Volume discrepancy between scilpy and tractometry_flow scripts

[cbedetti]

Hi,

I'm running scil_evaluate_bundles_individual_measures.py, scil_compute_bundle_volume.py and the last version of tractometry_flow (a42d1d7) inside a scilpy 1.6.0 container.

I tested it with the AF_R bundle from HCP 100206.

• scil_evaluate_bundles_individual_measures.py

"volume": [
    70150.390625
  ],

• scil_compute_bundle_volume.py

{
  "100206__AF_R": {
    "volume": 70150.390625
  }
}

• tractometry_flow

"100206": {
    "AF_R": {
      "volume": 69072.265625
    }
  }

Is there a reason for this difference?

[frheault]

Hi,

Yes, it is normal. To increase consistency, code in the script scripts/scil_compute_bundle_voxel_label_map.py ensures that a bundle is contiguous (no isolated voxels in the binary mask). Due to edge cases in the computation (in low-density areas, close to endpoints in fanning regions), some voxels are considered isolated (not touching the main bundle).

This result is mostly explainable because the function ndi.label uses a flood fill based using a cross structure (6 neighbors) rather than a ball structure (26 neighbors). @arnaudbore what do you think?

    bundle_disjoint, _ = ndi.label(binary_bundle)
    unique, count = np.unique(bundle_disjoint, return_counts=True)
    val = unique[np.argmax(count[1:])+1]
    binary_bundle[bundle_disjoint != val] = 0