An Automated and Accurate Methodology to Assess Ki-67 Labeling Index of Immunohistochemical Staining Images of Breast Cancer Tissues
Automatic scoring of Ki-67 with digital image analysis would improve the accuracy of the diagnostic. However, automatic Ki-67 assessment is very challenging due to complex variations of cell characteristics. In this paper, we propose an integrated framework for accurate Ki-67 scoring. The main contributions of our method are: a robust cell detection algorithm to detect all tumor and non-tumor cells; a clustering model for computerized classification of tumor and non-tumor cells and subsequent proliferation rate scoring by quantifying Ki-67, based on classified cells which appear in breast cancer immunohistochemical images. Unlike existing Ki-67 scoring techniques, our methodology works on whole slide images (WSI) using patches that are extracted from detected tissues. As the size of each sample is so large they can not be handled as a single image. Therefore, each slide is divided into small parts and on edge tiles merging is considered to preserve the continuity of nuclei. The proposed method has been extensively evaluated on tissue microarray (TMA) whole slides, and the cell detection performance is comparable to manual annotations and is very accurate compared with the estimation of an experienced pathologist.
Digital pathology, Ki-67, Immunohistology, Computer-aided diagnosis, Whole-slide imaging.
As described, the proposed method automatically excludes stromal and epithelial tissue. However, the operator can manually specify regions to process. Specifying smaller regions do not affect the overall results as for each region, different values would be considered to correct results and there would not be variances in the intra- and inter-regions. Scoring on all of the slides takes ∼8 min which is quite competitive as we deal with 7k square tiles with 256×256 sizes extracted from each slide. In order to merge each patch, we take neighborhood tiles which increases the number of images to 14k tiles. We chose 30 patient slides manually evaluated by an expert pathologist to test the method. The proposed method achieved 91% classification accuracy, 0.93 precision, 0.89 recall, and 0.91 F-score value as indicated in Table 1. Besides, in Fig 4 labeling indexes defined by ImmunoRatio  cell counting and by our method are shown. Our proposed method shows a great continuity in edge parts and detects almost all of the cells with a high accuracy automatically. In some cases the ImmunuRatio web-based tool could not segment positively stained cells and some negatively stained cells are not detected. Furthermore, ImmunoRatio works on just one area and does not consider critical areas that are edge parts. Our proposed method not only detects all stained cells but also considers edge cells to give overall result of specified region by pathologist.