A Study of Breast Tumor Detection from Textrue Features from Mammography Images Based on Deep Learning

Over the past decades, breast cancer has become the top one incidence rate in cancers in women in Taiwan. The standardized cancer incidence rate of breast cancer in women and the fatality rate are 69.1 and 12.0 (per 100 thousand population), respectively. More than 10 thousand women were diagnosed with breast cancer and over 2 thousand women died from it every year. Mammography is the significant tool to efficiently screen early stage breast cancer and improve prognosis of patients. It is used to detect calcification and tumor to find non-invasive breast cancer. Therefore, it is a very important research topic in artificial intelligence to detect early stage breast cancer. This project is a three-year project, and the main goal is to identify breast cancer from mammography images by using deep learning. Different to the general methods as training deep learning models from mammography images directly, a novel strategy, which is training deep learning models from texture features of mammography images, is proposed in this project. Meanwhile, an inference system on edge device is designed and developed for clinic room in project. A feedback mechanism and transfer learning method are used to tune the trained models. In the first year, the main goal is to collect mammography images and develop the texture feature extraction method. In the second year, we will focus on labeling images and training deep learning models of detection of breast tumor. In the third year, we will design and develop an inference system for use in clinic. We expect the proposed methods in this project can be great benefit to diagnostic techniques for breast cancer detection.

Project ID:PB10901-0697
External Project ID:MOST108-2221-E182-031-MY3