Automated CRC Tissue Classification Using Texture Features

Automatic tissue type recognition from histology images is significant for digital pathology. Texture analysis is used to recognize different types of tissues with frequency, notably estimating the ratio of tumor and stroma in histology samples. Still, the research available has concentrated mostly on binary distinctions while not accounting for the multiclassity nature of histological images. This work introduces a new set of 5,000 human colorectal cancer histological images of eight diverse types of tissues. With this dataset, we compare the classification performance of different texture descriptors and classifiers. Our results show that the best classification approach dramatically improves performance, boosting tumor-stroma separation accuracy from 96.9 to 98.6 and establishing a state-of-the-art multiclass tissue classification accuracy of 87.4. The dataset is being made publicly available under a Creative Commons license as a benchmark for future research in automated histological image analysis.