June 13, 2019 –
Title: Deep Learning Systems for Automated Lesion Detection, Segmentation and Classification in Mammography
PhD Candidate: Dina Abdelhafiz
Co-Major Advisors: Sheida Nabavi and Reda Ammar
Associate Advisors: Sanguthevar Rajasekaran and Clifford Yang
Date/Time: Thursday, June 13th, 2019 11:00 AM
Location: HBL 1947 Conference Room
The detection of breast masses in mammogram (MG) images using CAD systems is a challenging task due to the varying sizes, shapes, and textures of masses. The detection of masses in dense tissues and, more generally, in dense breasts is often considered more challenging due to the similar visual aspects of normal and abnormal dense tissues, which complicates the interpretation of mammographic images. Recently, Deep learning (DL) and in particular, convolutional neural networks (CNNs), has exhibited a state-of-the-art performance in various machine learning tasks including object detection and classification in medical applications.
In this proposal, my aim is to develop an automated DL system to precisely localize, segment, and classify mass lesions in MGs and address the above challenges. To develop a DL model with a competitive detection accuracy, First, I collected and pre-processed MG images that come with different resolutions from public repositories and from UCONN health center (UCHC). Secondly, I started conducting a detailed review of the strengths, limitations, and performance of the most recent CNNs applications in analyzing MG images. Also, I studied the effect of popular pre-processing techniques on breast cancer classification and detection using state-of-the-art CNNs. I have developed a novel residual attention DL system for automated mass segmentation, detection, and classification in MG images. The overall system precisely detects segments of mass lesions in MG images and classify the detected binary segmented lesions into benign or malignant. The developed DL system shows superior performance compared to the other state-of-the-art DL models, such as FCN, Dilated-Net, U-Net, SegNet, Faster R-CNN and YOLO, in detecting and segmenting masses, especially for heterogeneously dense and dense MG images.
In my ongoing work, I focus on improving the developed DL CNN model using attention breast density model which takes an MG image as input and generate a mask that estimates the breast density at the pixel level. The generated attention mask helps to identify and exploit the effective dense segments of MG images to support the proposed system in its detection and classification decision. Also, to increase the performance of the developed automated DL system, I want to incorporate previous years MG images with current MG images. This can help to reduce false detections.