X-Ray Breast Imaging Optimization: Towards Improved Cancer Screening and Diagnosis
Breast cancer is a significant global health concern, requiring continuous advancements in breast imaging to improve early detection and patient outcomes. Early diagnosis is crucial for effective breast cancer treatment and improved survival rates. Imaging techniques, such as digital mammography (DM), ultrasound, and magnetic resonance imaging, play a pivotal role in early detection. However, no single method can detect all cases or rule out those without the disease. This thesis attempts to address critical challenges in breast imaging, particularly for digital breast tomosynthesis (DBT) and contrast enhanced mammography (CEM), and enhance its effectiveness through the integration of computational models and artificial intelligence (AI)-based technologies. We investigated methods for enhancing image quality and interpretation in x-ray-based breast imaging modalities, contributing to the development of improved screening and diagnostic tools.
Chapter 2 introduced a three-dimensional breast shape model under compression, enabling the generation of realistic breast phantoms representing the breast cancer screening population. These phantoms can be of great aid in research and development, allowing for the simulation of various clinical scenarios to develop breast image processing algorithms without subjecting patients to unnecessary additional x-ray exposures.
Chapter 3 delved into developing a deep learning-based scatter estimation method for DBT and grid-less DM images, significantly reducing the computation time required for scatter correction. This approach improved the overall image quality by mitigating the presence of scatter radiation, setting the stage for enhanced post-processing and reconstruction methods.
Chapter 4 extended the scatter correction method to CEM, focusing on low- and highenergy images, while introducing a novel approach for estimating breast thickness through deep learning. While the impact of scatter correction was smaller than expected, this work provided a marked improvement in the quality of recombined CEM images, particularly for the purposes of enhancing lesion visibility and iodine quantification.
Chapter 5 addressed the need for a breast shape model in the medio-lateral oblique (MLO) view to complement the cranio-caudal (CC) view model introduced in Chapter 2. A novel scanning setup using smartphone-based infrared cameras allowed us to collect MLO breast surface data, contributing to the development of an MLO-based breast shape model. 166
Chapter 6 integrated AI into single-view DBT, aiming to assess its impact on radiologist performance. The study revealed that AI-based decision and navigation support can enhance radiologist accuracy, particularly in reducing false-positive recalls, while maintaining stable reading times. Overall, this thesis has laid the foundation for advancements in breast imaging by combining computational models and AI-based technologies. These innovations hold the potential to transform breast cancer screening and diagnosis, ultimately improving patient outcomes and the efficient utilization of healthcare resources.