Selection of proper artificial intelligence techniques developed for CT scan image analysis of liver cancer using fuzzy AHP-TOPSIS Authors: Geng Lu, Yuheng Wang, Chunying Li, Cong Wu Published: 2026 ABSTRACT ObjectiveComprehensive identification and prioritization of developed artificial intelligence methods for applications of radiology can help to select proper techniques. This study aimed to introduce the best artificial intelligence techniques developed for CT scan image analysis of liver cancer using fuzzy AHP-TOPSIS.Materials and methodsTo identify the artificial intelligence techniques developed, a systematic search was performed in five reliable databases. The developed methods were categorized into four groups based on their application type. After that, the Delphi method was applied in two rounds to determine the proper criteria for selecting the best artificial intelligence techniques. To estimate the relative weights of the criteria also, the fuzzy analytical hierarchy process (FAHP) method was used. In the next step, to prioritize the identified artificial intelligence techniques, the technique for order of preference by similarity to the ideal solution (TOPSIS) method was applied.Results260 artificial intelligence techniques were identified. Seven selection criteria of validity, accuracy, comprehensiveness, processing time, cost, simplicity, and executive capability were introduced. The deep learning-based image reconstruction model (weight = 0.896) in the group of detection and diagnosis, dual-energy CT deep learning radiomics model (weight = 0.862) in the group of prediction, prognosis, and registration, hybrid densely connected UNet (H-DenseUNet) technique (weight = 0.888) in the group of segmentation and classification, and deep learning, radiomics, and clinical (DLRC) model (weight = 0.956) in the group of treatment and therapy were selected as the proper methods.ConclusionsThese findings represent a total approach to the developed techniques which can be used for designing methods with better performance in the future.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12880-026-02506-w. Introduction Liver cancer remains a major global health burden [1]. It is introduced as the second cause of death due to cancer in the world, and its incidence and mortality are steadily increasing [1]. Tumors can be created in the liver and metastasize to the liver from other organs, like the colon, rectum, breast, and lung. A wide range of neoplasms, including hepatocellular carcinoma (HCC) and intrahepatic cholangiocarcinoma, are considered liver cancers [2]. Early diagnosis of this disease helps to prevent further liver damage. Computed Tomography (CT), as a non-invasive and flexible method, is recognized as an effective diagnostic instrument for detecting hepatic abnormalities [3]. Manual analysis of CT scan images by physicians is a time-consuming process because they must meticulously examine the sequence of slices in both directions for accurate identification of the lesions [4]. Furthermore, this classification process is inherently subjective, and diagnosis greatly depends on the expertise level of the physician [4]. Moreover, approximately two trillion medical images are annually produced in the world [5]. While there are not enough radiologists in numerous nations, particularly in developing countries [5]. An approach for solving these issues is the use of artificial intelligence tools, which provide a precise diagnosis and reduce the workload of radiologists [6]. In addition, the complexity of liver tumor imaging due to diverse tumor types, like hepatocellular carcinoma and cholangiocarcinoma, variations in tissue characteristics, and overlapping radiographic appearances, has driven significant research in automated diagnostic methods [7]. Artificial intelligence (AI) techniques have been widely studied for tasks such as lesion detection, segmentation, classification, and characterization in imaging [8, 9]. These techniques, as a useful product of computer science, are a pivotal element in improving clinical diagnosis [10]. It provides new possibilities in the healthcare field by incorporating novel tools [11]. In the area of oncology, AI has demonstrated its ability to influence multiple aspects of cancer treatment. Also, it has increased the precision and speed of diagnoses and facilitated more dependable clinical judgments [12]. AI has achieved levels of predictive accuracy that are better than those of traditional statistical experts for cancer and other diseases [13–15]. The accurate and rapid analysis of computed tomography (CT) scans is crucial for the diagnosis, classification, and treatment planning of liver cancer. In recent years, artificial intelligence (AI) has emerged as a transformative technology in this domain, offering powerful tools for automating and improving image interpretation. A significant body of research, particularly between 2018 and 2024, has focused on the use of various AI methodologies to liver cancer CT imaging [7, 16–20]. Deep Learning (DL) models, notably convolutional neural networks (CNNs) and their variants like U-Net for segmentation tasks, have demonstrated remarkable performance in the diagnosis of malignant nodules, classification of tumor types, and segmentation of lesions with precision that often rivals or surpasses human [7, 17, 19]. For instance, the results of the studies performed by Wei et al. [19] and Nisa et al. [20] have shown the efficacy of deep learning in early hepatocellular carcinoma detection from CT scans. Radiomics, which involves the extraction of several quantitative features from medical images, has also seen substantial integration with AI, especially DL, for tasks such as risk stratification, treatment response prediction, and prognosis in liver cancer [21]. Mokrane et al. highlighted how radiomic features, when analyzed by machine learning algorithms, can provide deeper insights into tumor heterogeneity and patient outcomes [22]. Systematic reviews have further underscored the breadth of AI’s impact, categorizing AI applications in liver cancer CT and MRI for diagnosis, classification, segmentation, and prediction [7, 23, 24]. These reviews frequently report the high accuracy rates achieved by AI models, often exceeding 90% in specific datasets [7, 23, 24]. However, as a critical gap in the literature, there is no structured, comprehensive, and comparative framework to select the optimal AI technique for CT analysis tasks of specific liver cancer. While many studies report the performance of individual AI models on specific datasets, they often lack direct comparison with other state-of-the-art techniques under standardized conditions. This makes it challenging for researchers and clinicians to determine which AI approach is most appropriate, given the multifaceted evaluation. Criteria such as model interpretability, computational cost, data requirements, generalization capabilities across diverse populations, are often overlooked in direct performance comparisons [25]. Multi-Criteria Decision-Making (MCDM) methods can solve this selection challenge. In these methods, multiple criteria were applied for the selection of the best items [26]. Relative weights of these criteria are determined based on their importance degree. Moreover, items were scored and finally prioritized by these criteria [27]. Among the available methods, classical AHP provides structured criteria weighting but does not handle uncertainty in expert judgments, while TOPSIS offers intuitive ranking based on closeness to ideal solutions but lacks inherent fuzziness handling. Other frameworks, such as SAW, VIKOR, ELECTRE, or PROMETHEE, either simplify ranking at the cost of accuracy or involve complex parameters that reduce interpretability in clinical contexts. Fuzzy AHP is applied to weight the selection criteria and TOPSIS is used to prioritize the items based on the weighted criteria [28]. By integrating these techniques, a robust, interpretable, and reliable hybrid framework is provided, particularly suitable for evaluating AI methods with multiple