Artificial Intelligence Applications in Hepatology

Over the past 2 decades, the field of hepatology has witnessed major developments in diagnostic tools, prognostic models, and treatment options making it one of the most complex medical subspecialties. Through artificial intelligence (AI) and machine learning, computers are now able to learn from complex and diverse clinical datasets to solve real-world medical problems with performance that surpasses that of physicians in certain areas. AI algorithms are currently being implemented in liver imaging, interpretation of liver histopathology, noninvasive tests, prediction models, and more. In this review, we provide a summary of the state of AI in hepatology and discuss current challenges for large-scale implementation including some ethical aspects. We emphasize to the readers that most AI-based algorithms that are discussed in this review are still considered in early development and their utility and impact on patient outcomes still need to be assessed in future large-scale and inclusive studies. Our vision is that the use of AI in hepatology will enhance physician performance, decrease the burden and time spent on documentation, and reestablish the personalized patient-physician relationship that is of utmost importance for obtaining good outcomes. Over the past 2 decades, the field of hepatology has witnessed major developments in diagnostic tools, prognostic models, and treatment options making it one of the most complex medical subspecialties. Through artificial intelligence (AI) and machine learning, computers are now able to learn from complex and diverse clinical datasets to solve real-world medical problems with performance that surpasses that of physicians in certain areas. AI algorithms are currently being implemented in liver imaging, interpretation of liver histopathology, noninvasive tests, prediction models, and more. In this review, we provide a summary of the state of AI in hepatology and discuss current challenges for large-scale implementation including some ethical aspects. We emphasize to the readers that most AI-based algorithms that are discussed in this review are still considered in early development and their utility and impact on patient outcomes still need to be assessed in future large-scale and inclusive studies. Our vision is that the use of AI in hepatology will enhance physician performance, decrease the burden and time spent on documentation, and reestablish the personalized patient-physician relationship that is of utmost importance for obtaining good outcomes. Technological advancements have created the unique opportunity to use artificial intelligence (AI) and more specifically machine learning (ML) in clinical medicine. With available computational power, AI has the potential to transform patient care without losing the patient-centric, physician-guided approach of traditional clinical medicine. This has become even more evident during the COVID-19 pandemic, which provided unprecedented advancements in technology acceptance and availability in all areas of society and in particular in the health care system. Although AI is considered the overarching term that details the rational exploitation of data by a machine, ML more specifically describes the building of models that learn from available data to improve the prediction or performance related to a specific task without actually programming.1Mintz Y. Brodie R. Introduction to artificial intelligence in medicine.Minim Invasive Ther Allied Technol. 2019; 28: 73-81Crossref PubMed Scopus (160) Google Scholar The ability of ML algorithms to predict outcomes can be exploited based on labeled (supervised) or unlabeled (unsupervised) data. By using the ML algorithm over time and providing more training, the desired output becomes progressively more accurate. Deep learning (DL) refines and narrows ML by using multiple neuronal networks that mimic the human neurologic system to analyze, identify, and learn from complex datasets.2Esteva A. Robicquet A. Ramsundar B. et al.A guide to deep learning in healthcare.Nat Med. 2019; 25: 24-29Crossref PubMed Scopus (1343) Google Scholar The neural networks are organized in multiple layers where the signal travels from the first layer (input) to the last layer (output) after going through multiple intervening layers. A few important issues have to be considered when aiming to implement AI in the clinical environment today (Figure 1). Beyond investments in technology in the health care sector, the quality of the data that are used to develop algorithms and predict outcome is most critical. In the field of hepatology research, several large prospective studies that are aimed to explore outcome are actively recruiting and will provide the quality and robustness of the data that are required.3Sanyal A.J. Shankar S.S. Calle R.A. et al.Non-invasive biomarkers of nonalcoholic steatohepatitis: the FNIH NIMBLE project.Nat Med. 2022; 28: 430-432Crossref PubMed Scopus (12) Google Scholar,4Hardy T. Wonders K. Younes R. et al.The European NAFLD Registry: a real-world longitudinal cohort study of nonalcoholic fatty liver disease.Contemp Clin Trials. 