Error in parcellated volumes prior to ca7ab2f

[cookpa]

Hi @BenBillot

Thanks for this amazing tool! I was looking into some discrepancies between the parcellated cortical volumes between the latest SynthSeg and some earlier versions.

EDIT: To be clear, I'm talking about the volumes output with --vol, not the labels themselves.

I think I found a bug that was fixed in commit ca7ab2f but I'm not sure it has been documented.

I believe the error was in this block

SynthSeg/SynthSeg/predict_synthseg.py

Lines 750 to 759 in 0369118

	# compute volumes
	volumes = np.sum(posteriors[..., 1:], axis=tuple(range(0, len(posteriors.shape) - 1)))
	if not v1:
	volumes = np.concatenate([np.array([np.sum(volumes)]), volumes])
	if post_patch_parc is not None:
	volumes_parc = np.sum(post_patch_parc[..., 1:], axis=tuple(range(0, len(posteriors.shape) - 1)))
	total_volume_cortex = np.sum(volumes[np.where((labels_segmentation == 3) | (labels_segmentation == 42))[0] - 1])
	volumes_parc = volumes_parc / np.sum(volumes_parc) * total_volume_cortex
	volumes = np.concatenate([volumes, volumes_parc])
	volumes = np.around(volumes * np.prod(im_res), 3)

Specifically, we have

    if not v1:
        volumes = np.concatenate([np.array([np.sum(volumes)]), volumes])

which inserts a new element into volumes. But then

        total_volume_cortex = np.sum(volumes[np.where((labels_segmentation == 3) | (labels_segmentation == 42))[0] - 1])

Only now labels 3 and 42 are not total cerebral cortex, they are total cerebral white matter.

In the change to per-hemisphere normalization, this problem is circumvented by computing the per-hemisphere totals before doing

    if not v1:
        volumes = np.concatenate([np.array([np.sum(volumes)]), volumes])

And so the parcellated volumes for [hemi] sum to the "[hemi] cerebral cortex" number in the CSV.

Prior to this, the total sum of all parcellated volumes (left and right) sum up to ("left cerebral white matter" + "right cerebral white matter").

I just wanted to make a note in case anyone else is confused by the changes, or has older output they may want to correct.