Integrating Explainable Artificial Intelligence into Histopathological Risk Assessment: A Scoping Review and Meta-analysis

Integrating Explainable Artificial Intelligence into Histopathological Risk Assessment: A Scoping Review and Meta-analysis | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Integrating Explainable Artificial Intelligence into Histopathological Risk Assessment: A Scoping Review and Meta-analysis Bandar Alshreef, Yousif Kariri This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9249572/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Explainable artificial intelligence is increasingly used to support cancer detection, grading, and prognosis from histopathology, yet clinical adoption remains limited by uncertainty about reliability, interpretability, and governance. We conducted a scoping review and meta-analysis of peer-reviewed studies applying explainable AI to histopathology-based cancer risk assessment. Searches of PubMed, Google Scholar, Scopus, and Web of Science identified 47 eligible studies. Post hoc visual methods, particularly CAM and Grad-CAM, dominated the field. Nine studies contributed to meta-analysis, yielding a pooled area under the curve of 0.962 (95% CI 0.909–0.985) with substantial heterogeneity (I² = 97.1%). Subgroup analysis showed higher and more consistent performance at the slide level than at the patient level, identifying unit of analysis as a major source of heterogeneity. These findings support a translational roadmap for clinically meaningful explanation, usability testing, prospective validation, and governance-aligned deployment. Biological sciences/Cancer Biological sciences/Computational biology and bioinformatics Health sciences/Medical research Health sciences/Oncology Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Histopathology continues to be the most widely accepted method for establishing a diagnosis and determining the grade of cancer. The method requires an experienced pathologist to examine numerous morphological features of the tumor to determine its type and aggressiveness. Although pathologists are experts in examining histologic material, the variability in interpretation among pathologists poses one of the greatest challenges to the establishment of a reliable method for establishing a diagnosis. A number of studies have reported variable levels of concordance among pathologists when they were asked to interpret the same histologic slides, suggesting the presence of diagnostic bias [ 1 ]. Digital pathology has introduced the possibility of using artificial intelligence (AI) to aid and enhance the evaluation of histopathological materials [ 2 ]. Recent studies have shown that AI-based algorithms can provide clinical-grade performance in identifying tumors, classifying them, and predicting prognoses in several types of cancers (e.g., prostate, breast, colon), and at least in some cases perform equally well, if not better, than human observers [ 35 – 38 , 45 ]. AI may be able to recognize small patterns in images that cannot be detected by the human eye and possibly even predict molecular mutations based upon the histological findings of a biopsy sample, thereby improving the accuracy of the stratification of patients by risk [ 5 , 39 ]. While there have been positive outcomes from the use of AI in evaluating histopathological materials, the translation of this technology into clinical practice is hindered by barriers. Systematic reviews of studies that compare the performance of AI-based models to those of clinicians generally report that the two are equivalent; however, the high risk of bias in many of these studies limits the ability to draw definitive conclusions regarding the superiority of AI-based models [ 46 ]. Additionally, the lack of clarity in how many deep-learning-based models arrive at their decisions (i.e., the "black box" nature of many of these models) makes it difficult for pathologists to understand, validate, and rely upon the outputs of these models [ 6 , 25 ]. Limited interpretability of many AI models is a significant obstacle to the safe and effective integration of these models into clinical settings. As a result, the subfield of Explainable AI (XAI) has developed to improve the transparency, accountability, and reliability of AI models [ 7 , 40 ]. In histopathology, XAI could enable pathologists to verify the logic used by an AI model to arrive at its conclusion(s) and potentially identify possible sources of error or bias [ 8 ]. Therefore, while the objectives of XAI in histopathology extend beyond simply explaining an AI model (explainability) to include supporting and enhancing the pathologist's diagnostic reasoning through providing causal explanations (causality) [ 21 , 32 ]. In addition to the need for XAI, the integration of AI models into clinical laboratory settings requires compliance with strict quality and safety regulations, such as ISO 15189 (Medical Laboratories—Requirements for Quality and Competence) and ISO 14971 (Application of Risk Management to Medical Devices) [ 9 ]. Compliance with these standards, along with reporting requirements for clinical trials such as CONSORT-AI and TRIPOD-AI, during all stages of the AI development cycle is necessary to ensure the responsible implementation of AI [ 30 , 41 , 42 ]. This review examines the critical relationship between XAI, risk stratification, and quality management. Through a systematic review of the existing literature, combined with a meta-analysis of the diagnostic performance of AI models, we will identify key obstacles and develop a structured framework to inform future research efforts and facilitate the transition of reliable and causal AI-based tools into clinical practice. Results Study selection and characteristics Our initial search yielded 857 studies. After removing duplicates, 642 articles were screened, and 123 full-text papers were assessed for eligibility. A final set of 47 studies met our inclusion criteria. The study selection process is detailed in the PRISMA flow diagram (Fig. 1 ). Characteristics of included studies The majority of the research studies contained herein were conducted on many different types of cancers; specifically, breast (34%), prostate (21%), lung (15%), and colorectal (11%) were the most prevalent (Fig. 2 ). It is probable that the high prevalence of studies concerning breast and prostate cancers (combined = 55%) is due to the existence of large public data sets from contests such as CAMELYON16. Additionally, it appears that the use of post-hoc visual explanations for image-based XAI methods (i.e., CAMs, Grad-CAM) in this type of image classification task has created a void in explanation methodologies for more complex and multi-modal predictive modeling. Key concepts and definitions Explainability: The capability to give a human-readable description of how an AI model works internally [ 7 , 21 ]. Interpretability: A measure of how well a human can identify why an AI model made a particular choice [ 26 ]. Causality: An evaluation of the level to which an explanation of a decision made by an AI model provides a level of causal insight to a human expert so that the expert has the information necessary for making decisions [ 3 , 21 ]. Transparency: The ability for a human to view how an AI model operates inside, usually only possible when the model is relatively simple (e.g., linear regression, decision tree) [ 4 ]. The concepts above are often confused or used interchangeably; however, they constitute a spectrum of technical transparency to clinical interpretation. As this review examines how XAI techniques can produce not only technical descriptions of an AI model but also clinically relevant insights into an AI model that assist in a pathologist’s diagnostic reasoning process. XAI techniques in histopathology The studies reviewed use a wide range of XAI techniques that fall into four categories: Post-hoc Visual Explanations These techniques provide a visual explanation after a model has been trained, and they show the areas of the image that had the largest impact on the decision made by the model. Techniques used for these types of visual explanations include: Class Activation Mapping (CAM): CAM projects the weight of the output layer backwards onto the convolutional feature maps that generated the output [ 33 ]. Gradient-weighted Class Activation Mapping (Grad-CAM): Grad-CAM is a generalized version of CAM, using the gradient flow through the final convolutional layer to create a coarse localization map [ 33 ] Saliency Maps: Saliency maps highlight the pixels with the highest contribution to the model’s prediction. [ 6 ] Layer-wise Relevance Propagation (LRP): LRP decomposes the prediction of a deep neural network by reversing the propagation of the relevance of the output back to the input features [ 6 ] Attention Mechanisms Unlike post-hoc explanations that are added to a pre-trained model after it has been developed, attention-based models incorporate the ability to focus on the most relevant aspects of the input data into the design of the model itself. Attention-based models have been widely adopted for MIL-based approaches to identifying the most informative regions in a whole-slide image of histological tissue [ 16 ], [ 17 ]. Concept-based Explanations Rather than explaining how a model makes a decision based on the low-level features present in the input data (e.g., pixel values), concept-based explanations describe how a model makes decisions based on higher-level concepts (e.g., presence of certain tissue structure or cellular pattern) that are meaningful to humans [ 8 ]. Inherently Interpretable Models There are some researchers who advocate for using inherently interpretable models (e.g., decision trees, linear models, etc.) in place of using post-hoc explanations for complex black-box models [ 4 ][ 29 ]. Applications of XAI in histopathological risk assessment The included studies applied XAI to a range of risk assessment tasks, including: Tumor Detection and Identification XAI is currently being applied to explain how Artificial Intelligence models are detecting and classifying tumors across different parts of the body (i.e. breast, prostate, lungs, colon). As one example, Grad-CAM is being used to display the exact regions of a histopathology slide where an Artificial Intelligence model determines benign