2020; 98106175Crossref PubMed Scopus (43) Google Scholar The enormous potential to account for a large number of variables in complex databases and determine the likelihood of specific outcomes in a very short time, will markedly outperform a single physician's capability that operates at the level of personal experience and medical education. Despite high expectations by many stakeholders and receptivity toward AI in the general society and among medical professionals, the level of implementation in clinical practice today is relatively low.5Xiang Y. Zhao L. Liu Z. et al.Implementation of artificial intelligence in medicine: status analysis and development suggestions.Artif Intell Med. 2020; 102101780https://doi.org/10.1016/j.artmed.2019.101780Crossref PubMed Scopus (33) Google Scholar There are several limitations that must be taken into account to allow for a safe application of AI and a higher penetration into clinical care. The assembly of high-quality representative data sets that eliminate unwanted and unconscious biases is a prerequisite for building ML models that do not perpetuate health care disparities. The inability of AI algorithms to account for information gained from a direct patient-physician interaction is an inherent limitation. The AI algorithms will never be able to substitute for a physician's direct interaction with their patients. We view AI as a complementary tool to significantly enhance patient-provider interaction and patient care. For AI's integration into hepatology clinical practice, multiple currently open questions need to be addressed including the quality of data synthesized, operational procedures, data and systems safety, and ethical challenges. Ethical challenges may arise from clinical decision making based on AI-generated diagnostic algorithms that are not readily recapitulated through medical reasoning. As seen with automated driving, a critical question arises around responsibility and liability in the context of a decision that is based on an AI algorithm. Similarly in medicine, the consequences of false-positive and false-negative results that are generated by AI are far reaching for patients and their providers. This is exacerbated by the fact that AI algorithms are comprised of complex interconnected structures with numerous parameters and a "black box" nature, offering little understanding of their inner working. Explainable AI is a set of processes that allows humans to comprehend the output created by ML algorithms, which help develop trust in the system and meet adherence to regulatory requirements.6Nazir S. Dickson D.M. Akram M.U. Survey of explainable artificial intelligence techniques for biomedical imaging with deep neural networks.Comput Biol Med. 2023; 156106668Crossref PubMed Scopus (3) Google Scholar The challenges in hepatology are to allow for a safe and evidence-based implementation of AI to support clinical decision making. Several AI approaches have been used in hepatology with a focus on identification of cases (through imaging and noninvasive tests [NIT]), augmentation of histologic analysis, and prediction of outcome (Table 1). This review article focuses on current developments in AI/ML with potential applications in hepatology and defines areas of research that should be addressed in the future. To prepare for this manuscript, a literature search was conducted by the authors using the electronic PubMed databased and the following search terms: "artificial intelligence," "machine learning," and "liver disease." The authors selected the most relevant English language articles to provide an overview of where AI may impact the practice of clinical hepatology.Table 1Summary of Potential AI Application in HepatologyAI application in hepatologyExamples of AI algorithmsLimitationsImaging-CNNs to diagnose hepatic steatosis based on ultrasound images.-CNNs for automated CT and MRI liver segmentation.-Variations in data acquisition by different scanners and imaging protocols, and image reconstruction methods.Histology-ML algorithms to enable quantitative measurement of NASH histologic features.-ML algorithms to predict response to NASH treatment.-ML algorithms to determine the presence of portal hypertension and predict outcomes.-Lack of universal standards for digitization of slides, data formatting, image data compression, and storage of meta-data.Identifying at-risk patients using NITs-Random forest ML model to predict the stage of fibrosis and identify patients with fibrotic NASH.-The AI-Cirrhosis-ECG score to detect cirrhosis.-The need for high-quality representative datasets to eliminate the potential for bias.Predicting outcomes-Cirrhosis Mortality Model to predict cirrhosis mortality.-ML models to predict graft failure within 30 d from liver transplantation.