[Paper Review] DeepCOVIDExplainer: Explainable COVID-19 Predictions Based on Chest X-ray Images

Md Rezaul Karim, Till Döhmen+4 more|arXiv (Cornell University)|Apr 9, 2020

COVID-19 diagnosis using AI25 references81 citations

TL;DR

DeepCOVIDExplainer proposes an explainable deep learning framework for automated COVID-19 detection from chest X-ray (CXR) images using a neural ensemble with Grad-CAM++ and LRP for interpretability. It achieves 96.12% positive predictive value and 94.6% F1 score for COVID-19 classification on a dataset of 15,959 CXR images, demonstrating high accuracy and clinical interpretability.

ABSTRACT

Amid the coronavirus disease(COVID-19) pandemic, humanity experiences a rapid increase in infection numbers across the world. Challenge hospitals are faced with, in the fight against the virus, is the effective screening of incoming patients. One methodology is the assessment of chest radiography(CXR) images, which usually requires expert radiologist's knowledge. In this paper, we propose an explainable deep neural networks(DNN)-based method for automatic detection of COVID-19 symptoms from CXR images, which we call DeepCOVIDExplainer. We used 15,959 CXR images of 15,854 patients, covering normal, pneumonia, and COVID-19 cases. CXR images are first comprehensively preprocessed, before being augmented and classified with a neural ensemble method, followed by highlighting class-discriminating regions using gradient-guided class activation maps(Grad-CAM++) and layer-wise relevance propagation(LRP). Further, we provide human-interpretable explanations of the predictions. Evaluation results based on hold-out data show that our approach can identify COVID-19 confidently with a positive predictive value(PPV) of 91.6%, 92.45%, and 96.12%; precision, recall, and F1 score of 94.6%, 94.3%, and 94.6%, respectively for normal, pneumonia, and COVID-19 cases, respectively, making it comparable or improved results over recent approaches. We hope that our findings will be a useful contribution to the fight against COVID-19 and, in more general, towards an increasing acceptance and adoption of AI-assisted applications in the clinical practice.

Motivation & Objective

• To address the challenge of rapid and accurate screening of COVID-19 patients in healthcare settings with limited radiologist availability.
• To develop a deep learning model that not only classifies CXR images into normal, pneumonia, and COVID-19 categories but also provides human-interpretable explanations.
• To improve clinical trust and adoption of AI in radiology by highlighting class-discriminating regions in CXR images using explainable AI techniques.
• To evaluate the model’s performance on a large, diverse dataset of 15,959 CXR images covering normal, pneumonia, and COVID-19 cases.

Proposed method

• The method uses a comprehensive preprocessing pipeline to standardize and enhance CXR images before augmentation.
• A neural ensemble model is trained on augmented CXR images to improve generalization and robustness across normal, pneumonia, and COVID-19 classes.
• Grad-CAM++ is applied to generate class-discriminating activation maps that highlight regions in the CXR images responsible for prediction decisions.
• Layer-wise relevance propagation (LRP) is used to further decompose and explain the contribution of each pixel to the final prediction.
• The integration of Grad-CAM++ and LRP enables human-readable, localized explanations of model predictions.
• The model is evaluated on hold-out test data to measure performance using precision, recall, F1 score, and positive predictive value.

Experimental results

Research questions

• RQ1 Can a deep neural network ensemble with explainable AI techniques achieve high accuracy in classifying CXR images into normal, pneumonia, and COVID-19 categories?
• RQ2 To what extent can Grad-CAM++ and LRP provide clinically meaningful and interpretable explanations of model predictions for CXR images?
• RQ3 How does the proposed method compare to recent approaches in terms of predictive performance and interpretability?
• RQ4 Can the model maintain high performance across diverse patient populations and imaging variations in a real-world clinical setting?

Key findings

• The model achieved a positive predictive value (PPV) of 91.6% for normal cases, 92.45% for pneumonia, and 96.12% for COVID-19 cases on hold-out test data.
• The precision, recall, and F1 score for COVID-19 classification were 94.6%, 94.3%, and 94.6%, respectively, indicating strong balanced performance.
• The integration of Grad-CAM++ and LRP enabled the model to highlight anatomically relevant regions in CXR images associated with disease, enhancing clinical interpretability.
• The model demonstrated performance comparable to or better than recent state-of-the-art approaches in both accuracy and explainability.
• The results suggest that the model can support clinical decision-making by providing reliable, interpretable predictions for early COVID-19 screening.