vs. malignant tissue [ 10 , 11 ]. Grading & Staging There have been attempts to develop AI models that will grade tumors automatically (for example, Gleason grading of prostate cancer) and stage diseases; XAI techniques may be able to help pathologists understand why they have received those results and where there may be errors [ 13 , 36 , 37 , 38 ]. The Tumor Micro-Environment (TME). XAI is being utilized to determine the complexities involved within the interactions of tumor cells with their micro-environmental surroundings. Included in this are the identification and quantification of tumor infiltrating lymphocytes (TILs), which have direct implications for predicting patient responses to immunotherapies. For instance, attention based models are capable of highlighting areas of high TIL density, that correlate with patient outcomes [ 19 ]. Predicting Molecular Markers An additional area where XAI is being utilized is in explaining models that predict molecular characteristics such as MSI or gene mutations directly from H&E stained slides. Heat maps provide an opportunity to show the morphological characteristics that a model associates with a particular molecular subtype and therefore provides a link between histology and genomics [ 14 , 39 ]. Quality safety and workflow integration Ensuring that the development of AI in pathology adheres to quality guidelines such as ISO 15189 and ISO 14971 is critical to ensure the safety of developing AI in pathology [ 9 , 24 ]. Documentation of XAI artifacts, e.g., as heat maps or conceptual attributions as part of an ISO 15189-compliant quality management system, and use as risk control elements in an ISO 14971-compliant risk assessment may provide assurance that risks associated with XAI have been mitigated. For example, a validation report documenting the agreement between a model's heatmap representation and a pathologist's annotations of a region of interest could help demonstrate that a model’s performance was robust when providing a diagnosis. The development of XAI systems, however, requires a user-centric approach. Explanations provided by poorly designed XAI systems can create additional cognitive burdens and disrupt clinical workflows [ 48 ]. Thus successful integration will require a focus on user-centric design, seamless integration into existing clinical workflows, minimal cognitive burden, and integration into existing hospital information systems. Considerations related to the display of XAI results (e.g., an overlay on a digital slide) and the interaction with XAI results (e.g., the ability to turn XAI explanations on or off) are key considerations, as well as the medicolegal ramifications of using AI to assist in diagnosis [ 48 ]. The roles of XAI across the AI lifecycle, aligned with relevant quality standards, are summarized in Table 1.” Table 1 Illustrative mapping of XAI roles to the AI lifecycle, aligned with quality standards. Lifecycle Phase XAI Role Quality Standard Development & Training Identify and mitigate bias; validate feature relevance; ensure model learns clinically meaningful patterns ISO 14971 (Risk Analysis) Validation & Verification Document concordance between model explanations and expert annotations; support internal validation ISO 15189 (QMS) Deployment & Monitoring Provide real-time explanations to support clinical decision-making; enable continuous performance monitoring ISO 15189 (QMS) Post-Market Surveillance Facilitate investigation of adverse events or unexpected outcomes; support root cause analysis ISO 14971 (Risk Control) Quality assessment In this evaluation of the diagnostic AI literature, we applied a quality assessment tool developed specifically for this area of research (see Supplementary File 5) to assess the 47 identified studies. In assessing the quality of these studies, we determined that 18 (38%) of them represented studies at low risk of bias, whereas 22 (47%) were considered to be at moderate risk, and 7 (15%) studies were classified as being at high risk of bias. The primary factors that contributed to the high risk of bias in many studies were the inclusion of non-sequential cases, the absence of an independent validation dataset, and the ambiguity surrounding how models were split into training datasets and test datasets. The results of our quality assessment further emphasize the need for studies evaluating diagnostic AI to have a higher level of study design rigor; this is consistent with the conclusions drawn from recent meta-analytic reviews [ 1 , 46 ]. Quantitative synthesis We selected 9 out of the 47 studies included in the scoping review that met the eligibility requirements for inclusion in the diagnostic meta-analysis. Those studies provided the necessary data to calculate the AUC with its respective 95% confidence interval. When all 9 studies were analyzed using a random-effects meta-analysis model, we observed a pooled AUC of 0.962 (95% CI: 0.909–0.985). However, the analysis indicated substantial heterogeneity among the studies (I² = 97.1%, τ² = 1.89, Q = 278.58, p < 0.001); this indicates substantial variability in model performance across the studies. The overall random-effects meta-analysis is presented in Fig. 3 . The corrected subgroup analysis by the unit of evaluation was a major cause of the heterogeneity noted above. Studies that evaluated performance at the individual slide level demonstrated higher and more consistent pooled AUC values of 0.987 (95% CI: 0.981–0.990) as well as less heterogeneity (I²=12.5%) than did those that evaluated performance at the patient level. Studies that evaluated performance at the patient level had lower pooled AUC values of 0.910 (95% CI: 0.830–0.955) and much greater heterogeneity (I²=93.1%). These findings are depicted in the revised forest plot (Fig. 4 ) and suggest that the method used for combining predictions made from the slide to the patient level has an impact on both how well a model performs and how consistently different studies report performance. The results were confirmed by a sensitivity analysis that excluded the four slide-level studies. The pooled estimate of the remaining five patient-level studies was 0.910 (95% CI: 0.830–0.955) and is consistent with the patient-level subgroup results. Discussion This systematic review has mapped the present state of XAI in histopathological risk assessment, which has revealed a promising prospect with several challenges. In addition, a quantitative meta-analysis was added to provide a performance standard that showed that, although pooled diagnostic accuracy is high (AUC = 0.962), the field is highly heterogeneous (I² = 97.1%) as the two most important sources of this variation are identified by our study as the unit of evaluation—slide level vs. patient level. We believe the greatest impact of our study is the distinction between the need for post-hoc explanations and the use of inherently interpretable models in high-risk clinical decision making. We found a growing consensus that causal interpretations must be the central focus of the development of XAI. Although AI has been very successful in many areas, it has also shown poor generalizability to external datasets, particularly when there has been a distribution shift from the training dataset (e.g., due to differences in patient populations or slide preparation) [ 1 , 46 , 47 ]. At the time of writing, there is no agreement on what metrics should be used to measure either the quality or the consistency of AI-generated explanations [ 22 , 23 , 33 , 40 ]. In fact, several studies have shown that some of the most popular methods (such as saliency maps) can be misleading [ 28 ], so that some researchers argue that current XAI methods create a false sense of security and therefore, inherently interpretable models should be the priority for all high-risk applications [ 29 , 34 ]. In addition, the clinical utility of XAI remains largely unproven. It is still unknown how explanations will be used in routine practice and if they will improve diagnostic accuracy or patient outcomes. One of the largest challenges will be to meet the multiple needs of different stakeholders; for example, pathologists may want explanations to increase their diagnostic confidence, while regulatory bodies will require auditable evidence of a system's safety and effectiveness [ 25 ]. To be clinically relevant, explanations must be framed within the context of domain-specific knowledge and not simply meet a technical requirement [ 25 ]. As an example, in prostate Gleason grading, a heatmap that highlights stromal regions without glandular relevance illustrates the gap between a technically correct explanation and clinical causality. A more causal explanation would highlight the specific atypical glandular formations that contributed to a higher Gleason score and link the AI's output to the pathologist's diagnostic criteria. These concerns are particularly relevant for clinical deployment, where a lack of causal explanations can undermine trust and hinder adoption. Our meta-analysis further highlights that methodological choices—such as slide-level versus patient-level evaluation—directly affect reported performance, underscoring the need for standardized evaluation protocols that reflect real-world clinical workflows. To serve as a guideline for future research, we have proposed a four-stage roadmap to move the application of Explainable AI (XAI) into histopathology: Pathologist-centric usability studies: Use usability study techniques such as cognitive walkthroughs or think aloud to measure how XAI will influence diagnostic efficiency, diagnostic confidence, and inter-rater reliability. Prospective clinical validation: Design and conduct prospective clinical trials to assess the real-world effects of using XAI to support the clinical workflow in terms of patient outcome. Standard evaluation metrics: Establish and validate standardization around metrics that can be used to compare the effectiveness of XAI explanations in terms of both quality and consistency. Publicly available benchmarking and dataset: Establish publicly available benchmarking and datasets for comparing XAI algorithms in histopathology. This review has several limitations. We restricted inclusion to English-language studies, which may have introduced language bias, and only nine studies contributed data to the primary meta-analysis. Because