Low diversity of MAISI when given few anatomies

[Masaaki-75]

I've been playing with MAISI for a few days and found that, if the anatomy list contains only a few labels (which is typical if i want to focus on 'head' region, e.g., ['brain', 'vertebrae C1', ...]), the inference may produce results with low diversity, even if i change the random seeds for each iteration (see below).

I am generating volumes with masks data provided by the MAISI team, and i guess the generation diversity highly depends on the diversity of such predefined mask database. So, i am wondering if there is a way to generate diverse head-region volumes using MAISI without reliance on the provided mask database?

[Can-Zhao]

Thanks for the feedback. The amount of head and neck training data is less then abdomen, which might be the reason. I was wondering if it would be possible to use your own head mask.