Stacked ensemble learning based on deep convolutional neural networks for pediatric pneumonia diagnosis using chest X-ray images

Abstract

Pneumonia is an acute respiratory infection caused by bacteria, viruses, or fungi and has become very common in children ranging from 1 to 5 years of age. Common symptoms of pneumonia include difficulty breathing due to inflamed or pus and fluid-filled alveoli. The United Nations Children's Fund reports nearly 800,000 deaths in children due to pneumonia. Delayed diagnosis and overpriced tests are the prime reason for the high mortality rate, especially in underdeveloped countries. A time and cost-efficient diagnosis tool: Chest X-rays, was thus accepted as the standard diagnostic test for pediatric pneumonia. However, the lower radiation levels for diagnosis in children make the task much more onerous and time-consuming. The mentioned challenges initiate the need for a computer-aided detection model that is instantaneous and accurate. Our work proposes a stacked ensemble learning of deep learning-based features for pediatric pneumonia classification. The extracted features from the global average pooling layer of the fine-tuned Xception model pretrained on ImageNet weights are sent to the Kernel Principal Component Analysis for dimensionality reduction. The dimensionally reduced features are further trained and validated on the stacking classifier. The stacking classifier consists of two stages; the first stage uses the Random-Forest classifier, K-Nearest Neighbors, Logistic Regression, XGB classifier, Support Vector Classifier (SVC), Nu-SVC, and MLP classifier. The second stage operates on Logistic Regression using the first stage predictions for the final classification with Stratified K-fold cross-validation to prevent overfitting. The model was tested on the publicly available pediatric pneumonia dataset, achieving an accuracy of 98.3%, precision of 99.29%, recall of 98.36%, F1-score of 98.83%, and an AUC score of 98.24%. The performance shows its reliability for real-time deployment in assisting radiologists and physicians.

Keywords: Chest X-rays; Computer-aided diagnosis; Deep learning; Pneumonia; Principal component analysis; Stacking classifier; Stratified K-fold; Transfer learning.

Fig. 1

Proposed architecture for pediatric pneumonia classification

Fig. 2

The architecture for xception deconstructed as (a) Entry flow, (b) Middle flow and (c) Exit flow

Fig. 3

Illustration of the working of a depthwise separable convolution network

Fig. 4

Samples of Normal x-rays and Pneumonia x-rays from the dataset in the first row and second row, respectively

Fig. 5

Confusion matrix. True Positive (TP)—number of pneumonia x-rays correctly predicted as pneumonia. False Negative (FN)—number of pneumonia x-rays wrongly predicted as normal. True Negative (TN)—number of normal x-rays correctly predicted as normal. False Positive (FP)—number of normal x-rays predicted wrongly as pneumonia

Fig. 6

Confusion matrix for xception predictions on the test data

Fig. 7

ROC curve for test data predictions made by the fine-tuned xception model

Fig. 8

Training and validation accuracy-loss history of the fine-tuned xception model

Fig. 9

Xception model performance on the validation set using different optimizers

Fig. 10

Xception model performance on the validation set using different learning rates with adam as the optimizer

Fig. 11

Class activation maps of misclassified X-rays (row 1: normal classified as pneumonia, row 2: normal classified as pneumonia, row 3: pneumonia classified as normal)

Fig. 12

Class activation maps of correctly classified X-rays (row 1: normal classified as normal, row 2: normal classified as normal, row 3: normal classified as normal)

Fig. 13

t-SNE feature representation of the test data extracted from the xception model

Fig. 14

Cumulative variance plot of the extracted xception features

Fig. 15

Confusion matrix for predictions made on the test dataset using the stacked classifier with kernel PCA

Fig. 16

Confusion matrix for predictions made on the test dataset using the stacked classifier without kernel PCA

Fig. 17

ROC curve for predictions made on the test dataset using the stacked classifier

Fig. 18

Confusion matrix for predictions made on the test dataset of normal vs pneumonia classification dataset [55]

Fig. 19

Confusion matrix for predictions made on the test dataset of normal vs pneumonia classification dataset [56]