ICAM: Interpretable Classification via Disentangled Representations and Feature Attribution Mapping

Feature attribution (FA), or the assignment of class-relevance to different\nlocations in an image, is important for many classification problems but is\nparticularly crucial within the neuroscience domain, where accurate mechanistic\nmodels of behaviours, or disease, require knowledge of all features\ndiscriminative of a trait. At the same time, predicting class relevance from\nbrain images is challenging as phenotypes are typically heterogeneous, and\nchanges occur against a background of significant natural variation. Here, we\npresent a novel framework for creating class specific FA maps through\nimage-to-image translation. We propose the use of a VAE-GAN to explicitly\ndisentangle class relevance from background features for improved\ninterpretability properties, which results in meaningful FA maps. We validate\nour method on 2D and 3D brain image datasets of dementia (ADNI dataset), ageing\n(UK Biobank), and (simulated) lesion detection. We show that FA maps generated\nby our method outperform baseline FA methods when validated against ground\ntruth. More significantly, our approach is the first to use latent space\nsampling to support exploration of phenotype variation. Our code will be\navailable online at https://github.com/CherBass/ICAM.\n