An integrated texton and bag of words classifier for identifying anaplastic medulloblastomas.

In this paper we present a combined Bag of Words and texton based classifier for differentiating anaplastic and non-anaplastic medulloblastoma on digitized histopathology. The hypothesis behind this work is that histological image signatures may reflect different levels of aggressiveness of the disease and that texture based approaches can help discriminate between more aggressive and less aggressive phenotypes of medulloblastoma. The bag of words approach attempts to model the occurrence of differently expressed image features. In this work we choose to model the image features via textons which can quantitatively capture and model texture appearance in the images. The texton-based features, obtained via two methods, the Haar Wavelet responses and MR8 filter bank, provide spatial orientation and rotation invariant attributes. Applying these features to the bag of words framework yields textural representations that can be used in conjunction with a classifier (κ-nearest neighbor) or a content based image retrieval system. Over multiple runs of randomized cross validation, a κ-NN classifier in conjunction with Haar wavelets and the texton, bag of words approach yielded a mean classification accuracy of 80, an area under the precision recall curve of 87 and an area under the ROC curve of 83 in distinguishing between anaplastic and non-anaplastic medulloblastomas on a cohort of 36 patient studies.