Image area extraction of biological objects from a thin section image by statistical texture analysis

The statistical method in texture image analysis was applied to area extraction of biological objects from thin section electron microscope images. Four standard estimators defined from the gray level co-occurrence matrix, called 'inverse difference moment,' 'angular second moment,' 'entropy' and 'contrast,' were especially examined using a test pattern consisting of 6 artificial textures and to practical biological thin section images. In the examination on the test pattern, among the estimators the 'contrast' discriminated all textures, but differences of some texture feature levels were small. To clearly discriminate textures, a modified estimator combining 'angular second moment' and 'contrast' was devised. As a result it discriminated all better than the 'contrast.' Electron microscope images used for the image processing are yeast morphological ones showing spherical autophagic bodies in the vacuole. Although the four standard estimators discriminated many organelles, they could not exactly extract images of the autophagic body. However, the modified estimator was able to extract all autophagic bodies from the vacuole image area except for minor points, some of which were not clearly detected by man observation. It as found out that the texture analysis-based method can be used to discriminate slight differences between image textures due to spatial staining granularity.