explainable AI evolves rapidly, relevant methods published after the search date may not have been captured. We also excluded grey literature and conference abstracts, and our modified quality assessment approach has not yet been formally validated for all explainability-focused pathology studies. Overall, explainable AI in histopathology shows strong but heterogeneous diagnostic performance and remains methodologically immature at the point of clinical translation. The next phase of the field should prioritize patient-level evaluation, clinically meaningful explanations, pathologist-centered usability studies, prospective validation, and closer alignment with quality and risk-management frameworks. Ultimately, this roadmap will accelerate the responsible translation of XAI into clinical practice—delivering trustworthy AI tools that support pathologists, improve diagnostic accuracy, and enhance patient outcomes. Methods Primary research studies published in an English-language journal that include: Histopathology (whole-slide images or otherwise) in an oncologic context XAI approaches applied to model outputs that are relevant to risk-assessment (e.g., prognosis, grading, stratification) Details about how they implemented their methodology to allow us to extract which XAI approach and task were utilized. Studies that do not relate to oncologic conditions or studies that are not research-based (editorial, review, etc.) are excluded. Study selection Two independent reviewers screened both title/abstracts and full-text versions of articles, with disagreements resolved through consensus. Data charting and synthesis The following information was abstracted: author/year, cancer type, task, AI architecture, XAI method, dataset, validation approach, and key findings, after which the evidence was synthesized thematically. Statistical analysis In order to quantify the diagnostic performance of AI models, a random effects meta-analysis was performed. Priority was given to studies that reported the area under the receiver operating characteristic curve (AUC) with 95% confidence intervals (CIs). Confidence interval bounds were used to derive logit-transformed AUCs and their variances. Pooled estimates and 95% CIs were calculated using the DerSimonian-Laird method with a restricted maximum likelihood estimator, then back-transformed to the AUC scale for interpretation. Heterogeneity was assessed using Cochran's Q test and the I² statistic. Pre-specified subgroup analyses compared slide-level and patient-level evaluation, and a sensitivity analysis excluding slide-level studies was also performed. All analyses were conducted in Python using the pandas and scipy libraries. Data availability All data analyzed in this study were derived from published articles and are reported in the manuscript and its Supplementary Information. The extracted study-level dataset used for the meta-analysis is available from the corresponding author upon reasonable request. Code availability The statistical code used for the quantitative synthesis was developed in Python and is available from the corresponding author upon reasonable request. Declarations Funding No specific funding was received for this work. Acknowledgments The authors would like to thank the Deanship of Scientific Research at Shaqra University for supporting this work. Author contributions B.A. conceived the study. B.A. and Y.A.K. designed the review, screened records, extracted and interpreted data, and revised the manuscript. B.A. performed the statistical analyses and drafted the manuscript. Y.A.K. critically reviewed the manuscript and approved the final version. All authors approved the submitted manuscript. Competing interests The authors declare no competing interests. References van der Velden BHM, Kuijf HJ, Gilhuijs KGA, Viergever MA. Explainable artificial intelligence (XAI) in deep learning-based medical image analysis. Med Image Anal. 2022;79:102470. doi: 10.1016/j.media.2022.102470 Bera K, Schalper KA, Rimm DL, Velcheti V, Madabhushi A. Artificial intelligence in digital pathology—new tools for diagnosis and precision oncology. Nat Rev Clin Oncol. 2019;16(11):703–715. doi: 10.1038/s41571-019-0252-y Holzinger A, Langs G, Denk H, Zatloukal K, Müller H. Causability and explainability of artificial intelligence in medicine. Wiley Interdiscip Rev Data Min Knowl Discov. 2019;9(4):e1312. doi: 10.1002/widm.1312 Rudin C. Stop explaining black box machine learning models for high-stakes decisions and use interpretable models instead. Nat Mach Intell. 2019;1(5):206–215. doi: 10.1038/s42256-019-0048-x Coudray N, Ocampo PS, Sakellaropoulos T, et al. Classification and mutation prediction from non–small cell lung cancer histopathology images using deep learning. Nat Med. 2018;24(10):1559–1567. doi: 10.1038/s41591-018-0177-5 Samek W, Wiegand T, Müller KR. Explainable artificial intelligence: understanding, visualizing and interpreting deep learning models. arXiv. 2017. doi: 10.48550/arXiv.1708.08296 Adadi A, Berrada M. Peeking inside the black-box: a survey on explainable artificial intelligence (XAI). IEEE Access. 2018;6:52138–52160. doi: 10.1109/ACCESS.2018.2870052 Chen Z, Bei Z, Rudin C. Concept whitening for interpretable image recognition. Nat Mach Intell. 2020;2:772–782. doi: 10.1038/s42256-020-00265-z International Organization for Standardization. ISO 15189:2022 Medical laboratories—Requirements for quality and competence. Geneva: ISO; 2022. Ehteshami Bejnordi B, Veta M, van Diest PJ, et al. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. JAMA. 2017;318(22):2199–2210. doi: 10.1001/jama.2017.14585 Campanella G, Hanna MG, Geneslaw L, et al. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nat Med. 2019;25(8):1301–1309. doi: 10.1038/s41591-019-0508-1 Echle A, Grabsch HI, Quirke P, et al. Clinical-grade detection of microsatellite instability in colorectal tumors by deep learning. Gastroenterology. 2020;159(4):1406–1416.e11. doi: 10.1053/j.gastro.2020.06.021 Bulten W, Kartasalo K, Chen PH, et al. Artificial intelligence for diagnosis and Gleason grading of prostate cancer: the PANDA challenge. Nat Med. 2022;28(1):154–163. doi: 10.1038/s41591-021-01620-2 Kather JN, Pearson AT, Halama N, et al. Deep learning can predict microsatellite instability directly from histology in gastrointestinal cancer. Nat Med. 2019;25(7):1054–1056. doi: 10.1038/s41591-019-0462-y Lundberg SM, Lee SI. A unified approach to interpreting model predictions. In: Guyon I, et al., eds. Advances in Neural Information Processing Systems 30. Curran Associates; 2017:4765–4774. Ilse M, Tomczak J, Welling M. Attention-based deep multiple instance learning. In: Proc 35th Int Conf Mach Learn. PMLR; 2018:2127–2136. Lu MY, Williamson DFK, Chen TY, et al. Data-efficient and weakly supervised computational pathology on whole-slide images. Nat Biomed Eng. 2021;5(6):555–570. doi: 10.1038/s41551-020-00682-w Naik N, Madani A, Esteva A, et al. Deep learning-enabled flash-frozen section analysis for intraoperative cancer diagnosis. Nat Med. 2020;26(10):1576–1580. doi: 10.1038/s41591-020-1042-9 Saltz J, Gupta R, Hou L, et al. Spatial organization and molecular correlation of tumor-infiltrating lymphocytes using deep learning on pathology images. Cell Rep. 2018;23(1):181–193.e7. doi: 10.1016/j.celrep.2018.03.086 Courtiol P, Maussion C, Verrill C, et al. Deep learning-based classification of mesothelioma improves prediction of patient outcome. Nat Med. 2019;25(10):1519–1525. doi: 10.1038/s41591-019-0583-3 Plass M, Kargl M, Kiehl TR, Regitnig P, Geißler C, Evans T, et al. Explainability and causability in digital pathology. J Pathol Clin Res. 2023;9(4):251–260. doi: 10.1002/cjp2.322 Adebayo J, Gilmer J, Muelly M, et al. Sanity checks for saliency maps. In: Advances in Neural Information Processing Systems 31. Curran Associates; 2018:9505–9515. Saporta A, Le T, Singh C, et al. A systematic evaluation of interpretable deep learning methods for medical imaging. IEEE Trans Med Imaging. 2022;41(10):2672–2684. doi: 10.1109/TMI.2022.3141566 International Organization for Standardization. ISO 14971:2019 Medical devices—Application of risk management to medical devices. Geneva: ISO; 2019. Ghassemi M, Oakden-Rayner L, Beam AL. The false hope of current approaches to explainable artificial intelligence in health care. Lancet Digit Health. 2021;3(11):e745–e750. doi: 10.1016/S2589-7500(21)00208-9 Lipton ZC. The mythos of model interpretability. In: ICML Workshop on Human Interpretability in Machine Learning. PMLR; 2016:96–100. Jaume G, Song A, Chen Y, et al. Quantifying the role of visual explanations in pathology. In: Medical Imaging with Deep Learning. PMLR; 2021:337–351. Adebayo J, Muelly M, Abelson H, Kim B. Post hoc explanations may be ineffective for detecting unknown spurious correlation. In: International Conference on Learning Representations (ICLR); 2022. Rudin C, Chen C, Chen Z, Huang H, Semenova L, Zhong C. Interpretable machine learning: fundamental principles and 10 grand challenges. arXiv. 2021. doi: 10.48550/arXiv.2103.11251 Liu X, Rivera SC, Moher D, et al. Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: CONSORT-AI. Nat Med. 2020;26(9):1364–1374. doi: 10.1038/s41591-020-1034-x Nagendran M, Chen Y, Lovejoy CA, et al. Artificial intelligence versus clinicians: systematic review. BMJ. 2020;368:m689. doi: 10.1136/bmj.m689 Holzinger A. From machine learning to explainable AI. In: World Symposium on Digital Intelligence for Systems and Machines (DISA). IEEE; 2018:55–66. Selvaraju RR, Cogswell M, Das A, et al. Grad-CAM: visual explanations from deep networks via gradient-based localization. In: Proc IEEE ICCV. IEEE; 2017:618–626. Ribeiro MT, Singh S, Guestrin C. "Why should I trust you?": explaining the predictions of any classifier. In: Proc ACM SIGKDD. ACM; 2016:1135–1144. Skrede OJ, De Raedt S, Kleppe A, et al. Deep learning for prediction of colorectal cancer outcome. Lancet. 2020;395(10221):350–360. doi: 10.1016/S0140-6736(19)32998-8 Ström P, Kartasalo K, Olsson H, et al. Artificial intelligence for diagnosis and grading of prostate cancer in biopsies. Lancet Oncol. 2020;21(2):222–232. doi: 10.1016/S1470-2045(19)30738-7 Bulten W, Balkenhol M, Bel-Bahar T, et al. Artificial intelligence assistance for prostate cancer grading. Lancet Oncol. 2021;22(8):1085–1086. doi: 10.1016/S1470-2045(21)00329-8 Nagpal K, Foote D, Liu Y, et al. Development and validation of a deep learning algorithm for Gleason grading. JAMA Oncol. 