Accurate OCT-based diffuse adult-type glioma WHO grade 4 tissue classification using comprehensible texture feature analysis

Maximum safe resection of diffuse adult-type glioma WHO grade 4 (glioma degrees 4) is a well-established procedure in order to prolong a patient's lifetime. However, differentiation between glioma and healthy brain tissue remains a challenge in an intraoperative setting. As shown previously, the attenuation coefficient extracted from OCT images does not provide sufficient discrimination between healthy gray matter and tumor infiltrated tissue. On the other hand, the majority of published approaches based on structural information in OCT images uses sophisticated machine learning approaches as e.g. deep convolutional neural networks which bring the need for large data sets for reasonable modeling. In this work, we propose a method, which is based on comprehensible texture features and provides reasonable classification performance for healthy gray and white matter against glioma degrees 4 samples, while requiring a relatively low number of samples for training. Our sample collective consists of 36 tissue samples from 27 different patients. Texture features based on the average contrast were optimized with respect to best discrimination of the tissue types of interest. Best discrimination could be obtained for the contrast of structure sizes of approximately 100 mu m. By using a linear discriminant analysis (LDA), we could achieve a sensitivity and specificity of 97.7 % and 91.7 %, respectively, when classifying white matter and samples with a mixture of white and gray matter against glioma degrees 4 samples. For an extended data set, also including pure gray matter samples, we determined a sensitivity of 86.7 % and a specificity of 86.3 %.

Automated summary

In this study, a method based on texture features is proposed, which can classify healthy gray and white matter against glioma degrees 4 samples with reasonable classification performance using a relatively low number of samples for training. The method achieves high classification performance without the need for large datasets and complex machine learning approaches.