Classification of Histopathology Images by Siamese and Triplet Network Embeddings

In computational pathology, one of the most investigated aspects concerns the analysis of patients’ tissue images. This type of image is often called a Whole Slide Image (WSI) and is characterised by a high resolution that makes impractical, if not impossible, the classical analysis methods that process the image in a single step. Automatic and Deep Learning methods fit into this context in an attempt to provide automatic medical decision support. In this paper, we present two networks capable, thanks to deep metric learning, of classifying tissue images using a low-dimensional embedding space. This projection makes it easy to visualise the elements of the training set, showing that the different classes are almost always well separated. The embedding is calculated by minimizing the distance between elements of the same class and maximising the distance between elements of different classes. Consequently, the proposed system offers an effective way to visualize the learned classification mechanism. Using four different datasets, we show how this system achieves results in line with and sometimes better than state-of-the-art methods, despite being a much simpler model. A further strength lies in the possibility of having a visualisation that makes the classification mechanism explicit. In addition, we provide a classification confidence score that can add useful information to the medical decision support system.