. 2023 Sep 21;36(6):2532–2553. doi: 10.1007/s10278-023-00854-5

Table 1.

Qualitative comparison of prior hippocampus segmentation studies

Key Method	Dataset	Modalities	Training Strategies	Average Dice(%)	Year
software data synthesis and multiple cohorts transferring with 3D CNN [14]	ADNI(135), HCP(547), AOBA(317), TRGY(112), OASIS(58), ABIDE(197)	3D T1 MRI	3D volumes (48 $\times$ 72 $\times$ 64)	$D S C_{ADNI}$ =85 $D S C_{HCP}$ =83 (left/right=82/84)	2018
L-CRRF with dense CRFs [15]	10 infant brain from UNCCH	3D T1w MRI (192 $\times$ 156 $\times$ 144)	3D volumes (13 $\times$ 13 $\times$ 13)	$D S C_{avg}$ =65.3 $D S C_{2 - w e e k}$ =58.8 $D S C_{3 - m o n t h}$ =63.6 $D S C_{6 - m o n t h}$ =66.1 $D S C_{9 - m o n t h}$ =68.0 $D S C_{12 - m o n t h}$ =67.0	2020
structured random forest with auto-context model [16]	8 private healthy subjects and 4 healthy subjects randomly from HCP	3T MRI(T1,T2), rs-fMRI	3D volumes (11 $\times$ 11 $\times$ 11)	$D S C_{avg}$ =69.0	2018
multimodel deep CNN jointly with 3D DenseNet [32]	ADNI(449)	3D T1w MRI	3D volumes (64 $\times$ 48 $\times$ 64)	DSC=87.0	2020
Tweaked U-Net with three alternative kernels of sizes 1 $\times$ 1, 3 $\times$ 3, and 5 $\times$ 5 [34]	ADNI (210)	3D T1w MRI	2D slices (256 $\times$ 256 $\times$ 1)	DSC=96.5	2022
nnU-Net with ViT [35]	HarP(270), Dryad(50), DecathHip(260)	3D T1w MRI	2D slices	$D S C_{DecathHip}$ =89.8 $D S C_{HarP}$ =+4.8 $D S C_{dryad}$ =+16.2	2022
GAN with modified U-Net generator [44]	Brain images from CIND center (MCI: 4, AD: 7, normal: 21)	3D T1w MRI, 3D T2w MRI	2D slices (128 $\times$ 128)	$D S C_{avg}$ =85.2	2019
dual-branch with improved SSA adapters [38]	MSD (Heart:20, Hippocampus:260, Pancreas:281, Spleen:41)	MRI, CT	2D slices, 3D patch	$D S C_{heart}$ =91.9 $D S C_{hipp}$ =69.9 $D S C_{pancreas}$ =74.5 $D S C_{spleen}$ =83.7	2022
GAN with 3D CNN generator and SVM discriminator [45]	Brain images from CIND center (MCI: 4, AD: 7, normal: 21)	3D T1w MRI, 3D T2w MRI	3D patch (128 $\times$ 128 $\times$ 128)	$D S C_{avg}$ =96.5	2020