A Physiological-Informed Generative Model for Improving Breast Lesion Classification in Small DCE-MRI Datasets.

Area of Science:

• Biomedical Image Processing
• Medical Imaging Physics
• Machine Learning in Medicine

Background:

• Deep learning (DL) in medical imaging requires large annotated datasets, which are difficult to obtain.
• Data augmentation using generative models is a common solution, but often lacks biological realism.
• Existing methods do not fully leverage the physical principles underlying medical imaging techniques.

Purpose of the Study:

• To develop a physiologically-aware data augmentation strategy for biomedical imaging.
• To improve the performance of deep learning classifiers in medical domains.
• To address the challenge of limited annotated data in medical image analysis.

Main Methods:

• A novel generative approach combining Physiologically Based Pharmacokinetic (PBPK) modeling and an Intrinsic Deforming Autoencoder (DAE).
• Application to breast Dynamic Contrast Enhanced-Magnetic Resonance Imaging (DCE-MRI) data.
• Testing on diverse datasets with varying acquisition protocols.

Main Results:

• The proposed physiologically-aware data augmentation significantly improves the performance of DL-based lesion classifiers.
• Demonstrated effectiveness across multiple datasets and acquisition protocols.
• Validated the importance of incorporating biological principles into generative models.

Conclusions:

• Physiologically-aware data augmentation is a promising strategy for enhancing DL in medical imaging.
• The PBPK-DAE approach offers a robust method for generating realistic synthetic medical images.
• This work advances the development of more accurate and data-efficient AI tools for disease diagnosis.