Lightweight Lung Ultrasound Video Analysis Model

Ultrasound, as a fast, convenient, non-radiation imaging technology, has been widely used in lung disease diagnosis and bedside monitoring in clinic. Due to the dynamic changes in lung tissue with respiratory movement, video-level based research has become a current research hotspot. Existed lung ultrasound (LUS) video model mostly have complex structures and numerous parameters, cannot achieve effective video compression. Thus, how to design a lightweight lung ultrasound (LUS) video analysis model is of great value for clinical application. In this paper, the proposed novel lightweight LUS video analysis model was mainly composed of four parts: similar frame filtering, channel aggregation, improved attention-based encoding analysis, and classification. Firstly, entropy difference characterization was employed to analyze and filter these adjacent frames with high similarity, implementing the selection of 20 initial keyframes in each LUS video. Then, the recursive convolutional layers with different kernel sizes were designed to further achieve channel aggregation (C=3). This will enable the model to achieve parameter reduction while retaining most video information. Next, after patch positional encoding, the improved self-attention module was used to analyze each patch. By using dilated convolution instead of linear layer to process original input X and matrix V, it can enrich input image information and comprehensively represent local and global information. Meanwhile, a spherical plane method was designed to uniformity normalize Q and K, enhancing the correlation representation effect. Finally, the MLP head was applied for the automatic LUS video scoring. 1672 LUS videos were collected to validate the proposed video scoring model, experimental results demonstrate that the model's accuracy achieved 90.91% and the size was 2.5Mb, with great application potential in clinic.