Grey level-based pre-processing procedure with a hybridized semi-automatic technique for LS centered upon level-set techniques utilizing manifold seed points [41]. The proposed hybridized technique encompassed a modified Fast Marching Level Set (FMLS) and a Threshold Centered Level Set (TCLS). A customized FMLS was employed to detect an optimum initial liver region as of manifold seed points chosen by the user. A TLS was utilized to extract the actual liver centered upon the initial liver region. The hybridized technique was favored for LS in pre-operative virtual liver surgery planning. Nevertheless, the automatic techniques sacrificed the liver extraction’s accuracy as their algorithm can distinguish the liver from the neighboring organs based on the resemblance of image intensity betwixt the organs [42].

A similar work was done on hybridized watershed segmentation for LT diagnosis. The nose was eliminated by a median filter [43]. The watershed segmentation ameliorated the region, indicating the presence of the needed objects. A solitary intensity threshold was returned, separating pixels into ‘2’ classes: foreground and background. The Wavelet transform determined the threshold values. It was simple and took very little time. An effectual system for detecting LC was deemed. The input was considered in the image or video form, and the output was the same as the inputted image. However, the method encompassed an over-segmentation issue.

Lakshmi et al. [44] introduced Kernelized Fuzzy C-Mean (KFCM) clustering with an adaptive LT segmentation threshold. A succession of test images on MICCAI 2008 LT segmentation datasets was taken to extract the tumor area. The tumor area was gauged. To reduce the effect of noise and enhance clustering, the KFCM introduced a kernel function on FCM. The result exhibited positive outcomes for the algorithm. Segmenting the LC area was effective, and it achieved a high peak signal-to-noise ratio (PSNR), lower Mean Squared Error (MSE) values, and higher consistency for detecting smaller changes in images. Nevertheless, the technique was not appropriate for noisy images.

The FCM approach for CT and LT segmentation was proposed, which analyzes altered intensity values and higher frequencies eradicated utilizing histogram equalization and a median filter scheme to increase the contrast of liver CT images [45]. The approach centered on neutrosophy handled indeterminacy, better uncertainty, lessened over-segmentation, and better accuracy and performance on uninformed and noisy images. The accuracy attained by the suggested approach was nearly 95% better than LS. However, the system was tested with limited CT images.

The Geodesic Actives Contour algorithm was aimed at generating pre-processing of the LT images [46,47]. After pre-processing, the tumor’s Region of Interest (ROI) was determined. The abnormal region’s boundary in a single slice was signified. Next, a generative model processed the remaining slices and was ameliorated by incorporating a restraint. A probabilistic scheme was employed to search for the tumor boundary, and the solution was attained by Bayesian inference. The Kullback-Leiblers divergence was employed to measure the outcomes’ consistency with the model’s restraint. However, simple training datasets were employed.

A novel algorithm recommended for identifying a unified level set model for integrating image gradient, area competition, and prior information was intended for CT-LT segmentation [47]. Unsupervised fuzzy clustering helped to assess LT’s probabilistic distribution [48,49]. It was employed to ameliorate the object indication function, state the directional balloon force, and regulate region competition. An inclusive and flexible platform was provided for balancing various forces, such as the dynamic interface, the image in the examination, and other previous information. The unified level set model was an effectual solution for LT segmentation on contrast-ameliorated CT images. Nevertheless, the model encompassed restrictions for these images with lower contrast, weaker boundaries, and field homogeneity.

Work analyzing structure-based approaches develops Principal Components Analysis for automatic liver image pre-processing [48–50]. Based on the created atlas, the utmost probable liver area of the test image was additionally determined via a probability map’s posterior classification approach, resulting in rough segmentation. Shapes intensity before the level set produced the final LS. ‘25’ test CT datasets as of the partner site were considered. Its outcomes were contrasted with two top-notch LS methods. Nevertheless, a smaller dataset was utilized for training [51–53].

Image rendered with hierarchical local region-centered Sparse Shape Composition (SSC) for LS on CT scans was found to be significant and promising [54,55]. In the initial training step, to augment the flexibility of shape prior models and devotedly capture the comprehensive local shape information, a multi-level local region-centered SSC model termed Multi Linear Regression (MLR)-SSC was described. The liver shapes were disintegrated into manifold areas in a multi-level fashion. The segmentation work was more effective and robust to local minima. The method rendered an inferior initial shape, which brought about a larger segmentation error on the last outcomes.

