Accurate region classification Clause Samples

Accurate region classification. Each of the ROIs identified in section 5.3.1 represent a meaningful anatomical structure. We employ a region-based classifier, to distinguish between DCIS and other normal or benign structures in the tissue. For each of the segmented ROIs five different types of texture features were extracted from the gray-scale intensities of the image. Statistics of gray level histogram, statistics calculated from the co-occurrence matrix, uniform local binary patterns for radii one and two, and texture features extracted from filter banks in particular Laplacian of Gaussian (LoG) at five scales, and Gabor filters at four scales and eight orientations. These texture features have shown strong discriminatory power in characterizing histopathology images. In total 256 features were extracted for each ROI. The dataset used in this study originate from 40 H&E stained WSIs of breast tissue sampled from 40 patients. 20 of the WSIs contain DCIS and 20 contain different types of benign abnormality. An expert annotated various regions containing DCIS in abnormal slides. In total, 206 DCIS regions were annotated. We evaluated the performance of our proposed DCIS detection system both at the slide level and in terms of detection and localization of the lesions in the WSI. The dataset was split into two independent subsets for training and testing. Both training and test sets each contain ten DCIS and ten benign abnormality slides. A logistic regression classifier was trained using the features extracted from the annotated data in the training set. All the ROIs in the test set were automatically detected and consequently classified and given a score which is the degree of suspicion that the region is a DCIS. To achieve a slide-based score, the highest scored region in a slide is used as the confidence that the case contains DCIS. At the slide level, all the cases were accurately classified. Free-response receiver operating characteristic (FROC) curve was used to assess CAD performance at the lesion level. The FROC curve is defined as the plot of lesion localization fraction versus the mean number of false positives per image. Table 1 summarizes the DCIS detection (sensitivity) levels at different average number of false positives per WSI. Table 1: Results for the proposed system. Sensitivity is provided at 4 levels of average numbers of false positives per WSI.