Artificial intelligence in gastroenterology and hepatology: Status and challenges

Abstract

Originally proposed by John McCarthy in 1955, artificial intelligence (AI) has achieved a breakthrough and revolutionized the processing methods of clinical medicine with the increasing workloads of medical records and digital images. Doctors are paying attention to AI technologies for various diseases in the fields of gastroenterology and hepatology. This review will illustrate AI technology procedures for medical image analysis, including data processing, model establishment, and model validation. Furthermore, we will summarize AI applications in endoscopy, radiology, and pathology, such as detecting and evaluating lesions, facilitating treatment, and predicting treatment response and prognosis with excellent model performance. The current challenges for AI in clinical application include potential inherent bias in retrospective studies that requires larger samples for validation, ethics and legal concerns, and the incomprehensibility of the output results. Therefore, doctors and researchers should cooperate to address the current challenges and carry out further investigations to develop more accurate AI tools for improved clinical applications.

Key Words: Artificial intelligence, Gastroenterology, Hepatology, Status, Challenges

Core Tip: Artificial intelligence (AI) technologies are widely used for medical image analysis in the gastroenterology and hepatology fields. Several AI models have been developed for accurate diagnosis, treatment, and prognosis based on images of endoscopy, radiology, pathology, achieving high performance comparable to experts. However, we should be aware of the certain constraints that limit the acceptance and utilization of AI tools in clinical practice. To use AI wisely, doctors and researchers should work together to address the current challenges and develop more accurate AI tools to improve patient care.

INTRODUCTION

Originally proposed by John McCarthy in 1955, artificial intelligence (AI) which involves machine learning (ML) and problem solving, has achieved a breakthrough and revolutionized the processing methods of clinical medicine with the increasing workloads of medical records and digital images. In clinical practice, AI consists of several overlapping technologies such as ML, artificial neural networks (ANNs), deep learning (DL), convolutional neural networks (CNNs), and recurrent neural networks[1,2] (Figure 1). Since the 1980s, ML has been performed to construct a mathematical model and predict outcomes based on input data, and it is roughly divided into supervised (labeled data), unsupervised (unlabeled data), and semi-supervised (both labeled and unlabeled data) learning techniques[3]. Recently, as a subset of ML, ANNs have received increased interest because they can identify and learn input data by themselves instead of being labeled by experts[4]. In the last decade, DL, a new model of ML, holds great promise in clinical medicine. DL is particularly suitable for enormous complex or highly dimensional medical image analysis and predictive modeling tasks using the multilayers of ANNs, including CNNs and recurrent neural networks[5,6]. Notably, given that convolutional and pooling layers can extract distinct features and fully connected layers can make a final classification, CNNs have demonstrated excellent performance in image recognition such as endoscopy, radiology, and pathology[7,8].

Figure 1 Timeline and related technologies of artificial intelligence. AI: Artificial intelligence; ANN: Artificial neural network; CNN: Convolutional neural network; RNN: Recurrent neural network.

In the fields of gastroenterology and hepatology, doctors are paying attention to AI technologies for the diagnosis, treatment, and prognosis of various diseases due to the heterogeneous expertise levels of doctors (majoring in endoscopy, radiology, and pathology), time-consuming procedures, and increasing workloads. Specifically, doctors usually assess medical images visually to detect and diagnose diseases based on personal expertise and experience. As the maturity of digitalization increases, a quantitative assessment of imaging information has become the reality instead of relatively inaccurate qualitative reasoning[9,10]. Although a lot of time is necessary to review and check image analysis traditionally, little information can be obtained. For example, using AI technologies to process pathology images can assess the histopathological classification and predict gene mutations in liver cancer[11], while only the mass nature can be identified by conventional pathology assessment. As a country with a high population, China has produced rapidly increasing medical records, which result in the high workloads[12]. Despite the progression of AI, gastroenterologists and hepatologists should always be aware of its limitations such as the retrospective manner of included studies and the utilization of not particularly suitable databases. In addition, it demands that doctors prepare for the effects and changes of AI on clinical practice in the real world.

In this review, we aim to (1) introduce how AI technologies process input data, learn from input data, validate the established model; (2) summarize the AI applications in endoscopy, radiology, pathology for accurate diagnosis, treatment, and prognosis; and (3) discuss the current limitations and future considerations of AI applications in the fields of gastroenterology and hepatology (Figure 2).

Figure 2 Artificial intelligence-assisted endoscopy, radiology, and pathology applications for medical image analysis in the fields of gastroenterology and hepatology, including detecting and evaluating lesions, facilitating treatment, and predicting treatment response and prognosis, and other potentials, using several deep learning models. CT: Computed tomography; MRI: Magnetic resonance imaging.

METHODS IN DEEP LEARNING

As the most suitable approach for medical image analysis, the DL approach does not require shaped regions of interest on images to complete feature selection and extraction based on a neural network structure[13,14]. After data collection and processing, the correct neural network is chosen to establish a model, followed by model validation to assess its true generalizability.

Data processing

Raw data are collected and analyzed, and corrupt data are identified and cleaned in the processing phase. Data selection methods are provided in Scikit-Learn[15], a Python machine learning library, which consist of univariate selection, feature importance, correlation matrix, and recursive feature elimination or addition. Other programming languages such as R Studio (http://www.r-project.org) or MATLAB software (University of New Mexico, New Mexico, United States) also offer a successful environment for AI, and they provide similar approaches to address specific tasks. Useful data and relevant variables from multiple data sources, which are applied to predict outcomes, are selected and divided into an initial training set and a testing set that allow training and internal validation of the model. Data in the training set should be different and nonredundant from that in the testing set. Notably, for small datasets, a higher proportion of data should be included in the testing set to measure the performance of the trained model accurately through cross-validation or in a bootstrapping procedure.

Modeling

After transforming the data into an appropriate format, different tools are developed for implementing ML. Although several programming tools such as Python, R Studio, and MATLAB vary among themselves, they provide similar options and algorithms to adjust the parameters based on specific tasks. The major classification algorithms for testing are Naive Bayes, Decision Trees, Support Vector Machine, K-Nearest Neighbor, and Ensemble Classifiers. Oversampling or undersampling of the unbalanced training data can be utilized to improve the representation of classes and prevent model bias during the modeling stage. Currently, as the calculation workload of the batch learning process is heavy, minibatch learning is more popularized with repeating epochs, which usually decreases errors for the training and testing phases. However, an early stopping technique would be adopted to address the overfitting problem if repeating epochs cannot ensure error reduction. Based on the evaluations of model performance, developers conduct feature engineering again to manipulate the features and approve the predictive values of the model. After the optimization phase, selection for the model is primarily based on trial-and-error and the best performance for specific problem-solving. Finally, model optimization is performed with adjusted parameters by testing different configurations.

Model validation

To evaluate the AI approaches, one of the most significant requirements is external validation, which is called the blind test. Any model developed within one dataset will merely reflect its idiosyncrasies, and will have poor performance in analyzing new settings. In addition, models can also be validated by internal data testing (e.g., k-fold cross-validation). In k-fold cross-validation, the dataset is separated into k subsets, including one subset for testing and the remaining (k-1) subsets for training a model. With all data used in both training and testing sets, the cross-validation process is repeated k times. The model performance is finally calculated as the average value of all k iterations. The k varies depending on the size of the dataset. For example, leave-one-out validation may be used in a small training set (< 200 data points), which means that k is equivalent to the dataset size. The appropriate and robust predictive model should have consistent performance between training and testing sets, preventing overfitting discrepancies.

ARTIFICIAL INTELLIGENCE IN ENDOSCOPY

With the advent and continuous improvement of fiberoptics, endoscopy has been playing a significant role in the diagnosis and treatment of gastrointestinal diseases. However, gastrointestinal diseases remain an enormous economic burden and lead to high mortality worldwide. AI is applicable in the gastroenterology fields within endoscopy[16-18], such as identification of esophageal and gastric neoplasia in esophagogastroduodenoscopy (EGD), detection of gastrointestinal bleeding in wireless capsule endoscopy (WCE), and polyp detection and characterization in colonoscopy, etc[19-63] (Table 1).

