Video-level Hierarchical Binary Classification of Lung Ultrasound Clinical Data

Deep learning approaches in lung ultrasound (LUS) imaging classification traditionally focus on the frame-level prediction, employing a severity score system ranging from score 0 to score 3. In clinical practice, LUS patterns are evaluated at the video level. Current threshold-based method aggregates scores from frame to video level by finding a single optimal threshold across all score levels. To better align with clinical evaluation practices, we proposed the first hierarchical binary classification (HBC) algorithm to find optimal thresholds for each score level. Top-to-bottom (score 3 to 0) and bottom-to-top (score 0 to 3) approaches are explored, where optimal thresholds serve as binary classifiers to separate videos of current score from non-current score groups. We evaluated the generalization capability of the HBC approach on 2 LUS datasets that contain different patient populations. Results show that hierarchical binary classification algorithm performs comparably to state-of-the-art, demonstrating its strong generalization ability across different patient populations.