Lightweight Deep Learning Model for Lung Ultrasound Image Scoring

This study introduces a lightweight deep learning model for qualitative scoring of lung ultrasound (LUS) images, aiming to improve the clinical applicability and deployment efficiency of intelligent ultrasound in the rapid and accurate assessment of pulmonary conditions. A total of 6,000 LUS images were utilized, with 5,000 allocated for model development and 1,000 reserved for independent external validation. The model architecture employs an improved FastViT as the backbone, incorporating the channel shuffle mechanism from ShuffleNet into the FastViT stem structure to strengthen shallow local detail representation and early-stage inter-channel interaction. This integration yields an end-to-end lightweight scoring network that balances inference efficiency and discriminative performance. Experimental results demonstrate that the proposed model contains merely 3.2M parameters with about 700M FLOPs, achieving an accuracy of 96.7% on the test set and 95.14% on the independent external test set, outperforming mainstream lightweight baselines such as ShuffleNet and FastViT. These findings indicate that the proposed LUS qualitative scoring model delivers efficient and precise assessment of disease severity, offering advantages of ease of deployment and rapid inference. It thus provides substantial support for clinical decision-making and early intervention, showcasing significant potential for real-world clinical application.