Table 1.

Nearest neighbor classification accuracies for the 502 images of 10 sub-cellular localizations.

	TAS+Haralick			TAS			Haralick
Statistics Dimension	Unmapped	3D	2D	Unmapped	3D	2D	Unmapped	3D	2D
1-neighbor classify	93.2%	79.1%	64.5%	91.8%	77.5%	59.6%	87.1%	68.3%	54.0%
3-neighbor classify	94.0%	83.2%	68.9%	90.6%	80.7%	64.3%	85.5%	70.9%	60.6%

Three types of image statistics vectors are tested: TAS with Haralick; TAS alone; Haralick alone. For a given image, the class of the image is compared to that of its nearest neighbors, nearest being defined using the Euclidean distance between the statistics vectors (unmapped), the statistics Sammon mapped into 3 dimensions (3D) and 2 dimensions (2D). Tests in which 1407 individual cells were automatically selected and cropped from the 502 images gave similar (in fact slightly higher) classification accuracies as those in the table (data not shown).