Solving the problem of imbalanced dataset with synthetic image generation for cell classification using deep...

Area of Science:

• Medical image analysis
• Machine learning applications in healthcare
• Computational pathology

Background:

• Class imbalance and limited annotated data are significant challenges in machine learning for medical image analysis.
• Clinical datasets, such as cervical smears from digitized Pap tests, often exhibit a majority of normal cells over abnormal ones.
• Manual examination of thousands of cells per smear is time-consuming and labor-intensive.

Purpose of the Study:

• To develop a method for generating synthetic clinical images to overcome data scarcity and class imbalance.
• To pretrain a subsequent classifier network using the generated synthetic data.
• To improve the accuracy of classifying abnormal cells in cervical smear images.

Main Methods:

• Utilized a custom variational autoencoder (VAE) for synthetic image generation.
• Applied the VAE to a clinical dataset of cervical cells.
• Compared the performance of the VAE-based approach with existing methods and modifications.
• Integrated VAE for pretraining a classifier network.

Main Results:

• Achieved a performance increase of 4.52% in classifying abnormal cells.
• Demonstrated the effectiveness of synthetic data generation in addressing class imbalance.
• The proposed method shows promise in enhancing the efficiency of automated cell analysis.

Conclusions:

• The developed variational autoencoder-based synthetic image generation effectively addresses class imbalance in clinical datasets.
• This approach enhances the performance of abnormal cell classification, contributing to more efficient diagnostic tools.
• The research advances the goal of creating a system to rank cervical smear samples by importance, aiding manual examination.