A self-supervised feature-standardization-block for cross-domain lung disease classification.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Convolutional Neural Networks (CNNs) excel in medical image classification but suffer performance degradation across different datasets.
• Current methods often require fine-tuning on target domain data, limiting their immediate applicability.
• Addressing the challenge of cross-domain generalization is crucial for robust medical AI.

Purpose of the Study:

• To develop a method for improving cross-domain medical image classification without access to target domain data.
• To introduce a self-supervised, plug-and-play feature standardization block (FSB) for robust feature extraction.
• To evaluate the effectiveness of the FSB in classifying lung diseases from chest X-rays.

Main Methods:

• Designed a feature standardization block (FSB) integrating image normalization (INB), contrast enhancement (CEB), and boundary detection (BDB).
• Applied the FSB as a plug-in module to established deep learning architectures (VGG, Xception, DenseNet).
• Evaluated performance on four diverse chest X-ray datasets for lung disease classification.

Main Results:

• The FSB demonstrated improved domain adaptation performance across all tested deep networks.
• Individual components (INB, CEB, BDB) yielded average accuracy improvements of 2%, 2%, and 5%, respectively.
• The combined FSB achieved an average accuracy improvement of 6% in cross-domain classification tasks.

Conclusions:

• The proposed feature standardization block (FSB) effectively enhances the domain generalization capabilities of deep learning models for medical image classification.
• FSB offers a valuable approach to mitigate performance degradation when applying models to new, unseen datasets without requiring target domain data.
• This method shows significant promise for improving the reliability and applicability of AI in diagnosing lung diseases from chest X-rays.