A Novel Hybrid Model Based on GAN and Stacking for Improving Minority Class Detection in an Imbalanced Dataset of Endotracheal Tubes Position Errors.

Authors

Abstract

Endotracheal tube (ETT) abnormal positioning diagnosis based on chest X-ray images can lead to errors, such as misclassifying malpositioned tubes as normal or normal tubes as abnormal. These two types of errors have unequal consequences. The first error, where malpositioned tubes are classified as normal, can result in fatal complications due to the failure to adjust the tube. The second error, where normal tubes are misclassified, is not harmful, as no intervention is required. However, the latter error may lead to unnecessary use of medical resources, such as physician time. Deep learning models often treat these errors equally when evaluated using global metrics such as accuracy, which provides an average view of normal and abnormal samples. Additionally, the datasets used for training ETT classification models are extremely imbalanced, with abnormal tubes as the minority class. This imbalance highly affects the classification performance for this critical class. In this work, we address this issue by balancing an ETT-tubes public dataset through the generation of synthetic features for the minority classes. We employ a two-stage method based on generative adversarial network (GAN) and oversampling to synthesize sample features for the first stage, then a stacking ensemble of multi-classifiers trained with those feature-balanced datasets for the second stage. The resulting model demonstrates improved performance for the minority class compared to existing approaches, achieving a recall of 75% for the abnormal class, while maintaining a strong overall performance with an average F1-score of 78% across all classes, while significantly reducing the computation cost by training the classifiers at the feature's level instead of the image's pixel level.

Tags

Topics