Skip to main content
. 2025 Jan 7;20(3):465–473. doi: 10.1007/s11548-024-03310-z

Table 1.

Cross-validation of point segmentation networks compared to nnU-Net

Model KPs ASSD [mm] SDSD [mm] HD [mm] n.a. [#]
PointNet [4] Förstner 5.96 ± 0.65 4.80 ± 0.40 25.86 ± 1.75 0
CNN 3.63 ± 0.64 3.24 ± 0.38 20.60 ± 1.80 1
DGCNN [5] Förstner 3.54 ± 0.47 3.24 ± 0.40 20.40 ± 2.04 0
CNN 3.07 ± 0.67 2.85 ± 0.38 18.37 ± 1.57 1
Point-Transformer [6] Förstner 3.25 ± 0.45 2.95 ± 0.34 17.52 ± 1.52 0
CNN 3.01 ± 0.62 2.83 ± 0.36 18.18 ± 1.54 2
nnU-Net [2] 2.27 ± 0.85 2.50 ± 0.40 16.62 ± 1.71 3

Bold denotes the best overall result

ASSD average symmetric surface distance, SDSD standard deviation of surface distances, HD Hausdorff distance.

Mesh reconstruction and surface distance computation are impossible when no keypoints are segmented for an object. We report these cases as the total number of non-assigned (n.a.) fissures