Table 3 Summary of available methods for liver tumor segmentation
From: Practical utility of liver segmentation methods in clinical surgeries and interventions
Reference | Method | Dataset | Performance |
|---|---|---|---|
Lin et al. [92] | Lucas-Kanade algorithm is used for discriminative training, followed by inference algorithm, which employs Lagrangian method and image sequence matching | LiTs | Accuracy: 0.8561 (SYSU-CT), 0.6571 (SYSU-US) |
et al. [88] | 2D-Slice Based U-Net and 3D Patch-Based CNN are employed for segmentation of liver and localization of tumor. Level-set method is used for tumor refinement | LiTs | Liver Dice: 96.31% ± 0.62% Liver RMSD: 1.99 mm ± 0.64 mm Tumor Dice: 72.45% ± 13.42% Tumor RMSD: 4.99 mm ± 2.18 mm |
Xi et al. [89] | Two Cascading U-ResNets for liver and tumor segmentation with a experimental study for measuring the impact of loss functions | LiTs | Liver Dice: 94.9% Liver VOE: 0.0095 Tumor Dice: 75.2% Tumor VOE: 0.379 |
Jiang et al. [93] | Cascaded Attention Hybrid Connection Network with a combination of soft and hard attention for liver and tumor segmentation | Training set: LiTS Test set: 3DIRCADb (20 patients), Clinical Dataset (117 cases) | 0.62 ± 0.07 (DSC) |
Seo et al. [94] | Modified U-Net (mU-Net) architecture with the residual path deconvolution over the skip-connections to prevent duplication of low-resolution information | LiTs | Liver Dice: 98.51% Liver VOE: 3.07% Tumor Dice: 89.72% Tumor VOE: 21.93% |
Vivanti et al. [95] | CNN trained with delineation of baseline CT scans and evaluated on follow up CT studies | Custom Dataset (67 Tumor in 21 scans) | VOE: 16.26% |
Bai et al. [90] | Multi-scale candidate generation method (MCG), 3D fractal residual network (3D FRN), and active contour model (ACM) are used in a coarse-to-fine manner for liver tumor segmentation | Training set: LiTS Test set: 3DIRCADb | Tumor Dice: 0.67 Tumor VOE: 0.324 Tumor MSD: 7.113 mm |