Table 1 Summary of key studies on artificial intelligence-assisted endoscopy in gastroenterology fields.

Ref.	Country	Disease studied	Design of study	Application	Number of cases	Type of machine learning algorithm	Outcomes (%)
							Accuracy	Sensitivity/Specificity
Esophagogastroduodenoscopy
Takiyama et al[19], 2018	Japan	Anatomical location of upper gastrointestinal tract	Retrospective	Recognition of the anatomical location of upper gastrointestinal tract	Training: 27335 images: 663 larynx, 3252 esophagus, 5479 upper stomach, 7184 middle stomach, 7539 lower stomach, and 3218 duodenum; Testing: 17081 images: 363 larynx, 2142 esophagus, 3532 upper stomach, 6379 middle stomach, 3137 lower stomach, and 1528 duodenum	CNNs	Larynx: 100; Esopha us: 100; Stomach: 99; Duodenum: 99	Larynx: 93.9/100; Esophagus: 95.8/99.7; Stomach: 98.9/93; Duodenum: 87/99.2
Wu et al[20], 2019	China	Diseases of upper gastrointestinal tract	Prospective	Monitor blind spots of upper gastrointestinal tract	Training: 1.28 million images from 1000 object classes; Testing: 3000 images for DCNN1, and 2160 images for DCNN2	CNNs	90.4	87.57/95.02
van der Sommen et al[21], 2016	Netherlands	EN-BE	Retrospective	Detection of EN in BE	21 patients with EN-BE (60 images), 23 patients without EN-BE (40 images)	SVM	NA	86/87
Swager et al[22], 2017	Netherlands	EN-BE	Retrospective	Detection of EN in BE	60 images: 40 with EN-BE and 30 without EN-BE	SVM	95	90/93
Hashimoto et al[23], 2020	United States	EN-BE	Retrospective	Detection of EN in BE	Training: 916 images with EN-BE; Testing: 458 images: 225 dysplasia and 233 non-dysplasia	CNNs	95.4	96.4/94.2
Ebigbo et al[24], 2020	Germany	EAC-BE	Retrospective	Detection of EAC in BE	Training: 129 images; Testing: 62 images: 36 EAC and 26 normal BE	CNNs	89.9	83.7/100
Horie et al[25], 2019	Japan	EAC and ESCC	Retrospective	Detection of EAC and ESCC	Training: 384 patients with 32 EAC and 397 ESCC (8428 images); Testing: 47 patients with 8 EAC and 41 ESCC (1118 images)	CNNs	98	98/79
Kumagai et al[26], 2019	Japan	ESCC	Retrospective	Detection of ESCC	Training: 240 patients (4715 images: 1141 ESCC and 3574 benign lesions); Testing: 55 patients (1520 images: 467 ESCC and 1053 benign)	CNNs	90.9	92.6/89.3
Zhao et al[27], 2019	China	ESCC	Retrospective	Detection of ESCC	165 patients with ESCC and 54 patients without ESCC (1383 images)	CNNs	89.2	87.0/84.1
Cai et al[28], 2019	China	ESCC	Retrospective	Detection of ESCC	Training: 746 patients (2438 images: 1332 abnormal and 1096 normal); Testing: 52 patients (187 images)	CNNs	91.4	97.8/85.4
Nakagawa et al[29], 2019	Japan	ESCC	Retrospective	Determination of invasion depth	Training: 804 patients with ESCC (14338 images: 8660 non-ME and 5678 ME); Testing: 155 patients with ESCC (914 images: 405 non-ME and 509 ME)	CNNs	SM1/SM2, 3: 91.0; Invasion depth: 89.6	SM1/SM2, 3: 90.1/95.8; Invasion depth: 89.8/88.3
Tokai et al[30], 2020	Japan	ESCC	Retrospective	Determination of invasion depth	Training: 1751 images with ESCC; Testing: 42 patients with ESCC (293 images)	CNNs	80.9	84.1/80.9
Ali et al[31], 2018	Pakistan	EGC	Retrospective	Detection of EGC	56 patients with EGC, 120 patients without EGC	SVM	87	91.0/82.0
Sakai et al[32], 2018	Japan	EGC	Retrospective	Detection of EGC	Training: 58 patients (348943 images: 172555 EGC and 176388 normal); Testing: 58 patients (9650 images: 4653 EGC and 4997 normal)	CNNs	87.6	80.0/94.8
Kanesaka et al[33], 2018	Japan	EGC	Retrospective	Detection of EGC	Training: 126 images: 66 EGC and 60 normal; Testing: 81 images: 61 EGC and 20 normal	SVM	96.3	96.7/95.0
Wu et al[34], 2019	China	EGC	Retrospective	Detection of EGC	Training: 9691 images: 3710 EGC and 5981 normal; Testing: 100 patients: 50 EGC and 50 normal	CNNs	92.5	94.0/91.0
Horiuchi et al[35], 2020	Japan	EGC	Retrospective	Detection of EGC	Training: 2570 images: 1492 EGC and 1078 gastritis; Testing: 285 images: 151 EGC and 107 gastritis	CNNs	85.3	95.4/71.0
Zhu et al[36], 2019	China	Invasive GC	Retrospective	Determination of invasion depth	Training: 245 patients with GC and 545 patients without GC (5056 images); Testing: 203 images: 68 GC and 135 normal	CNNs	89.2	76.5/95.6
Luo et al[37], 2019	China	EAC, ESCC, and GC	Prospective	Detection of upper gastrointestinal cancers	Training: 15040 individuals (125898 images: 31633 cancer and 94265 control); Testing: 1886 individuals (15637 images: 3931 cancer and 11706 control)	CNNs	91.5-97.7	94.2/85.8
Nagao et al[38], 2020	Japan	GC	Retrospective	Determination of invasion depth	1084 patients with GC (16557 images); Training: Testing = 4:1	CNNs	94.5	84.4/99.4
Wireless capsule endoscopy
Ayaru et al[39], 2015	United Kingdom	Small bowel bleeding	Retrospective	Prediction of outcomes	Training: 170 patients with small bowel bleeding; Testing: 130 patients with small bowel bleeding	ANNs	Recurrent bleeding 88; Therapeutic intervention: 88; Severe bleeding: 78	Recurrent bleeding: 67/91; Therapeutic intervention: 80/89; Severe bleeding: 73/80
Xiao et al[40], 2016	China	Small bowel bleeding	Retrospective	Detection of bleeding in GI tract	Training: 8200 images: 2050 bleeding and 6150 non-bleeding; Testing: 1800 images: 800 bleeding and 1000 non-bleeding	CNNs	99.6	99.2/99.9
Usman et al[41], 2016	South Korea	Small bowel bleeding	Retrospective	Detection of bleeding in GI tract	Training: 75000 pixels: 25000 bleeding and 50000 non-bleeding; Testing: 8000 pixels: 3000 bleeding and 5000 non-bleeding	SVM	91.8	93.7/90.7
Sengupta et al[42], 2017	United States	Small bowel bleeding	Retrospective	Prediction of 30-d mortality	Training: 4044 patients with small bowel bleeding; Testing: 2060 patients with small bowel bleeding	ANNs	81	87.8/90/9
Leenhardt et al[43], 2019	France	Small bowel bleeding	Retrospective	Detection of GIA	Training: 600 images: 300 hemorrhagic GIA and 300 non-hemorrhagic GIA; Testing: 600 images: 300 hemorrhagic GIA and 300 non-hemorrhagic GIA	CNNs	98	100.0/96.0
Aoki et al[44], 2020	Japan	Small bowel bleeding	Retrospective	Detection of small bowel bleeding	Training: 41 patients (27847 images: 6503 bleeding and 21344 normal); Testing: 25 patients (10208 images: 208 bleeding and 10000 non-bleeding)	CNNs	99.89	96.63/99.96
Yang et al[45], 2020	China	Small bowel polyps	Retrospective	Detection of small bowel polyps	1000 images: 500 polyps and 500 non-polyps	SVM	96.00	95.80/96.20
Vieira et al[46], 2020	Portugal	Small bowel tumors	Retrospective	Detection of small bowel tumors	39 patients (3936 images: 936 tumors and 3000 normal)	SVM	97.6	96.1/98.3
Colonoscopy
Fernández-Esparrach et al[47], 2016	Spain	Colorectal polyps	Retrospective	Detection of polyps	24 videos containing 31 different polyps	Energy maps	79	70.4/72.4
Komeda et al[48], 2017	Japan	Colorectal polyps	Retrospective	Detection of polyps	Training: 1800 images: 1200 adenoma and 600 non-adenoma; Testing: 10 cases	CNNs	70.0	83.3/50.0
Misawa et al[49], 2017	Japan	Colorectal polyps	Retrospective	Detection of polyps	Training: 1661 images: 1213 neoplasm and 448 non-neoplasm; Testing: 173 images: 124 neoplasm and 49 non-neoplasm	SVM	87.8	94.3/71.4
Misawa et al[50], 2018	Japan	Colorectal polyps	Retrospective	Detection of polyps	196631 frames: 63135 polyps and 133496 non-polyps	CNNs	76.5	90.0/63.3
Chen et al[51], 2018	China	Colorectal polyps	Retrospective	Detection of diminutive colorectal polyps	Training: 2157 images: 681 hyperplastic and 1476 adenomas; Testing: 284 images: 96 hyperplastic and 188 adenomas	DNNs	90.1	96.3/78.1
Urban et al[52], 2018	United States	Colorectal polyps	Retrospective	Detection of polyps	Training: 8561 images: 4008 polyps and 4553 non-polyps; Testing: 1330 images: 672 polyps and 658 non-polyps	CNNs	96.4	96.9/95.0
Renner et al[53], 2018	Germany	Colorectal polyps	Retrospective	Differentiation of neoplastic from non-neoplastic polyps	Training: 788 images: 602 adenomas and 186 non-adenomatous polyps; Testing: 186 images: 52 adenomas and 48 hyperplastic lesions	DNNs	78.0	92.3/62.5
Wang et al[54], 2018	United States	Colorectal polyps	Retrospective	Detection of polyps	Training: 5545 images: 3634 polyps and 1911 non-polyps; Testing: 27113 images: 5541 polyps and 21572 non-polyps	CNNs	98	94.4/95.9
Mori et al[55], 2018	Japan	Colorectal polyps	Prospective	A diagnose-and-leave strategy for diminutive, non-neoplastic rectosigmoid polyps	Training: 61925 images; Testing: 466 cases (287 neoplastic polyps, 175 nonneoplastic polyps, and 4 missing specimens)	SVM	96.5	93.8/91.0
Byrne et al[56], 2019	Canada	Colorectal polyps	Retrospective	Detection and classification of polyps	Training: 60089 frames of 223 videos (29% NICE type 1, 53% NICE type 2 and 18% of normal mucosa with no polyp); Testing: 125 videos: 51 hyperplastic polyps and 74 adenoma	CNNs	94.0	98.0/83.0
Blanes-Vidal et al[57], 2019	Denmark	Colorectal polyps	Retrospective	Detection of polyps	131 patients with polyps and 124 patients without polyps	CNNs	96.4	97.1/93.3
Lee et al[58], 2020	South Korea	Colorectal polyps	Retrospective	Detection of polyps	Training: 306 patients (8593 images: 8495 polyp and 98 normal); Testing: 15 patients (15 polyps videos)	CNNs	93.4	89.9/93.7
Gohari et al[59], 2011	Iran	CRC	Retrospective	Determination of prognostic factors of CRC	1219 patients with CRC	ANNs	Colon cancer: 89; Rectum cancer: 82.7	NA/NA
Biglarian et al[60], 2012	Iran	CRC	Retrospective	Prediction of distant metastasis in CRC	1219 patients with CRC	ANNs	82	NA/NA
Takeda et al[61], 2017	Japan	CRC	Retrospective	Diagnosis of invasive CRC	Training: 5543 images: 2506 non-neoplasms, 2667 adenomas, and 370 invasive cancers; Testing: 200 images: 100 adenomas and 100 invasive cancers	SVM	94.1	89.4/98.9
Ito et al[62], 2019	Japan	CRC	Retrospective	Diagnosis of cT1b CRC	Training: 9942 images: 5124 cTis + cT1a, 4818 cT1b, and 2604 cTis + cT1a; Testing: 5022 images: 2604 cTis + cT1a, and 2418 cT1b	CNNs	81.2	67.5/89.0
Zhou et al[63], 2020	China	CRC	Retrospective	Diagnosis of CRC	Training: 3176 patients with CRC and 9003 patients without CRC (464105 images: 28071 CRC and 436034 non-CRC); Testing: 307 patients with CRC and 1956 patients without CRC (84615 images: 11675 CRC and 72940 non-CRC)	CNNs	96.3	91.4/98.0

