Automated image registration for dynamic contrast-enhanced breast magnetic resonance imaging

Abstract

[en] This thesis presents an automated approach to image registration for dynamic contrastenhanced breast magnetic resonance imaging (MRI), a critical task in medical diagnostics that enhances the analysis and interpretation of sequential image data. Image registration, particularly within the domain of breast MRI, faces significant challenges due to the deformable nature of breast tissue and the high degree of accuracy required for effective diagnosis and treatment planning. The work employs advanced machine learning models to develop an efficient and robust system capable of aligning multiple MRI scans over time with high precision. The primary methodological contribution of this thesis is the integration of a convolutional neural network model, designed to adapt to the unique complexities presented by the high variability and dynamic changes in breast MRI scans. This approach facilitates improved diagnostic capabilities by enhancing the temporal analysis of contrast patterns in breast tissue, which is crucial for identifying and monitoring various pathological conditions. Experimental results demonstrate the effectiveness of the proposed system, which achieved a Dice score of 0.782 ± 0.009 in one of the models and demonstrate substantial improvements in alignment efficiency compared to traditional image registration techniques. The system’s ability to provide rapid and precise alignments promises significant benefits for clinical practices, including better monitoring of disease progression and more tailored treatment strategies for breast cancer patients.