Section: New Results

Image-based modeling

Participants : Marie-Odile Berger, Charlotte Delmas, Antoine Fond, Erwan Kerrien, Gilles Simon, Pierre-Frédéric Villard, Brigitte Wrobel-Dautcourt.

Finding Manhattan directions in uncalibrated images

Finding orthogonal vanishing points (VPs) in a photography has many potential applications in computer vision and computer graphics, including perspective correction, image-based reconstruction and texture extraction. Surprisingly, while this problem has been extensively studied in the literature, manual solutions are still used in most software. Existing algorithms generally follow two steps. First, lines are grouped into pencils, whose centers are considered as potential VPs. Then, an orthogonality measure is evaluated for every triplet of VPs and the most plausible triplet is used for camera calibration. A drawback of this approach is that complex and time-consuming techniques have to be used to solve the general problem of VP detection, while only three particular VPs are finally used. By contrast, we propose an effective and easy-to-implement algorithm that estimates the zenith and the horizon line before detecting the VPs, using simple properties of the central projection and exploiting accumulations of oriented segments around the horizon. Our method is fast and yields an accuracy comparable, and even better in some cases, to that of state-of-the-art algorithms. This work was submitted to Eurographics 2016.

Tools reconstruction for interventional neuro-radiology

Minimally invasive techniques impact surgery in such ways that, in particular, an imaging modality is required to maintain a visual feedback. Live X-ray imaging, called fluoroscopy, is used in interventional neuroradiology. Such images are very noisy, and cannot show but the vasculature and no other brain tissue. In particular, since at most only two projective fluoroscopic views are available, containing absolutely no depth hint, the 3D shape of the micro-tool (guidewire, micro-catheter or micro-coil) can be very difficult, if not impossible to infer, which may have an impact on the clinical outcome of the procedure.

In collaboration with GE Healthcare, we aim at devising ways to reconstruct the micro-tools in 3D from fluoroscopy images. Charlotte Delmas has been working as a PhD CIFRE student on this subject since April 2013. She first devised a solution in a two-view reconstruction context. A paper reporting on this work was published and an oral presentation was made at SPIE Medical Imaging 2015 [19] . The focus was made this year on reconstructing the first coil, a single wire that tangles into a complex pattern when deployed in an aneurysm. The challenge is to produce a 3D reconstruction with as little X-ray dose and acquisition time as possible. Two paths were simultaneously followed this year: 1) design and compare various configurations to rapidly shoot 6 X-ray views from different viewpoints with a biplane (stereo) system; 2) compensate the lack of information (small number of images) with a prior, incorporated in the tomographic reconstruction algorithm, to express the sparsity in space of the shape to be reconstructed. Preliminary work sets forward a simultaneous fast rotation of both planes around the patient's head. A database is currently being acquired for a full validation in a view to submitting this work for publication early next year.

Individual-specific heart valve modeling

Mitral valve surgical repair is a complex procedure where the outcome largely depends on the surgeons’ experience. Predicting a good coaptation of the two leaflets post-operatively is one of the most difficult sub-tasks in the procedure. We worked on a biomechanical simulation framework [25] that computes the leaflet deformation and coaptation based on individual-specific microtomography data as an initial step toward patient-based mitral valve surgical repair assistance through simulation. Results from FEM analysis on three explanted porcine hearts showed that it is possible to obtain the real shape of the leaflets during systolic peak. We also measured the influence of the positions of chordae tendineae on the simulation and showed that marginal chordae have a greater influence on the final shape than intermediary chordae.

Quasi-periodic noise removal

Images may be affected by quasi-periodic noise. This undesirable feature manifests itself by spurious repetitive patterns covering the whole image, well localized in the Fourier domain. While notch filtering permits to get rid of this phenomenon, this however requires to first detect the resulting Fourier spikes, and, in particular, to discriminate between noise spikes and spectrum patterns caused by spatially localized textures or repetitive structures. Several approaches have been investigated. First, we have reviewed the available methods, most of them requiring expert tuning, with applications to experimental mechanics in view [11] . We have also proposed two automated patch-based approaches. A parametric approach has been investigated in [14] (more information available in [26] ), based on the detection of noise spikes as statistical outliers to the distribution expected from natural non-noisy patches, which is known to follow the inverse of a power of the frequency. A non-parametric approach, based on a-contrario detection, was also proposed in [22] .