Pseudo-labeling generative adversarial networks for medical image classification.

Area of Science:

• Medical Imaging
• Machine Learning
• Computer Vision

Background:

• Semi-supervised learning is increasingly vital in medical image analysis.
• Limited labeled data poses a significant challenge in medical image classification.
• Generative Adversarial Networks (GANs) offer potential for data augmentation.

Purpose of the Study:

• To introduce Pseudo-Labeling Generative Adversarial Networks (PLGAN), a novel semi-supervised algorithm for medical image classification.
• To leverage a small set of labeled images and numerous unlabeled images to enhance classification performance.
• To address the limitations of existing methods in data-scarce medical imaging scenarios.

Main Methods:

• PLGAN combines MixMatch for pseudo-label generation on synthetic and unlabeled images.
• Contrastive learning and self-attention mechanisms are integrated to focus on relevant image details.
• Cyclic consistency loss is employed to mitigate mode collapse issues in contrastive learning.
• Global and local classifiers are designed to capture complementary classification information.

Main Results:

• PLGAN demonstrates high learning performance on four medical image datasets with limited labeled and unlabeled data.
• On the OCT dataset, PLGAN achieved an 11% higher classification accuracy compared to MixMatch using 100 labeled and 1000 unlabeled images.
• Additional experiments confirmed the overall effectiveness and robustness of the proposed PLGAN algorithm.

Conclusions:

• PLGAN effectively enhances semi-supervised medical image classification by utilizing synthetic data generation.
• The algorithm shows significant improvements in accuracy, particularly in data-limited settings.
• PLGAN offers a promising solution for improving diagnostic capabilities through advanced machine learning techniques in healthcare.