Unsupervised Domain Adaptation to Classify Medical Images Using Zero-Bias Convolutional Auto-Encoders and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Supervised deep learning for image classification requires large labeled datasets, which are scarce in medical imaging due to annotation complexities.
• Current methods like fine-tuning deep convolutional neural networks (CNNs) still rely on limited labeled medical data.

Purpose of the Study:

• To develop a hierarchical unsupervised feature extractor to minimize the dependency on annotated training data in medical imaging.
• To improve the accuracy and robustness of medical image classification using limited labeled data.

Main Methods:

• Utilized a multi-layer zero-bias convolutional auto-encoder for unsupervised feature extraction, transforming generic features into locally relevant ones for medical images.
• Implemented a context-based feature augmentation scheme to enhance the discriminative power of learned representations.
• Leveraged pre-trained CNNs for natural images to transfer knowledge.

Main Results:

• The unsupervised approach demonstrated superior accuracy compared to conventional unsupervised methods.
• Achieved better performance than baseline fine-tuned supervised CNNs on three public medical image datasets.
• Successfully reduced reliance on annotated medical imaging data.

Conclusions:

• The proposed hierarchical unsupervised feature extractor is effective in improving medical image classification accuracy.
• This method offers a promising solution for scenarios with limited labeled medical data, advancing AI in healthcare.
• Unsupervised learning can be a viable alternative to supervised methods when data is scarce.