2019;5(12):1730–1736. doi: 10.1001/jamaoncol.2019.3854 Fu Y, Jung AW, Torne RV, et al. Pan-cancer computational pathology by combining protein and gene expression. Nat Commun. 2020;11(1):6097. doi: 10.1038/s41467-020-19911-8 Arrieta AB, Díaz-Rodríguez N, Del Ser J, et al. Explainable Artificial Intelligence (XAI): concepts, taxonomies, opportunities and challenges. Inf Fusion. 2020;58:82–115. doi: 10.1016/j.inffus.2019.12.012 Sounderajah V, Ashrafian H, Rose S, et al. QUADAS-AI: a quality assessment tool for AI diagnostic studies. Nat Med. 2021;27(10):1663–1665. doi: 10.1038/s41591-021-01519-z Collins GS, Dhiman P, Andaur Navarro CL, et al. TRIPOD-AI protocol. BMJ Open. 2021;11(7):e048008. doi: 10.1136/bmjopen-2020-048008 Lu MY, Chen TY, Williamson DFK, Zhao M, Shady M, Lipkova J, et al. AI-based pathology predicts origins for cancers of unknown primary. Nature. 2021;594(7861):106–110. doi: 10.1038/s41586-021-03512-4 Marra G, van der Laak J, Ciompi F. A survey of deep learning in histopathology. Med Image Anal. 2022;76:102301. doi: 10.1016/j.media.2021.102301 Aung TN, Liu M, Su D, et al. Pathologist-Read vs AI-Driven Assessment of Tumor-Infiltrating Lymphocytes in Melanoma. JAMA Netw Open. 2025;8(7):e2518906. doi: 10.1001/jamanetworkopen.2025.18906 McGenity C, Re-Sugiura L, Schultheiss M, et al. Artificial intelligence for diagnostic and prognostic image analysis in cancer. Lancet Digit Health. 2024;6(8):e648–e658. doi: 10.1016/S2589-7500(24)00105-4 Zech JR, Badgeley MA, Liu M, et al. Variable generalization performance of a deep learning model. PLoS Med. 2018;15(11):e1002683. doi: 10.1371/journal.pmed.1002683 Beede E, Baylor E, Hersch F, et al. A human-centered evaluation of a deep learning system deployed in clinics for the detection of diabetic retinopathy. Proc CHI Conf Hum Factors Comput Syst. ACM; 2020:1–12. Additional Declarations No competing interests reported. Supplementary Files SupplementaryInformationnpjDM.pdf Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9249572","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":615942466,"identity":"ff0389aa-7239-436f-956e-0c3016238045","order_by":0,"name":"Bandar Alshreef","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA7ElEQVRIiWNgGAWjYBACAwYGxgOMDUAW8+EDIAEwm5AWBogWtrQEkrXkGBCnxZz/8IODP3fck+Nn4/n8mYfBRnbDAfaHH/BpsZyRZnCY90yxsWQb7zZpHoY04w0HeIwl8DrsBoPBYca2hMQN93u3MfMwHE4EamHAr+X88Q8Hf7Yl1O8/xvMY6LD/QC3sj3/g1XIgx+AAb1tCggEbDwPQYQeAWhjM8NpiOSOn4DBQi+GMY2xmknMMko1nHuYxs8CnxZz/+MaHQIfJ87cxP/7wpsJOtu94++Mb+LSguxOImUlQPwpGwSgYBaMAOwAAUUtONoi+Vw8AAAAASUVORK5CYII=","orcid":"","institution":"Shaqra University","correspondingAuthor":true,"prefix":"","firstName":"Bandar","middleName":"","lastName":"Alshreef","suffix":""},{"id":615942467,"identity":"65d00e3e-f1ff-457f-b071-0d28fc2d358d","order_by":1,"name":"Yousif Kariri","email":"","orcid":"","institution":"Shaqra University","correspondingAuthor":false,"prefix":"","firstName":"Yousif","middleName":"","lastName":"Kariri","suffix":""}],"badges":[],"createdAt":"2026-03-28 04:24:20","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9249572/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9249572/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":105985820,"identity":"19eeea48-608a-4bef-b11a-61bb773b3470","added_by":"auto","created_at":"2026-04-02 07:26:11","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":803314,"visible":true,"origin":"","legend":"\u003cp\u003ePRISMA-ScR flow diagram showing the study selection process\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-9249572/v1/0745b29f9565ce4a353bc420.png"},{"id":105985834,"identity":"805d62c1-3c9d-4d40-a159-e69e81445bdb","added_by":"auto","created_at":"2026-04-02 07:26:14","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":50319,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of included studies by cancer type (n=47).\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-9249572/v1/5056f9cde291af0e9e3dd78d.png"},{"id":105986054,"identity":"c75bcb6b-400a-4e6f-b576-f8d15191efdb","added_by":"auto","created_at":"2026-04-02 07:27:28","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":40806,"visible":true,"origin":"","legend":"\u003cp\u003eForest plot of the overall random-effects meta-analysis of diagnostic performance (AUC).\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-9249572/v1/42b5dc4bf26a2472ee19dc86.png"},{"id":105985889,"identity":"eede4df6-7c9a-4abf-8b79-4be287b3caa0","added_by":"auto","created_at":"2026-04-02 07:26:26","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":717725,"visible":true,"origin":"","legend":"\u003cp\u003eForest plot of the subgroup meta-analysis based on the unit of analysis (slide-level vs. patient-level).\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-9249572/v1/2fc1e7496b226225848ff0a4.png"},{"id":107428921,"identity":"e63b5d99-b62d-4005-a3ec-4310d7cc972d","added_by":"auto","created_at":"2026-04-21 11:57:01","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1928882,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9249572/v1/33a2768f-c8a0-4866-aa29-b18db8d84f71.pdf"},{"id":105985896,"identity":"d2d4320c-21b6-4824-95b1-9db91be01aaa","added_by":"auto","created_at":"2026-04-02 07:26:28","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":50757,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryInformationnpjDM.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9249572/v1/342670c2ae833fe9dac58e45.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Integrating Explainable Artificial Intelligence into Histopathological Risk Assessment: A Scoping Review and Meta-analysis","fulltext":[{"header":"Introduction","content":"\u003cp\u003eHistopathology continues to be the most widely accepted method for establishing a diagnosis and determining the grade of cancer. The method requires an experienced pathologist to examine numerous morphological features of the tumor to determine its type and aggressiveness. Although pathologists are experts in examining histologic material, the variability in interpretation among pathologists poses one of the greatest challenges to the establishment of a reliable method for establishing a diagnosis. A number of studies have reported variable levels of concordance among pathologists when they were asked to interpret the same histologic slides, suggesting the presence of diagnostic bias [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Digital pathology has introduced the possibility of using artificial intelligence (AI) to aid and enhance the evaluation of histopathological materials [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eRecent studies have shown that AI-based algorithms can provide clinical-grade performance in identifying tumors, classifying them, and predicting prognoses in several types of cancers (e.g., prostate, breast, colon), and at least in some cases perform equally well, if not better, than human observers [\u003cspan additionalcitationids=\"CR36 CR37\" citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e, \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e]. AI may be able to recognize small patterns in images that cannot be detected by the human eye and possibly even predict molecular mutations based upon the histological findings of a biopsy sample, thereby improving the accuracy of the stratification of patients by risk [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]. While there have been positive outcomes from the use of AI in evaluating histopathological materials, the translation of this technology into clinical practice is hindered by barriers. Systematic reviews of studies that compare the performance of AI-based models to those of clinicians generally report that the two are equivalent; however, the high risk of bias in many of these studies limits the ability to draw definitive conclusions regarding the superiority of AI-based models [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e]. Additionally, the lack of clarity in how many deep-learning-based models arrive at their decisions (i.e., the \"black box\" nature of many of these models) makes it difficult for pathologists to understand, validate, and rely upon the outputs of these models [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eLimited interpretability of many AI models is a significant obstacle to the safe and effective integration of these models into clinical settings. As a result, the subfield of Explainable AI (XAI) has developed to improve the transparency, accountability, and reliability of AI models [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e]. In histopathology, XAI could enable pathologists to verify the logic used by an AI model to arrive at its conclusion(s) and potentially identify possible sources of error or bias [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Therefore, while the objectives of XAI in histopathology extend beyond simply explaining an AI model (explainability) to include supporting and enhancing the pathologist's diagnostic reasoning through providing causal explanations (causality) [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn addition to the need for XAI, the integration of AI models into clinical laboratory settings requires compliance with strict quality and safety regulations, such as ISO 15189 (Medical Laboratories\u0026mdash;Requirements for Quality and Competence) and ISO 14971 (Application of Risk Management to Medical Devices) [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Compliance with these standards, along with reporting requirements for clinical trials such as CONSORT-AI and TRIPOD-AI, during all stages of the AI development cycle is necessary to ensure the responsible implementation of AI [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e, \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThis review examines the critical relationship between XAI, risk stratification, and quality management. Through a systematic review of the existing literature, combined with a meta-analysis of the diagnostic performance of AI models, we will identify key obstacles and develop a structured framework to inform future research efforts and facilitate the transition of reliable and causal AI-based tools into clinical practice.