In this chapter we have presented a neural network approach to automatic segmentation of MRI data sets of the human brain by neural network computation. The GRBF architecture enables an effective combination of unsupervised and supervised learning procedures.
Although the results presented in the preceding section are very encouraging, several critical issues still need to be discussed with regard to the data, the aim of segmentation, and the validation of the results. These remarks can provide an outlook for further research activities:
(1) Validation. As there is no "gold standard" for evaluating the quality of segmentation results, careful considerations have to be made with regard to defining the objectivity, reliability, and validity of the procedure. (a) Objectivity. This has two aspects: On one hand, more detailed investigations have to be made with regard to the interobserver variability of the segmentation results that are influenced by human interaction: In the unsupervised learning approach there may be interindividual variations in cluster assignment decisions, whereas in the supervised classification approach there may be different segmentation results due to different choices of the training data that are interactively labeled by a human observer performing manual contour tracing. A second aspect that should be subject to more detailed investigation is the reproducibility of the GRBF neural network segmentation results compared to other semiautomatic segmentation procedures.
(b) Reliability. A closer examination of the intraobserver variability of the segmentation results has to be made in cases where the same data are reprocessed by the same individual at different times. The segmentation results may differ because of those parts of the procedure that require human interaction, as explained previously. A different aspect is the variability of the segmentation results in repetitive MRI examinations of the same subject at different times. This is especially important for studies that focus on the temporal monitoring of size and growth of pathological processes.
(c) Validity. The semiquantitative evaluation by an experienced neuroradiologist as presented in the previous section can only be a first step to critical analysis of segmentation quality. In further studies, this should be accompanied by objective measurements such as the planimetric analysis in a series of cross-sections after neuropathological preparation of animal brains.
(2) Data. Segmentation results can be influenced by the choice of the MRI sequences as well as different ways of extracting information from these input data.
(a) MRI sequences. Further studies are required with regard to which MRI sequences should contribute to the multispectral data set serving as raw data for the segmentation procedure. These may vary according to the focus of interest in specific clinical situations. One might expect, for example, that proton density weighted sequences are not very helpful when a detailed white/gray matter segmentation is to be performed. On the other hand, additional sequences such as FLAIR maybe useful for providing better segmentation results for various types of pathological tissue. Such considerations should be guided by the attempt to achieve a reasonable trade-off between two complementary optimization criteria: segmentation quality and MRI sequence acquisition time. The latter is a critical issue when examining patients instead of healthy volunteers.
(b) Product space. Because of local inhomogeneities of imaging properties, a specific tissue class may have different location-dependent gray-level appearance. Thus, it could be helpful to extend the segmentation procedures explained in this chapter in order to account for such gray-level shift effects. A possible solution could be to operate in the product space of gray-level and spatial coordinates, serving as a new input feature space. This could provide a more comprehensive description of the data set and, thus, may improve the segmentation results.
(3) Aim of segmentation. The three tissue types—gray matter, white matter, and CSF—are just a preliminary choice for defining brain tissue segmentation classes, although they turn out to be sufficient for most clinical applications. However, they may be completed by introducing additional classes for venous blood, meninges, etc., thus providing a more fine-grained tissue separation. This is extremely important when segmentation is to be used for defining the spatial extent of pathological lesions. Here, one could specifiy classes for types of pathological tissue, e.g., for defining focal lesions such as tumors or multiple sclerosis plaques.
In summary, a wide range of further research topics has to be covered in future studies based on the results presented in this article. Such investigations can help to introduce automatic neural network segmentation as a cost-effective and reliable tool for routine medical image processing according to the growing importance of quantitative image analysis techniques for clinical decision making.
Was this article helpful?