Table 3 Domain-specific (DS) pre-processing pipeline, designed to preserve anatomical structure.

From: The impact of pre-processing techniques on deep learning breast image segmentation

Order

Parameter

2D mammography models

3D breast MRI models

1st

Orientation

All breasts pointed in the same direction

Right Anterior Superior (RAS)

2nd

Pixel Spacing

Already isotropic

Median value of each axes

3rd

Pixel Intensity

Minimum-Maximum, Percentiles,

Z-Score, Histogram Normalization

Minimum-Maximum, Percentiles,

Z-Score, Histogram Normalization

4th

Image Size

Padding and resizing to

256 × 256, 512 × 512, 1024 × 1024

256 × 256 × 64, 512 × 512 × 64

  1. The table shows the order of application and corresponding methods for each step, including orientation standardization, pixel/voxel spacing harmonization, pixel intensity normalization (Min-Max, P1-P99, Z-Score, HistNorm), and resizing. For (a) 2D mammography, resizing was applied either with padding to the maximum original image size or directly without padding, while for (b) 3D breast MRI, volumetric resizing was performed.
Back to article page