Abstract
In evaluating histological images obtained from pathological samples, taking advantage of the fact that each tissue component exhibits specific staining characteristics. We propose a method for classification of hematoxylin and eosin (HE)-stained breast tumor tissues as benign or malignant type. Two types of neural networks are used in the two-step process consisting of the extraction of features from HE staining images and evaluation of their values. In the feature extraction stage, regions in HE staining images are divided into four categories, nucleus, cytoplasm, interstitium and background, on the basis of color information. In the evaluation stage, values of the extracted features based on stainability (percentage area of nucleus, cytoplasm and interstitium, frequency distribution and size) are used. The method developed was applied to 30 test images from 17 cases consisting of 10 different types of tissues and to 26 images from 13 cases of blind samples. The accuracy of classification of breast tumors as benign or malignant was excellent, and the effectiveness of the method was confirmed. The proposed method has two advantages: (1) it is effective in the evaluation of staining images that exhibit considerable variability such as in contrast (variation in input images) and (2) applications to other types of HE-stained tumors are possible by the use of the algorithm proposed in this study.
