FastSurferCNN: segstats.py¶

segstats.py is a script that is equivalent to FreeSurfer’s mri_segstats. However, it is faster and (automatically) scales very well to multi-processing scenarios.

Full commandline interface of FastSurferCNN/segstats.py¶

Script to calculate partial volumes and other segmentation statistics of a segmentation file.

usage: python segstats.py (-norm|-pv) <input_norm> -i <input_seg>
                   -o <output_seg_stats> [optional arguments]
                   [{measures,mri_segstats} ...]

Named Arguments¶

--pvfile, -pv

Path to image used to compute the partial volume effects (default: the file passed as normfile). This file is required, either directly or indirectly via normfile.

-norm, --normfile

Path to biasfield-corrected image (the same image space as segmentation). This file is used to calculate intensity values. Also, if no pvfile is defined, it is used as pvfile. One of normfile or pvfile is required.

-i, --segfile

Segmentation file to read and use for evaluation (required).

-o, --segstatsfile

Path to output segstats file.

--excludeid

List of segmentation ids (integers) to exclude in analysis, e.g. –excludeid 0 1 10 (default: None).

Default: []

--ids

List of exclusive segmentation ids (integers) to use (default: all ids in –lut or all ids in image).

--merged_label

Add a ‘virtual’ label (first value) that is the combination of all following values, e.g. –merged_label 100 3 4 8 will compute the statistics for label 100 by aggregating labels 3, 4 and 8. With –robust, the values are dropped from the union of these labels.

Default: []

--robust

Whether to calculate robust segmentation metrics. This parameter expects the fraction of values to keep, e.g. –robust 0.95 will ignore the 2.5% smallest and the 2.5% largest values in the segmentation when calculating the statistics (default: no robust statistics == –robust 1.0). With –legacy_freesurfer, one value less is ignored at the top, like mri_segstats does.

--measure_only

Only calculate the Measures in the header, no PV table.

Default: False

Suboptions¶

subparser: Possible choices: measures

Advanced options (not shown in -h)¶

--threads

Number of threads to use (defaults to number of hardware threads: 4)

Default: 4

--patch_size

Patch size to use in calculating the partial volumes (default: 32).

Default: 32

--empty

Keep ids for the table that do not exist in the segmentation (default: drop).

Default: False

--device

Device to run inference on: auto (default), cpu, cuda (NVIDIA, or AMD with ROCm), a specific gpu (e.g. cuda:1), or mps (Apple silicon gpu). auto uses cuda if available, else mps, else cpu.

Default: 'auto'

--sid

The subject id to use, if not passed we try to extract the subject id from the path passed to –t1. For multi-subject processing, use –remove_suffix if sid is not the second to last element of input file passed to –t1.

--sd

Directory in which evaluation results should be written. Will be created if it does not exist.

--lut

Path and name of LUT to use.

--legacy_freesurfer

Reproduce FreeSurfer mri_segstats numbers (default: off). Please note, that exact agreement of numbers cannot be guaranteed, because the condition number of FreeSurfers algorithm (mri_segstats) combined with the fact that mri_segstats uses ‘float’ to measure the partial volume corrected volume. This yields differences of more than 60mm3 or 0.1% in large structures. This uniquely impacts highres images with more voxels (on the boundary) and smaller voxel sizes (volume per voxel).

Default: False

--mixing_coeff

Save the mixing coefficients (default: off).