Computerized classification method for histological classifications of masses using objective features based on clinicians' subjective impressions

Our purpose of this study is to develop a computerized classification method for histological classifications of masses using objective features based on clinicians' subjective impressions. An observer study is first conducted to obtain clinicians' subjective impression for nine image features on mass. Nine image features are selected by taking into account image features that clinicians' commonly used for describing masses on ultrasonographic images. In the proposed method, the location and the area of mass are determined by an experienced clinician. We define some extraction methods for each of nine image features. The extraction method is selected such that the correlation coefficient would become the highest between objective features and average of clinicians' subjective ratings. An artificial neural network (ANN) with the nine objective features is employed for distinguishing among four different types of histological classifications on masses. The classification accuracies of the proposed method were evaluated by using 216 ultrasonographic images. The classification accuracies of the proposed method were 85.9% (55/64) for invasive carcinomas, 83.3% (30/36) for noninvasive carcinomas, 89.1% (49/55) for cysts, and 82.0% (50/61) fibroadenomas, respectively.