Two-stage neural network for volume segmentation of medical images

We present a system to segment and label CT/MRI brain slices using feature extraction and unsupervised clustering. In this technique, each voxel is assigned a feature pattern consisting of a scaled family of differential geometrical invariant features. The invariant feature pattern is then assigned to a specific region using a two-stage neural network system. The first stage is a self-organizing principal components analysis (SOPCA) network that is used to project the feature vector onto its leading principal axes found by using principal components analysis. This step provides an effective basis for feature extraction. The second stage consists of a self-organizing feature map (SOFM) which will automatically cluster the input vector into different regions. The optimum number of regions (clusters) is obtained by a model fitting approach. Finally, a 3D connected component labeling algorithm is applied to ensure region connectivity. Implementation and performance of this technique are presented. Compared to other approaches, the new system is more accurate in extracting 3D anatomical structures of the brain, and can be adapted to real-time imaging scenarios.