Development of a clinically applicable matching tool between MRI and close-up X-ray mammograms for the purpose of MRI-matching guided stereotactic breast biopsy
Final Report Abstract
The aim of this project was to develop and investigate a novel method for patient specific matching of magnetic resonance imaging (MRI) and close-up stereotactic biopsy X-ray to fit a mammogram-based lesion workup and thereby reduce the need for costly and time-consuming MRI biopsies. Within the project we developed the fundamental methods to transfer positions of lesion, which are only visible in MRI, onto spot mammograms for therapy planning and evaluation. The method is based on a two-step image registration which first matches the MRI to a full mammogram using a biomechanical simulation, then matches the full mammogram with the spot mammogram using an image-based registration. The matching based on the biomechanical model tackles the huge deformations of the breast during X-ray mammography. To automatically create the patient-specific model, we developed a robust image segmentation method based on unsupervised learning to extract the breast geometry from MRI volumes. Using a meshing, adding a material model and boundary conditions enables us to formulate and solve the deformation simulation with the Finite Element Method (FEM). Since the computation with FEM is demanding we successfully developed and investigated in depth the possibility to compute deformation simulations of the breast using a machine learning approach. Once trained it can predict the deformed state of the breast during mammography in a few seconds with only a small error of approx. 3 mm compared to FEM. Finally we developed a GPU-based ray tracing to synthesize X-ray mammograms. This approach is used to either synthesize full mammograms or spot mammogram, which can in the second step be registered to the real world equivalent. For matching full mammograms and spot mammograms we developed and validated a method based on image similarity: after finding a region of interest in the full mammogram, the best overlap of spot mammogram and full mammogram is computed by optimizing image similarity metrics. We combined several of those to get a robust matching result. Investigations with respect to patient characteristics revealed data characteristics influencing the matching accuracy: lesions centrally or at the tip of the breast can e.g. be matched with higher accuracy than lesions close to the chest muscle. The described method was initially investigated on clinical data using a simulation model for stereotactic guided biopsy. The aim of this simulation was to determine the success rate (correct sampling of the target lesion) using the area predicted by the model to center a stereotactic guided biopsy. After setting up a clinical database with a total of 47 patients, for which cases with suspicious calcifications subsequently biopsied under stereotactic guided biopsy were included, we performed several evaluations. Using 14 datasets in which the compression angle in full mammogram and spot mammogram was both cranio-caudal, the median TRE of the matching between MRI and full mammogram was 20.9 mm, the TRE of the matching between full mammogram and spot mammogram was 14.5 mm and the TRE of the complete process of transferring the lesion from the MRI to the spot mammogram was 27.8 mm. From 14 cases investigated in this way, in 6 cases an X-ray guided biopsy could be considered successful since the simulated biopsy area overlaps with the lesion position in the spot mammogram. Additionally, we investigated the matching medio-lateral oblique mammograms. In summary, we demonstrated in an initial feasibility study that biopsy planning based on our results would lead to a successful outcome in several cases. Hence, the present work is an essential first step to allow X-ray guided lesion work-up, even if the lesion is only visible in MRI. Beside fulfilling the aims of the proposed project, the developed methods and results such as the matching of diffusion weighted MRI and conventional MRI are highly relevant for direct clinical use.
Publications
-
Image registration between MRI and spot mammograms for X-ray guided stereotactic breast biopsy: preliminary results. Medical Imaging 2021: Image-Guided Procedures, Robotic Interventions, and Modeling (2021, 2, 15), 45. American Geophysical Union (AGU).
Said, Sarah; Clauser, Paola; Ruiter, Nicole V.; Baltzer, Pascal A. T. & Hopp, Torsten
-
X-ray Synthesis Based on Triangular Mesh Models Using GPU-Accelerated Ray Tracing for Multi-modal Breast Image Registration. Lecture Notes in Computer Science (2021), 87-96. American Geophysical Union (AGU).
Maul, J.; Said, S.; Ruiter, N. & Hopp, T.
-
Image based registration between full x-ray and spot mammograms for x-ray guided stereotactic breast biopsy. Medical Imaging 2022: Image-Guided Procedures, Robotic Interventions, and Modeling (2022, 4, 4), 92. American Geophysical Union (AGU).
Said, Sarah; Clauser, Paola; Ruiter, Nicole; Baltzer, Pascal & Hopp, Torsten
-
MRI breast segmentation using unsupervised neural networks for biomechanical models. 16th International Workshop on Breast Imaging (IWBI2022) (2022, 7, 13), 18. American Geophysical Union (AGU).
Said, Sarah; Meyling, Michael; Huguenot, Rémi; Horning, Marcel; Clauser, Paola; Ruiter, Nicole; Baltzer, Pascal & Hopp, Torsten
-
Image based registration between full x-ray and spot mammograms: analysis of registration accuracy in subgroups. Medical Imaging 2023: Image-Guided Procedures, Robotic Interventions, and Modeling (2023, 4, 3), 63. American Geophysical Union (AGU).
Said, Sarah; Clauser, Paola; Ruiter, Nicole; Baltzer, Pascal & Hopp, Torsten
-
Image registration of diffusion weighted and conventional breast MRI. Medical Imaging 2023: Image Processing (2023, 4, 3), 66. American Geophysical Union (AGU).
Hopp, Torsten; Tabet, Ibrahim; Said, Sarah; Clauser, Paola; Baltzer, Pascal & Ruiter, Nicole
