Combining quantitative and qualitative knowledge for scoring pleural
line in lung ultrasound
Abstract
With the development of lung diseases and the wide application of lung
ultrasound (LUS), the independent analysis of various indicators in LUS
images are of great significance in clinic. In this paper, we proposed a
quantitative and qualitative method for extracting, analyzing, and
scoring pleural lines with different lesions in LUS images. The
extraction module was composed of the customized cascaded localization
and segmentation models based on convolution and multilayer perceptron.
The analysis module used eight textural and three morphological
parameters to quantitatively analyze the features of two different
output images from the localization and segmentation models,
respectively. The scoring module adopted four supervised machine
learning classifiers including support vector machine, k-nearest
neighbor, random forest, and decision tree to qualitatively evaluate
pleural lines with different severities in LUS images. The experiments
were performed on the 5390 LUS images (i.e., segmentation: 1620,
scoring: 3770) acquired from Coronavirus Disease 2019 pneumonia patients
with convex ultrasound probes. Experimental results showed that the
proposed line extraction method can accurately achieve the localization
and segmentation of pleural lines. The support vector machine classifier
with combining texture and morphological features as input achieved
optimal scoring performance with the accuracy, sensitivity, specificity,
F1 score, and AUC being 94.47%, 97.31%, 94.50%, 0.9457, and 0.9822,
respectively. Compared with other models, it also proves that the
proposed method has better scoring performance. Thus, the proposed
method has great application potential for clinical application.