A Novel Weighted Majority Voting-Based Ensemble Framework for Lung Ultrasound Pattern Classification in Pneumonia Patients

Towards automating semi-quantitative analysis of lung ultrasound (LUS) data, various deep learning-based (DL) classification models have been developed to detect LUS patterns among pneumonia patients. These patterns include horizontal artifacts, vertical artifacts, and small to large consolidations. These models showed overall promising results, however, did struggle to obtain satisfactory performance in classifying all these patterns correctly. In this paper, we propose an ensemble framework in which multiple classification models are employed and their contributions to the final prediction are adjusted based on a weighed voting mechanism. The weight of each model's prediction is assigned based on the confidence level for that prediction and the model's overall performance for that predicted class. As a result, the final prediction does not only depend on the majority predicted class, but also on the class with the maximum weight, among the predicted classes. As a proof of concept, we...

Towards automating semi-quantitative analysis of lung ultrasound (LUS) data, various deep learning-based (DL) classification models have been developed to detect LUS patterns among pneumonia patients. These patterns include horizontal artifacts, vertical artifacts, and small to large consolidations. These models showed overall promising results, however, did struggle to obtain satisfactory performance in classifying all these patterns correctly. In this paper, we propose an ensemble framework in which multiple classification models are employed and their contributions to the final prediction are adjusted based on a weighed voting mechanism. The weight of each model’s prediction is assigned based on the confidence level for that prediction and the model’s overall performance for that predicted class. As a result, the final prediction does not only depend on the majority predicted class, but also on the class with the maximum weight, among the predicted classes. As a proof of concept, we employed three different classification models (RegNetX, ResNet50, and ResNet18 with spatial attention) in the ensemble framework. The proposed framework is evaluated to classify patterns in LUS frames from ICLUS-DBv1 dataset. When comparing individual performances of the models with the ensemble approach, a noticeable improvement in overall classification is found resulting in an F1-Score of 0.685. Moreover, pattern-wise classification is also enhanced in the proposed approach in-comparison to that of the individual models. On comparison with the existing state-of-the-art techniques, our proposed weighted majority voting-based ensemble approach gives a comparable performance in classifying LUS patterns.