Feature fusion-based ensemble CNN learning optimization for automated detection of pediatric pneumonia

Abstract

Detecting pediatric pneumonia accurately and rapidly is crucial for timely treatment, especially considering its association with seasonal changes and potentially fatal outcomes. However, medical image analysis using convolutional neural network (CNN) models faces challenges such as limited labeled data, image noise, class imbalance, and overfitting. Regularization techniques are often insufficient, necessitating advanced approaches for successful pneumonia detection. Our study aims to accurately detect pneumonia by proposing an ensemble CNN framework that incorporates optimal feature fusion. A novel image preprocessing algorithm has been developed that applies hierarchical template-matching to reduce image noise and improves the learning of relevant features. Instead of relying solely on a few pre-defined CNN models combined through majority voting, multiple CNN models with different architectures are trained on the pneumonia dataset using fine-tuning and transfer learning techniques. To obtain an optimal feature set, Chi-Square and mRMR methods iteratively eliminate irrelevant features from the fully connected layer of each CNN model. These optimal feature sets are then concatenated to enhance feature vector diversity for classification. The study’s results illustrate that when compared to state-of-the-art approaches, this framework achieves exceptional accuracy (98.94%) and F1 score (99.12%) on a public test dataset. These findings strongly support the conclusion that the proposed feature fusion-based model outperforms individual models and the majority voting ensemble method in terms of performance. The significance of this successful approach to pneumonia detection lies in its potential to provide clinicians and healthcare professionals with an effective solution for accurate and rapid diagnosis, particularly in pediatric cases.