Iteratively Pruned Deep Learning Ensembles for COVID-19 Detection in Chest X-rays

Abstract

We demonstrate use of iteratively pruned deep learning model ensembles for detecting pulmonary manifestation of COVID-19 with chest X-rays. This disease is caused by the novel Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) virus, also known as the novel Coronavirus (2019-nCoV). A custom convolutional neural network and a selection of ImageNet pretrained models are trained and evaluated at patient-level on publicly available CXR collections to learn modality-specific feature representations. The learned knowledge is transferred and fine-tuned to improve performance and generalization in the related task of classifying CXRs as normal, showing bacterial pneumonia, or COVID-19-viral abnormalities. The best performing models are iteratively pruned to reduce complexity and improve memory efficiency. The predictions of the best-performing pruned models are combined through different ensemble strategies to improve classification performance. Empirical evaluations demonstrate that the weighted average of the best-performing pruned models significantly improves performance resulting in an accuracy of 99.01% and area under the curve of 0.9972 in detecting COVID-19 findings on CXRs. The combined use of modality-specific knowledge transfer, iterative model pruning, and ensemble learning resulted in improved predictions. We expect that this model can be quickly adopted for COVID-19 screening using chest radiographs.

Keywords: COVID-19; Convolutional neural network; Deep learning; Ensemble; Iterative pruning.

FIGURE 1.

Graphical abstract of the proposed study.

FIGURE 2.

CXRs showing (A) clear lungs, (B) bacterial pneumonia manifesting as consolidations in the right upper lobe and retro-cardiac left lower lobe, and (C) COVID-19 pneumonia infection manifesting as peripheral opacities in the left lung.

FIGURE 3.

The segmentation approach showing U-Net based mask generation and Lung ROI cropping.

FIGURE 4.

Architecture of the customized CNN model. (I/P = Input, CONV = Convolution, GAP = Global average pooling, DO = Dropout, D = Dense with Softmax activation, N = Normal predictions, A = Abnormal Predictions).

FIGURE 5.

Architecture of the pretrained CNNs. (I/P = Input, PCNN = truncated model, ZP = Zero-padding, CONV = Convolution, GAP = Global Average Pooling, DO = Dropout, D=Dense with Softmax activation, O/P = Output).

FIGURE 6.

Grad-CAM Visualizations showing salient ROI detection by different pruned models. (A) CXR showing COVID-19 viral pneumonia-related opacities with GT annotations, (B) VGG-16 pruned model, (C) VGG-19 pruned model, and (D) Inception-V3 pruned model. Bright red corresponds to the pixels carrying higher importance and hence weights for categorizing the test sample to the COVID-19 viral pneumonia category.

FIGURE 7.

Confusion matrix obtained with the weighted-average pruned ensemble.

FIGURE 8.

ROC curves showing micro/macro-averaged and class-specific AUC obtained with the weighted-average pruned ensemble.