Pre-Processing Methods

Preprocessing Function Used Definition Advantage of Using This

Method Method
Background [Link](), Removes unnecessary black Reduces irrelevant information
Removal [Link]() regions and image borders, and computational complexity. It
retaining only the breast allows the model to focus only on
tissue region for analysis. diagnostically important breast
tissue instead of learning from
background pixels.
Thresholding [Link]() Converts the image into Simple, fast, and highly effective
foreground and background for mammograms. It provides
regions based on pixel efficient breast extraction without
intensity values, helping requiring computationally
separate the breast area expensive segmentation
from the surrounding techniques.
background.
Contour-Based [Link](), Identifies the largest Automatically localizes the most
ROI Localization [Link]() contour corresponding to relevant anatomical region,
the breast region and crops reducing noise and improving
it as the Region of Interest classification accuracy without the
(ROI). need for manual annotation.
Gaussian Filtering [Link]() Applies Gaussian Produces cleaner images while
smoothing to reduce preserving important structural
random high-frequency information, making feature
noise present in extraction and deep learning more
mammographic images. reliable.
Median Filtering [Link]() Replaces each pixel value Particularly effective for removing
with the median of its speckle and impulse noise while
neighboring pixels to preserving lesion boundaries and
suppress noise. tissue edges better than averaging
filters.
CLAHE (Contrast [Link](), Enhances local contrast by Improves visibility of masses,
Limited Adaptive [Link]() applying adaptive tumors, and microcalcifications
Histogram histogram equalization to without excessively amplifying
Equalization) small image regions. image noise. It is widely
considered one of the best contrast
enhancement techniques for
mammograms.
Min-Max [Link]() or Scales pixel intensity values Ensures consistent input
Normalization mathematical scaling to a fixed range, usually distributions for EfficientNet-B1,
between 0 and 1. leading to faster convergence,
improved training stability, and
better model performance.
Data ImageDataGenerator(), Generates additional Increases dataset diversity,
Augmentation RandomRotation(), training samples through reduces overfitting, improves
RandomFlip(), image transformations such generalization capability, and
RandomZoom() as rotation, flipping, enables the model to perform well
zooming, and brightness on unseen mammographic images.
adjustment.
Feature Engineering Methods:

Feature Function Used Definition Advantage of Using This

Engineering Method
Method
Mean Intensity [Link]() Computes the average Provides information about
pixel intensity within the overall brightness of
the breast region. breast tissue and helps
identify abnormal dense
regions.
Standard [Link]() Measures the variation Helps quantify tissue
Deviation or dispersion of pixel heterogeneity, which is often
intensity values. associated with suspicious
lesions and tumors.
Skewness [Link]() Measures the Captures irregular intensity
asymmetry of the pixel patterns that may indicate
intensity distribution. abnormal breast structures.
Kurtosis [Link]() Measures the Helps identify unusual
peakedness or flatness intensity concentrations that
of the intensity may be associated with
distribution. lesions or
microcalcifications.
GLCM Contrast graycoprops(glcm,'contrast') Measures the local Useful for detecting texture
intensity variation differences between normal
between neighboring tissue and tumor regions.
pixels.
GLCM graycoprops(glcm,'correlation') Measures the Helps characterize structural
Correlation dependency between patterns and tissue
neighboring pixel organization within
intensities. mammograms.
GLCM Energy graycoprops(glcm,'energy') Measures textural Higher values indicate
uniformity by homogeneous tissue, while
calculating repeated lower values may indicate
pixel patterns. abnormalities.
GLCM graycoprops(glcm,'homogeneity') Measures the closeness Helps distinguish smooth
Homogeneity of pixel distribution to breast tissue from
the GLCM diagonal. heterogeneous cancerous
regions.
Area [Link]() Calculates the total size Provides important
of the detected lesion or morphological information
ROI. for differentiating benign and
malignant abnormalities.
Perimeter [Link]() Measures the length of Helps analyze lesion
the lesion boundary. margins, which are often
irregular in malignant
tumors.
Circularity Derived from Area and Perimeter Quantifies how closely Benign lesions are often
the lesion resembles a more circular, while
perfect circle. malignant lesions tend to
have irregular shapes.
Compactness Derived from Area and Perimeter Measures how densely Useful for identifying
packed the lesion shape spiculated or irregular tumor
is. boundaries associated with
malignancy.
Eccentricity regionprops() Measures the elongation Helps distinguish round
of a lesion. benign masses from
elongated or irregular
malignant lesions.
Edge Density [Link]() Calculates the Captures boundary
proportion of edge complexity and helps
pixels within the ROI. identify suspicious lesion
structures.
Sobel Mean [Link]() + [Link]() Computes the average Highlights tissue transitions
gradient magnitude in and edge information useful
the image. for lesion characterization.
Sobel Standard [Link]() + [Link]() Measures variation in Helps quantify texture
Deviation image gradients. complexity and structural
irregularities in breast tissue.
Density Ratio Pixel-based density calculation Measures the proportion Breast density is an
of dense tissue within important clinical indicator
the breast region. and is strongly associated
with breast cancer risk.
Density Category Threshold-based categorization Classifies breast tissue Provides clinically
into density levels. meaningful information
regarding tissue composition
and screening difficulty.
Left-Right Image comparison operations Measures differences Asymmetry is a significant
Asymmetry between the left and radiological indicator of
right breast regions. potential abnormalities.
Top-Bottom Image comparison operations Measures differences Helps identify localized
Asymmetry between upper and structural changes and
lower breast regions. suspicious tissue patterns.
Microcalcification Connected component analysis Counts the number of Microcalcifications are
Count detected among the earliest indicators
microcalcification of breast cancer and improve
clusters. diagnostic sensitivity.
Mean Region measurement functions Calculates the average Provides additional
Microcalcification size of detected information about lesion
Size microcalcifications. characteristics and potential
malignancy.
Deep Feature EfficientNetB1(weights='imagenet') Automatically learns Captures complex patterns,
Extraction hierarchical image textures, and lesion
features using a deep characteristics that may not
convolutional neural be represented by
network. handcrafted features,
significantly improving
classification performance.