heterogeneous performance criteria [28]. Piegat and Salabun performed a comparative analysis of MCDM methods for assessing the severity of chronic liver disease [29]. Jana et al. also applied decision-making in selecting diagnostic methods for early-stage cancer using a hybrid MCDM-fuzzy approach [30]. Moreover, Mahmood et al. used multi-criteria decision making for the evaluation and ranking of cancer information [31]. However, none of these studies used multi-criteria decision-making methods to select the best artificial intelligence techniques developed for CT scan image analysis of liver cancer. In total, artificial intelligence can facilitate human work and decrease some human errors, particularly in medicine. For this purpose, many techniques have been developed in different areas. However, the high number of these methods makes it difficult for the user to choose the best and most suitable ones. Comprehensive identification and prioritization of developed artificial intelligence methods for CT scan image analysis of liver cancer can help to select proper techniques. Finally, liver cancer imaging presents practical and clinical relevance for decision support systems because CT is widely used in clinical practice to evaluate liver lesions due to its high spatial resolution and accessibility. Taken together, the clinical urgency of liver cancer diagnosis and the extensive body of existing AI research in this domain justify its selection as the application focus for our comparative evaluation of AI techniques using a multi-criteria decision-making framework. Regarding the importance of this issue, this study aimed to introduce the best artificial intelligence techniques developed for CT scan image analysis of liver cancer using fuzzy AHP-TOPSIS. The fuzzy AHP–TOPSIS framework provides a practical method for evaluating and prioritizing AI techniques for liver cancer CT analysis. By weighting and ranking seven criteria, it translates complex performance data into clear, interpretable results. Fuzzy AHP captures expert judgments on selection criteria under uncertainty, while TOPSIS ranks techniques relative to an ideal solution. Clinicians can use the framework to identify AI tools that balance accuracy and efficiency for routine workflow integration; radiologists can assess usability and segmentation quality; and researchers can systematically compare emerging models, standardize evaluations, and guide further development. By combining transparency, adaptability, and rigorous multi-criteria assessment, this framework enables informed, evidence-based decisions and practical adoption of AI techniques in both clinical and research settings. Materials and methods This study, as descriptive-analytical and retrospective research, was conducted in four stages as follows. The goal of this study was to determine the best artificial intelligence techniques developed for CT scan image analysis of liver cancer using fuzzy AHP-TOPSIS. Identification of relevant technique To identify the artificial intelligence techniques developed for CT scan image analysis of liver cancer, a systematic search was performed in five reliable databases (ISI, PubMed, Scopus, Medline, and Embase), up to 15 April 2024. This search utilized combinations of three sets of keywords found in the titles, abstracts, and keywords sections of the articles. The first group of keywords included “liver cancer*” OR “liver neoplasm*” OR “liver tumor*” OR “liver tumour*” OR “liver malignan*” OR “hepatic tumor*” OR “hepatic tumour*” OR “hepatic malignan*” OR “Liver lesion*” OR “liver nodule*” OR “hepatic nodule*” OR “hepatic neoplasm*” OR “hepatic Cancer*”. The second group of keywords was “AI” OR “Artificial Intelligence” OR “Machine Intelligence” OR “Computational Intelligence” OR “Computer-aided” OR “Computer-Assisted” OR “image processing” OR “image analysis” OR “Machine Learning” OR “deep learning”. The third group of keywords was “CT” OR “Computed Tomography” OR “computerized tomography”. Search strategy is available in Table S1. As inclusion criteria, the systematic review included all research that developed artificial intelligence techniques developed for CT scan image analysis of liver cancer. It encompassed the different types of studies published in English. However, this review excluded review articles, meta-analyses, conference papers, editorial letters, book chapters, notes, short surveys, and data papers. Additionally, as exclusion criteria, studies on other diseases except liver cancer, on other tools except CT scans, and in other languages except English were removed. Classification of identified techniques In this step, first, the papers for the development of the novel techniques were studied, and the application type of them was determined. Then, the developed methods were categorized into four groups based on their application type. Determination of criteria relative weights At this stage, firstly, the Delphi method was applied in two rounds to determine the proper criteria for selecting the best artificial intelligence techniques [32]. To estimate the relative weights of the criteria, the fuzzy analytical hierarchy process (FAHP) method was used. In the first round of the Delphi method, experts were asked to state the appropriate criteria in selecting proper AI methods developed for CT Scan image analysis of liver cancer. This information was then collected and summarized. The list of criteria extracted from the first round was then sent to the experts in the second round, and they were asked to add to the list if they had any other criteria in mind or to remove a criterion from the list if it was inappropriate. After that, the information from the second round was also collected and summarized. Operational definitions for each criterion are provided in the results section to ensure their clarity. In the FAHP method, firstly, a hierarchical structure was developed, including the main goal and identified criteria. Expert judgments were then collected through pairwise comparisons using linguistic terms, which were converted into triangular fuzzy numbers to account for ambiguity in human perception. Fuzzy synthetic extent values were calculated, and the resulting fuzzy weights were defuzzified and normalized to obtain crisp priority weights for each criterion by the fuzzy method of Chang et al. [33]. In the FAHP method, pairwise comparisons were completed by 80 experts. These participants included 22 professors, 27 associate professors, 24 assistant professors, and 7 PhD candidates with research experience in the application of artificial intelligence in medical sciences, including the prediction, prevention, diagnosis, and treatment of liver cancer from various countries. To select these individuals, the research team first compiled a list of 150 experts with relevant expertise by searching Google Scholar. This list was then evaluated in terms of inclusion and exclusion criteria. The inclusion criteria required experts to have research experience in artificial intelligence applied to medical sciences and to have published at least five articles in this field within the last three years. The exclusion criteria eliminated authors who had contributed to papers already included in the study to avoid bias. Based on these criteria, 125 individuals were selected to participate. These individuals were invited to join the study via email, which explained the study’s purpose and stages. Ninety-two people accepted the invitation, and ultimately, 80 completed the questionnaires. These people were asked to compare the criteria two by two by linguistic vocabulary based on a five-point Likert scale from equally important to extremely important. In the next step, the linguistic vocabulary was transformed into triangular fuzzy numbers. Triangular fuzzy membership functions were adopted to represent expert judgments in the FAHP process because they effectively capture the uncertainty and vagueness inherent in linguistic assessments, while remaining computationally simple and transparent. Their structure supports stable aggregation and defuzzification and preserves the reciprocal properties required for pairwise comparisons. Moreover, triangular fuzzy numbers are widely used in FAHP-based healthcare and engineering decision-making studies, ensuring methodological consistency and comparability with existing literature. Then, the relative weight of each criterion was computed by the fuzzy method of Chang et al. as follows [33]. Definition of a triangular fuzzy number A triangular fuzzy number is defined as: 1M = (l, m, u) Where l, m, and u denote the lower, modal, and upper bounds, respectively. Formation of paired comparison matrix. A decision matrix was comprised, as presented in Eq. 1.