MRI segmentation using histogram based active contour method: application to prostate MR image analysis
Résumé
Purpose Image segmentation of MRI data is a challenging problem due to the specificity of the acquisition process and more precisely to the particularity of the associated corrupting noise that can be modeled by a Rician distribution. This particular distribution leads to multimodal area within MR image where foreground and target organs are characterized by different probability density functions (pdf) (mainly Rayleigh distribution and Gaussian distribution). Due to this particularity, classical automatic segmentation technique like Chan-Vese often failed in obtaining satisfying results of delineations in MR images where pdf of organs could overlap with background. This is the case for example in MR prostate image analysis, where bladder, prostate, and rectum can be hardly segmented thanks to classical approaches. Method In the framework of active contour segmentation [2], we propose to evaluate for MRI segmentation performances of histogram based method [1,3]. Main idea of this approach is to define, in a variationnal framework of region based active contours segmentation, evolution criteria accounting for the distances between probability density functions (pdf) of the inner and outer regions of the active contour and given pdfs of reference of objects and background. To achieve the performances evaluation, first a statistical study of the accuracy of the technique is proposed on synthetic MR data for different levels of corrupting noise and for different distance criteria between two given pdf and, second, a prospective study made on prostate MRI is shown for bladder, prostate and rectum segmentation, which respective delineation is of primary interest in the radiotherapy treatment of prostatic adenocarcinoma. Results A statistical study performed on synthetic MR images shows that histogram based active contour method outperform classical approach like the Chan-Vese one in terms of accuracy and that depending on the distance criterion selected, an adaptation to the level of corrupting noise is possible to reach the best accuracy possible. The prospective study made on prostate MRI shows that the proposed approach suits well to the particular MR distribution function: mainly, bladder, prostate and rectum segmentation are better performed by the proposed approach than with more classical active contour approaches.