NiftyNet meeting 3rd September 2018

Attendance: Wenqi L, Jorge C., Zach ER, Carole S.

Version 0.4 to be released by end of September:

• Includes tf.dataset IO and callbacks
• Dataset API: Multi threads wrapper for python generator. Now feasible to wrap python functions and work on np.arrays
• New loss functions: Tavosky loss Mixture of Dice and Jaccard
• Group normalisation
• Random affine on the fly
• squeeze and excitation unit

WIP

• CSV reader still missing - 2 different situations:
  • Unique ID multiple fields
  • Same ID used for multiple instances (e.g. multi places of sampling)

Current problem: Appropriate matching by user for Subject ID Need to add a remove field in the config for handling subject names. By default remove all the ones in the contain field. If remove fields are indicated, will only remove these. Action Carole

• Merge bias field random augmentation (at image level using np arrays)

• Brainstorming on batch norm problem

  • Change of statistics due to differences in batch statistics