Tissue classification in microscope imaging based on a discrete-band spectroscopy

Abstract

Automatic classification of tissue types has a potential use in endoscopic or microscopic imaging. For this study, the microscope images from composite tissue samples were measured. Different shades of red meat such as beef and pork and bloodless tissues such as chicken breast and fat were chosen to provide with similar and contrasted colors. We applied a partial least squares discriminant analysis (PLS-DA) to classify the tissue type of the image pixels. With the RGB color images that are usually available in the hospital, we could classify beef, pork and chicken with an accuracy of only 85.7% and 80%, 60% and 11.4%, respectively. There was an apparent limitation in differentiating the tissue type due to the spectral overlapping of RGB colors. To increase the classification accuracy, band-pass filtered images were taken at the center frequencies of 414, 542, 655 and 832 nm without any spectral overlapping. Using these discrete-band spectral images, the classification accuracy reached to all 100% except for beef that was 96%. In capsule endoscopy where the amount of image data is prohibitively large, automatic detection of bleeding or cancerous region is of great interest and we believe that this method can be applicable in real time monitoring.