-Random forest ML model to predict incident HCC.-Primary sclerosing cholangitis risk estimate tool (PREsTo) to predict outcomes in patients with PSC.-Lack of prospective AI-based randomized clinical trials that demonstrate the added value of AI models in improving clinical outcomes for patients.AI, artificial intelligence; CNN, convolutional neural network; CT, computer tomography; HCC, hepatocellular carcinoma; ML, machine learning; MRI, magnetic resonance imaging; NASH, nonalcoholic steatohepatitis; NIT, noninvasive test; PSC, primary sclerosing cholangitis. Open table in a new tab AI, artificial intelligence; CNN, convolutional neural network; CT, computer tomography; HCC, hepatocellular carcinoma; ML, machine learning; MRI, magnetic resonance imaging; NASH, nonalcoholic steatohepatitis; NIT, noninvasive test; PSC, primary sclerosing cholangitis. Imaging modalities represent a cornerstone in the assessment of liver disease. The ability to assess liver morphology and perfusion in addition to masses or hepatic steatosis as point-of-care testing is unprecedented. Although ultrasound is the first-line technology in clinics today, high-end imaging modalities include magnetic resonance imaging (MRI) proton density fat fraction (PDFF) to assess hepatic steatosis, magnetic resonance elastography to assess liver and spleen stiffness, and phase contrast MRI-enhanced imaging to allow for assessment of blood flow. With these technologies, a complete assessment of liver disease stage is possible and the value of more invasive assessment through liver biopsy has declined. Emerging data link these NITs to clinically relevant outcomes.7Gidener T. Ahmed O.T. Larson J.J. et al.Liver stiffness by magnetic resonance elastography predicts future cirrhosis, decompensation, and death in NAFLD.Clin Gastroenterol Hepatol. 2021; 19: 1915-1924Abstract Full Text Full Text PDF PubMed Scopus (44) Google Scholar The wealth of data that are acquired from these imaging modalities make these technologies particularly suitable for postacquisition processing using AI. Only a fraction of the available data is actually used to build an image that informs the clinical decision making. Most available research data were generated on the use of AI in ultrasound assessment of liver disease. Convolutional neural networks (CNNs) have shown a very high accuracy of replicating the diagnosis of hepatic steatosis made based on ultrasound B-mode images.8Zamanian H. Mostaar A. Azadeh P. et al.Implementation of combinational deep learning algorithm for non-alcoholic fatty liver classification in ultrasound images.J Biomed Phys Eng. 2021; 11: 73-84Crossref PubMed Scopus (16) Google Scholar In a more recent analysis, DL of raw ultrasound data reached an area under the receiver operating curve (AUROC) of 0.98 when compared with the reference standard MRI-PDFF in detecting hepatic steatosis, even in the absence of phantoms to train imaging acquisition.9Han A. Byra M. Heba E. et al.Noninvasive diagnosis of nonalcoholic fatty liver disease and quantification of liver fat with radiofrequency ultrasound data using one-dimensional convolutional neural networks.Radiology. 2020; 295: 342-350Crossref PubMed Scopus (51) Google Scholar Detection of significant (≥F2) or advanced (≥F3) fibrosis by ultrasound-based elastography was explored in a study that used 3392 images from 328 cases at the Massachusetts General Hospital. Augmentation of shear wave elastography using a CNN improved the AUROC of conventional shear wave elastography from 0.74 to 0.89 for detecting the histologic stages ≥F2 by improving image quality, the selection of a region of interest, and classifying the region of interest.10Brattain L.J. Telfer B.A. Dhyani M. et al.Objective liver fibrosis estimation from shear wave elastography.Annu Int Conf IEEE Eng Med Biol Soc. 2018; 2018: 1-5PubMed Google Scholar When considering ultrasound for the detection of hepatocellular carcinoma (HCC), a deep CNN demonstrated an AUROC of 0.92 for distinguishing benign from malignant liver lesions.11Yang Q. Wei J. Hao X. et al.Improving B-mode ultrasound diagnostic performance for focal liver lesions using deep learning: a multicentre study.EBioMedicine. 2020; 56102777Abstract Full Text Full Text PDF Scopus (40) Google Scholar Importantly, this was superior to the diagnostic sensitivity and specificity of experienced radiologists, comparable with contrast-enhanced computed tomography (CT), and only slightly inferior to contrast-enhanced MRI.11Yang Q. Wei J. Hao X. et al.Improving B-mode ultrasound diagnostic performance for focal liver lesions using deep learning: a multicentre study.EBioMedicine. 2020; 56102777Abstract Full Text Full Text PDF Scopus (40) Google Scholar The high-performance metrics of AI and ML in refining diagnostic accuracy for liver disease do not overcome the inherent limitations that specific imaging modalities have. Traditionally, the use of CT imaging to detect hepatic steatosis did not exhibit a high accuracy for mild hepatic steatosis.12Lawrence D.A. Oliva I.B. Israel G.M. Detection of hepatic steatosis on contrast-enhanced CT images: diagnostic accuracy of identification of areas of presumed focal fatty sparing.AJR Am J Roentgenol. 