AI: Artificial intelligence; CNN: Convolutional neural network; EN: Early-stage neoplasia; BE: Barrett’s esophagus; SVM: Support vector machine; NA: Not available; EAC: Esophageal adenocarcinoma; ESCC: Esophageal squamous cell carcinoma; EGC: Early-stage gastric cancer; GC: Gastric cancer; ANN: Artificial neural network; GI: Gastrointestinal; GIA: Gastrointestinal angioectasia; DNN: Deep neural network; CRC: Colorectal cancer.

EGD

Inadequate examination of the upper gastrointestinal tract is one of the reasons for misdiagnosing several EGD diseases. Based on AI-assisted EGD, the upper gastrointestinal tract can be divided into the pharynx, esophagus, stomach (upper, middle, lower), and duodenum with high values of the area under the curve (AUC)[19]. Furthermore, several AI technologies have classified the images of the stomach within EGD to significantly monitor the blind spots, and their accuracy has reached the ability of experienced endoscopists[20,34].

Endoscopic surveillance for Barrett’s esophagus (BE) is the potential risk factor for esophageal adenocarcinoma (EAC), of which the prognosis is related to disease staging[64,65]. However, accurate detection of esophageal neoplasia and early EAC remains difficult for experienced endoscopists[66]. An AI system developed by Ebigbo et al[24] enabled early detection of EAC with high sensitivity and specificity, they subsequently designed a real-time system for neoplasia classification in magnification. Both accurate detection of early EAC is important in BE images and the novel system of high invasion accuracy also deserves clinical attention[23]. In esophageal squamous cell carcinoma (ESCC), these tumors are often diagnosed at advanced stages, while early ESCC seems to be detected based on endoscopists’ experience because they are almost impossible to visualize with white light endoscopy. Fortunately, with AI technologies, small esophageal lesions (< 10 mm) are recognized, and there is an AI system showing diagnostic accuracy of 91.4%, which is higher than that of high-level (with experience of > 15 years, 88.8%), mid-level (with experience of 5-15 years, 81.6%), and junior-level (with experience of < 5 years, 77.2%) endoscopists[28]. In addition, the prognosis of ESCC can be proved by differentiating tumor invasion depth[29,30].

The prognosis of gastric cancer (GC) mainly depends on the early detection and invasion depth of the disease. It is extremely difficult for endoscopists to recognize early gastric cancer (EGC), which is often accompanied by gastric mucosal inflammation, and the false-negative rate of EGC in EGD has reached nearly 25.0%[67,68]. AI-assisted EGD has the potential to address tough tissues. However, the first reported CNNs-based AI system for detection of EGC had a low positive predictive value of 30.6%, leading to misdiagnosis of gastritis and misinterpretation of the gastric angle as GC[69]. In 2019, Wu et al[34] examined the detection of GC by AI and validated 200 endoscopic images, with increased accuracy, sensitivity, and specificity values (92.5%, 94%, and 91%, respectively). Furthermore, the AI system, named GRAIDS, has achieved diagnostic sensitivity close to that of expert endoscopists (94.2% vs 94.5%), and it demonstrated a robust performance showing high diagnostic accuracy in a multicenter study[37]. Besides detection, one of the most important criteria for curative resection is the invasion depth. The invasion depth prediction of GC by AI was first developed by Kubota et al[70], and the model showed the accuracy of T-stages (T1 = 77%, T2 = 49%, T3 = 51%, T4 = 55%, respectively). Considering that endoscopic mucosal resection is appropriate for intramucosal cancers (M) and submucosal cancers (invasion < 500 μm) (SM1), a more detailed classification is urgently needed. Therefore, an AI system was developed to differentiate the depths of M or SM1 and SM2 (submucosal invasion ≥ 500 μm) for GC with significantly higher sensitivity, specificity, and accuracy than those of skilled endoscopists[38].

