convolutional neural networks for mesh-based … · "convolutional neural networks on graphs...

19
CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED PARCELLATION OF THE CEREBRAL CORTEX Guillem Cucurull, Konrad Wagstyl, Arantxa Casanova, Petar Veličković, Estrid Jakobsen, Michal Drozdzal, Adriana Romero, Alan Evans, Yoshua Bengio 1

Upload: others

Post on 04-Oct-2020

22 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED … · "Convolutional neural networks on graphs with fast localized spectral filtering." In Advances in Neural Information Processing

CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED PARCELLATION OF THE CEREBRAL CORTEX

Guillem Cucurull, Konrad Wagstyl, Arantxa Casanova, Petar Veličković, Estrid Jakobsen, Michal Drozdzal, Adriana Romero, Alan Evans, Yoshua Bengio

�1

Page 2: CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED … · "Convolutional neural networks on graphs with fast localized spectral filtering." In Advances in Neural Information Processing

CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED PARCELLATION OF THE CEREBRAL CORTEX

CORTEX PARCELLATION

▸ Different areas of the cerebral cortex are involved in different cognitive processes:

• Visual processing, language comprehension

▸ Mapping these areas helps us understand how the cortex is organized

Insert Cortex Parcellation Example Image

�2

Cortex Parcellation. From Glasser et al. 2016

Page 3: CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED … · "Convolutional neural networks on graphs with fast localized spectral filtering." In Advances in Neural Information Processing

CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED PARCELLATION OF THE CEREBRAL CORTEX

WHAT IS A CORTICAL MESH

▸ Common coordinate system

▸ Can represent multiple imaging modalities and features

▸ Can be used to coregister cortical surfaces between different individuals

Insert Cortical Mesh Example Image

�3

Zoom of a cortical mesh

Page 4: CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED … · "Convolutional neural networks on graphs with fast localized spectral filtering." In Advances in Neural Information Processing

CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED PARCELLATION OF THE CEREBRAL CORTEX

PREVIOUS WORK

▸ Glasser et al. 2016

Used an MLP classifier to recognize each cortical area, trained on semi-automatic labels of cortex parcellation.

▸ Jakobsen et al. 2016

Compared connectivity templates of each vertex of the cortical surface to parcellate Broca’s region.

�4

Page 5: CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED … · "Convolutional neural networks on graphs with fast localized spectral filtering." In Advances in Neural Information Processing

CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED PARCELLATION OF THE CEREBRAL CORTEX

WHY CONVOLUTIONS

▸ Convolutional models achieve state of the art on image segmentation, which is analogous to this task.

▸ Important properties:

▸ Parameter sharing

▸ Exploits data structure

�5

Image from http://petar-v.com/GAT/

Page 6: CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED … · "Convolutional neural networks on graphs with fast localized spectral filtering." In Advances in Neural Information Processing

CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED PARCELLATION OF THE CEREBRAL CORTEX

GRAPH CONVOLUTIONS

▸ Graph Convolutional Networks [Kipf 2017], ChebNets [Defferrard 2016]:

▸ Graph Attention Networks [Veličković 2018]:

�6

h′�i = σ( ∑j∈𝒩i

1cij

Whj)

h′�i = σ( ∑j∈𝒩i

αijWhj)

From the normalized

adjacency matrix

Attention coefficient

Page 7: CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED … · "Convolutional neural networks on graphs with fast localized spectral filtering." In Advances in Neural Information Processing

CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED PARCELLATION OF THE CEREBRAL CORTEX

BASELINES▸ NodeMLP

• Independently classify each node

▸ NodeAVG

• Predict the most frequent label at each node, computed in the training set

▸ MeshMLP

• Process the full mesh at once, concatenating all nodes together

�7

Page 8: CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED … · "Convolutional neural networks on graphs with fast localized spectral filtering." In Advances in Neural Information Processing

CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED PARCELLATION OF THE CEREBRAL CORTEX

DIFFERENCES BETWEEN MODELS

�8

Method

Properties NodeAVG NodeMLP MeshMLP GraphConv

Neigh Info

Global Info

Features

Page 9: CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED … · "Convolutional neural networks on graphs with fast localized spectral filtering." In Advances in Neural Information Processing

CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED PARCELLATION OF THE CEREBRAL CORTEX

BROCA’S REGION PARCELLATION

▸ Broca’s region has an important role in language processing

It can be subdivided into two adjacent cytoarchitectonic areas: 44 and 45.

�9

▸ Dataset

▸ Manually labeled parcellation of Broca’s area from 100 different subjects.

