NiftyNet meeting 5th October 2018

Attendance: Tom Va., Dan, Carole, Tom V., Jose, Fernando, Pritesh, Zach, Wenqi

- hack day

Zach and Tom Va. organised a 'NiftyNet hack day' -- to jointly work on a new feature for niftynet or improve an old one; implementing probabilistic u-net.

- I/O for image classification/regression

sampler to do partial loading according to the coordinates (Tom Va., Wenqi)

- create a sampler folder

sampler_*.py currently in niftynet/engine/. image_window_dataset and sampler modules are generic, they can be used without niftynet's application, good to have them in a folder as a standalone module (Wenqi)

- event handlers/callbacks

demo to be done to show the usage (Wenqi)

- evaluation

essential blocks ready -- the inference engine needs refactoring, implementing multiple metrics and losses

- network layers

a generic probabilistic layer to do sampling from distributions e.g. those in prob. unet, GAN generator, variational autoencoder.

- training data augmentation

possible features: scale augmentation for vessel segmentation (Fernando), artefact simulation in k-space

- cross entropy input type casting issue

double check the relevant loss function (Wenqi, Zach)

- 'no new-net' implementation

https://arxiv.org/abs/1809.10483 find out any building blocks that are missing from niftynet, the goal is to replicate the results (Dan)

- preparations for the RSNA Quantitative Imaging Reading Room Showcase

poster/slides (Zach)

- new logo

Vote for a new logo (Tom Va.)

- tutorial on image classification

example config/notebook demo (Pritesh)

- config file refactoring

to follow up the progress (Dan, Dzhosh)

- model zoo file storage and licensing

permanent URLs for fetching the files, version control?, each entry should have a license (Fernando, Wenqi)