Table 4.
Research on abdomen segmentation.
| Authors (Year) | Method | Medical Image | Performance | Notes |
|---|---|---|---|---|
| Abdominal organ | Differs between organs | |||
| Landman et al. [85] (2018) | FCN | Magnetic resonance images | DSC 1: 0.56–0.93 | |
| Gibson et al. [86] (2018) | Dense V-Networks | Computed tomography | DC 2: 0.76–0.9 | High-resolution activation maps; batch-wise spatial dropout |
| Kim et al. [73] (2020) | 3D U-Net Atlas-based |
Computed tomography | DC: 0.60–0.96 DC: 0.15–0.81 |
Multi-organs were tested; the U-Net result could be comparable to that of an interobserver |
| Kart et al. [74] (2021) | nn-UNet | Magnetic resonance images | DC: 0.82–0.9 | |
| Abdominal tumor | ||||
| Abdel-Massieh et al. [87] (2010) | CV 3 method | Computed tomography | Overlap error: 0.22 | Gaussian blurring; isodata threshold |
| Abd-Elaziz et al. [88] (2014) | CV method | Computed tomography | Error rate: 0.002–0.012 | Regional pixel growing and morphological processing |
| Manjunath et al. [75] (2021) | ResUNet | Computed tomography | DSC: 0.96 | Replacing convolutional blocks with residual blocks |
| Vorontsov et al. [76] (2019) | FCN | Computed tomography | DSC per lesion: (automated) <10 mm: 0.14 10–20mm: 0.53 >20 mm: 0.68 |
Two-step segmentation: the first is FCN for livers, and the second is FCN for lesions in livers |
| Liang et al. [77] (2020) | Square-window based CNN | Magnetic resonance images | DSC: 0.73 on the test set | |
| Pellicer-Valero et al. [78] (2021) | Retina U-Net | Magnetic resonance images | DSC: (prostate) 0.915 | Two 3D CNNs: the first one takes a T2-weighted MRI as the input, and the second one takes an MRI and the output from the first one as inputs |
| Chen et al. [79] (2020) | 3D AlexNet | Magnetic resonance images | DSC: 0.97 | |
| Li et al. [83] (2020) | MRBS-U-Net | Endoscopic ultrasound | DSC: 0.92 |
1 Dice similarity coefficient, 2 Dice coefficient, 3 computer vision.