LUS (QLUS) approaches aim at mitigating these drawbacks through the spectral characterization of imaging patterns. However, so far they have been implemented by manual pattern segmentation. Here, we present a novel, automatic and real-time QLUS approach to support the clinical diagnosis of four lung diseases. Radiofrequency in-vivo data were acquired by means of a multifrequency approach (3–6 MHz) from two hospitals. Data from 73 patients, for a total of 10 231 images, were collected. The images are automatically segmented by a novel algorithm into three regions—pre-pleura, pleura, and post-pleura—and quantified, respectively, into and . Additionally and were derived as the ratios and, respectively. Successively, three spectral features were derived for each : center frequency, bandwidth, and . Each patient was characterized by the mean value of each feature. Multiclass classification was tested using four classifiers. Results show the potential of our automatic and real-time QLUS approach in supporting clinical diagnosis among four lung diseases (pneumonia, pulmonary fibrosis, COPD, CPE) with accuracies up to 74%.