Classification of breast and colorectal tumors based on percolation of color normalized images

•We propose an approach that associates color normalization with percolation features.•The method was applied in colorectal and breast histological images.•Relevant AUC rates were obtained for classifying malignant and benign tumor images.•Color normalization improved the results obtained for colorectal images.•We evaluate the effect of local and global percolation features on each dataset. [Display omitted] Percolation is a fractal descriptor that has been applied recently on computer vision problems. We applied this descriptor on 58 colored histological breast images, and 165 colored histological colorectal images, both stained with Hematoxylin and Eosin, in order to extract features to differentiate between benign and malignant cases. The experiments were also performed over normalized images, aiming to analyze the influence of different color normalization techniques on percolation-based features and whether they can provide better classification results. The feature sets obtained from the application of the method on the original images and on the normalized images with three different techniques were tested using 12 different classifiers. We compared the obtained results with other relevant methods in the area and observed significant contributions, with AUC rates above 0.900 in both normalized and non-normalized images. We also verified that color normalization does not contribute to the classification of breast tumors when associated with percolation features. However, color normalized images from the colorectal tumor’s dataset provided better results than the original images.