2012; 199: 44-47Crossref PubMed Scopus (49) Google Scholar In a recent analysis using a fully automated volumetric hepatosplenic segmentation algorithm and 3-dimensional CNNs with MRI-PDFF as reference standard, the AUROC to detect mild, moderate, and advanced hepatic steatosis exhibited values of 0.669, 0.854, and 0.962, respectively.13Pickhardt P.J. Blake G.M. Graffy P.M. et al.Liver steatosis categorization on contrast-enhanced CT using a fully automated deep learning volumetric segmentation tool: evaluation in 1204 healthy adults using unenhanced CT as a reference standard.AJR Am J Roentgenol. 2021; 217: 359-367Crossref PubMed Scopus (19) Google Scholar Thus, even CNN and high-end CT imaging lack the accuracy to detect mild degrees of hepatic steatosis. The more advanced imaging modalities outperform ultrasound in the diagnostic accuracy of hepatic steatosis and detection of advanced fibrosis. Whole-liver segmentation is an automated method that uses CNN for imaging biomarkers. In a recent study, MRI-PDFF detecting hepatic steatosis and transverse relaxometry (R2∗) detecting iron overload showed an excellent agreement with the histologic lesions in 165 participants of whom 61% had nonalcoholic fatty liver disease (NAFLD).14Martí-Aguado D. Jiménez-Pastor A. Alberich-Bayarri Á. et al.Automated whole-liver MRI segmentation to assess steatosis and iron quantification in chronic liver disease.Radiology. 2022; 302: 345-354Crossref PubMed Scopus (12) Google Scholar In a smaller study on 62 participants, texture analysis–derived parameters on non-contrast-enhanced T1-weighting was comparable with magnetic resonance elastography to detect advanced versus early fibrosis.15Schawkat K. Ciritsis A. von Ulmenstein S. et al.Diagnostic accuracy of texture analysis and machine learning for quantification of liver fibrosis in MRI: correlation with MR elastography and histopathology.Eur Radiol. 2020; 30: 4675-4685Crossref PubMed Scopus (37) Google Scholar Using a generalized CNN automated liver segmentation was feasible even across CT and MRI for automated liver biometry.16Wang K. Mamidipalli A. Retson T. et al.Automated CT and MRI liver segmentation and biometry using a generalized convolutional neural network.Radiol Artif Intell. 2019; 1180022Crossref Scopus (64) Google Scholar An area of special interest in AI-supported imaging is the augmentation of radiology reports to routinely include aspects of liver health and disease, even if the indication to perform the radiologic examination is not in the context of liver disease. The ability of MRIs to detect changes in the nodularity of the liver surface correlates well with the presence of advanced fibrosis on liver histology.17Dioguardi Burgio M. Sartoris R. Beaufrere A. et al.Liver surface nodularity on non-contrast MRI identifies advanced fibrosis in patients with NAFLD.Eur Radiol. 2022; 32: 1781-1791Crossref PubMed Scopus (3) Google Scholar Therefore, one clinical application where AI can run in the backend of an imaging server will be to highlight the presence of increased surface nodularity to trigger the radiologist to include the suspicions of cirrhosis in the structured reports.18Schattenberg J.M. Emrich T. Refining imaging tools to detect advanced fibrosis: could liver surface nodularity address an unmet need in the NAFLD epidemic?.Eur Radiol. 2022; 32: 1757-1759Crossref PubMed Scopus (1) Google Scholar Despite the promising results of AI algorithms in liver imaging, several issues limit their widespread use including variations in data acquisition by different scanners, imaging protocols, and image reconstruction methods to the final selection of radiomic features.19Nam D. Chapiro J. Paradis V. et al.Artificial intelligence in liver diseases: improving diagnostics, prognostics and response prediction.JHEP Rep. 2022; 4100443PubMed Google Scholar Several concrete steps need to be taken to standardize the measurement and analysis of imaging biomarkers. The details of algorithm development including the datasets and computer source code should be shared to ensure transparent translation into the clinical workflow. Liver biopsy is still considered the gold standard for diagnosing nonalcoholic steatohepatitis (NASH) and fibrosis, although the semiquantitative evaluation of the key histologic features of steatosis, inflammation, ballooning, and fibrosis by the pathologists has been shown to be subjective and prone to major intraobserver and interobserver variability.20Gawrieh S. Knoedler D.M. Saeian K. et al.Effects of interventions on intra- and interobserver agreement on interpretation of nonalcoholic fatty liver disease histology.Ann Diagn Pathol. 