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rm{{\rm{\tilde A}}}}\, = \left[ \matrix{ {\rm{1 }}\,\,{\rm{ }}\,\,\,\,{{\rm{M}}_{{\rm{12}}}}\,\,...\,\,\,\,{{\rm{M}}_{{\rm{1n}}}} \hfill \cr {{\rm{M}}_{{\rm{21}}}}\,\,\,\,\,\,{\rm{1}}\,\,\,\,\,\,\,...\,\,\,\,{{\rm{M}}_{{\rm{2n}}}} \hfill \cr \vdots \,\,\,\,\,\,\,\,\,\,\,\,\,\, \vdots \,\,\,\,\,\,\, \ddots \,\,\,\,\, \vdots \hfill \cr {{\rm{M}}_{{\rm{n1 }}}}\,\,\,\,{{\rm{M}}_{{\rm{n2}}\,\,}}...\,\,\,\,\,\,1 \hfill \cr} \right]\,\,$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\sim}{a}}_{ij}\:$$\end{document}represents the fuzzy preference of criterion i over criterion j. Calculation of fuzzy synthetic extent value. The fuzzy synthetic extent value (Si) for each criterion \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i\:$$\end{document}is calculated as:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\>{S_i} = \>\sum\limits_{j = 1}^m {M_{gi}^J\> \times } \>{\left[ {\sum\limits_{i = 1}^n {\sum\limits_{j = 1}^m {M_{gi}^J} } } \right]^{ - 1}}$$\end{document} Where i, j, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{M}_{gi}^{j}\:$$\end{document}are the column number, row number, and fuzzy numbers of the paired matrix, respectively. Equation 3 to 5 were also applied to estimate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sum\:_{j=1}^{m}{M}_{gi}^{j}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sum\:_{i=1}^{n}\sum\:_{j=1}^{m}{M}_{gi}^{i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\left[\sum\:_{i=1}^{n}\sum\:_{j=1}^{m}{M}_{gi}^{i}\right]}^{-1}$$\end{document}. 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sum\limits_{j=1}^{m}{M}_{gi}^{j}=\:\left(\sum\limits_{j=1}^{m}{l}_{j}\sum\limits_{j=1}^{m}{m}_{j}\sum\limits_{j=1}^{m}{u}_{j}\right)$$\end{document} 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sum\limits_{i=1}^{n}\sum\limits_{j=1}^{m}{M}_{gi}^{j}=\:\left(\sum\limits_{i=1}^{m}{l}_{i}\sum\limits_{i=1}^{m}{m}_{i}\sum\limits_{i=1}^{m}{u}_{i}\right)$$\end{document} 6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\left[\sum\limits_{i=1}^{n}\sum\limits_{j=1}^{m}{M}_{gi}^{j}\right]}^{-1}=\:\left(\frac{1}{\sum\:_{i=1}^{m}{u}_{i}}\:\frac{1}{\sum\:_{i=1}^{m}{m}_{i}}\:\frac{1}{\sum\:_{i=1}^{m}{l}_{i}}\right)$$\end{document} Estimation of criterion weight. The degree of possibility of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\sim}{S}}_{i}\ge\:{\stackrel{\sim}{S}}_{k}$$\end{document} is computed, and the normalized fuzzy weight vector is obtained as:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:W=({w}_{1},{w}_{2},\dots\:,{w}_{n}{)}^{T}$$\end{document} Defuzzification of criterion weight. To obtain crisp weights, defuzzification is performed using the centroid (center of gravity) method:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{i}^{*}=\frac{{l}_{i}+{m}_{i}+{u}_{i}}{3}$$\end{document} Normalization of crisp weights. Finally, the normalized crisp weights are calculated as:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{i}=\frac{{w}_{i}^{*}}{\sum\:_{i=1}^{n}{w}_{i}^{*}}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sum\:}_{i=1}^{n}{w}_{i}=1$$\end{document}. Computation of consistency index. As the final step in FAHP analysis, the consistency index was calculated to confirm the reliability of the obtained weights. In the present study, the Gogus and Boucher method was used to compute this index. Its value must be below 0.1 [34]. Prioritization of identified techniques To prioritize the identified artificial intelligence techniques for CT scan image analysis of liver cancer based on the weighted criteria using FAHP, the technique for order of preference by similarity to ideal solution (TOPSIS) method was used. TOPSIS is one of the multi-criteria decision-making (MCDM) techniques. The process in the TOPSIS method begins by constructing and normalizing the decision matrix, followed by applying the FAHP-derived weights to obtain the weighted normalized decision matrix. The ideal (best) and negative-ideal (worst) solutions were then identified for all criteria. Finally, the Euclidean distance of each alternative from these two reference points was computed, and a closeness coefficient was derived to indicate proximity to the ideal solution. Alternatives were ranked in descending order of this coefficient, with higher values representing better performance. For performing the steps of the TOPSIS technique, the identified methods, along with their short descriptions, were sent to 80 experts. Extracted primary information, including type/name of method, features of method, application type, application place, population, performance values (such as accuracy, sensitivity, specificity, precision, ROC (AUC), dice score, recall value, F1- score capability, and Jaccard index), and quality level of the papers were extracted from the studies in the form of a table. This table was sent to the experts as guidance (Tables S2 to S5). To assess the quality of the papers, appraisal instruments from the Joanna Briggs Institute (JBI) designed for cross-sectional and cohort studies were utilized. The JBI checklist serves as an effective instrument to examine the quality of different studies [35]. Upon filling out these checklists, the sum of affirmative responses was calculated to derive a total score. Subsequently, the papers were classified into three groups based on this score: low quality, moderate quality, and high quality. Criteria of quality assessment for the cross-sectional studies include details of criteria inclusion as well as the studied subjects and settings, validity and reliability related to measurement, regular criteria applied for condition measurement, confounding factor recognition, used strategies of managing the confounding factors, validity and reliability of outcome measurement, and statistical analysis appropriateness. Criteria of quality assessment for cohort studies also are similarity in groups, exposure validity and similarity in groups, identification and adjustment of confounding factors, validity and reliability of outcomes measurement, sufficiency and implementation of follow-up time, and statistical analysis appropriateness. The full text of the papers was also sent to them on request. They were asked to select a score between 1 and 10 for each method in terms of each criterion. The scores assigned by the experts were aggregated using the arithmetic mean to obtain a single representative value for each of the identified methods before applying the TOPSIS technique. The computational steps of the TOPSIS method are explained as follows [36]. Formation of decision matrix. In this step, a decision matrix was drawn by m methods and n criteria.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left[ {\matrix{ {{x_{11}}} & {{x_{12}}} & \cdots & {{x_{1j}}} \cr {{x_{21}}} & {{x_{22}}} & \cdots & {{x_{2j}}} \cr \vdots & \vdots & \vdots & \vdots \cr {{x_{i1}}} & {{x_{i2}}} & \cdots & {{x_{ij}}} \cr } } \right]$$\end{document} Computation of the normalized decision matrix. The values of the decision matrix were normalized by Eq. 11.