Charlotte Delmas, Phd student in Magrit team will defend her thesis on the 10th of November at 1:30 pm in room C005. The thesis is entitled :”3D reconstruction of curvilinear micro-devices in interventional neuroradiology”.
This CIFRE thesis was co-supervised by Yves Trousset, Cyril Riddell (GE Healthcare), Marie-Odile Berger and Erwan Kerrien (INRIA Nancy) in partnership with the CHU of Nancy.
The clinical practice for intra-cranial endovascular treatments using a biplane image guiding system makes for two situations where a significantly faster and less irradiating alternative to tomography is possible : the 3D follow-up of a guidewire moving within cerebral vascular twists, and the position and configuration assessment of a coil unfurled inside an aneurysm. Those two curvilinear devices we aim to reconstruct in 3D may be elongated in 3D (guidewire) or spiraled into a ball shape (coil).
We propose a segmentation-based stereoscopic reconstruction method to reconstruct a guidewire from two orthogonal views.
The task of segmenting a guidewire from fluoroscopic images was boiled down to the succession of a denoising step, a simple thresholding followed by skeletonization. We propose a single original filter designed from standard diffusion filters. It is optimized for denoising a guidewire modeled as a 2D curvilinear structure appearing not very contrasted against a uniform and noisy background.
The reconstruction phase relies on 3D hypothesis generation. They are expressed as 3D curve fragments obtained from the stereoscopic matching of easily parameterized 2D curve fragments using the input skeleton. Since some hypothesis may have risen from geometrical configurations with no physical reality, a search performed in a graph modeling hypothesis (nodes) and constraints (edges) allows for the identification of a subset of hypothesis that, once linked together, describes the support for a smooth 3D curve corresponding to the guidewire reconstruction.
However a curvilinear structure as complex as a coil cannot be precisely reconstructed using stereoscopy.
We propose a tomographic reconstruction method necessitating 6 views uniformly distributed over 180 degrees. Because of its curvilinear nature, we can make use of the coil’s sparsity to constrain the reconstruction problem. The resulting reconstruction algorithm is iterative and alternates between soft-thresholding and 3D directional filtering (by extension of the diffusion filter previously developed in 2D) which allows for a significant reduction of angular subsampling artifacts.
Those two reconstruction methods were validated on clinical data in collaboration with experienced neuroradiologists.