WCE

AI-assisted WCE enables endoscopists to highlight suspicious regions on examination of the digestive tract noninvasively, including detection of small bowel bleeding, ulcers, and polyps, celiac disease, etc. Based on specific AI classifiers and validation techniques (mainly k-fold cross-validation), these models utilized still frames, pixels, or real-time videos to identify patients with small bowel bleeding with accuracy above 90% for most studies[40,41,43,44]. A CNNs-based algorithm, established in a retrospective analysis of 10000 WCE images (8200 and 1800 in the training and testing set, respectively) and validated by 10-fold cross-validation, was proposed for automatic detection of small bowel bleeding. The model was performed with a high F-1 score of 99.6% and precision of 99.9% for both active and inactive bleeding frames[40]. Besides detection, several emerging AI tools have been developed to stratify patients for the possibility of recurrent bleeding, treatment requirement, and mortality estimate to prevent repeated endoscopies in a significant proportion of patients with potential recurrent upper or lower gastrointestinal bleeding[39,42].

Colonoscopy

Colorectal polyp detection and appropriate polypectomy during colonoscopy is the standard way to prevent colorectal cancer (CRC). Since missed colorectal polyps can potentially progress into CRC, AI-assisted colonoscopy has been developed for polyp detection and characterization, and predicting the prognosis of CRC. In terms of polyp detection, an automated AI system using an energy map was developed in 2016, and it showed barely satisfactory performance[47]. Urban et al[52] used 8641 labeled images and 20 colonoscopy videos as the training and testing set to establish a CNNs model to identify colonic polyps, and the model had an accuracy of 96.4%[52]. Notably, the models should be validated to improve accuracy. After validating the model developed by Wang et al[54] with 27113 newly collected images from 1138 patients, the model showed acceptable performance (sensitivity = 94.38%, specificity = 95.2%, and AUC = 0.984). In addition, polyp characterization with magnifying endoscopic images is useful for identifying pit or vascular patterns to improve performance. AI tools with narrow-band imaging[51] or endoscopic videos[56] can be used to differentiate diminutive hyperplastic polyps and adenomas with high accuracy. Specifically, diminutive polyps (≤ 5 mm) may also be identified during colonoscopy[55]. In addition, AI may assist doctors in predicting the prognosis of CRC. An ANNs model, which was developed from a dataset of 1219 CRC patients, may predict patient survival and influential factors more accurately than a Cox regression model[59], and it also enables doctors to predict the risk of distant metastases[60].

ARTIFICIAL INTELLIGENCE IN RADIOLOGY

There is a disproportionate growing rate between radiological imaging data and the number of trained radiologists, and it has forced radiologists to compensate by increasing productivity[71]. The emergence of AI technologies has eased the current dilemma and dramatically advanced radiology image analysis, including ultrasound, computed tomography (CT), and magnetic resonance imaging (MRI) in the fields of gastroenterology and hepatology. In addition, the rise of radiomics, which is a new technology in radiology and cancer, can extract abundant quantifiable objective data to evaluate surgical resection and predict treatment response[72-112] (Table 2).

Table 2 Summary of key studies on artificial intelligence-assisted radiology in hepatology fields.

