ICAM: Interpretable Classification via Disentangled Representations and Feature Attribution Mapping

TY - CHAP

T1 - ICAM: Interpretable Classification via Disentangled Representations and Feature Attribution Mapping

AU - Bass, Cher

AU - da Silva, Mariana

AU - Sudre, Carole H.

AU - Tudosiu, Dan

AU - Smith, Stephen M

AU - Robinson, Emma

PY - 2020/12/6

Y1 - 2020/12/6

N2 - Feature attribution (FA), or the assignment of class-relevance to different locationsin an image, is important for many classification problems but is particularly crucialwithin the neuroscience domain, where accurate mechanistic models of behaviours,or disease, require knowledge of all features discriminative of a trait. At the sametime, predicting class relevance from brain images is challenging as phenotypesare typically heterogeneous, and changes occur against a background of significantnatural variation. Here, we present a novel framework for creating class specificFA maps through image-to-image translation. We propose the use of a VAE-GANto explicitly disentangle class relevance from background features for improvedinterpretability properties, which results in meaningful FA maps. We validate ourmethod on 2D and 3D brain image datasets of dementia (ADNI dataset), ageing(UK Biobank), and (simulated) lesion detection. We show that FA maps generatedby our method outperform baseline FA methods when validated against groundtruth. More significantly, our approach is the first to use latent space sampling tosupport exploration of phenotype variation. Our code will be available online athttps://github.com/CherBass/ICAM.

AB - Feature attribution (FA), or the assignment of class-relevance to different locationsin an image, is important for many classification problems but is particularly crucialwithin the neuroscience domain, where accurate mechanistic models of behaviours,or disease, require knowledge of all features discriminative of a trait. At the sametime, predicting class relevance from brain images is challenging as phenotypesare typically heterogeneous, and changes occur against a background of significantnatural variation. Here, we present a novel framework for creating class specificFA maps through image-to-image translation. We propose the use of a VAE-GANto explicitly disentangle class relevance from background features for improvedinterpretability properties, which results in meaningful FA maps. We validate ourmethod on 2D and 3D brain image datasets of dementia (ADNI dataset), ageing(UK Biobank), and (simulated) lesion detection. We show that FA maps generatedby our method outperform baseline FA methods when validated against groundtruth. More significantly, our approach is the first to use latent space sampling tosupport exploration of phenotype variation. Our code will be available online athttps://github.com/CherBass/ICAM.

M3 - Conference paper

VL - 33

BT - Advances in Neural Information Processing Systems

ER -