Left & right ventricle detection using AAMs

Automated left and right ventricular contour detection in cardiac MR and CT using Active Appearance

Background

Active appearance models (AAMs) have been successfully used for image segmentation in cardiac MRI. AAM segmentation consists of two different phases: During a training phase, it learns the relationship between model parameter displacements and the residual errors induced between a training image and a synthesized model. During the matching (detection) phase, it attempts to find the best fit of the model to the data in a new image. Capturing both shape and texture from a training set of manually segmented images the AAM segmentation is highly specific to the task given. However, the cumbersome manual annotation of the images in the training sets makes it difficult to implement the method in clinical routine. In clinical practice there is a large variation in image characteristics as a result of variation in MR hardware and software (MR systems from different vendors, different surface coils, and different pulse sequences) and the large spectrum of cardiac pathologies. Application of an AAM based-algorithm is therefore challenging, since the method should be reliable under varying conditions.

Goals

The goal of this project is to make the AAM approach applicable in short axis MR clinical practice. We addressed two different questions:

  • What is the optimal number of images to be included in the training set?
  • What is the optimal mixture of images from healthy and pathological cases in the training set?
  • Is it necessary to construct separate models for different vendors, or is it sufficient to combine data from multiple vendors in a (larger) vendor independent model?
  • What is the impact of image inhomogeneity on the accuracy of AAM segmentation?

Approach

This project consists of two different phases: first the implementation of the AAM algorithm to be suitable to automatically segment short axis cardiac MR images. Second, the optimization of the algorithm that consists of describing an optimal protocol for the use in clinical routine.

Preliminary results

We demonstrated that AAM based contour detection can be used in cardiac MR imaging studies in clinical practice. In other words, defining an appropriate training set of at least 180 data sets is a crucial step towards obtaining high quality results of the segmentation based on AAM. Furthermore, the best training set distribution of images from normal and pathological ventricles seems to be 80%-20%. Finally, in case MR scanners from multiple vendors are used, it is essential to define different models for each of the vendors. The inclusion of low quality images in the training set should be avoided.

AAM Result
Figure 1
Result of endocardial and epicardial contour detection in an end-diastolic (left) and end-systolic (right) image.

AAM Matching Movie
Movie 1 (2.7Mb)
Example of AAM matching process for a mid-ventricular image. It shows successful detection of the endocardial and epicardial contours even when the initial location of the AAM shape is initialized far away from center of the left ventricular cavity.

Status

Finished.

Contact

Rob J. van der Geest, PhD
Division of Image Processing
Department of Radiology, 1-C2S
Leiden University Medical Center
P.O. Box 9600
2300 RC Leiden
The Netherlands
Tel. +31 (0)71 526 2138
Fax. +31 (0)71 526 6801
e-mail: R.J.van_der_Geest@lumc.nl

Publications

  1. Mitchell SC, Lelieveldt BPF, van der Geest RJ, Bosch JG, Reiber JHC, Sonka M. Segmentation of cardiac MR images: An active appearance model approach. proc. SPIE Medical Imaging 2000 Image Processing 2000;3979:224-234.
  2. Mitchell SC, Lelieveldt BPF, van der Geest RJ, Bosch JG, Reiber JHC, Sonka M. Multistage hybrid active appearance model matching: Segmentation of left and right ventricles in cardiac MR images. IEEE Trans Med Imag 2001;20:415-423.
  3. Mitchell SC, Bosch JG, Lelieveldt BPF, van der Geest RJ, Reiber JHC, Sonka M. 3-D Active appearance models: Segmentation of cardiac MR and ultrasound images. IEEE Med Imag 2002;21(9):1167-1178.
  4. Lelieveldt BPF, Uzümcü M, van der Geest RJ, Reiber JHC, Sonka M. Multi-view active appearance models for consistent segmentation of multiple standard views: application to long- and short-axis cardiac MR images. In: Computer Assisted Radiology and Surgery - CARS 2003. HU Lemke, MW Vannier, K Inamura, AG Farman, K Doi, JHC Reiber (Eds.). Elsevier Science BV 2003: 1141-1146.
  5. Üzümcü, M, van der Geest RJ, Sonka M, Lamb HJ, Reiber JHC, Lelieveldt BPF. Multiview active appearance models for simultaneous segmentation of cardiac 2- and 4-chamber long-Axis magnetic resonance images. Invest Radiol 2005;40:195-203
  6. van der Geest RJ, Lelieveldt BPF, Angelié E, Danilouchkine M, Swingen C, Sonka M, Reiber JHC. Evaluation of a new method for automated detection of left ventricular boundaries in time series of magnetic resonance images using an Active Appearance Motion Model. J Cardiovasc Magn Reson 2004;6(3) 609-617.
  7. Angelié E, Oost ER, Hendriksen D, Lelieveldt BP, Van der Geest RJ, Reiber JH. Automated contour detection in cardiac MRI using active appearance models: the effect of the composition of the training set. Invest Radiol. 2007;42(10):697-703.