Towards the Generation of Medical Imaging Classifiers Robust to Common Perturbations

Area of Science:

• Medical Imaging
• Machine Learning
• Artificial Intelligence

Background:

• Machine learning (ML) and artificial intelligence (AI) classifiers show promise for diagnosing diseases from medical images.
• However, a lack of robustness limits their clinical application.

Purpose of the Study:

• To enhance the robustness of AI/ML imaging classifiers.
• To evaluate the impact of data perturbations on classifier performance.

Main Methods:

• Simultaneously applied common perturbations (noise, contrast, blur, rotation, tilt) to training and test images.
• Compared performance with classifiers trained using adversarial noise.
• Utilized the PneumoniaMNIST and Breast Ultrasound Images (BUSI) datasets.

Main Results:

• Classifiers trained with perturbed data showed comparable performance on unperturbed test images.
• Significantly improved performance on perturbed test data (single and multiple perturbations) compared to unperturbed classifiers.
• Achieved statistically significant improvements (p-values < 0.05) for various perturbations.

Conclusions:

• Training AI/ML classifiers with perturbed data enhances robustness to image distortions.
• No performance decrease was observed on unperturbed test images.
• Perturbed data training offers benefits without significant downsides for medical imaging AI.