Generic affine in-variance shape parameterized together with graph cuts intended for the segmentation utilizes training sets with regions of atypical local shape that were determined as a preliminary segmentation. The geodesics active contour locally corrected the organ segmentation on abnormal images [56–58]. Utilizing shape and amelioration constraints, optimized graph cuts were employed for segmenting the vasculature in addition to hepatic tumors. A significant reduction in the LS errors could be seen. All tumors were detected, and the trouble was anticipated with a 0.9% error. The technique’s robustness in examining livers as complex clinical cases to permit temporal monitoring of patients having hepatic cancer was demonstrated. Nevertheless, a high false-positive rate was present [59–61].

Image segmenting cascaded with Enhanced Deep Convolution Neural Network (EDCNN) aims at effective LT segmentation. For cascade segmenting of the liver in addition to lesions on CT images with restricted image quantity, deep EDCNN was built and trained. The liver image was segmented using the EDCNN and rendered for the EDCNN training [62]. Next, the EDCNN segmented the tumor areas on the liver ROI areas as envisaged utilizing the EDCNN. The false positives were significantly reduced by partitioning the hepatic tumor on the liver ROI. A public dataset was taken for testing. In addition, many metrics were utilized to quantitatively evaluate its performance. A DICE score of 95.22% was produced aimed at the test set of CT images. However, the system was trained with restricted datasets [63–65].

Li et al. [66] presented probability in addition to the local restraint level set model intended for LT segmentation as of CT volumes. The target’s Density Distributions (DD) and the multi-modal DD of the backdrop can minimize the probability energy approximated manifold areas. The ramp related to the edges for weak boundaries was preserved via the edge detector. The scheme to the Chans-Vese and the geodesics level set models and the manual segmentation performed by specialists were compared. Concerning segmenting the hepatictumors, the Chan-Vese model was not triumphant. The model trounced the geodesic level set model. However, the system was trained with restricted image features [67–69].

Another work with gradient ameliorated level-set centered segmentation is intended for LT as of CT Images [70]. It was truncated to keep the CT image intensity values in a fixed range to enhance the image contrast surrounding the liver and LT. These were done during the pre-processing step. Firstly, the Convolutional Neural Network (CNN) segmented the liver in a coarse-to-fine manner intended for removing non-liver tissues aimed at succeeding TS. For roughly localizing the liver [71,72], a 2D slice-centered U-net was utilized. 3D patch-centered full CNN was employed to refine the LS and roughly localize the LT. The better segmentation performance of the pipeline was identified over top-notch methods has illustrated in Fig. 3. Nevertheless, the system encompassed a high computational expense.

Lu et al. [73] formed a 3D CNN for automatic LS. In addition, a liver probability map was attained to generate a preliminary segmentation. Next, the learned probability map was incorporated into the graph cut energy function for additional segmentation alteration. The detiled image segmentation process has been illustated in the Fig. 4. It does not need any user interaction aimed at initialization, which was the main benefit. Therefore, non-experts could perform the method. In addition, it was an early endeavor to engage DL algorithms for 3D-LS. Nevertheless, poor classification accuracy was present.

Chi et al. [74] aimed at automatic liver vasculature segmentation on CT images centered on the vessel context. Voxels were grouped hierarchically to liver vasculature. Initially, they were locally grouped into vessel branches with the benefit of a vessel junction gauge. Next, utilizing manifold feature point voting mechanisms, they were globally grouped into vasculature. ‘10’ clinical CT datasets were estimated. The below 3 min are taken by segmenting 3rd-order vessel trees as CT images of the portal venous phase on a PC with a 2.0 GHz dual-core processor. The average segmentation accuracy was 98%. However, the system encompassed higher segmentation errors.

A fully Convolutional Network (FCN) was developed to address the LS task and detect liver metastases on CT examinations [75]. Although considering semantic segmentation, FCN was demonstrated to be an enormously powerful tool. A relatively smaller dataset was taken where the FCN performance was explored and contrasted with patch-centered CNN and sparsity-centered classification schemes. True positive rates of 0.86 and 0.6 false positives per case were achieved utilizing the fully automatic algorithm, resulting in propitious and clinically pertinent outcomes. Nevertheless, the system needed a huge dataset.