▸ Each mesh has 1195 nodes, with 9 features per node, and 3 possible labels.

▸ Augment the features with x,y,z absolute position of the vertex

From Geranmayeh, 2014

Page 10: CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED … · "Convolutional neural networks on graphs with fast localized spectral filtering." In Advances in Neural Information Processing

CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED PARCELLATION OF THE CEREBRAL CORTEX

EXPERIMENTAL SETUP

▸ Given the limited size of the dataset we use 10-fold cross validation to report results:

- 8 folds for training, 1 for validation, 1 for test

- The models are trained using Adam optimizer

- The optimized loss function is either weighted cross entropy or Dice loss

�10

Page 11: CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED … · "Convolutional neural networks on graphs with fast localized spectral filtering." In Advances in Neural Information Processing

CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED PARCELLATION OF THE CEREBRAL CORTEX

RESULTS

�11

Page 12: CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED … · "Convolutional neural networks on graphs with fast localized spectral filtering." In Advances in Neural Information Processing

CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED PARCELLATION OF THE CEREBRAL CORTEX

RESULTS

�12

Page 13: CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED … · "Convolutional neural networks on graphs with fast localized spectral filtering." In Advances in Neural Information Processing

CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED PARCELLATION OF THE CEREBRAL CORTEX

RESULTS

�13

Page 14: CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED … · "Convolutional neural networks on graphs with fast localized spectral filtering." In Advances in Neural Information Processing

CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED PARCELLATION OF THE CEREBRAL CORTEX

RESULTS

�14

Page 15: CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED … · "Convolutional neural networks on graphs with fast localized spectral filtering." In Advances in Neural Information Processing

CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED PARCELLATION OF THE CEREBRAL CORTEX

RESULTS

�15

Using node

position

as features

Page 16: CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED … · "Convolutional neural networks on graphs with fast localized spectral filtering." In Advances in Neural Information Processing

CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED PARCELLATION OF THE CEREBRAL CORTEX

RESULTS

�16

Subject 1

Subject 2

Subject 1

Subject 2

Page 17: CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED … · "Convolutional neural networks on graphs with fast localized spectral filtering." In Advances in Neural Information Processing

CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED PARCELLATION OF THE CEREBRAL CORTEX

SUMMARY

▸ Investigated the use of graph CNNs for cortical mesh parcellation

▸ Evaluated on Broca’s area segmentation

▸ Graph CNNs models outperform the baselines

�17

Page 18: CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED … · "Convolutional neural networks on graphs with fast localized spectral filtering." In Advances in Neural Information Processing

CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED PARCELLATION OF THE CEREBRAL CORTEX

REFERENCES‣ Glasser, Matthew F., Timothy S. Coalson, Emma C. Robinson, Carl D. Hacker, John Harwell, Essa Yacoub, Kamil Ugurbil et

al. "A multi-modal parcellation of human cerebral cortex." Nature 536, no. 7615 (2016): 171-178. ‣ Jakobsen, Estrid, Franziskus Liem, Manousos A. Klados, Şeyma Bayrak, Michael Petrides, and Daniel S. Margulies.

"Automated individual-level parcellation of Broca's region based on functional connectivity." Neuroimage (2016). ‣ Defferrard, Michaël, Xavier Bresson, and Pierre Vandergheynst. "Convolutional neural networks on graphs with fast

localized spectral filtering." In Advances in Neural Information Processing Systems, pp. 3844-3852. 2016. ‣ Kipf, Thomas N., and Max Welling. "Semi-supervised classification with graph convolutional networks." International

Conference on Learning Representations, ICLR, (2017). ‣ Veličković, Petar, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. "Graph Attention

Networks.” International Conference on Learning Representations, ICLR, (2018). ‣ Geranmayeh, Fatemeh, Sonia LE Brownsett, and Richard JS Wise. "Task-induced brain activity in aphasic stroke patients:

what is driving recovery?." Brain 137, no. 10 (2014): 2632-2648.

�18

Page 19: CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED … · "Convolutional neural networks on graphs with fast localized spectral filtering." In Advances in Neural Information Processing

Guillem Cucurull, Konrad Wagstyl, Arantxa Casanova, Petar Veličković, Estrid Jakobsen, Michal Drozdzal, Adriana Romero, Alan Evans, Yoshua Bengio

�19

THANKS!CONVOLUTIONAL NEURAL NETWORKS FOR MESH-BASED PARCELLATION OF THE CEREBRAL CORTEXhttps://openreview.net/forum?id=rkKvBAiiz