# Engineering the Image Representation for Deep Learning in Contrast-Enhanced Mammography: A Systematic Analysis of Preprocessing and Anatomical Masking

## TL;DR

This paper shows that preprocessing images in contrast-enhanced mammography significantly improves deep learning model performance and stability.

## Contribution

A systematic analysis of preprocessing as a first-class design variable in medical AI for contrast-enhanced mammography.

## Key findings

- Anatomically constrained preprocessing improves discrimination performance and training stability across CNN architectures.
- Breast mask-based representations show substantial gains in AUROC and AUPRC compared to raw inputs.
- Preprocessing enhances model robustness and generalization, independent of network complexity.

## Abstract

Deep-learning models applied to contrast-enhanced mammography (CEM) are known to be highly sensitive to the input image representation. However, preprocessing is often treated as a secondary step and rarely analyzed as an independent design variable. In this work, we present a systematic engineering analysis of a deterministic, label-independent preprocessing pipeline for CEM images. The pipeline integrates intensity normalization, global histogram matching, local contrast enhancement, denoising, and anatomically constrained breast masking. Using a controlled experimental design, identical deep-learning architectures were trained under different input representations to isolate the impact of preprocessing on classification performance and stability. Across convolutional neural network architectures, anatomically constrained preprocessing consistently improves discrimination performance, reduces variability across cross-validation folds, and enhances training stability. Breast mask-based representations demonstrate substantial gains in AUROC and AUPRC compared to raw DICOM inputs. These findings highlight image preprocessing as a first-class engineering component in medical AI pipelines. Breast masking significantly improves robustness and generalization, independently of network architecture complexity. From a clinical perspective, improving model robustness and sensitivity to malignant lesions may contribute to more reliable AI-assisted decision support in contrast-enhanced mammography, particularly in settings characterized by acquisition variability and heterogeneous patient populations.