spie sprint

[massich]

This Issue will be used document and drive the experimentation discussed in SPIE

Description

A dirty implementation of our prostate methods using deep learining. The volumes would be sliced (as sub volume) in a sliding manner and a decision (cancer / non-cancer) would be made for each subvolume.

Tasks

• basic caffe knowledge
• Initial caffe integration
  • Segment only the midle slice of the volume
  • volume slicing

Extra- things to try

• Incorporate a position prior. Maybe you can have T1, DCE, ADC (+ a modality position or atlas) see issue Add position information to the deep learning #5
  • use Atlas Prior
  • use position image [X, Y, Z] of size (MxNxRx3) where the central pixel has a 0,0,0

[massich]

@glemaitre I think this is what we need https://github.com/BVLC/caffe/wiki/Model-Zoo#fcn

[massich]

https://www.youtube.com/watch?v=rvMVqPsXL10 @ min 40, injecting rgb images + depth, either at the first level of the network (or more fancy, inject one of the images at some deeper layer within the network). We can use this to inject the DCE and the ADC at its corresponding resolution