Enhanced Digital Breast Tomosynthesis diagnosis using 3D visualization and automatic classification of lesions

Breast cancer represents the main cause of cancer-related deaths in women. Nonetheless, the mortality rate of this disease has been decreasing over the last three decades, largely due to the screening programs for early detection. For many years, both screening and clinical diagnosis were mostly done through Digital Mammography (DM). Approved in 2011, Digital Breast Tomosynthesis (DBT) is similar to DM but it allows a 3D reconstruction of the breast tissue, which helps the diagnosis by reducing the tissue overlap. Currently, DBT is firmly established and is approved as a stand-alone modality to replace DM. The main objective of this thesis is to develop computational tools to improve the visualization and interpretation of DBT data. Several methods for an enhanced visualization of DBT data through volume rendering were studied and developed. Firstly, important rendering parameters were considered. A new approach for automatic generation of transfer functions was implemented and two other parameters that highly affect the quality of volume rendered images were explored: voxel size in Z direction and sampling distance. Next, new image processing methods that improve the rendering quality by considering the noise regularization and the reduction of out-of-plane artifacts were developed. The interpretation of DBT data with automatic detection of lesions was approached through artificial intelligence methods. Several deep learning Convolutional Neural Networks (CNNs) were implemented and trained to classify a complete DBT image for the presence or absence of microcalcification clusters (MCs). Then, a faster R-CNN (region-based CNN) was trained to detect and accurately locate the MCs in the DBT images. The detected MCs were rendered with the developed 3D rendering software, which provided an enhanced visualization of the volume of interest. The combination of volume visualization with lesion detection may, in the future, improve both diagnostic accuracy and also reduce analysis time. This thesis promotes the development of new computational imaging methods to increase the diagnostic value of DBT, with the aim of assisting radiologists in their task of analyzing DBT volumes and diagnosing breast cancer.