\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eStudy selection and characteristics\u003c/h2\u003e \u003cp\u003eOur initial search yielded 857 studies. After removing duplicates, 642 articles were screened, and 123 full-text papers were assessed for eligibility. A final set of 47 studies met our inclusion criteria. The study selection process is detailed in the PRISMA flow diagram (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eCharacteristics of included studies\u003c/h3\u003e\n\u003cp\u003eThe majority of the research studies contained herein were conducted on many different types of cancers; specifically, breast (34%), prostate (21%), lung (15%), and colorectal (11%) were the most prevalent (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). It is probable that the high prevalence of studies concerning breast and prostate cancers (combined\u0026thinsp;=\u0026thinsp;55%) is due to the existence of large public data sets from contests such as CAMELYON16. Additionally, it appears that the use of post-hoc visual explanations for image-based XAI methods (i.e., CAMs, Grad-CAM) in this type of image classification task has created a void in explanation methodologies for more complex and multi-modal predictive modeling.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003eKey concepts and definitions\u003c/h3\u003e\n\u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eExplainability: The capability to give a human-readable description of how an AI model works internally [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e].\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eInterpretability: A measure of how well a human can identify why an AI model made a particular choice [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e].\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eCausality: An evaluation of the level to which an explanation of a decision made by an AI model provides a level of causal insight to a human expert so that the expert has the information necessary for making decisions [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e].\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eTransparency: The ability for a human to view how an AI model operates inside, usually only possible when the model is relatively simple (e.g., linear regression, decision tree) [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e].\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eThe concepts above are often confused or used interchangeably; however, they constitute a spectrum of technical transparency to clinical interpretation. As this review examines how XAI techniques can produce not only technical descriptions of an AI model but also clinically relevant insights into an AI model that assist in a pathologist\u0026rsquo;s diagnostic reasoning process.\u003c/p\u003e\n\u003ch3\u003eXAI techniques in histopathology\u003c/h3\u003e\n\u003cp\u003eThe studies reviewed use a wide range of XAI techniques that fall into four categories:\u003c/p\u003e\n\u003ch3\u003ePost-hoc Visual Explanations\u003c/h3\u003e\n\u003cp\u003eThese techniques provide a visual explanation after a model has been trained, and they show the areas of the image that had the largest impact on the decision made by the model.\u003c/p\u003e \u003cp\u003eTechniques used for these types of visual explanations include:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eClass Activation Mapping (CAM): CAM projects the weight of the output layer backwards onto the convolutional feature maps that generated the output [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e].\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eGradient-weighted Class Activation Mapping (Grad-CAM): Grad-CAM is a generalized version of CAM, using the gradient flow through the final convolutional layer to create a coarse localization map [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eSaliency Maps: Saliency maps highlight the pixels with the highest contribution to the model\u0026rsquo;s prediction. [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eLayer-wise Relevance Propagation (LRP): LRP decomposes the prediction of a deep neural network by reversing the propagation of the relevance of the output back to the input features [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eAttention Mechanisms\u003c/h2\u003e \u003cp\u003eUnlike post-hoc explanations that are added to a pre-trained model after it has been developed, attention-based models incorporate the ability to focus on the most relevant aspects of the input data into the design of the model itself. Attention-based models have been widely adopted for MIL-based approaches to identifying the most informative regions in a whole-slide image of histological tissue [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e], [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eConcept-based Explanations\u003c/h3\u003e\n\u003cp\u003eRather than explaining how a model makes a decision based on the low-level features present in the input data (e.g., pixel values), concept-based explanations describe how a model makes decisions based on higher-level concepts (e.g., presence of certain tissue structure or cellular pattern) that are meaningful to humans [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e\n\u003ch3\u003eInherently Interpretable Models\u003c/h3\u003e\n\u003cp\u003eThere are some researchers who advocate for using inherently interpretable models (e.g., decision trees, linear models, etc.) in place of using post-hoc explanations for complex black-box models [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e][\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e].\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eApplications of XAI in histopathological risk assessment\u003c/h2\u003e \u003cp\u003eThe included studies applied XAI to a range of risk assessment tasks, including:\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eTumor Detection and Identification\u003c/h2\u003e \u003cp\u003eXAI is currently being applied to explain how Artificial Intelligence models are detecting and classifying tumors across different parts of the body (i.e. breast, prostate, lungs, colon). As one example, Grad-CAM is being used to display the exact regions of a histopathology slide where an Artificial Intelligence model determines benign vs. malignant tissue [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eGrading \u0026amp; Staging\u003c/h2\u003e \u003cp\u003eThere have been attempts to develop AI models that will grade tumors automatically (for example, Gleason grading of prostate cancer) and stage diseases; XAI techniques may be able to help pathologists understand why they have received those results and where there may be errors [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e, \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e, \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003cb\u003eThe Tumor Micro-Environment (TME).\u003c/b\u003e \u003c/p\u003e \u003cp\u003eXAI is being utilized to determine the complexities involved within the interactions of tumor cells with their micro-environmental surroundings. Included in this are the identification and quantification of tumor infiltrating lymphocytes (TILs), which have direct implications for predicting patient responses to immunotherapies. For instance, attention based models are capable of highlighting areas of high TIL density, that correlate with patient outcomes [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003ePredicting Molecular Markers\u003c/h2\u003e \u003cp\u003eAn additional area where XAI is being utilized is in explaining models that predict molecular characteristics such as MSI or gene mutations directly from H\u0026amp;E stained slides. Heat maps provide an opportunity to show the morphological characteristics that a model associates with a particular molecular subtype and therefore provides a link between histology and genomics [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eQuality safety and workflow integration\u003c/h2\u003e \u003cp\u003eEnsuring that the development of AI in pathology adheres to quality guidelines such as ISO 15189 and ISO 14971 is critical to ensure the safety of developing AI in pathology [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. Documentation of XAI artifacts, e.g., as heat maps or conceptual attributions as part of an ISO 15189-compliant quality management system, and use as risk control elements in an ISO 14971-compliant risk assessment may provide assurance that risks associated with XAI have been mitigated. For example, a validation report documenting the agreement between a model's heatmap representation and a pathologist's annotations of a region of interest could help demonstrate that a model\u0026rsquo;s performance was robust when providing a diagnosis.\u003c/p\u003e \u003cp\u003eThe development of XAI systems, however, requires a user-centric approach. Explanations provided by poorly designed XAI systems can create additional cognitive burdens and disrupt clinical workflows [\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e]. Thus successful integration will require a focus on user-centric design, seamless integration into existing clinical workflows, minimal cognitive burden, and integration into existing hospital information systems. Considerations related to the display of XAI results (e.g., an overlay on a digital slide) and the interaction with XAI results (e.g., the ability to turn XAI explanations on or off) are key considerations, as well as the medicolegal ramifications of using AI to assist in diagnosis [\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e]. The roles of XAI across the AI lifecycle, aligned with relevant quality standards, are summarized in Table 1.\u0026rdquo;\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eIllustrative mapping of XAI roles to the AI lifecycle, aligned with quality standards.