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{ij}=\:\frac{{x}_{ij}}{\sqrt[]{\sum\:_{i=1}^{n}{{x}_{ij}}^{2}}}\:,\:i=1,\:2,\:\dots\:,\:m,\:j=1,\:2,\:\dots\:,\:n$$\end{document} Computation of the weighted normalized decision matrix. The normalized values in the previous step were weighted using Eq. 12. By multiplying the weight of each criterion (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{j}$$\end{document}) at the score of each method (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{ij}$$\end{document}), the weighted score of each method (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{V}_{ij}$$\end{document}) was obtained.12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{V}_{ij}=\:{w}_{j}\:\times\:\:{r}_{ij}\:,\:i=1,\:2,\:\dots\:,\:m,\:j=1,\:2,\:\dots\:,\:n\:$$\end{document} Construction of positive and negative ideal solutions. In this step, the worst methods (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{V}_{j}^{-}$$\end{document}) and best methods (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{V}_{j}^{+}$$\end{document}) were determined by Eqs. 13 and 14. In these equations, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{J}_{+}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{J}_{-}$$\end{document} are criteria with positive impact and negative impact, respectively.13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\eqalign{ & \>V_j^ + = \left( {{\rm{max}}{V_{ij}},\>j \in \>{J_ + }\>;{\rm{min}}{V_{ij}},\>j \in \>{J_ - }\>} \right), \cr & \,\,\,\,\,\,\,\,\,\,\,\,\,\>i = 1,\>2,\> \ldots \>,\>m,\>j = 1,\>2,\> \ldots \>,\>n \cr}$$\end{document}14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\eqalign{ & \>V_j^ - = \left( {{\rm{min}}{V_{ij}},\>j \in \>{J_ + }\>;{\rm{max}}{V_{ij}},\>j \in \>{J_ - }\>} \right), \cr & \,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\>i = 1,\>2,\> \ldots \>,\>m,\>j = 1,\>2,\> \ldots \>,\>n \cr}$$\end{document} Computation of distance from worst and best methods. In this step, the distance of the weighted score of each method (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{V}_{ij}$$\end{document}) from the worst method (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{V}_{j}^{-}$$\end{document}) and the best method (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{V}_{j}^{+}$$\end{document}) are estimated by Eqs. 15 and 16, respectively.15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{S}_{i}}^{-}=\:\sqrt[]{\left[\sum\:_{j=1}^{n}{({V}_{ij}-{{V}_{j}}^{-})}^{2}\right]}\:i=1,\:2,\:\dots\:,\:m,\:j=1,\:2,\:\dots\:,\:n$$\end{document}16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{S}_{i}}^{+}=\:\sqrt[]{\left[\sum\:_{j=1}^{n}{({V}_{ij}-{{V}_{j}}^{+})}^{2}\right]}\:i=1,\:2,\:\dots\:,\:m,\:j=1,\:2,\:\dots\:,\:n$$\end{document} Ranking process. The closeness coefficient index is the similarity rate of each method to the best condition, which was calculated as follows. 17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${S_i}w = {S_i}^ - /({S_i}^ - + {S_i}^ + ),\,\,0 \le {S_i}w1,\,\,i = 1,2,..,m$$\end{document} The method with the highest coefficient closeness index is the best technique. Data analysis Descriptive analysis is used for qualitative data. Moreover, MATLAB software (version 2018a) was applied for calculating and analyzing data. To verify judgment validity in FAHP, the aggregated fuzzy pairwise comparison matrix was defuzzified using the center of area method and subjected to Saaty’s consistency test. The maximum eigenvalue was λmax = 7.42, yielding a consistency index CI = 0.070 and a consistency ratio CR = 0.053, which is below the acceptable threshold of 0.10. This confirms that the expert judgments are consistent and that the resulting FAHP weights are reliable. Furthermore, to examine the robustness of the results of TOPSIS, a sensitivity analysis was conducted by varying the elements of the pairwise comparison matrices by ± 10%. The crisp weights were recalculated under these perturbations to observe any potential ranking changes. Several measures were implemented to ensure consistency and minimize bias in expert judgments throughout the FAHP process. First, a diverse panel of 80 experts was selected, including radiologists, AI researchers, and clinical engineers, to reduce disciplinary bias and ensure balanced perspectives. Second, experts were provided with standardized evaluation guidelines and clear definitions of all criteria, ensuring uniform understanding before completing pairwise comparisons. Third, linguistic scales were used instead of precise numerical values, allowing experts to express judgments naturally and reducing forced precision bias. To improve consistency, each expert’s pairwise comparison matrix was checked individually using the consistency ratio (CR), and matrices exceeding the acceptable threshold (CR > 0.10) were reviewed or excluded. Subsequently, expert judgments were aggregated using the geometric mean method, which mitigates the influence of extreme or outlier opinions. Finally, the aggregated matrix was subjected to a global consistency test, confirming that the final judgments were reliable. These steps collectively ensured that the derived weights are both consistent and minimally biased. The reporting quality of the study was assessed using the Checklist for Artificial Intelligence in Medical Imaging (CLAIM), which consists of 42 items covering the title, abstract, introduction, methods, results, discussion, and other study information. Each item was evaluated as reported (score = 1) or not reported (score = 0). The total CLAIM score for each study was calculated by summing the reported items. To facilitate interpretation, CLAIM compliance was expressed as a percentage of the maximum possible score. The reporting quality was categorized as follows: excellent (≥ 80%), good (60–79%), moderate (40–59%), poor (20–39%), and very poor (< 20%). Results Identification of relevant technique Figure 1 shows the PRISMA flowchart. In the literature review, 22,748 papers were found by the selected keyword. Among them, 260 papers were entered into the study. The complete list of these studies is presented in Table S6. Figure 2 depicts the distribution of studies performed in the various countries, and Fig. 3 illustrates the distribution of studies conducted in the different years. Most of the studies were performed in China (97 articles; 37.31%), India (39 articles; 15.00%), the United States (28 articles; 10.77%), Germany (12 articles; 4.62%), Japan (11 articles; 4.23%), South Korea (10 articles; 3.85%), and Canada (9 articles; 3.46%), respectively. Also, most of the studies were conducted in 2023 (78 articles; 30.00%), 2022 (42 articles; 16.15%), 2021 (41 articles; 15.77%), 2020 (27 articles; 10.38%), and 2019 (20 articles; 7.69%). Fig. 1PRISMA flow diagram PRISMA flow diagram Fig. 2Distribution of studies performed in different countries Distribution of studies performed in different countries Fig. 3Distribution of studies performed in different years Distribution of studies performed in different years Classification of identified techniques Based on the application type, the techniques were categorized into four groups, including detection and diagnosis (group I), prediction, prognosis, and registration (group II), segmentation and classification (group III), and treatment and therapy (group IV). Group I (detection and diagnosis) includes methods aimed at identifying liver lesions or supporting diagnostic decisions, such as deep learning models for automatic tumor detection or benign–malignant differentiation. Group II (prediction, prognosis, and registration) comprises techniques focused on outcome prediction, disease progression, survival analysis, or spatial alignment of images, for example, radiomics-based survival prediction models or deep learning approaches for multi-phase CT registration. Group III (segmentation and classification) contains methods designed for pixel-level tumor or liver delineation and tumor type classification, such as U-Net-based segmentation networks or CNN-based tumor classifiers. Group IV (treatment and therapy) includes AI techniques supporting treatment planning, response assessment, or therapy optimization, such as models predicting treatment response or guiding radiotherapy and interventional procedures. The number of the identified papers in groups I, II, III, and IV were 89, 72, 114, and 17, respectively. Methods presented in some studies had multiple applications. Prioritization of identified techniques Seven