Robustness evaluation against corruptions for Optical Diffraction Tomography-based classifiers.

Area of Science:

• Biomedical Imaging
• Machine Learning Safety
• Computational Imaging

Background:

• Optical Diffraction Tomography (ODT) offers advanced 3D imaging capabilities.
• Real-world noise significantly impacts ODT classifier performance, necessitating robustness testing.
• Machine learning safety in ODT is crucial but underexplored.

Purpose of the Study:

• To establish the first comprehensive robustness testing protocol for ODT-based classifiers.
• To introduce and evaluate a new data augmentation strategy, CutPix, for enhancing ODT classifier robustness and accuracy.
• To address the lack of research on ODT classifier robustness against various noise factors.

Main Methods:

• Developed a robustness testing protocol simulating 16 corruption scenarios for ODT data.
• Introduced CutPix, a data augmentation technique combining fractal pattern mixing and cut-and-concatenate methods.
• Evaluated CutPix's performance against existing methods on corrupted ODT datasets.

Main Results:

• CutPix significantly enhances the robustness of ODT classifiers across diverse corrupted environments.
• The proposed method demonstrates superior performance, especially against pattern-based noises.
• Achieved notable improvements in accuracy and resilience compared to current augmentation techniques.

Conclusions:

• CutPix represents a significant advancement in improving the reliability of ODT classifiers in noisy conditions.
• The developed protocol and CutPix strategy provide a foundation for safer and more robust ODT applications.
• Publicly available code and scripts facilitate further research in ODT robustness evaluation.