In the present study we evaluated a texture analysis approach using LBP texture features in combination with a machine learning method to identify tissue types in a large series of digitized colorectal cancer TMAs. Segmentation of tumor tissue into epithelium and stroma facilitates automated assessment of protein expression within the respective tissue compartments. Protein expression quantification can be performed as a sequential process in which a primary algorithm performs the segmentation and a secondary algorithm calculates the area and intensity of an immunohistochemical staining. Computerized tissue type-specific interpretation of immunohistochemical staining has the potential to produce more reliable and reproducible results as compared to visual quantification methods by a human observer . In addition, an algorithm that identifies tumor epithelium could be utilized for the purpose of identifying regions of interest to be punched from the donor block in the process of TMA construction or for laser capture micro dissection of specific cells of interest .
The tumor epithelium can exhibit a range of textures, from an appearance close to the normal tissue in well-differentiated cancer to the lack of organizational features in poorly differentiated tumors. Epithelial tissue texture is different from stromal texture which is organized in specific directions and is loosely arranged . Several powerful pattern recognition methods have emerged during the last few years, especially within texture classification . As a rule many of these techniques assume that the textures are uniformly presented and captured in the same orientation. In the analysis of tumor tissue, samples are cut in various planes and positioned at different angles on slides for analysis. Thus, uniform orientation is not possible to achieve and analysis of tissue texture should be invariant to orientation. Also, the algorithm should be robust with regard to variations in image contrasts due to tissue processing and factors related to image acquisition.
The LBP operator is a rotation and grayscale invariant texture descriptor and is therefore interesting in the context of tissue texture analysis. LBP has been successfully used in various applications. For example, the LBP algorithm is used for face recognition  and other applications within biometrics, including iris recognition  and fingerprint identification . The LBP operator has been proven to be highly discriminative and its key advantages are computational efficiency and invariance to monotonic gray level changes .
Texture-based algorithms for classification of tumor tissue have, to some extent, been previously studied, but generally included only small specimen series. In one study, image texture analysis was used for mapping dysplastic fields in colorectal tissue . Another study showed that identification of normal vs. abnormal prostatic tissue components in large-scale histological scenes was feasible using Haralick's co-occurrence texture features . For classifying breast histology images, texture-based operators using supervised learning have been employed [34, 35]. In a recent publication, a wavelet-based, multiscale framework for texture-based color image segmentation was used to differentiate various tissue compartments in ovarian carcinoma. In that study an average of 71.5% of pixels were assigned to the correct class by the algorithm i.e. five tissue types manually annotated in the images by the human observer . A direct comparison with our results is not feasible, since we focused on the discrimination between two tissue types. Also, in the current study we analysed the accuracy on an image-block level in comparison to studies that report pixel-level results. We argued that obtaining a ground truth with regard the tissue categories (stroma and epithelium) as defined by a human observer on a pixel-level would not be possible without substantial inter-observer variability.
In a few earlier reports, the LBP algorithm has been adapted for tissue classification. In a previous report that compared different histogram-based feature sets for tissue images, the LBP obtained the highest classification accuracy . Another approach using LBP was employed to determine tissue as either stroma-rich or stroma-poor from digitized whole-slide neuroblastoma slides. The approach was tested on 43 whole-slide samples and provided an overall classification accuracy of 88% . The LBP/C algorithm described here discriminates between epithelium and stroma with a higher accuracy (99%) than the method presented by Sertel et al. This may be partly due to differences in the tissue architecture in neuroblastoma as compared to the morphology of the stroma in colorectal cancer tissue. In addition, differing LBP parameters, classifier selections (SVM vs. k-nearest neighbor), and incorporation of the contrast information might explain part of the difference in accuracy. In another study LBP features were used for classification of sub-cellular protein localization and also, the algorithm has been applied on pap smears to classify cervix cells as either normal or abnormal [38, 39].
In the present study the accuracy of the LBP/C texture classifier for assigning the correct histological class was significantly higher with the LBP/C operator as compared to Haralick features and Gabor filters. The LBP operator can be seen as a unifying method to the traditionally divergent statistical and structural models of texture analysis. The rotation invariance and tolerance against illumination changes of the LBP operator may be factors that have an impact on the outcome in our setting. Regarding Gabor filters, it has been suggested they have a tendency to over-represent low frequency components and under-represent higher-frequency components and thus may not always be suitable for texture analysis of natural images . Since the discriminative accuracy of all three descriptors was excellent (AUC > 0.95) no firm conclusions on the superiority of one single approach can be drawn and performance results may vary according to the analyzed tissue type.
As mentioned previously, the current method is based on image blocks, with a size of approximately 40 micrometers. Thin rows of tumor cell or non-stromal cells interspersed with stroma might therefore be wrongly classified as stroma. Future studies are needed to assess resolution requirements for segmentation of specific tissue types or disease states (e.g. infiltrating inflammatory cells).
We used colorectal cancer as a model to test the ability of the texture algorithm to differentiate the two histological tissue types. Whether our results will be applicable to other cancer types than colorectal cancer, needs to be explored in further research. In this study we analyzed a series of tissue samples immunoassayed for analysis of the EGFR protein and visualized by the DAB chromogen. It cannot be ruled out that immunohistochemical staining process influenced the results, although the algorithm should be invariant to color/image intensity. Also, the methods used for antigen retrieval may modify tissue architecture and thus the texture of the tissue. The reason for us to choose the EGFR protein staining was that the staining was of good quality, i.e. there was only little cross reactivity between the epithelial and stromal compartments. We analyzed immunohistochemically stained tissue sections and not haematoxylin-eosin stained tissue, because our aim was to test the performance of the algorithm on samples prepared for tissue protein expression analysis.
In future studies, it will be of interest to apply texture analysis on other cancers, e.g. breast- and prostate tumor samples. A computerized segmentation into tumor epithelium and stroma would be of relevance in studies regarding the tumor microenvironment, especially when applied to large series on digitized whole slides samples. Stromal cells and their roles in cancer prognosis  and response-prediction  have been increasingly recognized. It has been proposed that induction or loss of certain proteins in the stroma may be critical in promoting the metastatic phenotype in cancers .
In addition to the segmentation of tumor tissue to specific compartments described in this study, texture classifiers for cancer tissue in combination with clinical and bio-molecular data may act as prognostic markers . By probing large sample areas and thousands of tissue specimens, previously undiscovered texture patterns for cancer with clinical and prognostic relevance could potentially be identified. Texture-based algorithms also have the potential to be used for more general tissue segmentation and image quality assessment in whole-slide images [42, 43]. Texture features combined with color information might be of interest and is currently a highly investigated topic in computer vision .