search

SyneRBI / SIRF-Exercises

SIRF Training and demonstration material
http://www.ccpsynerbi.ac.uk
Apache License 2.0
18 stars 21 forks source link

Fix mapem explode #109

Closed ashgillman closed 3 years ago

ashgillman commented 3 years ago

Fix #96

No more NaNs

image

KrisThielemans commented 3 years ago

I guess this needs to be rebased

ashgillman commented 3 years ago

Whoops, I guess I was in a rush. Fixed

KrisThielemans commented 3 years ago

thanks. It could be useful to draw some attention to this in a doc-cell. I don't think that the original paper has this "epsilon". I believe it is necessary because of OSEM.

ashgillman commented 3 years ago

I'm not 100% sure why subsets would impact this update, which is in image space (that's in no way to say I doubt it!). But logically, if locally x_reg = 1 and x_em = 0 then you'd get the denominator = 0, so it seems pretty reasonable that that could occur and that the eps is a good idea. I've added a fairly generic explanation - do you think it's adequate?

KrisThielemans commented 3 years ago

sure.

I believe that if you initialise the algo with non-zero, and you use 1 subset, it can never get to this case, but it's ages ago I've checked that or indeed read the paper. Happy with your modification.