TABLE 3.

First cross-cohort evaluation.

Region of interest	Segmentation method	Dice score	Precision	Recall	Hausdorff distance (in voxels)
Left hippocampus	MRI U-Net	79.09 ± 2.63%	74.72 ± 4.27%	84.23 ± 3.15%	3.44 ± 0.74
	Cropped MRI U-Net	84.44 ± 2.32%	78.47 ± 4.17%	91.60 ± 2.47%	3.19 ± 0.64
	Shape MRI U-Net	84.92 ± 2.56%	79.46 ± 5.03%	91.57 ± 3.60%	3.16 ± 0.77
	Tissue MRI U-Net	84.32 ± 2.16%	79.04 ± 4.12%	90.59 ± 2.90%	3.33 ± 0.85
	Autocontext MRI U-Net	80.55 ± 2.61%	73.99 ± 4.23%	88.67 ± 3.50%	3.33 ± 0.72
	FreeSurfer 6.0	79.41 ± 3.77%	78.89 ± 5.46%	80.20 ± 4.39%	4.24 ± 1.25
Right hippocampus	MRI U-Net	80.15 ± 2.25%	74.54 ± 3.12%	86.80 ± 3.08%	3.92 ± 1.14
	Cropped MRI U-Net	82.85 ± 2.52%	75.97 ± 3.91%	91.27 ± 2.31%	3.80 ± 1.05
	Shape MRI U-Net	84.19 ± 2.50%	77.94 ± 4.49%	91.83 ± 3.28%	3.62 ± 1.04
	Tissue MRI U-Net	82.88 ± 2.35%	76.86 ± 3.60%	90.08 ± 2.71%	3.68 ± 1.11
	Autocontext MRI U-Net	80.51 ± 2.20%	73.08 ± 3.27%	89.79 ± 3.03%	3.88 ± 1.16
	FreeSurfer 6.0	79.57 ± 3.54%	77.71 ± 5.53%	81.78 ± 3.40%	4.61 ± 1.11

The performance of the proposed methods (in terms of Dice score, precision, recall, and Hausdorff distance) was tested on a new unseen dataset from a different cohort (i.e., AddNeuroMed cohort) than the one used for training. The performance is reported also for the segmentations obtained using FreeSurfer 6.0 on the same data. All evaluation metrics are expressed as mean ± standard deviation.