Uncertain-CAM: Uncertainty-Based Ensemble Machine Voting for Improved COVID-19 CXR Classification and Explainability

Abstract

The ongoing coronavirus disease 2019 (COVID-19) pandemic has had a significant impact on patients and healthcare systems across the world. Distinguishing non-COVID-19 patients from COVID-19 patients at the lowest possible cost and in the earliest stages of the disease is a major issue. Additionally, the implementation of explainable deep learning decisions is another issue, especially in critical fields such as medicine. The study presents a method to train deep learning models and apply an uncertainty-based ensemble voting policy to achieve 99% accuracy in classifying COVID-19 chest X-rays from normal and pneumonia-related infections. We further present a training scheme that integrates the cyclic cosine annealing approach with cross-validation and uncertainty quantification that is measured using prediction interval coverage probability (PICP) as final ensemble voting weights. We also propose the Uncertain-CAM technique, which improves deep learning explainability and provides a more reliable COVID-19 classification system. We introduce a new image processing technique to measure the explainability based on ground-truth, and we compared it with the widely adopted Grad-CAM method.

Keywords: COVID-19; deep learning; explainable AI; intelligent signal processing; uncertainty.

Figure 1

Framework of the proposed method. In the preprocessing stage, data were prepared for training phase wherein we generate m models using cross-validation for each CNN network, and for each model, we generate s snapshots models using cyclic cosine annealing. Snapshots models were used to obtain optimal voting weight.

Figure 2

Process of computing IoU from Uncertain-CAM Output and Ground Truth. The generated heatmap is masked based on color threshold and is compared with ground truth masked to generate IoU, which describes how much generated masked is aligned with the ground truth masked.

Figure 3

Samples of dataset used in the study.

Figure 4

Proposed network performance output on unseen data.

Figure 5

IoU scores with and without uncertainty.