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLifecycle Phase\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eXAI Role\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eQuality Standard\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eDevelopment \u0026amp; Training\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIdentify and mitigate bias; validate feature relevance; ensure model learns clinically meaningful patterns\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eISO 14971 (Risk Analysis)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eValidation \u0026amp; Verification\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDocument concordance between model explanations and expert annotations; support internal validation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eISO 15189 (QMS)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eDeployment \u0026amp; Monitoring\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eProvide real-time explanations to support clinical decision-making; enable continuous performance monitoring\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eISO 15189 (QMS)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePost-Market Surveillance\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFacilitate investigation of adverse events or unexpected outcomes; support root cause analysis\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eISO 14971 (Risk Control)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eQuality assessment\u003c/h2\u003e \u003cp\u003eIn this evaluation of the diagnostic AI literature, we applied a quality assessment tool developed specifically for this area of research (see Supplementary File 5) to assess the 47 identified studies. In assessing the quality of these studies, we determined that 18 (38%) of them represented studies at low risk of bias, whereas 22 (47%) were considered to be at moderate risk, and 7 (15%) studies were classified as being at high risk of bias. The primary factors that contributed to the high risk of bias in many studies were the inclusion of non-sequential cases, the absence of an independent validation dataset, and the ambiguity surrounding how models were split into training datasets and test datasets. The results of our quality assessment further emphasize the need for studies evaluating diagnostic AI to have a higher level of study design rigor; this is consistent with the conclusions drawn from recent meta-analytic reviews [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eQuantitative synthesis\u003c/h2\u003e \u003cp\u003eWe selected 9 out of the 47 studies included in the scoping review that met the eligibility requirements for inclusion in the diagnostic meta-analysis. Those studies provided the necessary data to calculate the AUC with its respective 95% confidence interval. When all 9 studies were analyzed using a random-effects meta-analysis model, we observed a pooled AUC of 0.962 (95% CI: 0.909\u0026ndash;0.985). However, the analysis indicated substantial heterogeneity among the studies (I\u0026sup2; = 97.1%, τ\u0026sup2; = 1.89, Q\u0026thinsp;=\u0026thinsp;278.58, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001); this indicates substantial variability in model performance across the studies. The overall random-effects meta-analysis is presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe corrected subgroup analysis by the unit of evaluation was a major cause of the heterogeneity noted above. Studies that evaluated performance at the individual slide level demonstrated higher and more consistent pooled AUC values of 0.987 (95% CI: 0.981\u0026ndash;0.990) as well as less heterogeneity (I\u0026sup2;=12.5%) than did those that evaluated performance at the patient level. Studies that evaluated performance at the patient level had lower pooled AUC values of 0.910 (95% CI: 0.830\u0026ndash;0.955) and much greater heterogeneity (I\u0026sup2;=93.1%). These findings are depicted in the revised forest plot (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e) and suggest that the method used for combining predictions made from the slide to the patient level has an impact on both how well a model performs and how consistently different studies report performance.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe results were confirmed by a sensitivity analysis that excluded the four slide-level studies. The pooled estimate of the remaining five patient-level studies was 0.910 (95% CI: 0.830\u0026ndash;0.955) and is consistent with the patient-level subgroup results.\u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis systematic review has mapped the present state of XAI in histopathological risk assessment, which has revealed a promising prospect with several challenges. In addition, a quantitative meta-analysis was added to provide a performance standard that showed that, although pooled diagnostic accuracy is high (AUC = 0.962), the field is highly heterogeneous (I² = 97.1%) as the two most important sources of this variation are identified by our study as the unit of evaluation—slide level vs. patient level.\u003c/p\u003e \u003cp\u003eWe believe the greatest impact of our study is the distinction between the need for post-hoc explanations and the use of inherently interpretable models in high-risk clinical decision making. We found a growing consensus that causal interpretations must be the central focus of the development of XAI. Although AI has been very successful in many areas, it has also shown poor generalizability to external datasets, particularly when there has been a distribution shift from the training dataset (e.g., due to differences in patient populations or slide preparation) [\u003cspan class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e46\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e47\u003c/span\u003e]. At the time of writing, there is no agreement on what metrics should be used to measure either the quality or the consistency of AI-generated explanations [\u003cspan class=\"CitationRef\"\u003e22\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e33\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e40\u003c/span\u003e]. In fact, several studies have shown that some of the most popular methods (such as saliency maps) can be misleading [\u003cspan class=\"CitationRef\"\u003e28\u003c/span\u003e], so that some researchers argue that current XAI methods create a false sense of security and therefore, inherently interpretable models should be the priority for all high-risk applications [\u003cspan class=\"CitationRef\"\u003e29\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e34\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn addition, the clinical utility of XAI remains largely unproven. It is still unknown how explanations will be used in routine practice and if they will improve diagnostic accuracy or patient outcomes. One of the largest challenges will be to meet the multiple needs of different stakeholders; for example, pathologists may want explanations to increase their diagnostic confidence, while regulatory bodies will require auditable evidence of a system's safety and effectiveness [\u003cspan class=\"CitationRef\"\u003e25\u003c/span\u003e]. To be clinically relevant, explanations must be framed within the context of domain-specific knowledge and not simply meet a technical requirement [\u003cspan class=\"CitationRef\"\u003e25\u003c/span\u003e]. As an example, in prostate Gleason grading, a heatmap that highlights stromal regions without glandular relevance illustrates the gap between a technically correct explanation and clinical causality. A more causal explanation would highlight the specific atypical glandular formations that contributed to a higher Gleason score and link the AI's output to the pathologist's diagnostic criteria. These concerns are particularly relevant for clinical deployment, where a lack of causal explanations can undermine trust and hinder adoption. Our meta-analysis further highlights that methodological choices—such as slide-level versus patient-level evaluation—directly affect reported performance, underscoring the need for standardized evaluation protocols that reflect real-world clinical workflows.\u003c/p\u003e \u003cp\u003eTo serve as a guideline for future research, we have proposed a four-stage roadmap to move the application of Explainable AI (XAI) into histopathology:\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003ePathologist-centric usability studies: Use usability study techniques such as cognitive walkthroughs or think aloud to measure how XAI will influence diagnostic efficiency, diagnostic confidence, and inter-rater reliability.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eProspective clinical validation: Design and conduct prospective clinical trials to assess the real-world effects of using XAI to support the clinical workflow in terms of patient outcome.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eStandard evaluation metrics: Establish and validate standardization around metrics that can be used to compare the effectiveness of XAI explanations in terms of both quality and consistency.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003ePublicly available benchmarking and dataset: Establish publicly available benchmarking and datasets for comparing XAI algorithms in histopathology.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003cp\u003e\u003c/p\u003e \u003cp\u003eThis review has several limitations. We restricted inclusion to English-language studies, which may have introduced language bias, and only nine studies contributed data to the primary meta-analysis. Because explainable AI evolves rapidly, relevant methods published after the search date may not have been captured. We also excluded grey literature and conference abstracts, and our modified quality assessment approach has not yet been formally validated for all explainability-focused pathology studies.\u003c/p\u003e \u003cp\u003eOverall, explainable AI in histopathology shows strong but heterogeneous diagnostic performance and remains methodologically immature at the point of clinical translation. The next phase of the field should prioritize patient-level evaluation, clinically meaningful explanations, pathologist-centered usability studies, prospective validation, and closer alignment with quality and risk-management frameworks. Ultimately, this roadmap will accelerate the responsible translation of XAI into clinical practice—delivering trustworthy AI tools that support pathologists, improve diagnostic accuracy, and enhance patient outcomes.