Statistical Prior Knowledge for Robust Medical Image Segmentation by Level Set

Region-based active contours give interesting results when applied for objects detection with poor contrasting edges and noise, but for medical image segmentation, where objects of interest are often present with inhomogeneity, these models fail. In [16], we presented an original approach that integrate prior knowledge on the target shape into active contours with free registration based on a complete and stable set of invariant descriptors. In this work, we intend to present a novel method that incorporate statistical prior knowledge into level set model for robust medical image segmentation resulting from machine learning algorithms. In fact, for medical imaging, in many situations like tumors segmentation, prior knowledge on shape is insufficient or unavailable because it vary from a patient to another. The adopted approach is demonstrated with two unsupervised learning methods. The first one use the Fuzzy C-Means (FCM) and the second one is based on the Hidden Markov Random Field Expectation-Maximization algorithm (HMRF-EM). Both of the two are used to classify the pixels of the target image into two classes: the object and the background to construct our reference shape to be used as prior knowledge. A weighting schema is then considered to guide the active contour model under both forces: the traditional and the proposed one. We validate the robustness and the accuracy of our approach using ground truths given from experts segmentation and standard data bases. Experiments and comparative studies with several recent works are highlighted and show that the integrated method achieves good performance results.