Candidate: Jose Rincon
Title: Facilitating Breast-conserving Surgery Using Presurgical MRI
Date: April 16, 2018
Time: 10:00am
Place: EIT 3142
Supervisor(s): Michailovich, Oleg
Abstract:
Magnetic resonance (MR) breast images captured before a patient's surgery provide great detail between breast tissues and their spatial information in 3D. Commonly, preoperative MR breast images are taken in a prone position, however, the breast surgery for tumour extraction is performed in a supine position. If one could properly align the preoperative breast image with the position of the patient during surgery, then surgeons could effectively remove breast cancer with adequate margins. Thus, saving the life of the patient while minimizing aesthetic damage to their breast. Due to the significant shape deformation of a breast between prone to supine positions, previous studies have shown that either non-rigid intensity-based image registration or biomechanical modelling using finite elements (FE) alone are able to model such great deformations. Previous studies deform the breast using FE based biomechanical models to estimate the deformation of the breast due to gravity. Since FE analysis uses a simplification of the data by triangulating the MRI data into a mesh, as well as using approximations not only about the biomechanical properties of the breast but also of the external forces that act upon the breast, then a non-rigid intensity-based registration is required to recover the remaining deformation from the supine image.
The state-of-the-art technique requires a triangulated mesh consisting of nodes with different biomechanical properties depending on if the tissue belongs to either fibro-glandular or fatty tissue. Thus, a two-class image segmentation is required for this task. The main problem with breast MRI segmentation is that there is not a clear boundary between the pectoralis muscle and the breast. In recent contributions, authors have opted for localizing the breast boundary manually. This is a very tedious and time-consuming task to do. They also assume that MRI data is a realization of a Gaussian Mixture Model (GMM). In reality, a Rician Mixture model (RMM) is a better model for noisy breast MRI data.
This thesis proposes an approach for detecting the boundary between the pectoralis muscle and the breast based on Dijkstra's algorithm. A novel method for two-class image segmentation under the assumption of an RMM is presented as well.
The prone-to-supine image registration pipe-line was implemented and tested on three patients in order to show how our two contributions can be incorporated into this problem.