1 CSIRO Mathematics Informatics & Statistics, Locked Bag 17, North Ryde, NSW 1670, Australia
2 University of New South Wales in Canberra, Australia
3 Research Center for Advanced Information Science and Technology, Aizu Research Cluster for Medical Engineering and Informatics, The University of Aizu, Aizu-Wakamatsu, Fukushima 965-8580, Japan
Our method has two steps: pixel classification and post-processing. In the pixel classification step, we firstly build an initial classifier on pixels. The initial classifier is built by segmenting the input image into a binary image with an adaptive threshold based on edge strength after Gaussian filtering. This classifier performs very well except misclassifying the pixels which are in the dark blobs among the neural structures. Two characters of these dark blobs are: (1) the dark blobs are close to each other in the neighboring images of the 3D stack; (2) most neural structures are narrow, while the dark blobs are not. Thus, we detect these dark blobs from the binary image by an opening operation using a round mask and checking the co-occurrence in the neighboring images. The dark blobs are eliminated after they are detected.
Then, we extract more features such as the Hessian magnitude, the difference of Gaussians, the mean and roundness of mean-shift segmentation, and the distance transformations of the binary images obtained from the first step. Multiple scales are used in the feature extraction. We use the results from the initial classifier together with all the other features (34 features in total) to train a SVM which is used to obtain a probability map. Some neural structures near the dark blobs may be incorrectly eliminated in the first step. As a result, they may have a low probability to be regarded as the neural structures by the SVM classifier. Post-processing is employed to recover these structures. We use a threshold on the probability map to select pixels as potential neural structures, and then the connected regions which have no holes are eliminated by attribute opening. The average probability value over a connected region is taken as the final probability.
For detailed description, please see this link.