2011; 15: 19-24Crossref PubMed Scopus (65) Google Scholar In 2011, the Food and Drug Administration provided a road map of drug approval including the achievement of 1 of 2 histologic end points as surrogates for outcomes: resolution of NASH without worsening of fibrosis, or regression of fibrosis by 1 stage or more without worsening of NASH.21Sanyal A.J. Brunt E.M. Kleiner D.E. et al.Endpoints and clinical trial design for nonalcoholic steatohepatitis.Hepatology. 2011; 54: 344-353Crossref PubMed Scopus (537) Google Scholar Unfortunately, the reliance on histologic end points has made it difficult to find suitable patients for trials and to reliably assess response to different treatments with clear examples from clinical trials documenting the lack of agreements among expert hepatopathologists.22Davison B.A. Harrison S.A. Cotter G. et al.Suboptimal reliability of liver biopsy evaluation has implications for randomized clinical trials.J Hepatol. 2020; 73: 1322-1332Abstract Full Text Full Text PDF PubMed Scopus (159) Google Scholar This has created an opportunity to use AI/ML algorithms to develop methods for quantification of the main histologic features of NASH that are less prone to variability by training the algorithm on digitized slides that are annotated by expert pathologists (Figure 2). In 2014, Gawrieh et al23Gawrieh S. Sethunath D. Cummings O.W. et al.Automated quantification and architectural pattern detection of hepatic fibrosis in NAFLD.Ann Diagn Pathol. 2020; 47151518Crossref PubMed Scopus (24) Google Scholar published one of the first studies on using supervised ML classifiers to automatically classify white regions in liver biopsies as a method to provide continuous quantitative measurement of macrosteatosis. The ML algorithm performed well with 89% overall accuracy when compared with consensus reading by 2 expert pathologists. The same group developed an AI-based model to quantify liver fibrosis and determine its pattern in patients with NASH achieving good to excellent correlation between the automatically generated collagen proportionate area and the pathologist fibrosis staging with a coefficient of determination ranging from 0.60 to 0.86.23Gawrieh S. Sethunath D. Cummings O.W. et al.Automated quantification and architectural pattern detection of hepatic fibrosis in NAFLD.Ann Diagn Pathol. 2020; 47151518Crossref PubMed Scopus (24) Google Scholar Another group used data from 246 patients with biopsy-proven NASH to develop a high-throughput ML-based quantification of steatosis, ballooning, inflammation, and fibrosis with high interclass correlation coefficient between the manual annotation and the software ranging from 0.97 for steatosis to 0.92 for fibrosis.24Forlano R. Mullish B.H. Giannakeas N. et al.High-throughput, machine learning-based quantification of steatosis, inflammation, ballooning, and fibrosis in biopsies from patients with nonalcoholic fatty liver disease.Clin Gastroenterol Hepatol. 2020; 18: 2081-2090Abstract Full Text Full Text PDF PubMed Google Scholar In the largest study to date, the PathAI team (Boston, MA) used liver biopsy samples from 3 large NASH clinical trials to build and validate a deep CNN to enable quantitative measurement of NASH histologic severity.25Taylor-Weiner A. Pokkalla H. Han L. et al.A machine learning approach enables quantitative measurement of liver histology and disease monitoring in NASH.Hepatology. 2021; 74: 133-147Crossref PubMed Scopus (62) Google Scholar The ML parameters also predicted clinical outcomes, such as progression to cirrhosis and hepatic decompensation, and by quantification of more complex features, such as portal inflammation and the ratio area of steatosis to ballooning. Furthermore, a Deep Learning Treatment Assessment (DELTA) Liver Fibrosis score was developed to capture changes in fibrosis severity from baseline to the end of treatment and showed correlation with other noninvasive fibrosis markers, such as the enhanced liver fibrosis score and liver stiffness by transient elastography. The same team at PathAI developed a new ML score to predict hepatic vein pressure gradient (ML-HVPG score) by using biopsies and HVPG measurements from a phase 2b trial. The ML-HVPG score had a stronger correlation with traditional HVPG than collagen proportionate area by morphometry and was able to identify patients with clinically significant portal hypertension with good accuracy (AUROC of 0.85 and 0.76 in the training and test sets, respectively).26Bosch J. Chung C. Carrasco-Zevallos O.M. et al.A machine learning approach to liver histological evaluation predicts clinically significant portal hypertension in NASH cirrhosis.Hepatology. 