Minutes Data Working Group 18 Jun 2020 - Project-MONAI/MONAI GitHub Wiki
Agenda
(All) Review discussions from other workgroups on requirements and additional tasks
(Brad) Finish reviewing requirement document
(Stephen) Check in on working group chat
Meeting Minutes
Group reviewed the requirements for passing data into MONAI
We need to strike the right balance between flexibility in supporting data types of all kinds (DICOM, HL7, other data inputs), but also flexible in supporting data structures of all kinds (BIDS, MSD, Clara, NiftyNet
It isn’t necessarily MONAI’s role to “pick a specification”
Researchers shouldn’t have to create/translate into a data format that forces them to do a lot more work with no value add
“As simple as possible, but no simpler”
Challenge with datasets today (BIDS, MSD) is that
They have specified the dataset as part of the experiment run
The structure is simplified as it is geared to a specific use case, not a generic case
Instead, let’s:
Separate “data set definition” and “experiment definition”, so that they can be supplied separately
Create a “transformer development framework” that defines how to take a dataset to be internally consumable
Create a set of “canned transformers”, that supports one for BIDS, one for MSD (supported by MONAI 0.2), ones for Clara Train and NiftyNet, so that researchers can bring their own data
And, Kaggle challenges etc can provide datasets in this supported format and gives the participants a place to start
The “experiment definition” will reference the data set definition, and include parameters to support specific runs, e.g.:
“These studies are for training, and these are for validation”
“Take a random seed of 80% for training, 20% for validation, for each label”
Next Steps
(All) Continue to iterate on this design
(Brad) Check with steering committee on outreach to BIDS