An enhanced intelligent image segmentation approach for brain magnetic resonance images.

摘要

Magnetic resonance imaging (MRI) is a widely used approach to obtain high quality clinical images. Post-processing the image data with segmentation methods can further aid in the visualization and recognition of soft tissues and lesions in brain.; In the past twenty years, different brain MR image segmentation algorithms have been developed, but the accuracy is not satisfying and most of them are sensitive to noise. Also, only few of them can perform probabilistic segmentation that is highly desirable for MR image quantitative analysis. Moreover, due to the complexity of MR imaging process and brain anatomical structure, conventional segmentation algorithms are not able to distinguish areas corrupted by inhomogeneity effect, especially putamen.; In order to solve the above problems, a novel hybrid segmentation algorithm for brain MR images is proposed in this dissertation. Beside relatively high accuracy and the ability to perform probabilistic segmentation, it can also detect putamen area and is robust to noise. This algorithm is based on: (1) multi-scale feature extraction, (2) hierarchical labeling structure, which includes expectation maximization (EM), fuzzy C-means (FCM) and self-organizing map (SOM) neural networks, (3) fuzzy rule based (FRB) system for putamen area, and (4) weighted probabilistic neural network (WPNN), which is a novel neural network structure capable of dealing with partial volume effect.; The segmentation process can be divided to three steps. First, a modified multidimensional input is sent to EM, FCM and SOM neural network simultaneously. The generated reference vectors from SOM neural network are then labeled in a "soft way" by the hierarchical labeling structure, which can generate soft label factors to indicate the likelihood that a reference vector from SOM belongs to a target class. Next, the quantized image generated by SOM neural network is passed to a FRB system integrated with edge detection and region growing approaches for putamen segmentation. Ultimately, WPNN adopts the reference vectors from SOM and weighting factors from hierarchical labeling structure to estimate the probability density function (pdf) and perform probabilistic classification. Putamen segmentation result is then added to this probabilistic classification result to finish the algorithm.; Three kinds of data sets were used for evaluation purpose, which are random distributed data sets, brain shaped phantom, and simulated MR image. Comparisons with ground truth and comparison between different algorithms were performed, and the effectiveness and robustness of the proposed algorithm were demonstrated. Finally, applications on real MR images were presented with satisfied results.