Role of artificial intelligence in gastrointestinal surgery

Abstract

Artificial intelligence is rapidly evolving and its application is increasing day-by-day in the medical field. The application of artificial intelligence is also valuable in gastrointestinal diseases, by calculating various scoring systems, evaluating radiological images, preoperative and intraoperative assistance, processing pathological slides, prognosticating, and in treatment responses. This field has a promising future and can have an impact on many management algorithms. In this minireview, we aimed to determine the basics of artificial intelligence, the role that artificial intelligence may play in gastrointestinal surgeries and malignancies, and the limitations thereof.

Key Words: Machine learning; Gastrointestinal surgery; Radiology; Artificial neural network; Deep learning

Core Tip: Artificial intelligence (AI), a term coined by John McCarthy in 1955, is the new-kid-on-the-block in the medical arena, with immense potential to revolutionize how patients may be managed in coming years. It is a science of creating machines with capability to reason and solve problems like human beings. The principal subfields of AI are: (1) Machine learning; (2) artificial neural networks; (3) natural language processing; and (4) computer vision. AI is improving the fields of gastrointestinal surgery and medicine immensely, assisting in diagnoses as well as management of gastrointestinal diseases. As with any new technology, AI has limitations and drawbacks.

INTRODUCTION

The platform of artificial intelligence is revolutionizing the medical field and changing approaches to reach a definitive diagnosis. Furthermore, artificial intelligence is expanding its horizons not only in clinical diagnoses, but also in the operative management of patients. Leonard et al[1] and John Hopkins University have shown this by developing the Smart Tissue Autonomous Robot (STAR) to perform ex vivo and in vivo anastomosis of the small bowel in porcine models[2]. Artificial intelligence can be vaguely defined as the study of sets of algorithms giving machines the ability to perform reasoning and solve problems, along with image, object, and word identification[3,4]. Another innovation in the surgical field was Gestonurse, which is a robotic scrub nurse used to hand instruments to the operating surgeon[5].

The index application of artificial intelligence in gastrointestinal surgery was documented in 1976, when it was introduced to aid computer analysis for acute abdominal pain[6]. The term artificial intelligence was coined by McCarthy et al[7] in 1955 during his summer research project. However, its origin can be traced back to 1950, when Turing[8], a British mathematician, used a computer to display human-like behavior in the Turing test.

Most of the initial advancements of artificial intelligence are depicted in the field of dermatology, radiology, and pathology, noted by an increase in Food and Drug Administration approvals for artificial intelligence devices in these fields[9]. Also, in pathology, a reduction in error rates from 3.4% to 0.5% was seen with the help of artificial intelligence[10]. Artificial intelligence has also seeped deeply into our lives in the form of daily assistance devices such as Siri, Alexa, chatbots, Google assistant, smart homes, navigation apps, autonomous vehicles, etc. Similarly, machine and deep learning is assisting physicians in diagnosing various gastrointestinal diseases, liver tumors, pancreatic neoplasms, and various infections.

Applying artificial intelligence to large and complex data is also helping to identify new variables and their associations, shaping changes in today’s clinical practices. In the surgical field, the surgeon must associate with data scientists to unfold the role of artificial intelligence. In the future, human roles may be limited to only supervision and authentication of different models of artificial intelligence.

METHODOLOGY

Search strategy

We performed a PubMed search for relevant articles, and then searched the article reference lists for additional appropriate studies. The keywords and combinations included in the search were as follows: “Artificial intelligence” and “radiology”; “artificial intelligence” and “pathology”; “artificial intelligence” and “endoscopic management”; “artificial intelligence” and “gastrointestinal surgery”; “artificial intelligence” and “gastric cancer”; “artificial intelligence” and “colorectal malignancy”; “artificial intelligence” and “pancreatic cancer”; and “artificial intelligence” and “robotic surgery”. The search was limited to publications in English. All the authors agreed that the articles selected for review were relevant.

FUNDAMENTALS OF ARTIFICIAL INTELLIGENCE

The basis and advances of artificial intelligence relies on machine learning and deep learning. Artificial intelligence has found many uses in surgical diseases, from diagnosis and scoring systems to preoperative and intraoperative planning and assistance, as well as in surgical training and robotics[11-13]. Principal subfields of artificial intelligence are: (1) Machine learning; (2) artificial neural networks; (3) natural language processing; and (4) computer vision (Figure 1).

Figure 1 Principal subfields of artificial intelligence.

