Open agitter opened 7 years ago
Relatively standard architecture: 3D convolutional neural networks. Voxel input data. Single output node for age.
This work falls prey to some hazards for using normality assumptions to characterize distributions that must not be normal:
The average interval between each scan 68.17 ± 92.23 days, with eight participants being scanned in Amsterdam first, three in London first.
Results:
Performance was generally strong using existing methods and the neural network when features were extracted and GM/WM were input directly. When raw voxel data were used, the CNN performance was essentially unaffected, while the existing method became much less reliable.
In terms of impact for our review, this paper doesn't provide transformational results in terms of performance. It does, however, highlight the role of deep neural networks as a powerful feature constructor.
https://arxiv.org/abs/1612.02572