Mammography Image Preprocessing

The following describes the preprocessing pipeline applied to all mammography images prior to model training and evaluation.

1. Data Format Conversion

• All mammograms were originally stored in DICOM format
• Converted to 16-bit grayscale PNG images
• This conversion preserves high-intensity resolution while enabling efficient processing in deep learning pipelines

2. Image Preprocessing Pipeline

All images were preprocessed following the same protocol:

• Background removal to eliminate irrelevant black borders and artifacts
• Breast region segmentation to isolate the region of interest and remove non-breast content, including annotations and text commonly present in mammograms
• Resizing to a fixed resolution of 1664 × 2048 pixels
• Intensity normalization to standardize pixel value distributions across scans

3. Dataset Split

• Data split was performed at the patient level to avoid data leakage
• Splits were defined as:
  • Training: 50%
  • Validation: 20%
  • Test: 30%

4. Summary

This preprocessing pipeline ensures:

• Consistent image resolution and intensity scaling
• Removal of irrelevant background and annotation artifacts
• Proper separation of patient data across splits
• Compatibility with deep learning-based mammography analysis models