Interpretable multi-view deep networks reveal brain cognitive imaging-genetic associations

This video program is a part of the Premium package:

Interpretable multi-view deep networks reveal brain cognitive imaging-genetic associations


  • IEEE MemberUS $11.00
  • Society MemberUS $0.00
  • IEEE Student MemberUS $11.00
  • Non-IEEE MemberUS $15.00
Purchase

Interpretable multi-view deep networks reveal brain cognitive imaging-genetic associations

1 view
  • Share
Create Account or Sign In to post comments
Brain functional connectivity (FC) depicts the functional relations between different brain regions. Neuroimaging based research has emerged to study brain FC. As brain dysfunctions are genetic inheritable, imaging-genetic integration may help uncover hidden biological mechanisms. The integration of imaging and genetic data, however, is challenging due to the high dimensionality. Moreover, deep networks are composed of a large number of layers and each layer contains complex nonlinear operations, resulting in difficulties in result explanation and biological-mechanism analysis. In this work, we propose an interpretable convolutional network model to address the challenges. Our model, gradient class activation mapping guided convolutional collaborative learning (gCAM-CCL), integrates imaging-genetic data using interpretable deep networks. To make the model interpretable, gCAM-CCL incorporates Grad-CAM and guided Grad-CAM approaches into two convolutional neural networks, later fused using a collaborative layer. The proposed model, gCAM-CCL, can generate activation/contribution maps of the input images/genes using guided-backpropagation and gradient-weighted feature-map combination. GCAM-CCL calculates the weights of feature maps using the gradient of class-label w.r.t. each feature map, and uses global average pooling to merge the gradients so as to combine feature maps. As a result, gCAM-CCL not only can obtain discriminative brain regions and genes but also can generate class-specific results, which further promotes biological mechanism analysis. When applied to the Philadelphia Neurodevelopmental Cohort (PNC), our model shows that low cognitive subjects and high cognitive subjects exhibit different FCs. High cognitive subjects tend to have a small number of dominant FCs connections which activates gCAM-CCL's attention, while low cognitive subjects tend to have a large number of activated FCs. Further analysis on the identified brain FCs shows that lingual gyrus is a significant hub for high cognitive subjects; and gene enrichment analysis on the identified genes shows that pathway ""Regulation of neurotransmitter levels"" is related to high cognitive ability while low cognitive subjects may have problem in pathway ""Midbrain development"" and ""Growth cone"".
Brain functional connectivity (FC) depicts the functional relations between different brain regions. Neuroimaging based research has emerged to study brain FC. As brain dysfunctions are genetic inheritable, imaging-genetic integration may help uncover hidden biological mechanisms. The integration of imaging and genetic data, however, is challenging due to the high dimensionality. Moreover, deep networks are composed of a large number of layers and each layer contains complex nonlinear operations, resulting in difficulties in result explanation and biological-mechanism analysis. In this work, we propose an interpretable convolutional network model to address the challenges. Our model, gradient class activation mapping guided convolutional collaborative learning (gCAM-CCL), integrates imaging-genetic data using interpretable deep networks. To make the model interpretable, gCAM-CCL incorporates Grad-CAM and guided Grad-CAM approaches into two convolutional neural networks, later fused using a collaborative layer. The proposed model, gCAM-CCL, can generate activation/contribution maps of the input images/genes using guided-backpropagation and gradient-weighted feature-map combination. GCAM-CCL calculates the weights of feature maps using the gradient of class-label w.r.t. each feature map, and uses global average pooling to merge the gradients so as to combine feature maps. As a result, gCAM-CCL not only can obtain discriminative brain regions and genes but also can generate class-specific results, which further promotes biological mechanism analysis. When applied to the Philadelphia Neurodevelopmental Cohort (PNC), our model shows that low cognitive subjects and high cognitive subjects exhibit different FCs. High cognitive subjects tend to have a small number of dominant FCs connections which activates gCAM-CCL's attention, while low cognitive subjects tend to have a large number of activated FCs. Further analysis on the identified brain FCs shows that lingual gyrus is a significant hub for high cognitive subjects; and gene enrichment analysis on the identified genes shows that pathway ""Regulation of neurotransmitter levels"" is related to high cognitive ability while low cognitive subjects may have problem in pathway ""Midbrain development"" and ""Growth cone"".