Development of a patient-specific deformable image registration model for breasts using positron emission tomography combined with magnetic resonance imaging by biomechanical strategy

Abstract

The simulation of large deformations of the breast has great potential for applications in the medical field, such as breast cancer diagnosis, image guided surgery, surgery planning and breast image registration. However, the positioning of the patient body will differ during each screening modality. Large-scale deformations of the breast during movement mean that modeling of the breast is a difficult task. It is therefore necessary to formulate a mechanical model of the breast that can predict the deformations of the breast during scanning. In this thesis, I propose an individualized biomechanical model to predict large-scale deformations of the breast in the supine to prone positions. The model combines finite element analysis with affine transformation. The mechanical properties of the breast tissues are individually assigned by using an optimization process, which allows the model to be patient-specific. Image registration with the use of positron emission tomography (PET) and magnetic resonance imaging (MRI) has been extensively studied in the literature. The biomechanical model of the breast is thus evaluated by using MRI and PET/computed tomography images from Hong Kong and American samples. The differences in the breast volume and density are determined by the biomechanical model in this study. Deformations in the breast images of both the Asian and American samples due to the effect of gravity are successfully modeled by using the finite element method. The accuracy of the developed model is determined by using the target registration error (TRE) of the lesion. The TRE for the Hong Kong and American samples is 4.77±2.20 mm and 8.40±7.15 mm, respectively. The results show that this model is able to accurately predict deformations of the breast in the supine to prone positions for images from both populations. In addition, the TRE has been found to be correlated with the image density, which indicates that this model can more accurately predict deformations of breasts with less density. A decision tree has also been generated through data mining to predict the registration accuracy.