Machine vision in the detection of prostate lesions in histologic sections

To explore the utility of N-gram encoding for the automated detection and delineation of regions of histologic abnormality in tissue sections of prostate.

Digitized imagery of tissue sections from normal prostate glandular tissue, stroma and regions of well- and poorly differentiated lesions was recorded and successively subdivided into square subregions of 256 x 256 to 16 x 16 pixels. N-grams of N = 2 to N = 6 were computed, with each element assuming a value representing an optical density interval 0.30 units wide, covering the range from optical density = 0.0 to 1.80. Then, from a large database, prototype frequency histograms of the different N-grams were established. For each subregion the Euclidean distances to the different prototype histograms were computed and defined as "distance to prototype" features. Standard discriminant analyses and a nonparametric classifier were used to assign subregions to the different tissue categories.

Classification of subregions was achieved for most discrimination tasks at a correct recognition rate ranging from 85% to 100% on both training set and test set data, with a few exceptions. N-grams of N > 4 had considerable discriminatory power.

N-gram encoding has the potential to provide highly discriminating, texture-based characterization of subregions of digitized imagery of prostate lesions and may be very useful in the development of decision procedures for the automated detection of prostate lesions by a machine vision system.