Publications

2016

H.-E. Gueziri, L. Lakhdar, M. McGuffin and C. Laporte, FastDRaW – Fast Delineation by Random Walker: application to large images, Presented at MICCAI Workshop on Interactive Medical Image Computing, Athens, Greece.

Show Abstract

Random walker (RW) segmentation is widely recognized for its robustness in medical image segmentation tasks. However, its computational cost precludes real-time visual feedback in interactive settings. This paper introduces FastDRaW, a fast multi-scale segmentation approach using the RW algorithm. Our approach relies on a two-step procedure. First, on a low-resolution version of the image, a coarse RW segmentation is computed over a restricted region of interest. Second, the result is refined by applying the RW algorithm at full resolution over a narrow strip around the coarse contour. For a standard medical image size of 512 × 512 pixels, segmentation is thus achieved in 0.13 ± 0.003 s, allowing continuous visual updates as the user labels the image ; while keeping an interactive response time of 1.23 ± 0.03 s for images of size 1500 × 1500. In addition to computation time reduction, experiments show a significant decrease in the required number of labels to achieve the segmentation.

[pdf] [project]

H.-E. Gueziri, S. Tremblay, C. Laporte and R. Brooks, Graph-based 3D-Ultrasound reconstruction of the liver in the presence of respiratory motion, LNCS Vol. 10129, pp.48-57, Reconstruction, Segmentation, and Analysis of Medical Images: MICCAI 2016 Workshops, Athens, Greece.

Show Abstract

In this paper, we explore the feasibility of 3D ultrasound (US) reconstruction of the liver in the presence of respiratory motion using a minimally cumbersome acquisition protocol involving a commonly available tracked 2D+t wobbler US probe. We exploit measurements of the probe’s displacement against the skin to coarsely assign frames to their corresponding respiratory states. These assignments are refined using a graph representation of the spatial adjacency relationships and appearance continuity between the frames. Finally, frames providing the smallest motion variation, within a respiratory state, are first selected using a shortest path strategy, then passed to the reconstruction algorithm. Our method is fully based on tracked US imaging and does not require a pre-operative reference image. Moreover, no breath-control effort is required on the part of the patient, thereby limiting the complexity of the acquisition protocol. We tested our approach with an intercostal acquisition protocol, demonstrating enahncements in stability and the quality of liver reconstruction at different respiratory states, compared to a naive gating approach.

[pdf] [bibtex]

H.-E. Gueziri, M. J. McGuffin, C. Laporte, A Generalized Graph Reduction Framework for Interactive Segmentation of Large Images, Computer Vision and Image understanding, Vol. 150, pp. 44-57.

Show Abstract

The speed of graph-based segmentation approaches, such as random walker (RW) and graph cut (GC), depends strongly on image size. For high-resolution images, the time required to compute a segmentation based on user input renders interaction tedious. We propose a novel method, using an approximate contour sketched by the user, to reduce the graph before passing it on to a segmentation algorithm such as RW or GC. This enables a significantly faster feedback loop. The user first draws a rough contour of the object to segment. Then, the pixels of the image are partitioned into “layers” (corresponding to different scales) based on their distance from the contour. The thickness of these layers increases with distance to the contour according to a Fibonacci sequence. An initial segmentation result is rapidly obtained after automatically generating foreground and background labels according to a specifically selected layer; all vertices beyond this layer are eliminated, restricting the segmentation to regions near the drawn contour. Further foreground/background labels can then be added by the user to refine the segmentation. All iterations of the graph-based segmentation benefit from a reduced input graph, while maintaining full resolution near the object boundary. A user study with 16 participants was carried out for RW segmentation of a multi-modal dataset of 22 medical images, using either a standard mouse or a stylus pen to draw the contour. Results reveal that our approach significantly reduces the overall segmentation time compared with the status quo approach ( p < 0.01). The study also shows that our approach works well with both input devices. Compared to super-pixel graph reduction, our approach provides full resolution accuracy at similar speed on a high-resolution benchmark image with both RW and GC segmentation methods. However, graph reduction based on super-pixels does not allow interactive correction of clustering errors. Finally, our approach can be combined with super-pixel clustering methods for further graph reduction, resulting in even faster segmentation.

[pdf] [bibtex]

2015

H.-E. Gueziri, M. McGuffin and C. Laporte, User-guided graph reduction for fast image segmentation, IEEE International Conference on Image Processing, pp. 286-290.

Show Abstract

Graph-based segmentation methods such as the random walker (RW) are known to be computationally expensive. For high resolution images, user interaction with the algorithm is significantly affected. This paper introduces a novel seeding approach for graph-based segmentation that reduces computation time. Instead of marking foreground and background pixels, the user roughly marks the object boundary forming separate regions. The image pixels are then grouped into a hierarchy of increasingly large layers based on their distance from these markings. Next, foreground and background seeds are automatically generated according to the hierarchical layers of each region. The highest layers of the hierarchy are ignored leading to a significant graph reduction. Finally, validation experiments based on multiple automatically generated input seeds were carried out on a variety of medical images. Results show a significant gain in time for high resolution images using the new approach.

[pdf] [bibtex]

2014

H.-E. Gueziri, M. McGuffin and C. Laporte, Visualizing Positional Uncertainty in Freehand 3D Ultrasound, Proc. SPIE 90361H, Medical Imaging: Image-Guided Procedures, Robotic Interventions, and Modeling, San Diego, USA.

Show Abstract

The freehand 3D ultrasound technique relies on position sensors attached to the probe to register the location of each image to a 3D space. However, the imprecision of the position sensors reduces the reliability of estimated image locations. In this paper, we propose a novel method to compute the positional uncertainty of an image plane. First, we use rigid body point-based registration to compute the error produced by each pixel of the image during the tracking. The Target Registration Error (TRE ) is used to compute the covariance matrix of errors at each pixel position. This covariance matrix is then decomposed as a 3D orientation error, in the x, y and z directions. Considering a volume around the image, we introduce the Image Plane Crossing Probability (IPCP) to determine the probability that the plane passes through each voxel. The result is a point cloud probability around the image plane, where each voxel contains the crossing probability and the contribution of each direction of the error. Finally, a simple volume rendering technique is used to visualize the uncertainty of the plane position. The results are validated in two steps. The first step is a Monte Carlo simulation to validate the estimate of the TRE covariance for the tracking errors. The second step simulates TRE errors on a plane and validates the associated positional uncertainty.

[pdf] [bibtex]

2011

H.-E. Gueziri, Segmentation d’objet dans un volume échographique, Master of Science Thesis, 51 pages. Department (UFR) of Mathematics and Computer Science, University of Paris Descartes.

[pdf]

2010

D. Medjahed Gamaz, H.-E. Gueziri and N. Haouchine, Modelling Postures of Human Movements. CIARP 15th IberoAmerican Congress on Pattern Recognition, pp. 202-211, Brazil.

[pdf] [bibtex]

H.-E. Gueziri, N. Haouchine, Construction de modèle pour la reconnaissance du mouvement humain, Bachelor of Science Thesis, 100 pages. Department of Computer Science, University of Sciences and Technology Houari Boumediene.

[pdf]