Engineering the Image Representation for Deep Learning in Contrast-Enhanced Mammography: A Systematic Analysis of...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Deep Learning

Background:

• Deep learning models for contrast-enhanced mammography (CEM) are sensitive to input image representation.
• Image preprocessing is often secondary and not independently analyzed for CEM AI.

Purpose of the Study:

• To systematically analyze a deterministic, label-independent preprocessing pipeline for CEM images.
• To evaluate the impact of preprocessing on deep learning classification performance and stability.

Main Methods:

• Developed a preprocessing pipeline: intensity normalization, histogram matching, contrast enhancement, denoising, and breast masking.
• Trained identical deep learning architectures with different input representations under controlled conditions.

Main Results:

• Anatomically constrained preprocessing consistently improved discrimination performance across CNN architectures.
• Breast mask-based representations showed significant gains in AUROC and AUPRC compared to raw DICOM.
• Preprocessing reduced variability and enhanced training stability.

Conclusions:

• Image preprocessing is a critical engineering component in medical AI pipelines.
• Breast masking significantly improves model robustness and generalization, independent of network complexity.
• This approach can lead to more reliable AI-assisted decision support in CEM.