NiftyNet Dev Meeting 15th April 2019

Attendance:

Wenqi, Felix, Tom Va., Jorge, Pedro, Ben, Carole

Strategy forward:

• Discussion around ways to cut between modular parts - Particular issue of default activation layer with loss function / Pb of associated documentation
• Discussion around move towards tensorflow 2.0 - Details to be discussed on dedicated meeting 10th May 2019 2pm - Input from TomVa. Changes associated to Application driver (no session anymore), collection of variables (no call to get Collection), use of tf.layers instead of current template op.
• Move towards more abstracted use
• Need to integrate as much as possible existing feature recently developed
• Need for detailed "How to pages" with examples of actual code implementation for user-dependent features
• Need to find ways to smooth learning curve for new adopters of NiftyNet
• Need for combination of commented contrib systems with recent published applications - ALL to act on this

Currently developed features to be included

Current assignment based on main developers of associated features.

Network

• Hemis - Tom
• VGG - Felix
• Guotai's BRATS network

Custom layers

• Gaussian Sampler - Ben
• Stochastic filter group (Masking kernels) - Felix
• Probability layer (use different probability distribution) - Felix

Preprocessing

• Histogram regression to classification - Felix
• MR Augmentation: motion, blurring, noise, RF spike - Richard

Sampler

• Uniform sampler with pad authorization - Ben
• CSV reader - Carole/Tom
  • exists as a module but still difficult to be used directly (need many modifications)

Handler - Tom

• Performance based gradient adjustments
  • Learning rate decay
  • Early stopping performance based
  • Reduced learning rate on plateau
• Online update
• Whole volume validation (still needs review and testing)

Custom applications - to use as demo / example

• Customised application using CSV reader - Pedro / Ze upon publication
• Multi-task - Felix upon publication
• Reinforcement learning - Kerstin upon publication
• Distillation loss - Irme upon publication
• RCNN/object detection - Carole
• Counting - Zach upon publication
• VAE modification - Reuben upon publication

Losses

• Quality Driven Loss - Zach
• Cosine similarity gradient enforcement - Irme
• Cosine loss for direction regression - Carole
• Smooth L1 - Carole
• Volume consistency - Carole
• Volume existence - Carole
• Variability loss - Carole
• Weight batch - Carole

Post-processing / Outputs

• Model introspection - With standalone demo jupyter notebook - Felix
• Multiple outputs - Demo - Felix/Carole

Tensorboard

• Histogram viewer

Documentation strategy - how to pages

• Loading model from different graph - Tom
• Event handler - Tom
• Multi output - Carole
• Custom placeholders - Tom