College of Engineering, Design & Computing Events

CEDC seminar series: Breast cancer recognition and localization based on CornetNet-Lite and Generative Adversarial Neural Network

| 11:00 AM - 12:15 PM
Room Number : 2507
North Classroom
1200 Larimer Street
Denver, CO
Yan Pang, PhD Student
Dept. of Electrical Engineering, University of Colorado Denver

Mammography is one of the most widely used techniques today to screen for breast cancer. This was demonstrated in multitudinous researches that showed computer aided detection have a high achievement to detect early breast cancer with the aid of deep neural networks. In the existence of current studies, however, the lesion recognition has not received a high diagnostic accuracy. Due to the paucity and unbalance of the sample data for training, the exact location of the lesion cannot be indicated incorrectly or imprecisely either. In this paper, we present a two-stage neural networks to precisely distinguish normal and abnormal and feedback the lesion location based on the mammogram images in two stages, which efficiently leverages the training dataset with either less cancer labels or unbalanced lesion types to eradicate the reliance on rarely breast lesion annotations. By contrasting the precious studies, our approach to classify and detect the lesion in the whole large mammogram images achieved excellent performance.  We tested our methodology using the publicly available dataset DDSM and the In-Breast dataset. Our highest accuracy achieved was 98.3% and 99.1%, respectively. The mean average of precision of abnormal region detection was 47 and 40 on the dataset DDSM and In-Breast dataset. In addition, our end-to-end model could be easily transferred from DDSM to In-Breast dataset with a small number of less annotated lesions due to our self-balanced lesion images in 21 types. These attainments are benefits to alleviate the burden of the researchers to distinguish the normal images, but with a higher accuracy to ascertain the location of the lesion in the abnormal ones.

Email Address:
CMS Login