Data-driven approaches to decision making in automated tumor grading. An example of astrocytoma grading.

OBJECTIVE

To compare four data-driven approaches to automated tumor grading based on morphometric data. Apart from the statistical procedure of linear discriminant analysis, three other approaches from the field of neural computing were evaluated.

STUDY DESIGN

The numerical basis of this study was computed tomography-guided, stereotactically obtained astrocytoma biopsies from 86 patients colored with a combination of Feulgen and immunhistochemical Ki-67 (MIB1) staining. In these biopsies the cell nuclei in four consecutive fields of vision were evaluated morphometrically and the following parameters determined: relative nuclei area, secant lengths of the minimal spanning trees and relative volume-weighted mean nuclear volumes of the proliferating nuclei.

RESULTS

Based on the analysis of these morphometric features, the multivariate-generated HOM grading system provides the highest correct grading rates (> 90%), whereas the two widely employed qualitative histologic grading systems for astrocytomas yield correct grading rates of about 60%. For automated tumor grading all approaches yield similar grading results; however, back-propagation networks provide reliable results only following an extensive training phase, which requires the use of a supercomputer. All other neurocomputing models can be run on simple UNIX workstations (AT&T, U.S.A).

CONCLUSION

In contrast to discriminant analysis, backpropagation and Kohonen networks, the newly developed neural network architecture model of self-editing nearest neighbor nets (SEN3) provides incremental learning; i.e., the training phase does not need to be restarted each time when there is further information to learn. Trained SEN3 networks can be considered ready-to-use knowledge bases and are appropriate to integrating further morphometric data in a dynamic process that enhances the diagnostic power of such a network.