Ref.	Country	Disease studied	Design of study	Application	Number of cases	Type of machine learning algorithm	Outcomes (%)
							Accuracy	Sensitivity/Specificity
Ultrasound-based medical image recognition
Gatos et al[72], 2016	United States	Hepatic fibrosis	Retrospective	Classification of CLD	85 images: 54 healthy and 31 CLD	SVM	87	83.3/89.1
Gatos et al[73], 2017	United States	Hepatic fibrosis	Retrospective	Classification of CLD	124 images: 54 healthy and 70 CLD	SVM	87.3	93.5/81.2
Chen et al[74], 2017	China	Hepatic fibrosis	Retrospective	Classification of the stages of hepatic fibrosis in HBV patients	513 HBV patients with different hepatic fibrosis (119 S0, 164 S1, 88 S2, 72 S3, and 70 S4)	SVM, Naive Bayes, RF, KNN	82.87	92.97/82.50
Li et al[75], 2019	China	Hepatic fibrosis	Prospective	Classification of the stages of hepatic fibrosis in HBV patients	144 HBV patients	Adaptive boosting, decision tree, RF, SVM	85	93.8/76.9
Gatos et al[76], 2019	United States	Hepatic fibrosis	Retrospective	Classification of CLD	88 healthy individuals (88 F0 fibrosis stage images) and 112 CLD patients (112 images: 46 F1, 16 F2, 22 F3, and 28 F4)	CNNs	82.5	NA/NA
Wang et al[77], 2019	China	Hepatic fibrosis	Prospective	Classification of the stages of hepatic fibrosis in HBV patients	Training: 266 HBV patients (1330 images); Testing: 132 HBV patients (660 images)	CNNs	F4: 100; ≥ F3: 99; ≥ F2: 99	F4: 100.0/100.0; ≥ F3: 97.4/95.7; ≥ F2: 100.0/97.7
Kuppili et al[78], 2017	United States	MAFLD	Retrospective	Detection and characterization of FLD	63 patients: 27 healthy and 36 MAFLD	ELM, SVM	ELM: 96.75; SVM: 89.01	NA/NA
Byra et al[79], 2018	Poland	MAFLD	Retrospective	Diagnosis of the amount of fat in the liver	55 severely obese patients	CNNs, SVM	96.3	100/88.2
Biswas et al[80], 2018	United States	MAFLD	Retrospective	Detection and risk stratification of FLD	63 patients: 27 healthy and 36 MAFLD	CNNs, SVM, ELM	CNNs: 100; SVM: 82; ELM: 92	NA/NA
Cao et al[81], 2020	China	MAFLD	Retrospective	Detection and classification of MAFLD	240 patients: 106 healthy, 57 mild MAFLD, 67 moderate MAFLD, and 10 severe MAFLD	CNNs	95.8	NA/NA
Guo et al[82], 2018	China	Liver tumors	Retrospective	Diagnosis of liver tumors	93 patients with liver tumors: 47 malignant lesions (22 HCC, 5 CC, and 10 RCLM), and 46 benign lesions	DNNs	90.41	93.56/86.89
Schmauch et al[83], 2019	France	FLL	Retrospective	Detection and characterization of FLL	Training: 367 patients (367 images); Testing: 177 patients	CNNs	Detection: 93.5; Characterization: 91.6	NA/NA
Yang et al[84], 2020	China	FLL	Retrospective	Detection of FLL	Training: 1815 patients with FLL (18000 images); Testing: 328 patients with FLL (3718 images)	CNNs	84.7	86.5/85.5
CT/MRI-based medical image recognition
Choi et al[85], 2018	South Korea	Hepatic fibrosis	Retrospective	Staging liver fibrosis by using CT images	Training: 7461 patients: 3357 F0, 113 F1, 284 F2, 460 F3, 3247 F4; Testing: 891 patients: 118 F0, 109 F1, 161 F2, 173 F3, 330 F4	CNNs	92.1–95.0	84.6–95.5/89.9–96.6
He et al[86], 2019	United States	Hepatic fibrosis	Retrospective	Staging liver fibrosis by using MRI images	Training: 225 CLD patients; Testing: 84 patients	SVM	81.8	72.2/87.0
Ahmed et al[87], 2020	Egypt	Hepatic fibrosis	Retrospective	Detection and staging of liver fibrosis by using MRI images	37 patients: 15 healthy and 22 CLD	SVM	83.7	81.8/86.6
Hectors et al[88], 2020	United States	Liver fibrosis	Retrospective	Staging liver fibrosis by using MRI images	Training: 178 patients with liver fibrosis; Testing: 54 patients with liver fibrosis	CNNs	F1-F4: 85; F2-F4: 89; F3-F4: 91; F4: 83	F1-F4: 84/90; F2-F4: 87/93; F3-F4: 97/83; F4: 68/94
Vivanti et al[89], 2017	Israel	Liver tumors	Retrospective	Detection and segmentation of new tumors in follow-up by using CT images	246 liver tumors (97 new tumors)	CNNs	86	70/NA
Yasaka et al[90], 2018	Japan	Liver masses	Retrospective	Detection and differentiation of liver masses by using CT images	Training: 460 patients with liver masses (1068 images: 240 Category A, 121 Category B, 320 Category C, 207 Category D, 180 Category E); Testing: 100 images with liver masses: 21 Category A, 9 Category B, 35 Category C, 20 Category D, 15 Category E	CNNs	84	Category A: 71/NA; Category B: 33/NA; Category C: 94/NA; Category D: 90/NA; Category E: 100/NA
Ibragimov et al[91], 2018	United States	Liver diseases requiring SBRT	Retrospective	Prediction of hepatotoxicity after liver SBRT by using CT images	125 patients undergone liver SBRT: 58 liver metastases, 36 HCC, 27 cholangiocarcinoma, and 4 other histopathologies	CNNs	85	NA/NA
Abajian et al[92], 2018	United States	HCC	Retrospective	Prediction of HCC response to TACE by using MRI images	36 HCC patients treated with TACE	RF	78	62.5/82.1
Zhang et al[93], 2018	United States	HCC	Retrospective	Classification of HCC by using MRI images	20 patients with HCC	CNNs	80	NA/NA
Morshid et al[94], 2019	United States	HCC	Retrospective	Prediction of HCC response to TACE by using CT images	105 HCC patients received first-line treatment with TACE	CNNs	74.2	NA/NA
Nayak et al[95], 2019	India	Cirrhosis; HCC	Retrospective	Detection of cirrhosis and HCC by using CT images	40 patients: 14 healthy, 12 cirrhosis, 14 cirrhosis with HCC	SVM	86.9	100/95
Hamm et al[96], 2019	United States	Common hepatic lesions	Retrospective	Classification of common hepatic lesions by using MRI images	Training: 434 patients with common hepatic lesions; Testing: 60 patients with common hepatic lesions	CNNs	92	92/98
Wang et al[97], 2019	United States	Common hepatic lesions	Retrospective	Demonstration of a proof-of-concept interpretable DL system by using MRI images	60 common hepatic lesions patients	CNNs	NA	82.9/NA
Jansen et al[98], 2019	Netherlands	FLL	Retrospective	Classification of FLL by using MRI images	95 patients with FLL (125 benign lesions: 40 adenomas, 29 cysts, and 56 hemangiomas; and 88 malignant lesions: 30 HCC and 58 metastases)	RF	77	Adenoma: 80/78; Cyst: 93/93; Hemangioma: 84/82; HCC: 73/56; Metastasis: 62/77
Mokrane et al[99], 2020	France	HCC	Retrospective	Diagnosis of HCC in patients with cirrhosis by using CT images	Training: 106 patients: 85 HCC and 21 non-HCC; Testing: 36 patients: 23 HCC and 13 non-HCC	SVM, KNN, RF	70	70/54
Shi et al[100], 2020	China	HCC	Retrospective	Detection of HCC from FLL by using CT images	Training: 359 lesions: 155 HCC and 204 non-HCC; Testing: 90 lesions: 39 HCC and 51 non-HCC	CNNs	85.6	74.4/94.1
Alirr et al[101], 2020	Kuwait	Liver tumors	Retrospective	Segmentation of liver tumors	Training: 100 images with liver tumors;Testing: 31 images with liver tumors	CNNs	95.2	NA/NA
Zheng et al[102], 2020	China	Pancreatic cancer	Retrospective	Pancreas segmentation by using MRI images	20 patients with PDAC	CNNs	99.86	NA/NA
Radiomics
Liang et al[103], 2014	China	HCC	Retrospective	Prediction of recurrence for HCC patients who received RFA	83 patients with HCC receiving RFA as first treatment (18 recurrence and 65 non-recurrence)	SVM	82	67/86
Zhou et al[104], 2017	China	HCC	Retrospective	Characterization of HCC	46 patients with HCC: 21 low-grade (Edmondson grades I and II) and 25 high-grade (Edmondson grades III and IV)	Free-form curve-fitting	86.95	76.00/100.00
Abajian et al[105], 2018	United States	HCC	Retrospective	Prediction of response to intra-arterial treatment	36 patients undergone trans-arterial treatment	RF	78	62.5/82.1
Ibragimov et al[91], 2018	United States	Liver tumors	Retrospective	Prediction of hepatobiliary toxicity of SBRT	125 patients undergone liver SBRT: 58 metapatients, 36 HCC, 27 cholangiocarcinoma, and 4 other primary liver tumor histopathologies	CNNs	85	NA/NA
Morshid et al[94], 2019	United States	HCC	Retrospective	Prediction of HCC response to TACE	105 patients with HCC: 11 BCLC stage A, 24 BCLC stage B, 67 BCLC stage C, and 3 BCLC stage D	CNNs	74.2	NA/NA
Ma et al[106], 2019	China	HCC	Retrospective	Prediction of MVI in HCC	Training: 110 patients with HCC: 37 with MVI and 73 without MVI; Testing: 47 patients with HCC: 18 with MVI and 29 without MVI	SVM	76.6	65.6/94.4
Dong et al[107], 2020	China	HCC	Retrospective	Prediction and differentiation of MVI in HCC	Prediction: 322 patients with HCC: 144 with MVI and 178 without MVI; Differentiation: 144 patients with HCC and MVI	RF, mRMR	Prediction: 63.4; Differentiation: 73.0	Prediction: 89.2/48.4; Differentiation: 33.3/80.0
He et al[108], 2020	China	HCC	Prospective	Prediction of MVI in HCC	Training: 101 patients with HCC; Testing: 18 patients with HCC	LASSO	84.4	NA/NA
Schoenberg et al[109], 2020	Germany	HCC	Prospective	Prediction of disease-free survival after HCC resection	Training: 127 patients with HCC; Testing: 53 patients with HCC	RF	78.8	NA/NA
Zhao et al[110], 2020	China	HCC	Retrospective	Prediction of ER of HCC after partial hepatectomy	Training: 78 patients with HCC: 40 with ER and 38 without ER; Testing: 35 patients with HCC: 18 with ER and 17 without ER	LASSO	80.8	80.0/81.6
Liu et al[111], 2020	China	HCC	Retrospective	Prediction of progression-free survival of HCC patients after RFA and SR	RFA: Training: 149 HCC patients undergone RFA Testing: 65 HCC patients undergone RFA; SR: Training: 144 HCC patients undergone SR Testing: 61 HCC patients undergone SR	Cox-CNNs	RFA: 82.0; SR: 86.3	NA/NA
Chen et al[112], 2021	China	HCC	Retrospective	Prediction of HCC response to first TACE by using CT images	Training: 355 patients with HCC; Testing: 118 patients with HCC	LASSO	81	85.2/77.2

AI: Artificial intelligence; CLD: Chronic liver disease; SVM: Support vector machine; HBV: Hepatitis-B virus; RF: Random forests; KNN: K-nearest neighbor; CNN: Convolutional neural network; NA: Not available; MAFLD: Metabolic associated fatty liver disease; FLD: Fatty liver disease; ELM: Extreme learning machine; HCC: Hepatocellular carcinoma; CC: Cholangiocarcinoma; RCLM: Colorectal cancer liver metastases; DNN: deep neural network; FLL: Focal liver lesions; CLD: Chronic liver disease; SBRT: Stereotactic body radiation therapy; TACE: Transarterial chemotherapy; PDAC: Pancreatic ductal adenocarcinoma; RFA: Radiofrequency ablation; BCLC: Barcelona clinic liver cancer staging; MVI: Microvascular invasion; mRMR: Minimum redundancy maximum relevance; LASSO: Least absolute shrinkage and selection operator; ER: Early recurrence; SR: Surgical resection.

