A cytoskeletal injury classifier based on 'spectrum enhancement' and data fusion

摘要

The morphology of cytoskeletal microtubules has been analyzed by fractal, direct and spectral methods. Sets of images were obtained from the epifluorescence microscopy of primary cultures of rat hepatocytes treated with fungicide concentrations of SO and 25 μg/ml for 2h. The morphological descriptors extracted by said methods included contour and mass fractal dimension, total variation, the L~1-norm of the Laplacian and properties of the "enhanced spectrum". The latter is obtained by suitably processing the logarithm of power spectral density with the aim of separating image structure (low spatial frequency) from texture (high spatial frequency). Descriptors were fused by principal components analysis. A classification algorithm was trained to tell undisturbed (control) cytoskeletal structures from those treated at the higher dose. The eigenvector matrix of the trained classifier was used to rank structures treated at the lower dose: from regression on the set centroid coordinates a tentative relation between the first principal component (the "response") and dose has been obtained. The same ranking procedure was applied to structures recovering from injury (24h after exposure to the higher dose) and the extent of recovery has been quantified. The paper includes a possible interpretation of some morphological descriptors and their role in automatic classification.