Skip to content

Latest commit

 

History

History
43 lines (27 loc) · 3.92 KB

faq.md

File metadata and controls

43 lines (27 loc) · 3.92 KB
Raw

FAQ

What input MRI is required?

Requirements are similar to FreeSurfer: High quality 3D T1-weighted MRI with 1mm isotropic resolution acquired at 3T or 1.5T. Other resolutions are automatically re-sampled to 1mm isovoxels (inspect the results in this case). Recommended sequences are those with an excellent gray/white matter contrast like Siemens MP-RAGE (ideally with TI=1100ms) as proposed by ADNI, or MDEFT. Similar sequences should work as well, just try it out.

Note: MR sequence parameters have an influence on the GM/WM contrast which can affect the morphometric measures (see Rebsamen et al., 2023).

DL+DiReCT expects the input without skull (a.k.a. brain extracted / skull-stripped). You may use the --bet options to skull-strip the input using HD-BET before processing. If the input is already skull-stripped, ensure background voxels are zero.

Can I use an MP2RAGE sequence as input?

In principal yes, DL+DiReCT should work fine on a (brain extracted) MP2RAGE sequence. However, the characteristic "salt & pepper" noise that is a result of numerical instability that amplifies noise in areas with a very low SNR like background may require special treatment for brain extraction. For brain extraction using HD-BET, you may use the the 2nd inversion recovery (proton density-weighted) image from an MP2RAGE sequence to generate the brain mask which is then applied to the original unified image. This is available in dl+direct with the options

--bet --mp2rage-inv2 <path_to_inv2.nii.gz>

Note: A de-noised image can also be calculated by multiplying the unified image with the 2nd inversion recovery (Fujimoto et al., 2014). This alternative (bet.py --mp2rage-inv2x) might work for other brain extraction algorithms but seems not optimal for HD-BET. Visual inspection of the brain mask is recommended. DL+DiReCT should be run on the original (unified) image with the brain mask applied and not on the de-noised version.

What is the accuracy of DL+DiReCT?

Evaluating morphometry methods is difficult due to the lack of a gold-standard reference. In the related publications (1, 2, 3), experimental assessment of robustness and sensitivity yielded good results.

A benchmarking study (Rusak et al., 2022) quantified DL+DiReCT to be sensitive enough to measure subtle atrophy rates as low as approx. 0.01 mm.

How to extract radiomics shape features?

You may derive radiomics shape features of subcortical structures from the segmentations (see Rebsamen et al., 2022). Given results in the directory <output_dir>, run:

radiomics_extractor --results_csv dl_radiomics.csv <output_dir>

What are the system requirements to run DL+DiReCT?

A linux environment with a GPU (at least 8 GB memory) and approx. 16 GB free memory is recommended. If no GPU is available, it will fallback to the CPU and processing time increases. For environments with low memory, you may use the option --lowmem to reduce RAM consumption (might slightly alter the results).

Is there a batch mode to process multiple MRI?

You may use batch-dl+direct to process multiple subjects in parallel by distributing jobs on CPU cores and GPUs. Requires GNU parallel (sudo apt install parallel). Put the input data in a directory rawdata with a subdirectory for each subject and the T1.nii.gz inside. Example (using 16 cores and two jobs on the GPU):

batch-dl+direct --cpu 16 --gpu 2 --bet rawdata DL+DiReCT_results