criteria of validity, accuracy, comprehensiveness, processing time, cost, simplicity, and executive capability for selection of the proper method were introduced, and their relative weights were estimated using fuzzy AHP analysis based on their degree of importance. Operational definitions for each criterion are provided to ensure their clarity. Validity refers to the degree to which an AI technique’s results align with expert radiologist assessments and clinical ground truth. Accuracy measures the correct detection and segmentation of liver tumors relative to reference annotations. Comprehensiveness assesses the ability of the AI system to analyze multiple tumor characteristics simultaneously, including size, shape, and texture. Processing time captures the average computational time required to generate results from a CT scan. Cost considers the financial and computational resources needed for implementation and maintenance. Simplicity reflects ease of use, including the level of training required and integration into clinical workflow. Finally, executive capability evaluates the feasibility of deploying the AI system in real-world clinical settings, including compatibility with existing infrastructure and scalability. Figure 4 shows the relative weights of these criteria. As a result, the criteria of validity (0.199), accuracy (0.191), executive capability (0.166), and cost (0.122) had the highest relative weights according to the experts, respectively. Table S7 represents the aggregated fuzzy pairwise comparison matrix. The ranked AI techniques demonstrate clear differences in clinical applicability across these seven evaluation criteria. Higher-ranked methods exhibit greater validity and accuracy, providing reliable tumor detection and segmentation consistent with expert assessments. They were more comprehensive, analyzing multiple tumor characteristics simultaneously, and offered faster processing times, facilitating efficient workflow integration. While some high-ranking techniques required higher initial computational resources, their overall cost-effectiveness was superior due to reduced manual intervention. They were also simpler to use, requiring minimal training, and possessed strong executive capability, allowing seamless deployment within existing clinical systems. In contrast, lower-ranked techniques showed limitations in one or more of these areas, impacting their practical utility for clinicians. Fig. 4Relative weights of the criteria used for selecting appropriate AI techniques Relative weights of the criteria used for selecting appropriate AI techniques Table 1 reports the best artificial intelligence techniques developed for CT scan image analysis of liver cancer in four groups based on the application type. It is required to mention that the information represented in Table 1 is related to the three techniques with the highest relative weights in each group. The complete information on all developed techniques is available in Table S6 in the appendices. Moreover, Table S8 describes the top-10 ranks of the artificial intelligence techniques. Also, Fig. 5 indicates the score of each criterion for the AI techniques in four groups. In the group of detection and diagnosis, a deep learning-based image reconstruction model developed by Taehee Lee et al. in South Korea was selected as the proper method. The second rank also belongs to the three-way decision support method designed by Yufei Chen et al. in China. The machine learning (ML) algorithm represented by S. Lysdahlgaard et al. in Denmark was placed in the third rank. In the group of prediction, prognosis, and registration, deep learning (DL) radiomics models designed by Mengsi Li in China were placed in the first rank. Radiomics and machine learning models represented by Vincenza Granata et al. in Italy were in second rank. Artificial intelligence (AI)-based screening system (ALMSS) developed by Hongjie Xin et al. in China had the third rank. In the group of segmentation and classification, the Hybrid densely connected UNet (H-DenseUNet) technique developed by Xiaomeng Li et al. in Hong Kong was selected as the proper method. The CTumorGAN technique, designed by Shuchao Pang et al. in Australia, obtained the second rank. The third rank belonged to the split thresholding (ST) technique, represented by Amir H. Foruzan et al. in Iran. In the group of treatment and therapy, the deep learning, radiomics, and clinical (DLRC) model developed by Zhongqi Sun et al. in China was selected as the proper method. The second rank also belongs to the framework based on deep convolutional neural networks (DCNN), represented by Ahmad Maaref et al. in Canada. The convolutional neural networks (CNN) model designed by Chunfeng Zheng et al. in China was placed in third rank. Table 1The best artificial intelligence techniques developed for CT scan image analysis of liver cancer in four groups based on the application typeGroupFirst authorYearCountryType/name of the methodDiseaseType of CT scanFeatures of methodWeightRankDetection and diagnosisLee [37]2022South KoreaDeep learning-based image reconstruction modelHepatic metastasisContrast-enhanced dual-energy CTBetter image quality and comparable diagnostic performance0.8961Chen [4]2017ChinaThree-way decision support methodFocalliver lesionCTPrecise classification0.8852Lysdahlgaard [38]2022DenmarkMachine learning (ML) algorithmLiver tumorCTInvestigation and classification of Radiomics features0.8613Prediction, prognosis, and registrationLi [39]2023ChinaDeep learning (DL) radiomics modelsHepatocellular carcinomaContrast-enhanced dual-energy CTAccurate prediction0.8621Granata [40]2023ItalyRadiomics and machine learning modelColorectal liver metastasesCTGood performance in prognosis0.8592Xin [41]2024ChinaArtificial intelligence (AI)-based screening system (ALMSS)Hepatocellular carcinomaContrast-enhanced CTEffective identification and characterization of liver metastases0.8583Segmentation and classificationLi [42]2018Hong KongHybrid densely connected U-Net (H-DenseUNet)Liver tumor3D abdominalCTEfficiently extracting intra-slicefeatures- hierarchically aggregatingvolumetric contexts0.8881Pang [43]2020AustraliaCTumorGANLiver tumorContrast-enhanced CTHigher contrast- higher resolution0.8792Foruzan [44]2009IranSplit thresholding (ST) techniqueLiver cancerCTImage enhancement and simplification- liver boundary tuning and extraction0.8403Treatment and therapySun [45]2023ChinaDeep learning, radiomics, and clinical (DLRC) modelHepatocellularcarcinoma (hcc)Contrast-enhanced computed tomography (CECT)Prediction of transarterial chemoembolization (TACE) efficacy0.9561Maaref [46]2020CanadaFramework based on deep convolutional neural networks (DCNN)Colorectal liver metastasesCTClassification- differentiation of treated and untreated lesions0.6262Zheng [47]2021ChinaConvolutional neural networks (CNN) modelLiver cancerCTImprovement of the therapeutic effect0.6033 Table 1The best artificial intelligence techniques developed for CT scan image analysis of liver cancer in four groups based on the application typeGroupFirst authorYearCountryType/name of the methodDiseaseType of CT scanFeatures of methodWeightRankDetection and diagnosisLee [37]2022South KoreaDeep learning-based image reconstruction modelHepatic metastasisContrast-enhanced dual-energy CTBetter image quality and comparable diagnostic performance0.8961Chen [4]2017ChinaThree-way decision support methodFocalliver lesionCTPrecise classification0.8852Lysdahlgaard [38]2022DenmarkMachine learning (ML) algorithmLiver tumorCTInvestigation and classification of Radiomics features0.8613Prediction, prognosis, and registrationLi [39]2023ChinaDeep learning (DL) radiomics modelsHepatocellular carcinomaContrast-enhanced dual-energy CTAccurate prediction0.8621Granata [40]2023ItalyRadiomics and machine learning modelColorectal liver metastasesCTGood performance in prognosis0.8592Xin [41]2024ChinaArtificial intelligence (AI)-based screening system (ALMSS)Hepatocellular carcinomaContrast-enhanced CTEffective identification and characterization of liver metastases0.8583Segmentation and classificationLi [42]2018Hong KongHybrid densely connected U-Net (H-DenseUNet)Liver tumor3D abdominalCTEfficiently extracting intra-slicefeatures- hierarchically aggregatingvolumetric contexts0.8881Pang [43]2020AustraliaCTumorGANLiver tumorContrast-enhanced CTHigher contrast- higher resolution0.8792Foruzan [44]2009IranSplit thresholding (ST) techniqueLiver cancerCTImage enhancement and simplification- liver boundary tuning and extraction0.8403Treatment and therapySun [45]2023ChinaDeep learning, radiomics, and clinical (DLRC) modelHepatocellularcarcinoma (hcc)Contrast-enhanced computed tomography (CECT)Prediction of transarterial chemoembolization (TACE) efficacy0.9561Maaref [46]2020CanadaFramework based on deep convolutional neural networks (DCNN)Colorectal liver metastasesCTClassification- differentiation of treated and untreated lesions0.6262Zheng [47]2021ChinaConvolutional neural networks (CNN) modelLiver cancerCTImprovement of the therapeutic effect0.6033 The best artificial intelligence techniques developed for CT scan image analysis of liver cancer in four groups based on the application type Focal liver lesion 3D abdominal CT Efficiently extracting intra-slice features- hierarchically aggregating volumetric contexts Hepatocellular carcinoma (hcc) Fig. 5The criteria scores for best AI techniques across four application groups: (a) detection and diagnosis, (b) prediction, prognosis, and registration, (c) segmentation and classification, and (d) treatment and therapy The criteria scores for best AI techniques across four application groups: (a) detection and diagnosis, (b) prediction, prognosis, and registration, (c) segmentation and classification, and (d) treatment and therapy Figure 6 depicts statistical distribution (percentage) of relative weight categories related to the AI techniques in four groups. Based on the results, techniques developed for segmentation and classification, and methods designed for treatment and therapy, compared to other techniques, had the highest and lowest percentages of high and very high relative weights, respectively. Fig. 6Statistical distribution (percentage) of relative weight categories related to AI techniques across four groups: (a) detection and diagnosis, (b) prediction, prognosis, and registration, (c) segmentation and classification, and (d) treatment and therapy Statistical distribution (percentage) of relative weight categories related to AI techniques across four groups: (a) detection and diagnosis, (b) prediction, prognosis, and registration, (c) segmentation and classification, and (d) treatment and therapy Table 2 shows the results of the sensitivity analysis. A sensitivity analysis was performed to evaluate the robustness of the TOPSIS-based ranking of the AI techniques developed for CT scan image analysis of liver cancer. The criterion weights derived from Fuzzy AHP were perturbed by ± 10% to examine the effect of weight variation on the final rankings. Of the 89 AI techniques developed for CT scan image analysis of liver cancer in the group of detection and diagnosis, the results indicate that the top-ranked methods in the studies of Lee, Chen, Lysdahlgaard, Yang, Bilic, and Suganeshwari remained highly stable across all tested scenarios. Moderate rank changes were observed among mid-ranked techniques (approximately ranks 20–50), particularly when greater weight was assigned to processing time and cost. This suggests that these methods are more sensitive to operational and resource-related criteria than to diagnostic performance metrics. The lowest-ranked techniques consistently remained at the bottom of the ranking under all weight-variation scenarios, confirming that their low performance is not attributable to a specific weighting configuration. Of the 71 AI techniques developed for CT scan image analysis of liver cancer in the group of prediction, prognosis, and registration, the top-ranked methods in the studies of Li, Granata, and Xin remained stable across all scenarios, confirming their strong dominance and robustness. Minor ranking changes were observed among mid-ranked alternatives, indicating competitive performance. While the lowest-ranked methods consistently remained unchanged. These findings demonstrate that the TOPSIS results are stable and not overly sensitive to small variations in model scores. Of the 115 AI techniques developed for CT scan image analysis of liver cancer in the group of segmentation and classification, the top-ranked methods in the studies performed by Li, Pang, and Bilic demonstrated high stability, with substantial weight gaps separating them from mid-ranked alternatives. Moderate rank fluctuations were observed among mid-ranked methods, indicating some sensitivity to small weight changes. The lowest-ranked methods remained consistently at the bottom, confirming the robustness of the ranking framework. Of the 115 AI techniques developed for CT scan image analysis of liver cancer in the group of segmentation and classification, the top-ranked method in the studies performed by Huang was highly stable, indicating very low sensitivity to changes in criteria weights. mid-ranked methods displayed moderate to high sensitivity, suggesting that their positions could shift with variations in the weighting, while the lowest-ranked methods maintained relatively stable ranks. Overall, the analysis confirms that extreme positions are robust, whereas intermediate rankings are more susceptible to weight changes. Overall, the sensitivity analysis demonstrates that the TOPSIS results are robust and stable, and not overly dependent on a particular set of criterion weights. This confirms the validity of the identified top-performing AI techniques under a wide range of practical decision-making preferences. Table 2The results of the sensitivity analysisGroupRank groupWeight rangeSensitivity degreeDetection and diagnosisRanks 1–50.835–0.896Very lowRanks 6–200.716–0.835Low to moderateRanks 21–500.543–0.716ModerateRanks 51–700.411–0.543Moderate to highRanks 71–890.010–0.411Low to highPrediction, prognosis, and registrationRanks 1–50.857–0.861Very lowRanks 6–150.711–0.857Low to moderateRanks 16–300.562–0.711ModerateRanks 31–500.417–0.562Moderate to highRanks 51–710.011–0.417HighSegmentation and classificationRanks 1–50.824–0.888Very lowRanks 6–200.72–0.82Low to moderateRanks 21–500.55–0.72ModerateRanks 51–1150.02–0.55Low to highTreatment and therapyRank 10.955–0.956Very lowRanks 2–40.626–0.626LowRanks 5–70.518–0.488ModerateRanks 8–120.405–0.362Moderate to highRanks 13–170.264–0.005High Table 2The results of the sensitivity analysisGroupRank groupWeight rangeSensitivity degreeDetection and diagnosisRanks 1–50.835–0.896Very lowRanks 6–200.716–0.835Low to moderateRanks 21–500.543–0.716ModerateRanks 51–700.411–0.543Moderate to highRanks 71–890.010–0.411Low to highPrediction, prognosis, and registrationRanks 1–50.857–0.861Very lowRanks 6–150.711–0.857Low to moderateRanks 16–300.562–0.711ModerateRanks 31–500.417–0.562Moderate to highRanks 51–710.011–0.417HighSegmentation and classificationRanks 1–50.824–0.888Very lowRanks 6–200.72–0.82Low to moderateRanks 21–500.55–0.72ModerateRanks 51–1150.02–0.55Low to highTreatment and therapyRank 10.955–0.956Very lowRanks 2–40.626–0.626LowRanks 5–70.518–0.488ModerateRanks 8–120.405–0.362Moderate to highRanks 13–170.264–0.005High The results of the sensitivity analysis The reporting quality of the study was evaluated using the 42-item CLAIM checklist, as shown in Table S9. The study achieved a total CLAIM score of 27 out of 42 items, corresponding to a compliance rate of 64.3%. According to the predefined classification criteria, this represents good reporting quality. Most CLAIM items related to study rationale, objectives, eligibility criteria, data sources, statistical methods, sensitivity analysis, study limitations, ethical approval, registration, and funding disclosure were adequately reported. Lower scores were observed for items related to artificial intelligence model development and validation, including dataset partitioning, model architecture specification, parameter initialization, training procedures, external