Multi-Modal Image Classification Using Low-Dimensional Texture Features For Genomic Brain Tumor Recognition

In this paper, we present a multi-modal medical image classification framework classifying brain tumor glioblastomas in genetic classes based on DNA methylation status. The framework makes use of computationally efficient 3D implementations of short local image descriptors, such as LBP, BRIEF and HOG, which are processed by a Bag-of-Patterns model to represent image regions, as well as deep-learned features acquired by denoising auto-encoders and hand-crafted shape features calculated on segmentation masks. The framework is validated against a cohort of 116 brain tumor patients from the TCIA database and is shown to obtain high accuracies even though the same image-based classification task is hardly possible for medical experts.