Abstract
Class imbalance is a pervasive challenge in medical image segmentation, where certain anatomical structures or pathological regions may be significantly underrepresented in the training data. This imbalance often leads to overfitting, where neural networks memorize the training data rather than learning generalizable features. In this work, we present a comprehensive analysis of overfitting behavior under class imbalance conditions and propose effective improvements.
Our study investigates the relationship between class imbalance ratios and overfitting patterns across different network architectures and training strategies. We identify key factors that contribute to overfitting, including the degree of class imbalance, network capacity, and regularization techniques. Through extensive experiments on multiple medical imaging datasets, we demonstrate that overfitting manifests differently across classes, with minority classes being particularly susceptible.
We propose several strategies to mitigate overfitting under class imbalance, including adaptive learning rates, class-specific regularization, and curriculum learning approaches. Our analysis provides insights into the fundamental mechanisms driving overfitting in imbalanced medical image segmentation tasks and offers practical guidelines for model development.
The findings from this work contribute to a better understanding of neural network behavior under class imbalance conditions and provide a foundation for developing more robust segmentation models in medical imaging applications.
BibTeX
@inproceedings{li2018overfitting,
title={Overfitting of neural nets under class imbalance: Analysis and improvements for segmentation},
author={Li, Zeju and Kamnitsas, Konstantinos and Glocker, Ben},
booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI 2019)},
year={2019},
doi={10.1007/978-3-030-32248-9_45}
}