Segmentation with Neural Networks

Computational intelligence in biomedical imaging: multidimensional analysis of spatio-temporal patterns

“…An algorithmic approach has been presented in [46] that aims to combine UC and SC for image segmentation, where the information obtained during UC is not discarded, but is used as an initial step towards subsequent SC. This is achieved by applying Generalized Radial-BasisFunctions-(GRBF-) neural networks according to [41,42]: Here, in [46], in a first unsupervised learning step, the voxelspecific signal intensity spectra of co-registered multispectral MRI data sets are clustered by VQ.…”

“…This is achieved by applying Generalized Radial-BasisFunctions-(GRBF-) neural networks according to [41,42]: Here, in [46], in a first unsupervised learning step, the voxelspecific signal intensity spectra of co-registered multispectral MRI data sets are clustered by VQ. In contrast to previous approaches to biomedical image segmentation, the data partitioning obtained in this way is not only considered as an irrelevant preliminary intermediate step to finally perform some supervised classification task at reduced computational expense, but is acknowledged as an independent result that may be valuable for exploratory analysis by human experts, see e.g.…”

“…4.2 below. Systematic interactive visual classification and collection of clusters corresponding to specific tissue classes has thus been introduced and compared to plain voxel-specific supervised classification in [46].…”

“…Healthy volunteers (aged 24-32 years) were examined employing a standardized MRI sequence protocol: T1 weighted MP-RAGE, T2 weighted and proton density (PD) weighted spin echo, and Inversion-Recovery (IR) sequences. Details of the MRI acquisition procedure and sequence parameters can be found in [46]. Figure 1 presents an example of corresponding images of the four different MRI acquisition techniques.…”

“…In [46], and the application-oriented publications [47,48], UC image segmentation was performed by minimal free energy VQ [12][13][14][15], enabling unsupervised exploratory analysis of the image data on different scales of feature space resolution. For subsequent SC analysis, the UC results are not discarded, but can be re-utilized for the training of a GRBF neural network.…”