Abdominal ultrasound

AI technologies have been applied to abdominal ultrasound-based medical images for the assessment of liver diseases, such as hepatic fibrosis and mass lesions. A support vector machine-derived approach was developed by Gatos et al[72] to detect and classify chronic liver disease (CLD) based on abdominal ultrasound. After quantifying 85 ultrasound images (54 healthy and 31 with CLD), the proposed model showed superior results (accuracy = 87.0%, sensitivity = 83.3%, and specificity = 89.1%), which greatly improved the diagnostic and classification accuracy of CLD. Furthermore, CNNs are employed to identify and isolate regions of different stiffness temporal stability under ultrasound to explore the impact on CLD diagnosis. The updated detection algorithm has augmented the accuracy to 95.5% after excluding unreliable areas and reducing interobserver variability[76]. Detecting and classifying hepatic mass lesions as benign or malignant is equally important. Schmauch et al[83] performed supervised training (367 ultrasonic images together with the radiological reports) to build a DL model, and the resulting algorithm had high receiver operating characteristic curves of 0.93 and 0.916 in lesion detection and characterization, respectively. Although the model could increase the diagnostic accuracy and detect potential malignant mass lesions, it should be further validated. In addition, combining AI technologies with contrast-enhanced ultrasound may improve the performance to identify and characterize liver cancer. For example, after AI-assisted contrast-enhanced ultrasound was applied to detect liver lesions in the arterial, portal, and late phases, the accuracy, sensitivity and specificity of the examination were markedly increased[82]. Due to the misty demonstration of gastroenterology within ultrasound, an AI-assisted ultrasound tool was limited.

CT/MRI

Liver diseases often present indeterminate behaviors on abdominal CT, and a biopsy is recommended according to the European Association for the Study of the Liver guidelines[113]. Based on a large dataset of CT images (7461 patients diagnosed with liver fibrosis), a CNNs model was developed and it outperformed the radiologists’ interpretation[85]. Furthermore, depending on ANNs-based contrast-enhanced CT images from 460 patients, Yasaka et al[90] conducted a retrospective study to classify liver masses into five categories with high accuracy, including (1) primary hepatocellular carcinoma (HCC); (2) malignant tumors apart from HCC; (3) early HCC, indeterminate masses, or dysplastic nodules; (4) hemangiomas; and (5) cysts. For patients diagnosed with liver tumors or pancreatic cancer, it is crucial to complete the liver or pancreas segmentation to assess the lesions and make the ideal treatment plan. Instead of conventional manual segmentation, a CNNs model was proposed to segment liver tumors based on CT images, with an accuracy of more than 80.0%, favoring suitable decision-making[101]. Additionally, a CNNs model was also developed for pancreas localization and segmentation using CT images[102]. Furthermore, monitoring tumor recurrence plays an important role in follow-up CT. Vivanti et al[89] collected and integrated the initial appearance of tumors, CT behaviors, and quantification of the tumor loads throughout the disease course, and then they designed an automated detection model of tumor recurrence with an accuracy of 86%.

Besides depending on CT images, a DL approach for pancreas segmentation can also be designed from MRI images. Several AI-assisted studies have shown promising results in classifying MRI liver lesions with/without risk factors and patients’ clinical data, improving the accuracy and yields of reference models[93,96,98,102].

Radiomics

Currently, radiomics has received great interest from doctors because this AI-assisted technology can extract indiscoverable quantifiable objective data of the radiological images and reveal the association with potential biological processes[114,115]. Preoperative stratification of patients at different risk of recurrence and prediction of survival after resection is fundamental to improve prognosis. As an independent risk factor of recurrence, microvascular invasion (MVI) cannot be provided in conventional radiological techniques[116]. Several studies have managed to use radiomic algorithms based on ultrasound, CT, or MRI to elaborate radiomic signatures for preoperative prediction of MVI[106-108]. Besides the prediction of recurrence, radiomics may also be utilized to predict survival after surgical resection. However, compared to the excellent AI models based on pathologic images, radiomics-based predictive models merely attain a low value of 0.78[109].

Beyond recurrence and survival prediction purposes, radiomics can also be utilized for prediction of patients’ response to transarterial chemoembolization (TACE) and radiofrequency ablation (RFA), and post-radiotherapy hepatotoxicity. A CNNs model developed from 105 HCC patients’ CT images had higher accuracy in predicting response to TACE than the Barcelona Clinic Liver Cancer stages[94]. In addition, Chen et al[112] designed an excellent clinical-radiomic model to predict objective response to first TACE based on 595 HCC patients’ CT images, which could assist the selection of HCC patients for TACE. Another study used radiomics of MRI images with clinical data to perform prediction of TACE response[105]. For HCC in the early stages, RFA is a recommended option. Based on radiomics, Liang et al[103] designed a model to predict the RFA response and HCC recurrence after RFA, obtaining high AUC, sensitivity, and specificity. Additionally, post-radiotherapy hepatotoxicity should be monitored to adjust the position and dose of radiotherapy. A CNNs model not only identified that irradiation of the proximal portal vein was associated with poor prognosis, it also predicted post-radiotherapy hepatotoxicity with an AUC of 0.85[91]. Ibragimov et al[91] applied a CNNs model to determine the consistent patterns in toxicity-related dose plans, and the AUC of the model for dose planned analysis was increased from 0.79 to 0.85 after the combination with some pre-treatment clinical features, showing that the combined framework can indicate the accurate position and dose of radiotherapy.

ARTIFICIAL INTELLIGENCE IN PATHOLOGY

Pathological analysis is considered the gold standard for the diagnosis of diseases in the fields of gastroenterology and hepatology. Currently, there is a shortage of pathologists around the world, which has become an obstruction for maintaining the accuracy of pathological analysis[117]. With the development of the whole-slide imaging (WSI) scanner and AI technologies, a combination of both technologies can ease the medical burden, improve the diagnosis accuracy, and even predict gene mutations and prognosis[118-147] (Table 3).

Table 3 Summary of key studies on artificial intelligence-assisted pathology in the gastroenterology and hepatology fields.

