Lesson 11 in class

Part 2 (2017)
21

Is pseudo labeling projecting that line for incorrect labels in the other direction? That explains a lot about why it’s so effective.

22

Any thoughts on Xception (Francois Chollet’s network - conveniently added to Keras as a built in model)?

Short cut architecture like ResNet but uses separable convolutional layers.

https://arxiv.org/abs/1610.02357

2 Likes
23

Any tips for how to work with 3D convolution layers?

1 Like
24

@samwit
Same set of layers as for 2D, so can be used in similar ways… (Just no pretrained networks available as far as I know…):

Conv3D, MaxPooling3D, GlobalAveragePooling3D

Can still use BatchNormalization and Dropout layers, etc.

Also I really like the NIFTI format - automatically rescales DICOM pixel values and gives you affine matrix for transform from voxel to patient coordinates.

1 Like
25

V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation
https://arxiv.org/abs/1606.04797

3 Likes
26

any tips/changes for settings in those layers or basically same as 2d?

27

is there a cancer blob in the images jeremy is showing?

28

Same just with an extra dimension for kernel sizes, subsampling, etc…

29

I found it difficult to get the pre-processing right.

30

Thanks David

31

Question for anyone who’s worked with the CT data: were there strategies you had to apply to get it to fit in memory properly? Was it workable at all on a P2?

1 Like
32

@renjithmadhavan

Really recommend working with NIFTI (neuroimaging format) instead of DICOM. Can convert with dicom2nifti package or multiple applications available (e.g. dcm2nii).

Automatically scales to Hounsfield units, gives you a 3D dataset including affine transform, etc. etc.

I posted a bit about it somewhere else showing basics.

Rescaling to 1 mm voxels is as simple as:

# convert dicom to nifti with dicom2nifti first
fp = '/path/to/file.nii.gz'
nii  = nibabel.load(fp)
img = scipy.ndimage.zoom(nii.get_data(), abs(nii.affine.diagonal()))
2 Likes
33

Thanks You @davecg

34

Dye is not radioactive. Just dense. Isovue 370 or Isovue 300 usually at my institution. /radiologist :skull_crossbones:

Issues are contrast induced nephropathy, allergic reactions, etc. etc.

5 Likes
35

the voice from the other mics break up, and its hard to understand other people speak

36

its perfect now!

1 Like
37

thats good

1 Like
38

were you able to hear alright when I spoke just now? Jeremy tried adjusting

39

Sounds good.

1 Like
40

clear

1 Like