Leveraging Deep Embeddings for Explainable Medical Image Analysis

Machine learning techniques applied to the medical image analysis domain provide valuable tools that improve the diagnostic process. Among the proposed machine learning methodologies, deep neural networks are state-of-the-art in medical domain applications. However, they still have the disadvantage of being black-box methods. On the other hand, the medical field requires approaches that propose decisions based on an explainable mechanism, providing meaningful suggestions to physicians. In this chapter, we propose a general paradigm for an explainable classification of medical imaging data. This paradigm adopts deep metric learning to provide an embedding that enables the representation of images in either two or three dimensions. Metric learning plus dimensionality reduction in 2-3D introduces a first level of explainability. In particular, this is achieved by showing the training images closer to the test ones, consequently allowing for neighbour identification. A subsequent level of explainability is added by an interpretable classifier. This chapter will also present four use cases demonstrating the application of the proposed paradigm, each related to a specific kind of image dataset, such as histopathological or X-ray images.