Ref.	Country	Disease studied	Design of study	Application	Number of cases	Type of machine learning algorithm	Outcomes (%)
							Accuracy	Sensitivity/Specificity
Basic AI-based pathology: diagnosis
Tomita et al[118], 2019	United States	BE and EAC	Retrospective	Detection and classification of cancerous and precancerous esophagus tissue	Training: 379 images with 4 classes: normal, BE-no-dysplasia, BE-with-dysplasia, and adenocarcinoma; Testing: 123 images with 4 classes: normal, BE-no-dysplasia, BE-with-dysplasia, and adenocarcinoma	CNNs	Mean: 83; BE-no-dysplasia: 85; BE-with-dysplasia: 89; Adenocarcinoma: 88	Normal: 69/71 BE-no-dysplasia: 77/88; BE-with-dysplasia: 21/97; Adenocarcinoma: 71/91
Sharma et al[119], 2017	Germany	GC	Retrospective	Classification and necrosis detection of GC	454 patients (6810 WSIs: 4994 for cancer classification and 1816 for necrosis detection) (HER2 immunohistochemical stain and HE stained)	CNNs	Cancer classification: 69.90; Necrosis detection: 81.44	NA/NA
Li et al[120], 2018	China	GC	Retrospective	Detection of GC	700 images: 560 GC and 140 normal (HE stained)	CNNs	100	NA/NA
Leon et al[121], 2019	Colombia	GC	Retrospective	Detection of GC	40 images: 20 benign and 20 malignant	CNNs	89.72	NA/NA
Sun et al[122], 2019	China	GC	Retrospective	Diagnosis of GC	500 WSIs of gastric areas with typical cancerous regions	DNNs	91.6	NA/NA
Ma et al[123], 2020	China	GC	Retrospective	Classification of lesions in the gastric mucosa	Training: 534 WSIs (1616713 images: 544925 normal, 544624 chronic gastritis, and 527164 cancer) (HE stained) Testing: 153 WSIs (399240 images: 135446 normal, 125783 chronic gastritis, and 138011 cancer) (HE stained)	CNNs, RF	Benign and cancer: 98.4; Normal, chronic gastritis, and GC: 94.5	Benign and cancer: 98.0/98.9; Normal, chronic gastritis, and GC: NA/NA
Yoshida et al[124], 2018	Japan	Gastric lesions	Retrospective	Classification of gastric biopsy specimens	3062 gastric biopsy specimens (HE stained)	CNNs	55.6	89.5/50.7
Qu et al[125], 2018	Japan	Gastric lesions	Retrospective	Classification of gastric pathology images	Training: 1080 patches: 540 benign and 540 malignant; Testing: 5400 patches: 2700 benign and 2700 malignant	CNNs	96.5	NA/NA
Iizuka et al[126], 2020	Japan	Gastric and colonic epithelial tumors	Retrospective	Classification of gastric and colonic epithelial tumors	4128 cases of human gastric epithelial lesions and 4036 of colonic epithelial lesions (HE stained)	CNNs, RNNs	Gastric adenocarcinoma: 97; Gastric adenoma: 99; Colonic adenocarcinoma: 96; Colonic adenoma: 99	NA/NA
Korbar et al[127], 2017	United States	Colorectal polyps	Retrospective	Classification of different types of colorectal polyps on WSIs	Training: 458 WSIs; Testing: 239 WSIs	A modified version of a residual network	93	88.3/NA
Wei et al[128], 2020	United States	Colorectal polyps	Retrospective	Classification of colorectal polyps on WSIs	Training: 326 slides with colorectal polyps: 37 tubular, 30 tubulovillous or villous, 111 hyperplastic, 140 sessile serrated, and 8 normal; Testing: 238 slides with colorectal polyps: 95 tubular, 78 tubulovillous or villous, 41 hyperplastic, and 24 sessile serrated	CNNs	Tubular: 84.5; Tubulovillous or villous: 89.5; Hyperplastic: 85.3; Sessile serrated: 88.7	Tubular: 73.7/91.6; Tubulovillous or villous: 97.6/87.8; Hyperplastic: 60.3/97.5; Sessile serrated: 79.2/89.7
Shapcott et al[129], 2018	UnitedKingdom	CRC	Retrospective	Diagnosis of CRC	853 hand-marked images	CNNs	84	NA/NA
Geessink et al[130], 2019	Netherlands	CRC	Retrospective	Quantification of intratumoral stroma in CRC	129 patients with CRC	CNNs	94.6	91.1/99.4
Song et al[131], 2020	China	CRC	Retrospective	Diagnosis of CRC	Training: 177 slides: 156 adenoma and 21 non-neoplasm; Testing: 362 slides: 167 adenoma and 195 non-neoplasm	CNNs	90.4	89.3/79.0
Wang et al[132], 2015	China	Hepatic fibrosis	Retrospective	Assessment of HBV-related liver fibrosis and detection of liver cirrhosis	Training: 105 HBV patients; Testing: 70 HBV patients	SVM	82	NA/NA
Forlano et al[133], 2020	UnitedKingdom	MAFLD	Retrospective	Detection and quantification of histological features of MAFLD	Training: 100 MAFLD patients; Testing: 146 MAFLD patients	K-means	Steatosis: 97; Inflammation: 96; Ballooning: 94; Fibrosis: 92	NA/NA
Li et al[134], 2017	China	HCC	Retrospective	Nuclei grading of HCC	4017 HCC nuclei patches	CNNs	96.7	G1: 94.3/97.5; G2: 96.0/97.0;G3: 97.1/96.6; G4: 99.5/95.8
Kiani et al[135], 2020	United States	Liver cancer (HCC and CC)	Retrospective	Histopathologic classification of liver cancer	Training: 70 WSIs: 35 HCC and 35 CC Testing: 80 WSIs: 40 HCC and 40 CC	SVM	84.2	72/95
Advanced AI-based pathology: prediction of gene mutations and prognosis
Steinbuss et al[136], 2020	Germany	Gastritis	Retrospective	Identification of gastritis subtypes	Training: 92 patients (825 images: 398 low inflammation, 305 severe inflammation, and 122 A gastritis) (HE stained) Testing: 22 patients (209 images: 122 low inflammation, 38 severe inflammation, and 49 A gastritis) (HE stained)	CNNs	84	A gastritis: 88/89; B gastritis: 100/93; C gastritis: 83/100
Liu et al[137], 2020	China	Gastrointestinal neuroendocrine tumor	Retrospective	Prediction of Ki-67 positive cells	12 patients (18762 images: 5900 positive cells, 6086 positive cells, and 6776 background from ROIs) (HE and IHC stained)	CNNs	97.8	97.8/NA
Kather et al[138], 2019	Germany	GC and CRC	Retrospective	Prediction of MSI in GC and CRC	Training: 360 patients (93408 tiles); Testing: 378 patients (896530 tiles)	CNNs	84	NA/NA
Bychkov et al[139], 2018	Finland	CRC	Retrospective	Prediction of CRC outcome	420 CRC tumor tissue microarray samples	CNNs, RNNs	69	NA/NA
Kather et al[140], 2019	Germany	CRC	Retrospective	Prediction of survival from CRC histology slides	Training: 86 CRC tissue slides (> 100000 HE image patches); Testing: 25 CRC patients (7180 images)	CNNs	98.7	NA/NA
Echle et al[141], 2020	Germany	CRC	Retrospective	Detection of dMMR or MSI in CRC	Training: 5500 patients; Testing: 906 patients	A modified shufflenet DL system	92	98/52
Skrede et al[142], 2020	3R23 Song 2020	CRC	Retrospective	Prediction of CRC outcome after resection	Training: 828 patients (> 12000000 image tiles); Testing: 920 patients	CNNs	76	52/78
Sirinukunwattana et al[143], 2020	UnitedKingdom	CRC	Retrospective	Identification of consensus molecular subtypes of CRC	Training: 278 patients with CRC; Testing: 574 patients with CRC: 144 biopsies and 430 TCGA	Neural networks with domain-adversarial learning	Biopsies: 85; TCGA: 84	NA/NA
Jang et al[144], 2020	South Korea	CRC	Retrospective	Prediction of gene mutations in CRC	Training: 629 WSIs with CRC (HE stained) Testing: 142 WSIs with CRC (HE stained)	CNNs	64.8-88.0	NA/NA
Chaudhary et al[145], 2018	United States	HCC	Retrospective	Identification of survival subgroups of HCC	Training: 360 HCC patients’ data using RNA-seq, miRNA-seq and methylation data from TCGA; Testing: 684 HCC patients’ data (LIRI-JP cohort: 230; NCI cohort: 221; Chinese cohort: 166, E-TABM-36 cohort: 40, and Hawaiian cohort: 27)	DL	LIRI-JP cohort: 75; NCI cohort: 67; Chinese cohort: 69; E-TABM-36 cohort: 77; Hawaiian cohort: 82	NA/NA
Saillard et al[146], 2020	France	HCC	Retrospective	Prediction of the survival of HCC patients treated by surgical resection	Training: 206 HCC (390 WSIs); Testing: 328 HCC (342 WSIs)	CNNs (SCHMOWDER and CHOWDER)	SCHMOWDER: 78; CHOWDER: 75	NA/NA
Chen et al[11], 2020	China	HCC	Retrospective	Classification and gene mutation prediction of HCC	Training: 472 WSIs: 383 HCC and 89 normal liver tissue; Testing: 101 WSIs: 67 HCC and 34 normal liver tissue	CNNs	Classification: 96.0; Tumor differentiation: 89.6; Gene mutation: 71-89	NA/NA
Fu et al[147], 2020	UnitedKingdom	EAC, GC, CRC, and liver cancers	Retrospective	Prediction of mutations, tumor composition and prognosis	17335 HE-stained images of 28 cancer types	CNNs	Variable across tumors/gene alterations	NA/NA

AI: Artificial intelligence; BE: Barrett’s esophagus; EAC: Esophageal adenocarcinoma; CNN: Convolutional neural network; GC: Gastric cancer; WSI: Whole-slide image; NA: Not available; DNN: Deep neural network; RF: Random forests; RNN: Recurrent neural network; CRC: Colorectal cancer; HBV: Hepatitis-B virus; SVM: Support vector machine; MAFLD: Metabolic associated fatty liver disease; HCC: Hepatocellular carcinoma; CC: Cholangiocarcinoma; ROI: Region of interest; IHC: Immunohistochemistry; MSI: Microsatellite instability; dMMR: Mismatch-repair deficiency; TCGA: The Cancer Genome Atlas; DL: Deep learning.