SVM was found to be highly influential for the automated liver as well as TS [35]. Utilizing training sets, the regions of a characteristic local shape were ascertained as of a liver’s initial segmentation. The liver segmentation in abnormal images was locally corrected by the geodesic active contour. Utilizing shape and enhancement constraints, graph cuts segmented the hepatic tumors. Significant reduction in the LS errors and also every tumor were detected. SVM and feature selection were utilized to reduce the total false tumor detection. The tumor trouble was anticipated with a 0.9% error. Nevertheless, the technique was computationally costly.

SVM aims at automated LT detection [36]. An automated CAD of LT as of CT images was generated. Markov Random Fields (MRF) embedded level set technique segmented the liver. Robustness was rendered to noise as well as fast segmentation. Shape analysis techniques found the shape ambiguities of the liver (segmented), which utilized a training set intended for correction. The graph cut technique detected hepatic tumors as of the corrected LS. FE was employed to classify them utilizing an SVM. However, the system encompassed poor classification accuracy.

Das et al. [37] presented a fuzzy clustering with a DT classifier for LC detection. With adaptive thresholding, the liver and other body parts were separated initially. Next, with spatial fuzzy clustering, the cancer-affected lesions of the liver were segmented. As for segmented cancerous areas, the informative features were extracted. It was classified into Hepato Cellular Carcinoma (HCC) and MET utilizing MLP and C4.5 DT classifiers. An effectual approach intended to automatically recognize LC was the SFCM-centered segmentation with a C4.5 DT classifier. Nevertheless, a lower recognition rate was obtained.

Huang et al. [38] rendered an arbitrary feature subspace ensemble-centered Extreme Learning Machine (ELM) intended for LT detection and segmentation. The ELM autoencoder was executed as a pre-training step to increase testing accuracy. ELM was trained as a 1-class classifier with merely healthy liver samples in automatic LT detection. The performance was contrasted with ‘2’-class ELM. A semi-automatic approach was utilized to train the classifier to choose samples in 3D space in LT segmentation. The technique was tested and evaluated using a cluster of patients’ CT data, and the experiment showed propitious outcomes. Nevertheless, the segmentation process was tedious.

Kernel-centered ELM for LT detection together with segmentation trained as a 1-class classifier with merely healthy liver samples in training [39]. It brought about a technique of tumor detection centered upon novelty detection. It was contrasted with ‘2’-class ELM. The semi-automatic approach was adopted using arbitrarily choosing samples in 3D space inside a restricted ROI intended for classifier training to extract the tumor boundary. A cluster of patients’ CT data was considered. The experimentation showed better detection and encouraged segmentation outcomes. However, it limited its applications to higher dimensional and larger data.

DL models are utilized to diagnose the LT by disparate image acquisitions with the progression of AI. Disparate works are done for lever segmentation and LT detection utilizing DL approaches. The DL techniques for LT detection are listed in Table 2. The summary in Table 2 outlines several techniques that provide insight into the range of accuracy reported in several methods discussed in the study. Image pre-processing is crucial for a precise outcome of succeeding steps. Generally, noise elimination, contrast enhancement, and edge enhancement are the ‘3’ kinds of pre-processing. As the Wiener filter can polish the boundary and conserve the image’s interior information well, it was utilized by the models to perform image pre-processing. The cumulative distribution function was utilized by histogram equalization to normalize image intensity, thereby elevating image brightness. The pre-processing schemes utilized in the above models executed comparatively well in network training for automated LS.

The entire ‘3’ sorts of LS methods have their advantages. Generally, to handle the complex segmentation issue, the GL-centered models are often utilized together. Using prior knowledge, the structure-centered models handled the liver’s unclear boundary. This means that some issues that the GL-based methods cannot handle can be dealt with through this model. Owing to the capacity to resolve computer vision tasks, a primary focus for the research community in previous years is deep convolutional NN [74–77]. Vital attributes for detecting tumor regions as of the image are termed Features. A statistical approach and a structured approach are the two ways it is done. Major researchers utilized a statistical approach. Co-occurrence matrix, Fractals, Gabor filters, and wavelet transform were the several statistical techniques for gauging texture. GLCM, utilized by major researchers, captures numerical feature values by utilizing spatial relationships among neighborhood pixel features. Several autonomous approaches for brain tumor detection, namely NN and SVM, have been popular in the prior decades. As DL can signify complicated structures, own-learning, and process enormous MRI-centered image data efficiently, it has reached a central tract regarding automation of Brain tumor detection. It can present the current trends and achievements in ML and DL techniques as shown in Fig. 6. Due to their automated FE techniques, DL models have been identified as being able to execute better and make the system more effective and intelligent [78].