BYU

Abstract by Jacob Stern

Personal Infomation


Presenter's Name

Jacob Stern

Co-Presenters

Allison Lambert
Junseong Ahn
Wilson Redd
York Westenhaver

Degree Level

Undergraduate

Co-Authors

Allison Lambert
Brad Hatch
Wilson Redd
York Westenhaver
Junseong Ahn

Abstract Infomation


Department

Computer Science

Faculty Advisor

David Wingate

Title

Siamese Network Pre-training and Attention-based CNN for Whole-slide Medical Image Classification

Abstract

Currently, cervical cancer in India has a mortality rate of forty nine percent. With over one hundred and sixty million women at risk of cervical cancer, there is a great need for more and faster screening of women ages thirty to fifty-nine. One challenge to achieving this goal is a lack of qualified pathologists to review the Pap Smears. Obtaining, staining, shipping and returning the sample for diagnosis is a time consuming process. In an effort to decrease the time required to diagnose a patient, we are creating a convolutional neural network to detect cancerous cervical cells. The difficulty in this task is classification of data with low signal-to-noise ratio — in a 80,000 x 20,000 pixel medical slide image, the relevant cancerous area might be only a few pixels wide. Furthermore, we attempt to identify these areas using only whole slide labels. We use two techniques to train a neural network for this task. One, we are using a morphological siamese network for pre-training that is more sensitive to cancer.  Two, we use an attention mechanism to hone the attention of the CNN on the most important regions of the image.