GANterfactual-Counterfactual Explanations for Medical Non-experts Using Generative Adversarial Learning

Abstract

With the ongoing rise of machine learning, the need for methods for explaining decisions made by artificial intelligence systems is becoming a more and more important topic. Especially for image classification tasks, many state-of-the-art tools to explain such classifiers rely on visual highlighting of important areas of the input data. Contrary, counterfactual explanation systems try to enable a counterfactual reasoning by modifying the input image in a way such that the classifier would have made a different prediction. By doing so, the users of counterfactual explanation systems are equipped with a completely different kind of explanatory information. However, methods for generating realistic counterfactual explanations for image classifiers are still rare. Especially in medical contexts, where relevant information often consists of textural and structural information, high-quality counterfactual images have the potential to give meaningful insights into decision processes. In this work, we present GANterfactual, an approach to generate such counterfactual image explanations based on adversarial image-to-image translation techniques. Additionally, we conduct a user study to evaluate our approach in an exemplary medical use case. Our results show that, in the chosen medical use-case, counterfactual explanations lead to significantly better results regarding mental models, explanation satisfaction, trust, emotions, and self-efficacy than two state-of-the art systems that work with saliency maps, namely LIME and LRP.

Keywords: counterfactual explanations; explainable AI; generative adversarial networks; image-to-image translation; interpretable machine learning; machine learning.

Figure 1

Schematic overview of our approach. A CycleGAN architecture is extended with the classifier that shall be explained. Both the generators of the CycleGAN include the classifier's decisions for the generated data into their loss function.

Figure 2

Example images of the used dataset. The top row shows images that are labeled as Normal, while the bottom row shows images labeled as Lung Opacity, indicating lungs that are suffering from pneumonia.

Figure 3

Examples of counterfactual images produced with our proposed approach. In each pair, the left image shows the original image, while the right image shows the corresponding counterfactual explanation. The red boxes were added manually to point the reader to the regions that were altered the most. The original images in the top row were classified as normal, while the original images in the bottom row were classified as pneumonia. The shown counterfactual images were all classified as the opposite as their respective counterpart.

Figure 4

Computational evaluation results of the counterfactual image generation performance. The confusion matrices show the number of samples out of each subset (Normal, Pneumonia, Total) of the rsna dataset that the classifier predicted to be the respective class before (y-axis) and after (x-axis) the samples had been translated by either the original CycleGAN or by our approach.

Figure 5

An example x-ray image classified as Pneumonia, as well as the different XAI visualizations used in our study when the slider is fully on the right side. Best viewed in color.

Figure 6

A simplified schematic of our prediction task.

Figure 7

Results of the explanation satisfaction and trust questionnaires. Error bars represent the 95% Confidence Interval (CI).

Figure 8

Results of the prediction task, and the task reflection questions. Error bars represent the 95% Confidence Interval (CI).

Figure 9

Results of the emotion questionnaires. Participants in the counterfactual condition felt significantly less angry and more relaxed compared to the LRP saliency map condition. For LIME, no significant differences were found. Error bars represent the 95% CI.

Figure 10

Significant differences regarding self-efficacy and general confidence of the participants in their predictions of the AI between the counterfactual condition and the saliency map conditions (LRP and LIME). Error bars represent the 95% CI.

Figure 11

Confidence of the participants in correct and false predictions. The significant difference between the counterfactual condition and the saliency map conditions is based on the confidence in correct predictions, not in the incorrect ones. Error bars represent the 95% CI.