Multi-modal lung ultrasound image classification by fusing image-based features and probe information

Lung ultrasound is a widely used portable, cheap, and non-invasive medical imaging technology that can be used to identify various lung pathologies. In this work, we propose a multi-modal approach for lung ultrasound image classification that combines image-based features with information about the type of ultrasound probe used to acquire the input image. Experiments on a large lung ultrasound image dataset that contains images acquired with a linear or a convex ultrasound probe demonstrated the superiority of the proposed approach for the task of classifying lung ultrasound images as “COVID-19”, “Normal”, “Pneumonia”, or “Other”, when compared to simply using image-based features. Classification accuracy reached 99.98% using the proposed combination of the Xception pre-trained CNN model with the ultrasound probe information, as opposed to 96.81% when only the pre-trained EfficientNetB4 CNN model was used. Furthermore, the experimental results demonstrated a consistent improvement in classification performance when combining the examined base CNN models with probe information, indicating the efficiency of the proposed approach.