Niftynet meeting 9th april 2018

Minutes NiftyNet 09/04/2018 Attendance: Dzhoshkun, Tom Va., Wenqi, Sebastiano, Zach, Carole

Update on IO design

IO design

Numpy all until network Concatenated set of transformations Dataset to maintain sequence Customised function to map to other sequences Conversion of indices into filename Decomposition of the loading procedure More Flexible Concatenated arrays for multiple modalities Need to pass meta data through Pb of knowledge of shape of tensors beforehand - Possible workaround with forced maximal size and padding Wenqi to look into data generator API

Internal data representation

Need for metadata (could be nested dictionary) Need to be able to put data back into right space of images From Generator API to be looked into Pb of multi task multi level (image vs patch) Can the Data field be a dictionary? (Multi level data: ex - Classification Image wide and Segmentation patch wise)

Upgrading

• Tom - Tensor flow requirements - ok with 1.6 Cuda 9
• Dzhoshkun: Upgrade should be system wide - New Nvidia drivers Upgrade to CUDNN needed as well
• Zach Memory saving gradient useful and working in 1.4

Need for more tutorials

Tom to look into Classification tutorial based on Eli's work workaround (one voxel for image output)

Ticket for jorge for web transfer

Need to move everything to GitHub

Config file All agree -> to be moved forward

Submission RSNA annual meeting

• Wednesday 11th April Deadline
• If accepted, commitment to presence at meeting
• Carole to read
• New Slack channel in place

Added options/possibilities

• Zach Validation - Change preprocessing if augmentation not needed/needed for fair comparison.
• Tom Weighted combination of losses
• providing access to the current iteration number from an application by default
• choosing a default image loader, document the optional package requirements clearly