Pre-training on large-scale T2w healthy/pathology data #11

naga-karthik · 2024-03-24T14:20:35Z

This PR adds the code to do supervised (and self-supervised) pretraining on T2w MRI data from a collection of public/private, healthy/pathology datasets.

TODO:

Add code for self-supervised pre-training
Test single-gpu and multi-gpu training

…of ancdata'

… is learning

…lve number of workers issue Fixes: #8

It seems that the transform randomly selects a foreground point from image, then use it as center crop. This means that it can find the closest voxel that is just outside the SC and use it as the center (hence it includes the SC). Source: Project-MONAI/MONAI#452 (comment)

…or the RandCoarseDropoutd and RandCoarseShuffled transforms

number of holes to be used for the RandCoarseDropoutd and RandCoarseShuffled transforms

From contrast-agnostic SC seg transforms (which heavily deform the image) to simpler transforms more suitable for lesion seg. Context: #10 (comment) Co-authored-by: Naga Karthik Enamundram <[email protected]>

From MONAI doc (https://docs.monai.io/en/latest/transforms.html#randcoarsedropoutd): - if True, dropout the regions of holes and fill value specified by 'fill_value' - if False, keep the holes and dropout the outside and fill value specified by 'fill_value'

… are using two augmented samples per 3D Volume

…orking

… training/validation transforms

valosekj and others added 30 commits March 12, 2024 17:05

change 'num_workers' to 0 to prevent 'RuntimeError: received 0 items …

f5881aa

…of ancdata'

Update input arg description

7aa9858

Rerun the notebook

16691d7

Print hyper-parameters into the log file

56ca187

Add '--cuda' input arg

ce7315a

track and save epoch time

1ebf833

Plot and save input and output validation images to see how the model…

208b2c1

… is learning

Add comment

5a9e69a

Do not plot 'outputs_v2' as it is a hidden representation

eca6616

Include the epoch number as master title

8e55985

Add 'torch.multiprocessing.set_sharing_strategy('file_system')' to so…

e7575c5

…lve number of workers issue Fixes: #8

Create validation_figures directory if it does not exist

3b6df14

Use 3 leading zeros for the epoch number in the figures fname

59b061c

Link issue

f2d648b

Add 'number_of_holes' arg to specify the number of holes to be used f…

c1553d6

…or the RandCoarseDropoutd and RandCoarseShuffled transforms

typo

0e802df

batch_size = 8

9255855

NUM_WORKERS = batch_size

8296913

number_of_holes=5

93c5415

number of holes to be used for the RandCoarseDropoutd and RandCoarseShuffled transforms

Update transforms for training of the fine-tuned model

a4182b5

From contrast-agnostic SC seg transforms (which heavily deform the image) to simpler transforms more suitable for lesion seg. Context: #10 (comment) Co-authored-by: Naga Karthik Enamundram <[email protected]>

update comment, remove unused imports

2573578

Add notebook with RandCoarseDropoutd transform debug

cfb8f12

remove unused 'max_spatial_size' arg

74b2f2c

use 'fill_value=0' for 'RandCoarseDropoutd'

bdb2ac1

Plot also RandCoarseDropoutd dropout_holes=False fill_value=0

f7f3689

Add note that the batch size is actually doubled (8*2=16), because we…

af9c6d0

… are using two augmented samples per 3D Volume

Add '--cuda' input arg

b416135

Remove 'AsDiscrete'

0303776

naga-karthik and others added 30 commits April 8, 2024 16:14

move normalization up before calculating loss

6be6e64

Fix mode for transforms.RandAffined

aac5bb0

uncomment transforms.CropForegroundd

2d9a381

Save 'checkpoint.pth' after each epoch

5540fa4

Make "--resume-from-checkpoint" and "--run-dir" input args properly w…

783117a

…orking

limit the number of slices to plot to 10

04dfeb8

Change transforms.Spacingd to 1.0 iso

fadc058

comment transforms.CropForegroundd. Again :-D

3f23f35

uncomment transforms

ce81111

change crop_pad_size and epoch nb

3830cb0

do not hardcode 'config["model"]["monai-unet"]["spatial_dims"]'

85d1785

fix plotting logic

4ff571e

Run 'plot_slices' only if local_rank == 0

ff8673e

Add 'local_rank' to BackboneModel to do not duplicate logger.info

d3ab5ce

Fix 'local_rank' passing to 'load_data' to do not duplicate logger.info

61fdf48

Use 'local_rank == 0' to not duplicate logger.info

7f4538e

Copy the transforms.py file to the log directory to keep track of the…

e4803bb

… training/validation transforms

Use 'if local_rank == 0 else None' and add 'logger.info'

6c0fb6d

Rotate val samples by 90 degrees before plotting

3648d91

logg loss function

4b1234d

Add dropout to 'monai-unet'

fe3fb74

Add a comment why we use 320 channels

4800530

Copy datalists to the log directory

61f2908

Fix typo

d5be144

Add 'git annex get $(find . -name "*T2w_lesion-manual.nii.gz")'

e2e4c07

Rotate by 270 degrees instead of 90

0d8aac8

Add attentionunet

d9da21a

Add attentionunet

4f8415f

Fix suptitle position

4d429fc

remove num_res_units for AttentionUnet

394c3ef

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Pre-training on large-scale T2w healthy/pathology data #11

Pre-training on large-scale T2w healthy/pathology data #11

naga-karthik commented Mar 24, 2024 •

edited

Loading

Pre-training on large-scale T2w healthy/pathology data #11

Are you sure you want to change the base?

Pre-training on large-scale T2w healthy/pathology data #11

Conversation

naga-karthik commented Mar 24, 2024 • edited Loading

naga-karthik commented Mar 24, 2024 •

edited

Loading