Inference time for automatic GM segmentation

[Nilser3]

Description

This issue to explore the inference time taken by GM segmentation models.

Models to test

1. sct_deepseg_gm SCT v. 6.5
2. seg_gm_contrast_agnostic release: r20250204

• The seg_gm_contrast_agnostic inference was performed through sct_deepseg (SCT branch : nlm/add_gm_contrast_agnostic_model , commit : 311307e24ae4f9bebd98574569294ab93f45ebd3) and not through nnUNetv2_predict

Dataset to test

The split test contains 233 volumes, 7 contrasts, at different dimensions and from different sites (see: #2 (comment))

Computer resource

The tests were performed on a CPU : Intel(R) Xeon(R) CPU E5-2640 v4 @ 2.40GHz , 10 Cores using codes/compute_inference_time.sh script.

Results

• Since both methods are 2D segmentation models, it was convenient to make a figure showing the inference time according to the number of 2D slices of each volume.
• It is observed that sct_deepseg_gm is in general 5 times faster than seg_gm_contrast_agnostic, and that for volume with a large number of 2D axial slices (i.e. marseille-7T-MP2RAGE), the segmentation can take up to more than one minute per volume.

Related issues

#2

[jcohenadad]

Great figure Nilser!