Abstract
Attributing the pixels of an input image to a certain category is animportant and well-studied problem in computer vision, with applicationsranging from weakly supervised localisation to understanding hidden effects inthe data. In recent years, approaches based on interpreting a previouslytrained neural network classifier have become the de facto state-of-the-art andare commonly used on medical as well as natural image datasets. In this paper,we discuss a limitation of these approaches which may lead to only a subset ofthe category specific features being detected. To address this problem wedevelop a novel feature attribution technique based on Wasserstein GenerativeAdversarial Networks (WGAN), which does not suffer from this limitation. Weshow that our proposed method performs substantially better than thestate-of-the-art for visual attribution on a synthetic dataset and on real 3Dneuroimaging data from patients with mild cognitive impairment (MCI) andAlzheimer's disease (AD). For AD patients the method produces compellinglyrealistic disease effect maps which are very close to the observed effects.