The segmentation and tracking of the heart in ultrasound sequences is a challenging problem, which is still unsolved, in its full generality despite recent advances in this area. An overview is available in [Frangi01]. The main difficulties concern the presence of complex motion patterns of the heart and the low quality of ultrasound data mainly due to speckle noise, edge dropout effect caused by motion, the presence of shadows produced by the dense muscles, and the low signal to noise ratio. Most of the proposed solutions follow two trends: 1) the use of deformable model trackers based on low-level image features (e.g., edges) [Nascimento08] and 2) pattern recognition methods based on high-level visual features, automatically learned with the objective to minimize the probability of recognition errors. In the literature, the deformable models are usually called bottom-up approaches, while pattern recognition models are known as top-down approaches [Carneiro08]. This project proposes a combination of the bottom-up and top-down approaches for solving the problem of segmenting and tracking the left ventricle (LV) of the heart in 2-D ultrasound data. This will allow a significant improvement of previous methods [Paragios03] since we are combining edge and motion information from bottom-up with visual appearance models used in top-down [Zheng08]. Comparing with bottom-up, we expect to obtain improved robustness in the case of edge drop out and rapid LV motion as well as an automatic procedure for contour initialization. Comparing with top-down we will obtain a significant improvement since we will use additional sources of information given by the heart dynamic model and image edges. This approach will allow the reduction of the number of training images used in top-down, solving the major drawback of this class of techniques.
Objectives:
The objective of this work is the implementation of a new approach for the problem of automatic left ventricle segmentation from ultrasound data. This new approach is essentially a combination of the top-down and bottom-up methods, and we target a segmentation system that is at the same time more robust to ultrasound imaging artefacts and less dependent on a large set of training data.
Detailed description:
Typically, left ventricle (LV) segmentation from ultrasound data sequence in time involves the following three steps: 1) contour prediction, 2) image measurement and 3) filtering. The contour prediction is based on the detection done at the previous time step and on the motion model of the heart. Using these two pieces of information, it is possible to estimate the approximate segmentation of the LV in the next time step, which works as a segmentation prior. Using this prior, the system then starts the analysis of the ultrasound data at the current time step, which produces an estimate of the LV contour using edge information (bottom-up approach) and the ultrasound imaging global patterns (top-down approach). The global patterns will be used by a pattern recognition method to obtain estimates of the deformable contour and its uncertainty. Finally, the filtering step combines the information from different sources (motion prediction, edges and pattern recognition estimate of the contour) in a Bayesian setting. The use of top-down estimates of shape allows a reduction of the uncertainty and improves the rejection of outliers. It plays a key role if the edge information is missing and also allows an automatic initialization of the contour in the first frame. Specifically, the following steps are planned in the work:
1. Literature review in the area of LV segmentation from ultrasound data sequence.
2. Design and implementation of the bottom-up/top-down approach.
3. Experiments and comparison using publicly available databases.
References:
[Frangi01] - A. F. Frangi, W. J. Niessen, and M. A. Viergever, "Threedimensional modeling for functional analysis of cardiac images: A review", IEEE Trans. Med. Imag., 20(1):2-25, 1 2001.
[Nascimento08] - J. Nascimento, J. S. Marques, "Robust Shape Tracking with Multiple Models in Ultrasound Images", IEEE Transactions on Image Processing, pp. 392-406, vol. 17, no. 3, March 2008.
[Carneiro08] - Gustavo Carneiro, Bogdan Georgescu, Sara Good, and Dorin Comaniciu, "Detection of Fetal Anatomies from Ultrasound Images using a Constrained Probabilistic Boosting Tree", IEEE Transactions on Medical Imaging, 2008.
[Paragios03] - N. Paragios, "A level set approach for shape-driven segmentation and tracking of the left ventricle", pp. 773-776 IEEE Transactions Medical Imaging, 2003.
[Zheng08] - Y. Zheng, A. Barbu, B. Georgescu, M. Scheuring and D. Comaniciu, "Four-Chamber heart modeling and automatic segmentation for 3D cardiac CT volumes using marginal space learning and steerable features",IEEE Transactions Medical Imaging, 2008 Requirements (grades, required courses, etc):
Expected results:
At the end of the work, the students will have enriched their experience in computer vision, machine learning and medical image analysis. In particular the following goals are expected:
- Deep understanding of current state-of-the-art LV segmentation from ultrasound data sequence.
- Design and implementation of the bottom-up/top-down approach.
- State-of-the-art results in the area of LV segmentation.