Learning-based Pulmonary Disease Detection Using Weak Labels for Volume Sweep Lung Ultrasound Imaging

Pulmonary disease is prevalent worldwide resulting in significant morbidity and mortality, but most of the world lacks accessible medical imaging for its assessment. Lung ultrasound is a highly accurate and relatively inexpensive tool for diagnosing pulmonary illness, but its deployment is limited by a lack of trained sonographers. Volume sweep imaging (VSI) of the lung circumvents the problem by shifting image acquisition to individuals with minimal medical and ultrasound background by using standardized ultrasound probe sweeps of the chest (recorded as cine loops) based on external body landmarks. However, significant clinical expertise is still needed to correctly interpret and identify the presence of abnormalities in the scan limiting its impact.In this paper, a novel learning-based approach for automated image interpretation of lung ultrasound VSI has been developed to reduce this dependence on experts. Ultrasound B-mode cine loops of sweeps of the right and left posterior lung, acquired with a commercial ultrasound scanner during routine clinical VSI exams, were used for analysis. Two custom hybrid Convolutional Neural Network (CNN)-based deep learning architectures, including one combined with machine learning algorithms, were used to perform binary classification (abnormal/normal) and compared with the ground truth provided by expert radiologists. Moreover, this learning-based approach adopts weak labels during training, where the diagnosis of the entire cine loop is used as the ground truth which is more cost-effective and less time-consuming than labeling each frame individually. Two methods of aggregating the information from individual frames to obtain clip-level predictions were implemented. Experiments were performed to compare the architectures and the aggregation methods and validate their performance against the ground truth. The best performance (accuracy: 91.67%, precision: 92.31%, recall: 85.71%, and F1-score: 88.89%) was obtained for the combined CNN and random forest classifier. The source code is hosted at: https://github.com/timespace22/Lung-Ultrasound-ClassificationClinical Relevance-The results demonstrate potential use of such learning-based approaches for automated image interpretation in lung VSI applications. The method also demonstrates robustness to using the entire ultrasound sweep which may contain extraneous information beyond the desired anatomy as compared to requiring a clinical expert to segment the relevant portion prior to input to the learning-based classifier.