Stress Testing Pathology Models with Generated Artifacts.

Area of Science:

• Digital pathology
• Machine learning in healthcare
• Artificial intelligence in medicine

Background:

• Machine learning (ML) models offer significant potential for advancing healthcare.
• The "black-box" nature of ML models presents considerable risks and challenges in clinical applications.
• Ensuring the reliability and safety of ML models is crucial for their adoption in healthcare.

Purpose of the Study:

• To develop a framework for generating tunable and describable artifacts for testing ML models.
• To evaluate the impact of simulated artifacts on the performance of digital pathology and kidney tissue segmentation models.
• To enhance the trustworthiness and robustness of machine learning models in healthcare settings.

Main Methods:

• Developed a custom Python module for artifact generation within a testing framework.
• Analyzed a digital pathology classification model and a kidney tissue segmentation model.
• Simulated various artifacts including bubbles, tissue folds, uneven illumination, marker lines, uneven sectioning, altered staining, and tissue tears.

Main Results:

• Observed performance degradation in ML models when exposed to artifacts, particularly altered stains, marker lines, tissue folds, and uneven sectioning.
• Demonstrated that the response of deep learning models to these simulated artifacts can be nonlinear.
• Identified specific artifact types that significantly impact model performance.

Conclusions:

• Generated artifacts serve as a valuable tool for rigorously testing machine learning models.
• Understanding model performance degradation due to artifacts is key to building trust and ensuring reliability.
• This approach aids in identifying potential failure points of ML models in healthcare scenarios.