Noblis Researchers Apply Explainable Artificial Intelligence (XAI) to a COVID-19 X-Ray Detection Study

Background

Artificial Intelligence (AI) and Machine Learning (ML) are powerful methods for data processing and analysis but are complex to understand. XAI can be used to clarify the deep learning methods within AI and assures the algorithms are looking at the right features during their decision-making process. The solutions it provides are applicable to many areas including, fraud detection, text analytics, object detection and security-screening systems.

A Noblis team is applying XAI to improve x-ray screening processes that may help diagnose COVID-19. Recent studies have shown that abnormalities in chest x-rays (CXR) can be associated with classifying the virus [1].

The Challenge

A key battle in the fight against COVID-19 is an effective screening process. Deep learning methods based on convolutional neural networks (CNNs) have been applied to classify approximately 14,000 CXR from patients into three categories: normal, non-COVID-19 pneumonia and COVID-19. These methods have reported accuracies higher than 90% [1-2]. However, the accuracies presented can be misleading, making it vital to pinpoint from where in the x-ray the decision is coming.

The Solution

To help with this, a Noblis research team, led by Ajay Patrikar and Matthew Kersting, is applying the latest XAI techniques to analyze results produced by these networks. These XAI techniques generate heat maps to highlight the important regions of the image that drove the classification decision.

Based on observations, artifacts present in the image (e.g., text found at the corners of the images) influenced the classification decision in some models. Text posted in the top left corner stating “PORTABLE UPRIGHT” does not indicate COVID-19. Therefore, this artifact should not be considered in the classification decision, despite the heatmap showing that it was highly influential.

Conclusions

The Noblis team came to two conclusions. First, XAI should be used to audit decisions produced by deep learning models to ensure they are focusing on reasonable aspects of the data. The heat maps provide a second level of understanding towards the classification process. Second, the artifacts in the image should be removed from the training data and the networks should be retrained. The artifacts are immaterial, only serving as a distraction to the CNNs.