2021; 74: 3146-3160Crossref PubMed Scopus (12) Google Scholar The advantage of the ML-based score is the fact that HVPG can be estimated from a standard percutaneous liver biopsy without the need for highly specialized interventional radiology procedures and human expertise to interpret HVPG tracing. In a series of abstracts presented at international meetings, Noureddin et al developed histologic scores via ML in a post hoc analysis of patients from the belapectin phase 2a trial (NCT02462967). This trial provided a cohort of patients with NASH cirrhosis (n = 143) with liver biopsies and phenotype data including HVPG and clinical outcomes. This analysis consisted of discovery and validation cohorts. A second harmonic generation/2-photon excitation fluorescence imaging-based tool provided an automated quantitative assessment of histologic features related to cirrhosis: 252 features related to septa, 21 related to nodules, and 184 related to fibrosis (SNOF). The investigators developed a ML score, SNOF, which significantly correlated with HVPG as a continuous variable (r = 0.57 for training and r = 0.70 for validation; P < .05 for both) and significantly clinically significant portal hypertension (AUROC of 0.85 for training and 0.74 for validation). Investigators also created 2 companion scores: SNOF-V score, which significantly predicted the presence of varices (AUROC of 0.86 for discovery and AUROC of 0.73 for validation cohorts); and SNOF-C score, which identified patients who had >20% change in HVPG 12 months apart with an AUROC of 0.89. Collectively, these data offer a compelling proof-of-concept that ML tools can be applied to liver histology to derive clinically important data that are otherwise difficult to collect from patients.27Noureddin M. Tai D. Chng E. et al.Derivation of machine learning histologic scores correlating with portal pressures and the development of varices in NASH patients with cirrhosis.J Hepatol. 2022; 77: S623-S624Abstract Full Text PDF Google Scholar,28Noureddin M. Goodman Z. Tai D. et al.Machine learning liver histology scores correlate with portal hypertension assessments in nonalcoholic steatohepatitis cirrhosis.Aliment Pharmacol Ther. 2023; 57: 409-417Crossref PubMed Scopus (1) Google Scholar Since then, several AI/ML technologies have been described to assess liver histology in the context of conditional drug approval in the indication NASH. These include second harmonic generation/2-photon excitation to provide quantitative assessment of NASH histologic features on unstained liver histology (Histoindex, Singapore), automated fibrosis quantification from stained slides (Pharmanest, Princeton, NJ), and multiparametric image analysis using proprietary software tools on digitalized histologic slides of entire lobe sections (Biocellvia, Marseille, France). These examples assert the integration of AI in the interpretation of NASH histology given the decreased variability in interpretation, fast processing of samples, and decreased pathologist workload. It will be important to have universal standardization of digitized slides in terms of data formatting, image data compression, and storage of metadata that will enable future discovery of histopathologic biomarkers. Multiple ML algorithms have been generated recently that can diagnose HCC on liver histology and provide risk stratification for recurrence of HCC after surgical resection.29Calderaro J. Seraphin T.P. Luedde T. et al.Artificial intelligence for the prevention and clinical management of hepatocellular carcinoma.J Hepatol. 2022; 76: 1348-1361Abstract Full Text Full Text PDF PubMed Scopus (31) Google Scholar For example, Lal et al30Lal S. Das D. Alabhya K. et al.NucleiSegNet: robust deep learning architecture for the nuclei segmentation of liver cancer histopathology images.Comput Biol Med. 2021; 128104075Crossref PubMed Scopus (54) Google Scholar developed a DL network architecture called NucleiSegNet to grade HCC nuclei on hematoxylin-eosin-stained liver cancer histopathology, which yielded superior results compared with traditional nuclei segmentation methods. An important tool for HCC surgery evaluation is segmentation of hematoxylin-eosin-stained slides by pathologists to assess tumor load before surgical resection and monitor treatment response. Wang et al31Wang X. Fang Y. Yang S. et al.A hybrid network for automatic hepatocellular carcinoma segmentation in H&E-stained whole slide images.Med Image Anal. 2021; 68101914Abstract Full Text Full Text PDF Scopus (26) Google Scholar developed a neural network–based DL model for automatic HCC segmentation that produced high accuracy in 3 public databases. Unfortunately, 50%–70% of patients with HCC experience tumor recurrence at 5 years postsurgical resection. Saillard et al32Saillard C. Schmauch B. Laifa O. et al.Predicting survival after hepatocellular carcinoma resection using deep learning on histological slides.Hepatology. 2020; 72: 2000-2013Crossref PubMed Scopus (97) Google Scholar developed 2 DL models based on whole-slide digitized images for predicting survival after HCC surgical resection and demonstrated better performance of the DL models in comparison with composite scores that used various clinical, pathologic, and biologic factors. Similarly, Yamashita et al33Yamashita R. Long J. Saleem A. et al.Deep learning predicts pos