The purpose of machine learning is to equip machines to learn and predict by identifying similar patterns[14]. Machine learning is based on the ability to do a task with the help of a set of algorithms without programming, and is further subdivided into supervised, unsupervised, and reinforced learning[4]. Supervised learning is beneficial to train machine learning algorithms to differentiate an anticipated sequence, whereas unsupervised learning is beneficial in identifying patterns in the given data[14]. An advantage of machine learning is noted in finding minute patterns in large data, which may be missed by human analysis, with the aid of complex and indirect nonlinear relationships or multivariate effects[15]. In surgical site infection analysis, machine learning has surpassed logistic regression with the help of nonlinear regression patterns, including multiple data resources like diagnosis, treatment, and blood tests[16]. An analogy of supervised learning is equipping a device to identify gastrointestinal metaplasia utilizing a large data set of lesions formerly recognized by a pathologist as gastrointestinal metaplasia. Comparatively, an example of unsupervised learning is tissue sample clustering derived from similar gene expressions, as it is based on finding patterns rather than algorithms[14].

Reinforced learning is a type of machine learning wherein the device makes decisions to achieve the most optimal results. It is similar to the trial and error learning process that humans use to fulfill their goals. An example of such a module is an artificial pancreas device used to accurately measure and inject doses of insulin for diabetic patients[17]. Another use of machine learning is precisely predicting mortality, sepsis, and acute kidney injury with the help of an intraoperative database[18,19].

An artificial neural network applying binary threshold functions was conceived in 1943 by McCulloch and Pitts[20]. It is the subdivision of artificial intelligence persuaded by the human nervous system and applied in various medical applications. Deep learning is a new type of artificial intelligence based on artificial neural networks. The deep learning neural network is composed of multiple layers, which enables them to learn minuscule patterns from the given datasets in comparison to simple or two-layered neural networks[21]. In deep learning, every network subgroup of inputs leads to various concealed subgroups in response to distinct characteristics of inputs, which allows for an in-depth understanding of datasets similar to multiple neuronal connections occurring in the human brain[22,23]. In the arena of clinical applications, the artificial neural network has shown high accuracy and sensitivity and specificity in predicting in-hospital mortality following abdominal aortic aneurysms. Similarly, its application in acute pancreatitis to assess severity has surpassed conventional APACHE II scoring (sensitivity and specificity of 80% and 85%, respectively), with a sensitivity and specificity of 89% and 96%, respectively[24]. To analyze images, convolutional neural networks are a type of deep neural network used in the detection of gastrointestinal malignancy within endoscopic images, radiologic images, and pathologic specimen assessments[14,25].

Computer vision, a subtype of artificial intelligence, portrays the ability of a device or machine to obtain and interpret images or videos like humans. It has been widely used in axial imaging for virtual colonoscopy, computer aided diagnosis, and decision making in surgery[26]. Real time assessments of minimally invasive laparoscopic sleeve gastrectomy videos have identified unplanned or forgotten surgical steps with 93% accuracy using computer vision[27].

Natural language processing is the subfield of artificial intelligence dealing with interpretation of text form and voice data from electronic health records[28]. It has a role in both preoperative and postoperative assessments of surgical patients to predict risks, outcomes, and complication rates. It is used in the surgical field to follow up on patients and monitor complications[29]. One of the most frequent applications of natural language processing is to foretell the risk of surgical site infection in various surgical procedures[30-32].

ARTIFICIAL INTELLIGENCE IN GASTROINTESTINAL SURGERY

Artificial intelligence is improving the field of gastrointestinal surgery, and helping in the diagnosis and management in many gastrointestinal diseases where images (endoscopy, radiology, and pathology) are used such as in Barrett’s esophagus, esophageal cancers, gastric cancer, Helicobacter infection, pancreatic diseases, liver tumors, colonic polyps, and malignancies. It also helps in identifying precancerous lesions more efficiently and to reduce missing small lesions. In comparison to other fields, the diagnosis of surgical diseases require more dynamic and complex algorithms, sometimes involving real time decision making, similar to differences in artificial intelligence software used to recognize faces and to drive autonomous vehicles.