validation, explainability methods, and failure analysis. These items were largely not applicable because the present study was a multi-criteria decision-making analysis rather than a primary AI model development study. Discussion In the present study, 260 artificial intelligence techniques developed for CT scan analysis were identified. The deep learning–based image reconstruction technique proposed by Lee et al. achieved the highest overall weight in the detection and diagnosis group because it demonstrated a superior balance across multiple evaluation criteria compared with other reviewed methods. Key factors contributing to its high score include its high diagnostic accuracy and specificity (97.34%), strong diagnostic acceptability, and enhanced lesion detection, particularly for small metastases (< 10 mm) [37]. In terms of comprehensiveness, the method integrates dual-energy CT reconstruction, a vendor-neutral platform (ClariCT.AI), and U-net-based denoising into a unified pipeline, addressing multiple aspects of image formation and enhancement simultaneously, whereas many competing approaches focus on a single stage, such as denoising or contrast enhancement alone [37]. From an operational standpoint, the technique scored highly in processing time and executive capability, as the automated and vendor-neutral design allows seamless integration into existing clinical workflows, unlike more computationally intensive approaches such as radiomics-based models that require extensive preprocessing or feature engineering. While initial costs for deep learning reconstruction are higher, its simplicity in routine use—requiring minimal manual tuning—further enhances clinical applicability [48]. When compared with other popular AI approaches, including transformers, graph neural networks, and hybrid radiomics models, the Lee et al. method maintains a clear advantage in practical deployability and interpretability. Transformer- and graph-based models often require large datasets and high computational resources, limiting their immediate clinical adoption [37, 49]. Hybrid radiomics models, although potentially high-performing, involve multiple processing steps and complex feature engineering, reducing workflow simplicity and increasing processing time. In contrast, the top-ranked reconstruction model achieves a balanced performance across all seven FAHP-TOPSIS criteria, combining accuracy, reliability, and operational feasibility. This comprehensive strength explains its higher overall ranking and supports its suitability as a clinically viable AI technique for liver cancer CT analysis. Of the techniques identified for prediction, prognosis, and registration, the dual-energy CT deep learning radiomics model developed by Mengsi Li et al. was selected as the most suitable method for predicting the macrotrabecular-massive subtype of hepatocellular carcinoma (MTM-HCC). This model achieved high scores because it combines deep learning features and handcrafted radiomics features extracted from virtual monoenergetic and material decomposition images with feature selection via LASSO and multivariable logistic regression to construct a predictive nomogram, resulting in strong predictive performance, with higher AUC, sensitivity, and specificity compared with simpler clinical-radiologic models across training and external validation sets [39]. The integration of dual-energy CT data enhances the model’s ability to capture tumor heterogeneity and subtle imaging patterns, improving comprehensiveness and validity relative to models relying on single-phase imaging or hand-crafted features alone [39]. From a practical perspective, this approach balances predictive accuracy, interpretability, and workflow feasibility: it can be applied to existing dual-energy CT protocols and yields interpretable outcome probabilities without requiring prohibitive computational resources. In contrast, other cutting-edge architectures, such as transformer-based models, have shown promise in medical imaging by capturing long-range dependencies and richer contextual information, but they often require larger datasets and higher computational cost, which can limit clinical adoption, especially in data-limited environments [50]. Similarly, graph neural networks have demonstrated strengths for structured relational representations in imaging tasks but remain computationally intensive and less established for standard CT predictive workflows [51]. Hybrid radiomics models that fuse deep learning with handcrafted features may offer competitive performance, but they can entail complex preprocessing pipelines that increase processing time and reduce simplicity compared with the nomogram approach in Li et al.’s study. Overall, the combination of robust predictive performance, clinical interpretability, and workflow integration explains why the dual-energy CT deep learning radiomics model received higher overall scores within the FAHP-TOPSIS evaluation. Among the techniques designed for segmentation and classification, the hybrid densely connected U-Net (H-DenseUNet) developed by Xiaomeng Li et al. in Hong Kong was selected as the most suitable method for liver and tumor segmentation due to its superior performance across multiple FAHP-TOPSIS criteria, including accuracy, validity, comprehensiveness, processing time, and executive capability. This model employs a cascaded learning strategy in which a lightweight ResNet performs rapid coarse segmentation to reduce computational burden, followed by H-DenseUNet’s hybrid feature extraction that captures both intra-slice and inter-slice contextual information. These features are then optimized through Hybrid Feature Fusion (HFF) to achieve precise liver and lesion delineation [42]. The method achieved a Dice score of 93.7% and a root mean square error of 0.015, significantly outperforming many conventional CNN-based and 2D/3D segmentation methods, and demonstrating highly competitive performance on the LiTS benchmark. Several factors contributed to H-DenseUNet’s high overall ranking. Its hybrid architecture efficiently combines dense connectivity and 3D spatial context, enabling accurate segmentation of both liver and tumors, including small lesions. Its computational efficiency through cascaded initialization reduces total processing time, and its workflow is relatively simple to implement, improving executive capability and facilitating clinical integration [42]. Compared with other advanced methods, including transformer-based models, graph neural networks (GNNs), and hybrid radiomics approaches, H-DenseUNet maintains a favorable balance between accuracy, interpretability, and operational feasibility. Transformer-based models can capture long-range dependencies but typically require large datasets and high computational resources, limiting their immediate clinical use [52]. GNN-based approaches enhance relational reasoning but increase model complexity and processing time, while hybrid radiomics methods require multiple preprocessing steps that can reduce simplicity and deployability. By excelling across FAHP-TOPSIS criteria—particularly in segmentation accuracy, workflow efficiency, and clinical applicability—H-DenseUNet represents a robust, practical, and clinically viable solution for liver and tumor segmentation. Among techniques introduced for treatment and therapy, the deep learning, radiomics, and clinical (DLRC) model developed by Sun et al. was selected as the most appropriate method for predicting response to transarterial chemoembolization (TACE) in hepatocellular carcinoma (HCC) [45]. This model achieved a high overall score because it integrates multi-modal data, combining contrast-enhanced CT imaging features with statistically significant clinical variables to construct a predictive multivariable model that demonstrated robust performance in training and validation cohorts (e.g., predictive AUCs > 0.85 and effective risk stratification for TACE) [45]. The combination of imaging and clinical features enhances predictive validity and comprehensiveness, capturing both tumor characteristics and