Thesis - Open Access
Master of Science (MS)
Department / School
Electrical Engineering and Computer Science
Breast cancer is the most common cancer in women worldwide, and the mammogram is the most widely used screening technique for breast cancer. To make a diagnosis in the early stage of breast cancer, the appearance of masses and microcalcifications on the mammogram are two crucial indicators. Notably, the early detection of malignant microcalcifications can facilitate the diagnosis and the treatment of breast cancer at the appropriate time. Making an accurate evaluation on microcalcifications is a timeconsuming and challenging task for the radiologists due to the small size and the low contrast of microcalcification. Compared to the background and mammogram image with noises, it is tough to be discriminated. Computer-Aided Detection (CADe) have been deployed to support radiologists. However, most of current CADe systems need to have hand-crafted image features to be implemented. For improvement in the conventional approach, Convolutional Neural Network (CNN) with no hand-crafted image feature is used in this thesis. CNN with Class Activation Map (CAM) is deployed to implement the microcalcification detection in mammograms. GoogLeNet architecture with nine inception modules and one CAM layer is used to improve the localization capability of GoogLeNet in microcalcification detection while maintaining the local information. The network is trained and tested with Curated Breast Imaging Subset of Digital Database for Screening Mammography dataset (CBIS-DDSM). This approach demonstrates that the localization ability of CAM for abnormal microcalcification regions on the mammogram can be improved by restoring the last two inception modules that were removed in the paper  . For the CAM, CAM layer is inserted in the position of the second auxiliary layer that was used in the original GoogLeNet  for training. This allowed us to use the intermediate feature at the same location from   for localization while maintaining the depth of the GoogLeNet . The experimental result shows that this method achieved about 225.15% better result at localizing microcalcification in mammograms than the existing method.
Library of Congress Subject Headings
Breast -- Imaging.
Breast -- Radiography.
Diagnostic imaging -- Digital techniques.
Breast -- Calcification.
Breast -- Cancer -- Diagnosis.
Includes bibliographical references
Number of Pages
South Dakota State University
Jhang, Jieun, "Localization of Microcalcification on the Mammogram Using Deep Convolutional Neural Network" (2018). Electronic Theses and Dissertations. 2957.