Benchmarking augmentation to enhance A-line and B-line classification with sparse data

In real-life medical applications, there is often no curated data available for training machine learning solutions. In this article, we explore the efficacy of various ways of addressing this issue, such as extending the training set using historical data, data transformation, and data augmentation. We conducted experiments on lung ultrasound data for A-line and B-line artifact detection. The dataset used was acquired from real-life examinations done in a hospital where physicians labeled approximately every tenth frame, resulting in sparsely labeled data. We expanded this dataset using historical ultrasound scans from various probe types and evaluated two transformation strategies—standard resizing and edge padding—alongside several augmentation methods. The biggest improvement in F1-score was from 0.69 to 0.86 for A-line detection, and from 0.53 to 0.8 for B-line detection. These results prove the utility of carefully chosen data transformation and augmentation when training on sparsely labeled data.