Beyond fine-tuning: Classifying high resolution mammograms using function-preserving transformations.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Mammogram classification is challenging due to small lesions in high-resolution images.
• Current methods like fine-tuning convolutional neural networks (CNNs) have limitations with medical images.
• Differences between natural and medical images hinder performance gains from standard algorithmic approaches.

Purpose of the Study:

• To introduce MorphHR, a novel framework for mammogram classification that goes beyond traditional fine-tuning.
• To propose a new transfer learning scheme integrating function-preserving transformations for network regularization.
• To improve the classification performance of deep CNNs for medical image analysis.

Main Methods:

• Proposed MorphHR framework with a novel transfer learning scheme.
• Integrated function-preserving transformations to regularize deep CNNs.
• Modified both early and late layers of the CNN for domain-specific feature learning.
• Demonstrated hardware scalability for processing high-resolution images on standard GPUs.

Main Results:

• Achieved significant improvement in mammogram classification performance over state-of-the-art techniques.
• Demonstrated classification performance on par with radiology experts.
• Showcased effectiveness on the ChestX-ray14 dataset, surpassing existing methods.
• Preserved relevant information by utilizing high-resolution images.

Conclusions:

• MorphHR offers a significant advancement in mammogram classification accuracy.
• The framework enables effective learning of domain-specific features from high-resolution medical images.
• MorphHR provides a scalable and high-performing solution for medical image analysis.