I don’t think you can use the integration from Scipy directly. You should try to find a closed-form solution to the integration (or an approximation) and rewrite it in theano.
You can have a look at PR #2688, where @domenzain is implementing logcdf for distribution. There are quite a lot of integration involved.