\u003c/p\u003e "},{"header":"Methods","content":"\u003cp\u003ePrimary research studies published in an English-language journal that include:\u003c/p\u003e\n\u003col\u003e\n\u003cli\u003eHistopathology (whole-slide images or otherwise) in an oncologic context\u003c/li\u003e\n\u003cli\u003eXAI approaches applied to model outputs that are relevant to risk-assessment (e.g., prognosis, grading, stratification)\u003c/li\u003e\n\u003cli\u003eDetails about how they implemented their methodology to allow us to extract which XAI approach and task were utilized.\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eStudies that do not relate to oncologic conditions or studies that are not research-based (editorial, review, etc.) are excluded.\u003c/p\u003e\n\n\u003cp\u003eStudy selection\u003c/p\u003e\n\u003cp\u003eTwo independent reviewers screened both title/abstracts and full-text versions of articles, with disagreements resolved through consensus.\u003c/p\u003e\n\n\u003cp\u003eData charting and synthesis\u003c/p\u003e\n\u003cp\u003eThe following information was abstracted: author/year, cancer type, task, AI architecture, XAI method, dataset, validation approach, and key findings, after which the evidence was synthesized thematically.\u003c/p\u003e\n\u003cp\u003eStatistical analysis\u003c/p\u003e\n\u003cp\u003eIn order to quantify the diagnostic performance of AI models, a random effects meta-analysis was performed. Priority was given to studies that reported the area under the receiver operating characteristic curve (AUC) with 95% confidence intervals (CIs). Confidence interval bounds were used to derive logit-transformed AUCs and their variances. Pooled estimates and 95% CIs were calculated using the DerSimonian-Laird method with a restricted maximum likelihood estimator, then back-transformed to the AUC scale for interpretation. Heterogeneity was assessed using Cochran\u0026apos;s Q test and the I\u0026sup2; statistic. Pre-specified subgroup analyses compared slide-level and patient-level evaluation, and a sensitivity analysis excluding slide-level studies was also performed. All analyses were conducted in Python using the pandas and scipy libraries.\u003c/p\u003e\n\u003cp\u003eData availability\u003c/p\u003e\n\u003cp\u003eAll data analyzed in this study were derived from published articles and are reported in the manuscript and its Supplementary Information. The extracted study-level dataset used for the meta-analysis is available from the corresponding author upon reasonable request.\u003c/p\u003e\n\u003cp\u003eCode availability\u003c/p\u003e\n\u003cp\u003eThe statistical code used for the quantitative synthesis was developed in Python and is available from the corresponding author upon reasonable request.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eFunding\u003c/p\u003e\n\u003cp\u003eNo specific funding was received for this work.\u003c/p\u003e\n\n\n\u003cp\u003eAcknowledgments\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eThe authors would like to thank the Deanship of Scientific Research at Shaqra University for supporting this work.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAuthor contributions\u003c/p\u003e\n\u003cp\u003eB.A. conceived the study. B.A. and Y.A.K. designed the review, screened records, extracted and interpreted data, and revised the manuscript. B.A. performed the statistical analyses and drafted the manuscript. Y.A.K. critically reviewed the manuscript and approved the final version. All authors approved the submitted manuscript.\u003c/p\u003e\n\u003cp\u003eCompeting interests\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003evan der Velden BHM, Kuijf HJ, Gilhuijs KGA, Viergever MA. Explainable artificial intelligence (XAI) in deep learning-based medical image analysis. Med Image Anal. 2022;79:102470. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.media.2022.102470\u003c/span\u003e\u003cspan address=\"10.1016/j.media.2022.102470\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBera K, Schalper KA, Rimm DL, Velcheti V, Madabhushi A. Artificial intelligence in digital pathology\u0026mdash;new tools for diagnosis and precision oncology. Nat Rev Clin Oncol. 2019;16(11):703\u0026ndash;715. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41571-019-0252-y\u003c/span\u003e\u003cspan address=\"10.1038/s41571-019-0252-y\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHolzinger A, Langs G, Denk H, Zatloukal K, M\u0026uuml;ller H. Causability and explainability of artificial intelligence in medicine. Wiley Interdiscip Rev Data Min Knowl Discov. 2019;9(4):e1312. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/widm.1312\u003c/span\u003e\u003cspan address=\"10.1002/widm.1312\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRudin C. Stop explaining black box machine learning models for high-stakes decisions and use interpretable models instead. Nat Mach Intell. 2019;1(5):206\u0026ndash;215. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s42256-019-0048-x\u003c/span\u003e\u003cspan address=\"10.1038/s42256-019-0048-x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCoudray N, Ocampo PS, Sakellaropoulos T, et al. Classification and mutation prediction from non\u0026ndash;small cell lung cancer histopathology images using deep learning. Nat Med. 2018;24(10):1559\u0026ndash;1567. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41591-018-0177-5\u003c/span\u003e\u003cspan address=\"10.1038/s41591-018-0177-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSamek W, Wiegand T, M\u0026uuml;ller KR. Explainable artificial intelligence: understanding, visualizing and interpreting deep learning models. arXiv. 2017. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.48550/arXiv.1708.08296\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.1708.08296\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAdadi A, Berrada M. Peeking inside the black-box: a survey on explainable artificial intelligence (XAI). IEEE Access. 2018;6:52138\u0026ndash;52160. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/ACCESS.2018.2870052\u003c/span\u003e\u003cspan address=\"10.1109/ACCESS.2018.2870052\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen Z, Bei Z, Rudin C. Concept whitening for interpretable image recognition. Nat Mach Intell. 2020;2:772\u0026ndash;782. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s42256-020-00265-z\u003c/span\u003e\u003cspan address=\"10.1038/s42256-020-00265-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eInternational Organization for Standardization. ISO 15189:2022 Medical laboratories\u0026mdash;Requirements for quality and competence. Geneva: ISO; 2022.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEhteshami Bejnordi B, Veta M, van Diest PJ, et al. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. JAMA. 2017;318(22):2199\u0026ndash;2210. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1001/jama.2017.14585\u003c/span\u003e\u003cspan address=\"10.1001/jama.2017.14585\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCampanella G, Hanna MG, Geneslaw L, et al. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nat Med. 2019;25(8):1301\u0026ndash;1309. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41591-019-0508-1\u003c/span\u003e\u003cspan address=\"10.1038/s41591-019-0508-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEchle A, Grabsch HI, Quirke P, et al. Clinical-grade detection of microsatellite instability in colorectal tumors by deep learning. Gastroenterology. 2020;159(4):1406\u0026ndash;1416.e11. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1053/j.gastro.2020.06.021\u003c/span\u003e\u003cspan address=\"10.1053/j.gastro.2020.06.021\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBulten W, Kartasalo K, Chen PH, et al. Artificial intelligence for diagnosis and Gleason grading of prostate cancer: the PANDA challenge. Nat Med. 2022;28(1):154\u0026ndash;163. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41591-021-01620-2\u003c/span\u003e\u003cspan address=\"10.1038/s41591-021-01620-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKather JN, Pearson AT, Halama N, et al. Deep learning can predict microsatellite instability directly from histology in gastrointestinal cancer. Nat Med. 2019;25(7):1054\u0026ndash;1056. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41591-019-0462-y\u003c/span\u003e\u003cspan address=\"10.1038/s41591-019-0462-y\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLundberg SM, Lee SI. A unified approach to interpreting model predictions. In: Guyon I, et al., eds. Advances in Neural Information Processing Systems 30. Curran Associates; 2017:4765\u0026ndash;4774.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIlse M, Tomczak J, Welling M. Attention-based deep multiple instance learning. In: Proc 35th Int Conf Mach Learn. PMLR; 2018:2127\u0026ndash;2136.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLu MY, Williamson DFK, Chen TY, et al. Data-efficient and weakly supervised computational pathology on whole-slide images. Nat Biomed Eng. 2021;5(6):555\u0026ndash;570. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41551-020-00682-w\u003c/span\u003e\u003cspan address=\"10.1038/s41551-020-00682-w\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNaik N, Madani A, Esteva A, et al. Deep learning-enabled flash-frozen section analysis for intraoperative cancer diagnosis. Nat Med. 2020;26(10):1576\u0026ndash;1580. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41591-020-1042-9\u003c/span\u003e\u003cspan address=\"10.1038/s41591-020-1042-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSaltz J, Gupta R, Hou L, et al. Spatial organization and molecular correlation of tumor-infiltrating lymphocytes using deep learning on pathology images. Cell Rep. 2018;23(1):181\u0026ndash;193.e7. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.celrep.2018.03.086\u003c/span\u003e\u003cspan address=\"10.1016/j.celrep.2018.03.086\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCourtiol P, Maussion C, Verrill C, et al. Deep learning-based classification of mesothelioma improves prediction of patient outcome. Nat Med. 2019;25(10):1519\u0026ndash;1525. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41591-019-0583-3\u003c/span\u003e\u003cspan address=\"10.1038/s41591-019-0583-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePlass M, Kargl M, Kiehl TR, Regitnig P, Gei\u0026szlig;ler C, Evans T, et al. Explainability and causability in digital pathology. J Pathol Clin Res. 2023;9(4):251\u0026ndash;260. