Interpreting Age Predictions from Brain Maps via Deep Neural Activations and Tensor Decomposition
Keywords: neuroimaging, mri, non-linear regression, tensor decomposition, interpretability methods, convolutional neural networks.
TL;DR: tensor decomposition technique applied to activations maps to interpret convolutional models trained on brain maps.
Abstract: Deep learning models, while effective, often lack transparent interpretability, especially in critical areas like healthcare. This paper introduces a novel approach to interpret 3D convolutional neural networks trained to estimate one or more of the individual's clinically relevant attributes from 3D brain maps. In contrast to interpretability methods commonly applied to object classification in images, such as gradient approaches like GradCAM, which must rely on per-instance explanations due to the spatial variation of object location, brain maps have a common spatial registration and we propose to compute explanations at the dataset-level. After organizing the internal activations of 3D convolutional neural network across the training dataset into a tensor, we use a constrained tensor decomposition to reveal the key spatial patterns that highlight the specific regions of the brain the model focuses on during its predictions. We use reconstruction error to guide the selection of the rank of the tensor decomposition, and fit linear models to relate the decomposition of activations to the original target attributes. We apply the method to network trained to estimate an individual's chronological age using brain maps of volume and stiffness computed from magnetic resonance imaging (MRI) and T1-weighted magnetic resonance elastography (MRE) scans, respectively. The tensor decomposition's spatial factors predominantly emphasize areas of the brain known to vary with aging. Additionally, the linear model fit to the decomposition has only a slight decrease in performance. The proposed decomposition technique provides a mechanism to interpret convolutional models applied to brain maps, and offers potential insights into the age-related structural changes in the brain.
Primary Area: applications to neuroscience & cognitive science
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 6590
Loading