Automatic Liver Lesion Segmentation in CT Combining Fully Convolutional Networks and Non-negative Matrix Factorization

Automated liver tissues delineation based on machine learning techniques: A survey, current trends and future orientations

“…RMSD is significantly related to ASD definition described by equation (11). The root-mean-square of ASD is the summation of distances squared under the square-root as defined by equation (12).…”

“…In [74], cascaded 2D FCN (CFCN) is used for liver segmentation, where the first FCN coarsely segments the liver and the second one refines it. In [12], another 2D FCN is used for the liver segmentation followed by a 3D deformable model optimization (3D DMO), based on local cumulative spectral histograms and non-negative matrix factorization (NMF). In [75], a 2D multi-channel FCN (MC-FCN) takes six slices as an input from multi-phase MRI imagery, where the used structure outperforms the U-Net on the utilized dataset.…”

“…Some of the aforementioned liver segmentation work use the same model for tumor segmentation as well. For example, the 2D FCN aided by the 3D DMO and NMF in [12], the 2D FCN in [73], the 2D FCN VA Mask-RCNN in [96], the DeepLabV3 followed by Pix2Pix GAN in [97], the ensemble of the three U-Net-like 2D FCN in [44,85], the mU-Net in [86], the CU-ResNets [93], the 3D AH-Net in [59], the 3D U-Net with context-aware modules in [52], and the H-DenseUNet in [42], all segment the tumors along with the liver simultaneously. In [58], the first 2D CompNet, mentioned earlier in the liver segmentation techniques, also aids in segmenting large tumors, followed by a 3D CompNet to segment the small ones.…”

Automated liver tissues delineation techniques: A systematic survey on machine learning current trends and future orientations

“…However, the norm currently in clinical routines is to manually or semi-automatically segment the liver from CT and MRI modalities. Even though, in some scenarios, these techniques can be more accurate than automatic ones (Zheng et al, 2017a), the underlying issues of manual and semi-automatic techniques are represented by their subjectivity (i.e., dependency on the radiologists' experience), intraand inter-radiologist variance, and time-consumption (Hu et al, 2016), especially for experts whose time is extremely valuable. Thus, comes the importance of using automatic methods with high segmentation performance.…”

“…They can be categorized into statisticalbased and learning-based approaches, where the former can be represented by scans intensities' statistical distribution, including atlases, statistical shape models (SSM), active shape models (ASM), level-set methods (LSM), and graph-cut (GC) methods (Yang et al, 2017). Usually, these methods are challenged by boundary leakage and under-or over-segmentation (Zheng et al, 2017a). On the other hand, the latter depends on either hand-crafted features as in conventional machine learning (ML) algorithms or empirically-found features as in the case of convolutional neural network (CNN), which is a special structure of the artificial neural network (ANN).…”