Document Type

Honors Thesis

Major

Biomedical Engineering

Advisor(s)

Andre Khalil

Committee Members

Kendra Batchelder, Melissa Ladenheim, Peter Stechlinski, Karissa Tilbury

Graduation Year

May 2022

Publication Date

Spring 2022

Abstract

As part of a first of its kind analysis of longitudinal mammograms, there are thousands of mammograms that need to be analyzed computationally. As a pre- processing step, each mammogram needs to be converted into a binary (black or white) spatial representation in order to delineate breast tissue from the pectoral muscle and image background, which is called a mammographic mask. The current methodology for completing this task is for a lab member to manually trace the outline of the breast, which takes approximately three minutes per mammogram. Thus, reducing the time cost and human subjectivity when completing this task for all mammograms in a large dataset is extremely valuable. In this thesis, an automated breast segmentation algorithm was adapted from a multi-scale gradient-based edge detection approach called the 2D Wavelet Transform Modulus Maxima (WTMM) segmentation method. This automated masking algorithm incorporates the first-derivative Gaussian Wavelet Transform to identify potential edge detection contour lines called maxima chains. The candidate chains are then transformed into a binary mask, which is then compared with the original manual delineation through the use of the Sorenson-Dice Coefficient (DSC). The analysis of 556 grayscale mammograms with this developed methodology produced a median DSC of 0.988 and 0.973 for craniocaudal (CC) and mediolateral oblique (MLO) grayscale mammograms respectively. Based on these median DSCs, in which a perfect overlap score is 1, it can be concluded a wavelet-based automatic breast segmentation algorithm is able to quickly segment the pectoral muscle and produce accurate binary spatial representations of breast tissue in grayscale mammograms.

Download

Share