Towards Generalizability and Robustness in Biological Object Detection in Electron Microscopy Images.

Area of Science:

• Biomedical data analysis
• Machine learning in healthcare
• Electron microscopy imaging

Background:

• Biomedical data presents unique challenges like limited data, volatility, and shifts, compromising machine learning model robustness and generalizability.
• Deploying machine learning models without proper tuning and data management in the presence of corruptions leads to reduced or misleading performance.

Approach:

• This study explores techniques to enhance machine learning model generalizability through iterative adjustments.
• We investigated detection tasks using electron microscopy images, comparing models trained with different normalization and augmentation techniques.
• Evaluated performance across transformer- and convolution-based detection architectures.

Key Points:

• Models trained with Group Normalization or texture data augmentation demonstrated superior performance over other normalization and classical data augmentation techniques.
• These improvements were consistent even when models were trained and tested on disjoint datasets acquired through diverse protocols.
• Achieved a 29% boost in average precision, signifying substantial enhancements in model generalizability.

Conclusions:

• Group Normalization and texture data augmentation are effective strategies for improving machine learning model generalizability in biomedical applications.
• The developed techniques enhance model resilience and adaptability to diverse datasets, crucial for real-world deployment.
• This research highlights the potential for robust machine learning solutions in analyzing complex biomedical imaging data.