# Explainable AI for medical imaging: Explaining pneumothorax diagnoses with Bayesian Teaching

**Authors:** Tomas Folke, Scott Cheng-Hsin Yang, Sean Anderson, Patrick Shafto  
**arXiv:** [2106.04684](https://arxiv.org/abs/2106.04684)  
**Date:** 2021-06-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2106.04684v1/

## Key points

- Limited expert time is a significant bottleneck in medical imaging, causing delays in diagnosis and treatment.
- Recent advancements in image classification have opened the door for integrating Artificial Intelligence (AI) as decision-support tools for medical experts.
- Establishing and maintaining trust between human experts and AI agents is crucial for effective AI support in medicine.
- Explainable AI, such as Bayesian Teaching, plays a key role in building trust by providing insights into AI decision-making processes.
- A study using Bayesian Teaching explanations improved participants' understanding of an AI system diagnosing pneumothorax in x-ray images.
- The research demonstrates that explainable AI can enhance collaboration between human experts and AI systems in medical imaging.
- Incorporating Bayesian Teaching explanations can improve decision-making processes in radiology, leading to more efficient and accurate diagnoses and benefiting patient care outcomes.

## Summary

Limited expert time is a significant bottleneck in medical imaging, leading to delays in diagnosis and treatment. Recent advancements in image classification have paved the way for the integration of Artificial Intelligence (AI) as decision-support tools for medical experts. This has the potential to greatly enhance radiologist productivity and ultimately improve public health outcomes. In order for AI to effectively support medical professionals, it is crucial to establish and maintain trust between human experts and AI agents. Explainable AI, which provides insights into the decision-making processes of AI systems, plays a key role in building this trust. One approach to explainable AI is Bayesian Teaching, a formal method rooted in cognitive science that aims to help users understand how AI reaches diagnostic judgments. A study was conducted using Bayesian Teaching explanations integrated into an interface for diagnosing pneumothorax in x-ray images. The experiment involved eight participants, including medical school students and radiology residents with varying levels of experience. Participants were tasked with predicting the AI's diagnostic decisions based on the explanations provided and determining when to certify the AI's decisions. Results from the study demonstrated that explanations generated through Bayesian Teaching significantly improved participants' understanding of the AI system and facilitated the development of appropriate trust in its diagnostic capabilities. By combining explanation-by-examples and explanation-by-features methods, researchers were able to optimize the explanations provided to users. Overall, this research highlights the importance of explainable AI in facilitating collaboration between human experts and AI systems in medical imaging. The findings suggest that incorporating Bayesian Teaching explanations can enhance decision-making processes in radiology and contribute to more efficient and accurate diagnoses, ultimately benefiting patient care outcomes.