A Data Augmentation Methodology to Reduce the Class Imbalance in Histopathology Images

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning performance relies heavily on training data quality and quantity.
• Class imbalance in multi-class/multi-label classification skews neural networks towards majority classes.
• Object detection faces foreground-background and inter-class imbalance; this study focuses on the latter.

Purpose of the Study:

• To address class imbalance in object detection without worsening foreground-background imbalance.
• To improve the detection of minority classes in highly unbalanced datasets.
• To develop a strategy suitable for datasets with high instance density and potential overlap.

Main Methods:

• Modified copy-paste data augmentation technique.
• Integration of weight-balancing methods within the loss function.
• Application to a nuclei detection dataset characterized by high instance density and imbalance.

Main Results:

• The proposed hybrid approach successfully improved the classification of minority classes.
• Performance on majority classes was not significantly compromised.
• The method is effective for datasets with high instance density where overlap is a concern.

Conclusions:

• A combination of data augmentation and loss function weighting effectively mitigates class imbalance in object detection.
• This strategy offers a practical solution for improving deep learning model performance on specialized, imbalanced datasets.
• The findings contribute to more robust and equitable object detection systems.