patient-specific factors [53]. In terms of processing time and simplicity, using pre-extracted radiomics signatures and standard CT imaging protocols helps maintain workflow feasibility. Compared with other advanced techniques, such as transformer-based models, which have demonstrated high performance but require large datasets and intensive computational resources that may limit clinical deployment, and graph neural networks, which capture structural relationships but introduce additional model complexity, hybrid radiomics approaches like the DLRC model provide a more balanced and clinically deployable solution [54]. By excelling across criteria including accuracy, interpretability, clinical relevance, and workflow integration, the DLRC model offers a robust predictive solution for treatment outcomes in HCC patients undergoing TACE. In general, the results of these studies show that precise determination of the properties of tumors, the combination of personal and clinical variables with artificial intelligence, and the combination of physician and artificial intelligence diagnosis can increase the accuracy of the model. Also, it should be noted that although the increase of layers in the model may increase the accuracy, it may affect the processing speed and its execution capability. The findings of this study provide a roadmap for future AI development in radiology by identifying key factors that contribute to clinically effective models. By evaluating 260 techniques using FAHP-TOPSIS across criteria such as accuracy, validity, comprehensiveness, processing time, cost, simplicity, and executive capability, the study highlights the importance of balancing technical performance with practical clinical applicability. High-ranking models, including deep learning–based reconstruction and dual-energy CT radiomics approaches, demonstrate that integrating multi-modal data, optimizing workflow efficiency, and ensuring interpretability are critical for adoption in routine practice. These insights suggest that future AI tools should not only maximize predictive accuracy but also be designed for seamless integration into clinical workflows, efficient processing, and user-friendly deployment, providing a structured framework to guide iterative development of robust, clinically viable radiology AI systems. To show the practical applicability of the prioritized AI techniques, a short case example or workflow can be considered. For instance, in the context of liver cancer CT analysis, a patient undergoes dual-energy CT imaging, and the acquired images are processed through a deep learning–based reconstruction pipeline (e.g., 40-keV virtual monoenergetic images processed with U-net denoising). The reconstructed images are then evaluated by a radiologist using lesion detection and segmentation tools, supported by hybrid models such as H-DenseUNet for accurate liver and tumor delineation. For prognosis or treatment planning, relevant radiomics features and clinical parameters are extracted and integrated into a DLRC or dual-energy CT radiomics model to predict treatment response, such as transarterial chemoembolization efficacy. The resulting output provides a risk stratification or decision support report, enabling clinicians to prioritize interventions, optimize patient-specific treatment plans, and monitor outcomes. While the FAHP-TOPSIS framework provides a systematic and transparent approach to ranking AI techniques, certain limitations must be considered. First, expert-based MCDM approaches inherently involve subjectivity, as the pairwise comparisons and criterion weightings are influenced by the knowledge, experience, and perspectives of the participating experts. Although measures such as consistency checks and aggregation of responses from 80 experts were applied to mitigate individual biases, some degree of subjectivity remains unavoidable. So, the criteria weights reflect the priorities of the specific expert panel consulted, and a different panel may yield slightly different weightings. Second, the dataset of AI techniques included in this study is derived from published literature, which may be subject to publication bias. The clinical context, image sets, dataset sizes, and research conditions were different in various studies. Moreover, across the 260 included studies, there were notable variations in several aspects of dataset quality, including image resolution, scanner types, contrast protocols, patient demographics, and annotation reliability. These variations can directly influence model performance, generalizability, and reproducibility. Therefore, dataset quality differences may contribute to observed discrepancies in AI technique performance, highlighting the need for standardized imaging protocols and benchmark datasets in future research. In the present study, the performance of AI methods was not compared using integrated datasets because only some aspects, such as accuracy, speed, and time, can be compared. While using multi-criteria decision-making methods, it is possible to evaluate the methods in terms of different criteria. Also, due to the large number of studies identified, it is not possible to assess the validity using real datasets. It is suggested that the performance of top-ranked AI methods is evaluated using real and integrated datasets in the next studies. Finally, the field of AI in radiology is rapidly evolving, with continuous development of novel architectures, transformer-based models, and hybrid radiomics approaches. As a result, the rankings presented in this study are time-sensitive, reflecting the state of the field at the time of data collection, and may require periodic updates to maintain relevance as new evidence emerges. Recognizing these limitations provides context for interpreting the FAHP-TOPSIS results and underscores the need for ongoing evaluation and iterative refinement of AI prioritization frameworks. However, the artificial intelligence methods developed for CT scan image analysis of liver cancer were generally classified and prioritized, which can be a guide for the next studies. Conclusion In total, 260 artificial intelligence techniques developed for CT scan analysis of liver cancer were found in four groups of detection and diagnosis, prediction, prognosis and registration, segmentation and classification, and treatment and therapy. Deep learning-based image reconstruction model, dual-energy CT deep learning radiomics model, hybrid densely connected UNet (H-DenseUNet) technique, and deep learning, radiomics, and clinical (DLRC) model were selected as the proper methods in each of the mentioned groups. As an example, to show the practical applicability of the prioritized AI techniques, a CT scan of a liver-cancer patient is reconstructed with deep learning, analyzed with AI-assisted detection and segmentation tools, and combined with radiomics-based prediction models to generate a risk-stratification report that supports personalized treatment planning and outcome monitoring. For practical implementation, these models can be integrated into clinical workflows, including PACS systems, decision support tools, and pre-processing pipelines, to enhance diagnostic accuracy, risk stratification, segmentation, and personalized treatment. Moreover, to ensure clinical applicability, developers should assess AI techniques systematically against the seven evaluation criteria (validity, accuracy, comprehensiveness, processing time, cost, simplicity, and executive capability), validate models with prospective or multi-center datasets, incorporate clinician oversight for continuous improvement, and ensure regulatory compliance. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2 Supplementary Material 3 Supplementary Material 3 Supplementary Material 4 Supplementary Material 4 Supplementary Material 5 Supplementary Material 5 Supplementary Material 6 Supplementary Material 6 Supplementary Material 7 Supplementary Material 7 Supplementary Material 8 Supplementary Material 8 Supplementary Material 9 Supplementary Material 9 Supplementary Material 10 Supplementary Material 10