One of the main challenges of histological image analysis is the high dimensionality of the images. This can be addressed via summarizing techniques or feature engineering. However, such approaches can limit the performance of subsequent machine learning models, particularly when dealing with highly heterogeneous tissue samples. One possible alternative is to employ unsupervised learning to determine the most relevant features automatically. In this paper, we propose a method of generating representative image signatures that are robust to tissue heterogeneity. At the core of our approach lies a novel deep-learning based mechanism to simultaneously produce representative image features as well as perform dictionary learning to further reduce dimensionality. By integrating this mechanism in a broader framework for disease grading, we show significant improvement in terms of grading accuracy compared to alternative local feature extraction methods.