A graph-based method for detecting characteristic phenotypes from biomedical images

We propose a novel method for detecting characteristic informative phenotype patterns from biomedical images. By building a metric space quantifying the difference between images, we learn the distributions of different classes, and then detect the characteristic regions using graph partition. We show that the detected regions are statistically significant. Our approach can also be used for designing differentiating features for specific data set. We apply our method to a digital pathology problem and successfully detect two characteristic phenotypes pertaining to normal liver and hepatoblastoma nuclei. In addition to digital pathology, our method can be applied to other biomedical problems for detecting characteristic phenotypes (e.g. location proteomics, genetic screens, cell mechanics, etc.).