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/cjp2.322\u003c/span\u003e\u003cspan address=\"10.1002/cjp2.322\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAdebayo J, Gilmer J, Muelly M, et al. Sanity checks for saliency maps. In: Advances in Neural Information Processing Systems 31. Curran Associates; 2018:9505\u0026ndash;9515.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSaporta A, Le T, Singh C, et al. A systematic evaluation of interpretable deep learning methods for medical imaging. IEEE Trans Med Imaging. 2022;41(10):2672\u0026ndash;2684. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/TMI.2022.3141566\u003c/span\u003e\u003cspan address=\"10.1109/TMI.2022.3141566\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eInternational Organization for Standardization. ISO 14971:2019 Medical devices\u0026mdash;Application of risk management to medical devices. Geneva: ISO; 2019.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGhassemi M, Oakden-Rayner L, Beam AL. The false hope of current approaches to explainable artificial intelligence in health care. Lancet Digit Health. 2021;3(11):e745\u0026ndash;e750. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/S2589-7500(21)00208-9\u003c/span\u003e\u003cspan address=\"10.1016/S2589-7500(21)00208-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLipton ZC. The mythos of model interpretability. In: ICML Workshop on Human Interpretability in Machine Learning. PMLR; 2016:96\u0026ndash;100.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJaume G, Song A, Chen Y, et al. Quantifying the role of visual explanations in pathology. In: Medical Imaging with Deep Learning. PMLR; 2021:337\u0026ndash;351.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAdebayo J, Muelly M, Abelson H, Kim B. Post hoc explanations may be ineffective for detecting unknown spurious correlation. In: International Conference on Learning Representations (ICLR); 2022.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRudin C, Chen C, Chen Z, Huang H, Semenova L, Zhong C. Interpretable machine learning: fundamental principles and 10 grand challenges. arXiv. 2021. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.48550/arXiv.2103.11251\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2103.11251\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu X, Rivera SC, Moher D, et al. Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: CONSORT-AI. Nat Med. 2020;26(9):1364\u0026ndash;1374. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41591-020-1034-x\u003c/span\u003e\u003cspan address=\"10.1038/s41591-020-1034-x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNagendran M, Chen Y, Lovejoy CA, et al. Artificial intelligence versus clinicians: systematic review. BMJ. 2020;368:m689. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1136/bmj.m689\u003c/span\u003e\u003cspan address=\"10.1136/bmj.m689\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHolzinger A. From machine learning to explainable AI. In: World Symposium on Digital Intelligence for Systems and Machines (DISA). IEEE; 2018:55\u0026ndash;66.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSelvaraju RR, Cogswell M, Das A, et al. Grad-CAM: visual explanations from deep networks via gradient-based localization. In: Proc IEEE ICCV. IEEE; 2017:618\u0026ndash;626.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRibeiro MT, Singh S, Guestrin C. \"Why should I trust you?\": explaining the predictions of any classifier. In: Proc ACM SIGKDD. ACM; 2016:1135\u0026ndash;1144.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSkrede OJ, De Raedt S, Kleppe A, et al. Deep learning for prediction of colorectal cancer outcome. Lancet. 2020;395(10221):350\u0026ndash;360. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/S0140-6736(19)32998-8\u003c/span\u003e\u003cspan address=\"10.1016/S0140-6736(19)32998-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStr\u0026ouml;m P, Kartasalo K, Olsson H, et al. Artificial intelligence for diagnosis and grading of prostate cancer in biopsies. Lancet Oncol. 2020;21(2):222\u0026ndash;232. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/S1470-2045(19)30738-7\u003c/span\u003e\u003cspan address=\"10.1016/S1470-2045(19)30738-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBulten W, Balkenhol M, Bel-Bahar T, et al. Artificial intelligence assistance for prostate cancer grading. Lancet Oncol. 2021;22(8):1085\u0026ndash;1086. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/S1470-2045(21)00329-8\u003c/span\u003e\u003cspan address=\"10.1016/S1470-2045(21)00329-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNagpal K, Foote D, Liu Y, et al. Development and validation of a deep learning algorithm for Gleason grading. JAMA Oncol. 2019;5(12):1730\u0026ndash;1736. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1001/jamaoncol.2019.3854\u003c/span\u003e\u003cspan address=\"10.1001/jamaoncol.2019.3854\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFu Y, Jung AW, Torne RV, et al. Pan-cancer computational pathology by combining protein and gene expression. Nat Commun. 2020;11(1):6097. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41467-020-19911-8\u003c/span\u003e\u003cspan address=\"10.1038/s41467-020-19911-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eArrieta AB, D\u0026iacute;az-Rodr\u0026iacute;guez N, Del Ser J, et al. Explainable Artificial Intelligence (XAI): concepts, taxonomies, opportunities and challenges. Inf Fusion. 2020;58:82\u0026ndash;115. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.inffus.2019.12.012\u003c/span\u003e\u003cspan address=\"10.1016/j.inffus.2019.12.012\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSounderajah V, Ashrafian H, Rose S, et al. QUADAS-AI: a quality assessment tool for AI diagnostic studies. Nat Med. 2021;27(10):1663\u0026ndash;1665. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41591-021-01519-z\u003c/span\u003e\u003cspan address=\"10.1038/s41591-021-01519-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCollins GS, Dhiman P, Andaur Navarro CL, et al. TRIPOD-AI protocol. BMJ Open. 2021;11(7):e048008. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1136/bmjopen-2020-048008\u003c/span\u003e\u003cspan address=\"10.1136/bmjopen-2020-048008\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLu MY, Chen TY, Williamson DFK, Zhao M, Shady M, Lipkova J, et al. AI-based pathology predicts origins for cancers of unknown primary. Nature. 2021;594(7861):106\u0026ndash;110. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41586-021-03512-4\u003c/span\u003e\u003cspan address=\"10.1038/s41586-021-03512-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMarra G, van der Laak J, Ciompi F. A survey of deep learning in histopathology. Med Image Anal. 2022;76:102301. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.media.2021.102301\u003c/span\u003e\u003cspan address=\"10.1016/j.media.2021.102301\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAung TN, Liu M, Su D, et al. Pathologist-Read vs AI-Driven Assessment of Tumor-Infiltrating Lymphocytes in Melanoma. JAMA Netw Open. 2025;8(7):e2518906. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1001/jamanetworkopen.2025.18906\u003c/span\u003e\u003cspan address=\"10.1001/jamanetworkopen.2025.18906\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcGenity C, Re-Sugiura L, Schultheiss M, et al. Artificial intelligence for diagnostic and prognostic image analysis in cancer. Lancet Digit Health. 2024;6(8):e648\u0026ndash;e658. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/S2589-7500(24)00105-4\u003c/span\u003e\u003cspan address=\"10.1016/S2589-7500(24)00105-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZech JR, Badgeley MA, Liu M, et al. Variable generalization performance of a deep learning model. PLoS Med. 2018;15(11):e1002683. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1371/journal.pmed.1002683\u003c/span\u003e\u003cspan address=\"10.1371/journal.pmed.1002683\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBeede E, Baylor E, Hersch F, et al. A human-centered evaluation of a deep learning system deployed in clinics for the detection of diabetic retinopathy. Proc CHI Conf Hum Factors Comput Syst. ACM; 2020:1\u0026ndash;12.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-9249572/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9249572/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eExplainable artificial intelligence is increasingly used to support cancer detection, grading, and prognosis from histopathology, yet clinical adoption remains limited by uncertainty about reliability, interpretability, and governance. We conducted a scoping review and meta-analysis of peer-reviewed studies applying explainable AI to histopathology-based cancer risk assessment. Searches of PubMed, Google Scholar, Scopus, and Web of Science identified 47 eligible studies. Post hoc visual methods, particularly CAM and Grad-CAM, dominated the field. Nine studies contributed to meta-analysis, yielding a pooled area under the curve of 0.962 (95% CI 0.909\u0026ndash;0.985) with substantial heterogeneity (I\u0026sup2; = 97.1%). Subgroup analysis showed higher and more consistent performance at the slide level than at the patient level, identifying unit of analysis as a major source of heterogeneity. These findings support a translational roadmap for clinically meaningful explanation, usability testing, prospective validation, and governance-aligned deployment.\u003c/p\u003e","manuscriptTitle":"Integrating Explainable Artificial Intelligence into Histopathological Risk Assessment: A Scoping Review and Meta-analysis","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-02 07:24:13","doi":"10.21203/rs.3.rs-9249572/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"5908cd28-169e-4d19-af3b-34a2821ee1f0","owner":[],"postedDate":"April 2nd, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":65545056,"name":"Biological sciences/Cancer"},{"id":65545057,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":65545058,"name":"Health sciences/Medical research"},{"id":65545059,"name":"Health sciences/Oncology"}],"tags":[],"updatedAt":"2026-04-21T11:56:09+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-02 07:24:13","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9249572","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9249572","identity":"rs-9249572","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}