Basic AI-assisted pathology: diagnosis

The basic role of pathology is disease diagnosis. In the fields of gastroenterology, there is an increasing need for automatic pathological analysis and diagnosis of GC. Based on the virtual version of pathological slices, several studies were performed to identify and classify GC automatically with high AUCs[120-122,126]. For example, a CNNs model was developed to distinguish gastric mass lesions including gastric adenocarcinoma, adenoma and non-neoplastic lesions, and it has achieved the highest AUC of 0.97 for the identification of gastric adenocarcinoma[126]. With regard to colorectal lesions, Wei et al[128] trained an AI-assisted model to classify colorectal polyps on WSIs, and notably, the performance of the model was similar to that of local pathologists whether in a single institution or other institutions. Besides diagnosis, a model based on more than 400 WSIs was developed to differentiate five common subtypes of colorectal polyps with accuracy of 93%[127]. In CRC, Shapcott et al[129] performed a retrospective study to develop a CNNs model for diagnosis based on 853 hand-marked images with an accuracy of 84%.

In the fields of hepatology, AI-assisted pathology is applied in patients with hepatitis B virus (HBV), metabolic associated fatty liver disease, HCC, etc. An automated, stain-free AI system can quantify the amount of fibrillar collagen to evaluate the degree of HBV-related fibrosis with the AUC > 0.82[132]. For patients with metabolic associated fatty liver disease, AI-assisted pathology tools were used to identify and quantify pathological changes including steatosis, macrosteatosis, lobular inflammation, ballooning, and fibrosis[133], and the algorithm output scores for quantitative comparison with experienced pathologists achieved good agreement. However, limited AI-assisted pathology tools have been built for HCC diagnosis. Notably, the MFC-CNN-ELM program was designed for nuclei grading of biopsy specimens from HCC patients, which revealed high performance in classifying tumor cells of different differentiation stages[134].

Advanced AI-assisted pathology: prediction of gene mutations and prognosis

Apart from AI-assisted pathology tools in diagnosis, it is no surprise that many tools have been developed for the prediction of gene mutations and prognosis in the fields of gastroenterology and hepatology. In CRC, AI tools have shown great effectiveness in predicting prognosis across all tumor stages based on WSIs[139,140], and several prospective multicenter studies have further validated the high prognosis performance[142]. Notably, a subset of genetic defects occurring in gastroenterology is related to some morphological features detected on WSIs. Among screened genetic defects, microsatellite instability and mismatch-repair deficiency are associated with the survival of gastrointestinal and colorectal cancer patients receiving immunotherapy. Therefore, an AI tool was designed to predict microsatellite instability and mismatch-repair deficiency directly from pathology, and it finally showed reasonably good performance in assisting immunotherapy[138,141]. Notably, Kather et al[140] further validated the above model’s performance in predicting overall survival from CRC pathology slides with a hazard ratio of 2.29 in CRC-specific overall survival (OS) and an hazard ratio of 1.63 in OS, respectively. However, besides the above-mentioned studies that have focused on tumor detection of CRC, few studies were designed to predict gene mutations and prognosis due to the more complicated and heterogeneous histomorphology in gastric diseases than that in the colon[136,137].

In the fields of hepatology, AI tools are mainly used to predict gene mutations and prognosis in HCC. For example, a model has higher accuracy in predicting survival postoperatively than using a composite score of clinical and pathological factors in HCC. In addition, the model may generalize well after validating the performance in an external dataset with different staining and scanning methods[146]. Chen et al[11] investigated a CNN (Inception V3) for automatic classification (benign/malignant classification with 96.0% accuracy, and differentiation degree with 89.6% accuracy) and gene mutation prediction from WSIs after resection of HCC. It was found that CTNNB1, FMN2, TP53, and ZFX4 could be predicted from WSIs with external AUCs from 0.71 to 0.89. Currently, after integrating clinical data, biological data, genetic data, and pathological data, the novel model may also be a promising approach. The first multi-omics model combined ribonucleic acid (RNA) sequencing, miRNA sequencing and methylation data from The Cancer Genome Atlas, and then employed AI technologies to predict and differentiate survival of HCC patients[145]. Other attempts have been made to develop models that can predict gene mutations directly based on WSIs of HCC. Using AI-assisted pathology, some approaches can predict gene expression and RNA sequencing, which may have the potential for clinical translation[147]. Interestingly, some gene expression such as PD-1 and PD-L1 expression, inflammatory gene signatures, and biomarkers of inflammation did trend with improved survival and response in HCC patients[148].

LIMITATIONS AND FUTURE CONSIDERATIONS

This review retrospectively summarized some key and representative articles with the possibility of missing some publications in AI-related journals. Although various studies have shown promising results in the fields of AI-assisted gastroenterology and hepatology, there are still several limitations to be discussed and resolved. One of the major criticisms is the lack of high-quality training, testing, and validation datasets for the development and validation of AI models. Due to the retrospective manner of most studies, selection bias must be considered at the training stage, meanwhile, overfitting and spectrum bias may result in overestimation of the model accuracy and generalization. According to the rigorous “six-steps” translation pipelines[149], doctors and AI researchers should join the calls that advocate for developing interconnected networks of collecting raw acquisition data which was shifted from processed medical images over the world and training AI on a large scale to obtain robust and generalizable models. Furthermore, the black-box nature of AI technologies has become a barrier to clinical practice, because developers and users do not know the details about how computers output the conclusion. Explainable AI for reliable healthcare is worth investigating to reach clinical interpretability and transparency. In addition, from the perspective of ethics and legal liabilities, AI models may potentially cause errors and challenge the patient-doctor relationship despite the fact that they improve the clinical workflow with enhanced precision. Especially in the fields of gastroenterology and hepatology, cancer discrimination may mean a completely different treatment. If misdiagnosis occurs during AI application, who should take responsibility- the doctor, the programmer, the company providing the system, or the patient? Issues such as ethics and legal liabilities should be demonstrated in the early phase to maintain the balance between minimal error rates and maximal patient benefits[150,151].

There have been an increasing number of studies applying AI to gastroenterology and hepatology over the past decade. In the future, the trend will continue and larger studies will be carried out to compare the performance of medical professionals with AI vs professionals without AI to highlight the importance of AI assistance. AI technologies will be utilized to develop more accurate models to predict and monitor disease progression and potential complications, and these models may ameliorate the insufficiency of medical resources in the remote underserved or developing regions. Besides, AI-assisted personalized imaging protocols and immediate three-dimensional reconstruction may further improve the diagnostic efficiency and accuracy. Researchers will be able to realize the mechanism of disease progression and treatment response through the combination of multi-modality images or multi-omics data. In addition, there is an emerging trend applying AI to drug development, such as prediction of compound toxicity, physical properties, and biological activities, which may assist chemotherapy for digestive system malignancy. Furthermore, AI could be used to process the data generated from the tissue-on-a-chip platform which could better summarize the tumor microenvironment, thus reach precise and individual chemotherapy in gastroenterology and hepatology. As synthetic lethality becomes a promising genetically targeted cancer therapy[152,153], AI could also be used for the detection of target synthetic lethal partners of overexpressed or mutated genes in tumor cells to kill cancers. Finally, AI tools could not replace endoscopists, radiologists, and pathologists in the near and even distant future. Computers would make predictions and doctors would make the final decision, in other words, they would always work together to benefit patients.

CONCLUSION

AI is rapidly developing and becoming a promising tool in medical image analysis of endoscopy, radiology, and pathology to improve disease diagnosis and treatment in the fields of gastroenterology and hepatology. Nevertheless, we should be aware of the constraints that limit the acceptance and utilization of AI tools in clinical practice. To use AI wisely, doctors and researchers should cooperate to address the current challenges and develop more accurate AI tools to improve patient care.

ACKNOWLEDGEMENTS

We thank Yun Cai for polishing our manuscript. We are grateful to our colleagues for their assistance in checking the data of the studies.