MULTI-VIEW MAMMOGRAPHIC IMAGE REGISTRATION USING CONVOLUTIONAL NEURAL NETWORKS

Abstract

Breast cancer is one of the leading causes of death for women worldwide. Research has shown that annual screening exams can aid in early detection, which can be critical for saving lives or even for less invasive treatment. A key part of annual screening involves clinicians viewing multiple X-ray images of the breasts, where each image is taken from a different viewing angle. This helps with confirming potential lesions and also with assessing the need for follow-up. However, it can sometimes be difficult to see a lesion in subsequent image views due to obscurations, such as in dense tissue cases where the background breast tissue will have similar X-ray attenuation compared to the lesion. In this dissertation, we present research on deformation field-based convolutional neural network (CNN) image registration algorithms that can aid clinicians by automatically finding the location of the same lesion in multiple X-ray image views. We focus on the two standard mammographic viewing angles, the craniocaudal (CC) and the mediolateral oblique (MLO). A key contribution from our work is the development of a novel distance-based regularization (DBR) technique which proves to significantly increase the networksÕ ability to find the corresponding location for a lesion between two X-ray image views. We tested our algorithms on computer-simulated as well as real 2D and 3D X-ray images. We characterized the networksÕ performance from a variety of aspects including differences in breast density, lesion size, and Breast Imaging-Reporting and Data System (BI-RADS) categorizations for the lesions. We compared our networksÕ performance with other algorithms and showed that they outperformed both state-of-the-art CNN and non-CNN-based registration techniques. We also researched and experimented with techniques for estimating uncertainties (i.e., confidence values) for our networksÕ registration outputs, based on deep ensemble-based methods. The results showed that the approaches have potential for providing indications of uncertainty for the CC/MLO lesion mappings. Lastly, we demonstrated the performance of our registration algorithms as a part of a larger system which uses a multi-view, multi-modality fusion approach, called Upstream Data Fusion (UDF). The UDF system is designed to detect and diagnose lesions, with minimal false alarms. The CNN registration enhanced the systemÕs ability to provide improved lesion detection and false-alarm reduction. Our research shows that the proposed CNN registration techniques provide a useful means for co-locating lesions between the CC and MLO views, even in challenging cases. The methods show promise for aiding clinicians to establish lesion correspondence quickly and confidently in dual-view X-ray mammography, thereby, improving diagnostic capability.