Intelligent interpretation of four lung ultrasonographic features with split attention based deep learning model. (March 2023)

• Record Type: Journal Article
• Authors: Chen, Jiangang; Shen, Mengjun; Hou, Size; Duan, Xiaoqian; Yang, Minglei; Cao, Yucheng; Qin, Wei; Niu, Qiang; Li, Qingli; Zhang, Yi; Wang, Yin
• Is Part Of: Biomedical signal processing and control. Volume 81(2023)
• Publication Date: 2023-03
• Subjects: Lung -- Deep learning -- Pleural effusion -- A/B-line -- Pulmonary consolidation; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.104228
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: To the best of our knowledge, this is the first work to propose an end-to-end deep learning model to classify the most common four LUS features (A-line, B-line, pulmonary consolidation, and pleural effusion). The proposed Mish activation function and Split-Attention structure effectively solve multi-scale feature extraction in LUS images and significantly improve the performance of the Convolution Neural Network(CNN) model. An independent test showed that our model has good generalization ability. Grad-cam method was applied to visualize the model to improve the interpretability of LUS. Abstract: Objectives: To develop and validate a deep learning (DL) model based on multi-scale features of Lung ultrasound (LUS) and attention mechanism to detect A-line, B-line, pulmonary consolidation, and pleural effusion caused by pulmonary gas–liquid ratio variations. Methods: A total of 6000 LUS images were prospectively collected from 3966 patients, of which 5545 images were selected. All the images were randomly divided into the training set (4, 436 images) and the testing set (1, 109 images) with a ratio of 4:1. Faced on multi-scale features of LUS, an end-to-end deep learning model based on multi-scale split attention and Mish function was proposed to automatically identify the four LUS features. Results: The overall prediction AUC, accuracy, specificity, and sensitivity of the independent test set were 99.76%, 98.20%, 99.41%, and 98.27%, respectively, and achieved significant and consistent improvement as compared to other deep learning baselines. Conclusions: Our proposed model could interpret the four important LUS features intelligently and be adopted as a support system in the routine diagnosis of an emergency clinician. Significance: This study can not only assist clinicians in recognizing common lung lesions but also provide a new method for the realization of high-quality intelligent diagnosis.