Automatic characterization and classification of ganglion cells from the salamander retina

The classification of retinal ganglion cells according to their morphologic al features is addressed by using a comprehensive set of shape measures and several clustering strategies. The morphological features considered inclu de many common measures (such as dendritic radii and the number of dendriti c segments) and three new quantifiable measures: 1) the area of influence o f the dendritic tree as calculated in an operator-independent manner by usi ng Minkowski sausages; 2) the complexity of tortuousity along each dendriti c segment as represented by the 3D bending energy; and 3) the coverage fact or as calculated by using the Bouligand-Minkowski fractal dimension, which is more accurate than the commonly used box-counting algorithm. We evaluate d four clustering approaches including the k-means and Ward's hierarchical clustering methods. By using these highly quantifiable methods to group the cells into classes, the present work has extended and reassessed the analy sis of 68 ganglion cells from the tiger salamander previously classified by Toris et al. ([1995] J. Comp. Neurol. 352:535-559). Though substantiating the number of classes (5) previously proposed by Toris et al., the results obtained here indicate a number of discrepancies among the members of each class, especially regarding the border between two classes, originally call ed the medium simple and the medium complex cells. Such an effect has motiv ated the proposal of new names for the medium simple and medium complex cla sses, now called small highly complex and medium cells, respectively. Also included in the present article are comprehensive statistics of each class, correlations among all the adopted shape measures, and examples of the cel ls from each class. The resultant classes that emerged were compared using their electrotonic characteristics and physiological profiles. J. Comp. Neu rol. 404:33-51, 1999. (C) 1999 Wiley-Liss, Inc.