Domain Adaptation for M&Ms Challenge

Dataset specifications.

	Siemens	Philips	GE	Canon	Total
Label	A	B	C	D
Centres	1	2 & 3	4	5
Training	75	50 & 25	25	0	175
Validation	4	5 & 5	10	10	34
Testing	16	19 & 21	40	40	136
Overall	95	74 & 51	75	50	345

Training data for GE Vendor are unlabeled.

Requirements

!conda install -c conda-forge ffmpeg -y !conda install tsnecuda -c cannylab -y Pytoch >= 1.6

export SLACK_TOKEN='you_slack_token'

Data Preparation

./scripts/mms2d.sh only_data
python3 tools/nifti2slices.py --data_path data/MMs
python3 tools/testgt2phases.py

Guidelines

MMs dataset naming:

• _full Get all volumes (not only segmented 'ED' and 'ES' phases volumes).
• _unlabeled Get only unlabeled volumes (for 'ED' and 'ES' phases)
• _centre*xyz* Get volumes (for 'ED' and 'ES' phases) for selected centres. Example _centre1, _centre13. Last one picks centres 1 and 3. Available Centres from 1 to 5.
• _vendor*jkl* Get volumes (for 'ED' and 'ES' phases) for selected vendors. Example _centreC, _vendorAB. Last one picks vendors A and B. Available Vendors 'A', 'B', 'C', 'D'.
• _all Get all segmented (which is ED and ES phases) slices (Training Labeled + Training Unlabeled + Validation + Test)

Normalization:

• none
• reescale
• reescale_phase
• reescale_full_vol
• standardize
• standardize_phase
• standardize_full_vol

Coral Loss: The weight of the CORAL loss (λ) is set in such way that at the end of training the classification loss and CORAL loss are roughly the same. It seems be a reasonable choice as we want to have a feature representation that is both discriminative and also minimizes the distance between the source and target domains

https://discuss.pytorch.org/t/why-dont-we-put-models-in-train-or-eval-modes-in-dcgan-example/7422/8

ToDo