Combining Generative and Discriminative Representation Learning for Lung CT Analysis With Convolutional Restricted...

Area of Science:

• Medical imaging
• Machine learning
• Computer-aided diagnosis

Background:

• Tissue classification systems heavily rely on feature selection, but often use standard, non-optimized filter banks.
• Representation learning, like restricted Boltzmann machines, can learn features from data but may not be optimal for classification tasks.
• Unsupervised learning methods, while useful for unlabeled data, do not inherently produce classification-optimal features.

Purpose of the Study:

• To develop an improved feature learning method for medical image classification.
• To combine generative and discriminative learning objectives for enhanced feature representation.
• To optimize filters for both data description and classification accuracy.

Main Methods:

• Proposed the convolutional classification restricted Boltzmann machine (CC-RBM).
• CC-RBM integrates both generative and discriminative learning objectives.
• Evaluated feature learning for lung texture classification and airway detection in CT images.

Main Results:

• The combined learning approach outperformed purely generative or discriminative methods.
• Lung tissue classification accuracy increased by 1 to 8 percentage points.
• Demonstrated that discriminative learning enhances unsupervised feature learners for classification.

Conclusions:

• A hybrid learning approach (generative + discriminative) is superior for feature learning in medical imaging.
• Convolutional classification restricted Boltzmann machines offer improved performance over standard methods.
• Discriminative guidance can optimize unsupervised feature learning for specific classification tasks.