Using 3D sMRI to predict age(classification/regression) tends to fail on 3D CNN since -
- Sparse Kernels compared to 3-square(9) and 3-cube(27)
- Blank Voxels too many blanks
In order to tackle these problems, I was told to do this task with 2D CNNs and it has advantages of -
- More Data Since there are only 1,187 brains
- Human-like? Slice view for detecting symptoms is widely used method and this approach does that as well.
Not yet... :)