Classification of brain tissue with optical coherence tomography by employing texture analysis
Marcel Lenz,Robin Krug,Christopher Dillmann,Nils C. Gerhardt,Hubert Welp,Kirsten Schmieder,Martin R. Hofmann
DOI: https://doi.org/10.1117/12.2307701
2018-05-24
Abstract:Biomedical tissue classification is of great interest in many fields, e.g. for finding a clear boundary for cancer resections. At the moment, no sufficient tool has been found to fulfill the needs for intraoperative surgical guidance. For a precise tumor removal, the resolution has to be as high as possible and the delivered information should be in real time. Otherwise intraoperative guidance cannot be done accurately. Optical coherence tomography (OCT) has already demonstrated its benefits in ophthalmology, dermatology and endoscopy. Providing μm resolution for a penetration depth of 1-2 mm at acquisition rates in the MHz regime, OCT is a perfect tool for contactless investigations during surgery. Additional benefit can be provided if the obtained images are analyzed and the tissue type is immediately classified. Usually, the histopathological analysis of ex vivo samples directly after removal conforms the tissue classification. A common practice for that is the histopathological analysis, where the samples are embedded in paraffin, stained with, e.g. hematoxylin and eosin and then, classified by an experienced pathologist, who analyzes tissue slices of approximately 10 μm thickness. By employing OCT for classification, an entire three-dimensional image can be classified without further preparation of the tissue. Here we present a texture feature based approach by utilizing local binary patterns, run length analysis, Haralicks texture features and Laws texture energy measures. After applying all these texture features, a principal component analysis (PCA) was performed, which decreased the dimensionality of the data set. This step was necessary in order to enhance the performance of the employed support vector machines (SVM) classifier. To find the best parameters for the kernel, a grid search and a 10 fold cross validation were done. As a first step, the texture analysis based post processing approaches were applied on 13 ex vivo brain tissue samples, which were diagnosed as meningioma (8), healthy white tissue (3) and healthy gray tissue (2). The samples were imaged with a commercial OCT system (Thorlabs Callisto). On the raw OCT images, some structural differences between healthy tissue and meningioma may already be recognized. However, an automated approach that does not require interpretation of the result, would certainly help the surgeon during surgery. At the end, we trained a SVM classifier, which was able to differentiate between healthy tissue and meningioma, with an accuracy of nearly 98%. As the next logical step, these findings will be validated intraoperatively and further application for texture based classification will be investigated.