A framework for semantic analysis of histopathological images using nonnegative matrix factorization

This paper presents a novel and general framework for histopathology image analysis using nonnegative matrix factorization. The proposed method uses a collection-based image representation called Bag of Features (BOF) to represents the visual information of a histopathology image collection. Convex Nonnegative Matrix Factorization (CNMF) is applied to a training set of images to find a compact representation in a latent topic space. The latent representation has two important characteristics: first, CNMF is able to find representative clusters of images in the collection, second, clusters are represented by convex linear combinations of images in the training set. This latent representation is exploited in different ways by the proposed framework: concept labels can be assigned to clusters using the labels of the constituting images, representative images and visual words can be identified for each cluster, and new unlabeled images can be labeled by mapping them to the latent space. The proposed annotation model has an interesting property, it is easily interpretable since it is possible to trace those visual words present in the image which contribute the most to a given annotation. This implies that annotations in an image may be explained by identifying the regions that contributed to them. An exploratory experimentation was performed in a histopathology dataset used to diagnose a type of skin cancer called basal cell carcinoma. The preliminary results show that the combination of BOF and NMF is an interesting alternative for biomedical image collection analysis with a high level of interpretability.