For the NeuroLF scanner geometry, four iterations of scatter estimation take around 9 minutes of which the majority is spent doing the single scatter simulation. This is because the detectors per ring and the views are currently not downsampled (axial downsampling is the same than for cylindrical scanners). I would propose to add a downsampling in detectors and views by an integer factor, while maintaining the buckets. This would ignore gaps between blocks, but since the simulation result is quite smooth anyways that shouldn't introduce large errors.