[ITK] Level set with shape prior

[Timothee Evain]

Hello everyone,

I'm trying the level-set segmentation with shape guidance as described in the manual guide; the one based on the Leventon et al. paper "Statistical shape influence in geodesic active contours".
https://itk.org/Doxygen/html/Examples_2Segmentation_2GeodesicActiveContourShapePriorLevelSetImageFilter_8cxx-example.html

Unfortunately, I fail to make it works correctly : when I look at the shape prior during optimization, after the initialization with an usual centering helper, it "drifts" away of the feature image until it disappears, which obviously leads to weird segmentation behavior. I've tried to toy with parameters, setting high weight on gradient term (the only one that actually links the shape with the feature map) in MAP cost function, but nothing make it right. It feels like the shape prior make no sense. 
Also, the example include an optimization through the OnePlusOneEvolutionary, and I'm not sure random permutation is the best way to achieve good prior. In the paper they used a gradient ascent, so I tried the gradient descent, but alas the cost function does not support derivative-based optimization.

Have anyone applied this method successfully on their data ? Any hints welcomed 
Thanks in advance

Tim