Transfer learning approach for pediatric pneumonia diagnosis using channel attention deep CNN architectures

Abstract

Chest X-ray is the most commonly adopted non-invasive and painless diagnostic test for pediatric pneumonia. However, the low radiation levels for diagnosis make accurate detection challenging, and this initiates the need for an unerring computer-aided diagnosis model. Our work proposes stacking ensemble learning on features extracted from channel attention deep CNN architectures. The features extracted from the channel attention-based ResNet50V2, ResNet101V2, ResNet152V2, Xception, and DenseNet169 are individually passed through Kernel PCA for dimensionality reduction and concatenated. A stacking classifier with Support Vector Classifier, Logistic Regression, K-Nearest Neighbour, Nu-SVC, and XGBClassifier is employed for the final- Normal and Pneumonia classification. The stacking classifier achieves an accuracy of 96.15%, precision of 97.91%, recall of 95.90%, F1 score of 96.89%, and an AUC score of 96.24% on the publicly available pediatric pneumonia dataset. We expect this model to help the real-time diagnosis of pediatric pneumonia significantly.