After exploring different options to handle processing large series (800+ slices) using recommendations from TotalSegmentator repo https://github.com/wasserth/TotalSegmentator?tab=readme-ov-file#resource-requirements, I found that T4 GPU with high RAM (>50 GB) was sufficient and the most efficient way. Using --body_seg or --force_split options keeps RAM and GPU RAM low but inference alone was taking about 8hrs.
While inference part was figured out, radiomics features are extracted sequentially for each label and each series was taking about an hour or ~12 hrs for a batch of 12 series. So the preprocessing notebook is edited to assign a single series to a VM instead of 12 series, for all series with more than 800 slices.
Lastly, the PR also includes the query for non-NLST cohort.
After exploring different options to handle processing large series (800+ slices) using recommendations from TotalSegmentator repo https://github.com/wasserth/TotalSegmentator?tab=readme-ov-file#resource-requirements, I found that T4 GPU with high RAM (>50 GB) was sufficient and the most efficient way. Using
--body_seg
or--force_split
options keeps RAM and GPU RAM low but inference alone was taking about 8hrs.While inference part was figured out, radiomics features are extracted sequentially for each label and each series was taking about an hour or ~12 hrs for a batch of 12 series. So the preprocessing notebook is edited to assign a single series to a VM instead of 12 series, for all series with more than 800 slices.
Lastly, the PR also includes the query for non-NLST cohort.