Breast cancer is one of the leading causes of death in females. Early detection of breast tumors is critical to increasing the survival of women diagnosed with this disease. Accurate computer-aided detection of breast tumors could improve early detection but requires segmentation, a process that provides the precise tumor location, size, boundary, and shape. Existing breast tumor segmentation approaches are sensitive to small changes in image quality (e.g., intensity, contrast, noise, artifacts), limiting their application in early detection of breast cancer. The goal of the proposed project is to overcome current limitations by building tumor segmentation methodologies that are robust to variations of image quality. We will use breast ultrasound images due to the noninvasive, painless, nonradioactive, and cost-effective nature of the imaging procedure. We propose the following specific aims to achieve this goal. (1) Model human breast anatomy. In clinical examination, the knowledge of breast anatomy helps radiologists distinguish between breast tissues. In this aim, we will develop a graphical model to represent the spatial relationship of different breast layers and to help distinguish tumor regions from normal regions. We will develop a new mathematical tool called tissue connectedness for modeling breast anatomy in ultrasound images. Tissue connectedness allows for the identification of different breast tissues and helps distinguish a breast tumor from normal tumor-like regions (e.g., artifacts, fat). (2) Model the visual saliency of breast tumors. Visual saliency is a property that makes an object in images stand out from neighboring objects. We will overcome the invalid assumption made in previous approaches that there is at least one tumor in the image by developing a robust model for estimating visual saliency of breast tumors. With the help of this model, we will detect all possible tumor regions that would attract a radiologist?s attention, with no output of salient regions when no tumor exists in an image. (3) Develop a domain-enriched deep learning framework for tumor segmentation. A deep learning-based framework will be developed to integrate the output of models from Aims 1 and 2 and will lead to an overall model that segments breast tumors. We will train and test the approach using 1800 breast ultrasound images from four medical schools collected using five different ultrasound devices. Seven quantitative metrics will be applied to evaluate the performance of the proposed segmentation approach. Discrepancies between computational and manual tumor segmentation will be used to refine the models. Success of the proposed project will enhance methodologies for robust and reproducible breast ultrasound image segmentation and broaden the use of computer-aided diagnosis for early detection of breast cancer.
Showing the most recent 10 out of 29 publications