In abdominal wall hernia repair, artificial intelligence uses supervised machine learning algorithms and preoperative computed tomography images to anticipate risks and requirements of component separation techniques. In a study, this model was found to be superior to a panel of consultant surgeons[33]. Artificial intelligence has been used in convolutional neural network-based algorithms to find stomach malignancy with endoscopic images since 2018[34]. One multicenter study introduced the Gastrointestinal Artificial Intelligence Diagnostic System model to diagnose upper gastrointestinal cancers in late 2019, comparing its accuracy with a specialist and trainee endoscopist, and noticed it to be better than the trainee and similar to the specialist in diagnosing stomach carcinoma[35]. Convolutional neural networks using narrow band imaging in endoscopy have been used to differentiate stomach cancer and gastritis with an upper limit of sensitivity of 95% and positive predictive value of 91%[36]. In studies of Barrett’s esophagus, deep learning model algorithms have been used to find neoplastic changes based on datasets of images, showing superior accuracy to amateur endoscopists, along with the optimum site for biopsy in 97% cases[37]. Another study using a convolutional neural network-based model was found to be more accurate in detecting early gastric carcinoma than a specialist endoscopist[38]. Other studies have used deep learning modules along with biomarkers, histologic images, and radiologic images to anticipate liver and lymph node spread in stomach carcinoma[39,40]. Various studies have also used evolved models anticipating risk stratification of stomach carcinoma and its survival prediction[41,42]. In esophageal carcinoma, deep learning models have been developed to find squamous cell carcinoma of the esophagus using narrow band imaging as the input; models anticipating depth of esophageal malignancy have also been developed[43-45].

In the large bowel, artificial intelligence has been used to find polyps and differentiate hyperplastic or adenomatous polyps based on deep learning and convolutional neural networks[46,47]. Artificial intelligence is also helpful in segregating malignant or non-malignant colonic lesions, as well as predicting depth of invasion and endoscopic resection margins[48,49]. With the aid of convolutional neural network-based deep learning, we can find real time automatic surgical phases in laparoscopic sigmoidectomy, to teach purposes and real time workflow recognition for the development of Context-Aware Computer-Assisted Surgery systems[50]. Platforms using electronic health records can anticipate perioperative risk, to improve clinical assessments using MySurgeryRisk algoritms[51]. Another similar machine learning user friendly platform for risk stratification and mortality in emergency surgery is Predictive Optimal Trees in Emergency Surgery Risk, which is based on the ACS-NSQIP dataset and is fruitful in preoperative counseling of patients and relatives[19].

Anticipation of resection of peritoneal carcinomatosis necessitates laparotomy to optimally classify the disease burden. Random forest models based on machine learning can determine resectability in this condition with accuracy to the tune of 98%, preventing unwanted laparotomies[52]. Another machine learning based model has been developed to examine intrahepatic cholangiocarcinoma patients preoperatively using various parameters, and to anticipate which patients would benefit from surgery[53,54]. Using artificial neural networks, the uterine artery can be distinguished from the ureter in laparoscopic hysterectomy, as well as other surgeries posing risk of ureteral injury[55]. In hepatic surgery, artificial intelligence helps in preoperative planning, intraoperative assistance using radiological imaging, and assessing the intricate anatomy[56]. In cases of pancreatic surgery, artificial intelligence is useful in anticipating postoperative pancreatic fistula risk following pancreatic head resection, and to help in planning management for better results[57]. In cholangiocarcinoma, computer auxiliary diagnosis systems can aid in the diagnosis and help improve the treatment plan[58].

In various gastrointestinal surgeries, postoperative management can be augmented with the help of artificial intelligence by predicting surgical site infection and blood loss requiring transfusion following pancreatic, colorectal, or hepatic surgeries[59]. After gastrointestinal surgeries like colorectal and bariatric surgery, natural language processing aids the use of variables from health records to preoperatively predict patients who have higher risks of anastomotic leak[51,60]. Artificial intelligence can also support intraoperatively using computer vision analysis of intraoperative videos and generate recommendations or warnings[61]. It can also assist in transplant surgeries both pre and post operatively by anticipating graft rejection or failure, morbidity, and survival post liver transplantation[62,63].

ROLE OF ARTIFICIAL INTELLIGENCE AND RADIOLOGY IN GASTROINTESTINAL TRACT SURGERY

Radiology assists in gastrointestinal diseases using images generated through X-ray, computed tomography, and magnetic resonance imaging, which are used by deep learning models as data inputs to identify anatomy and differentiate malignant or benign tumors[64]. Radiomics has been used to assess and differentiate the malignant potential of various tumors by assessing images via deep learning of shapes, texture, and histograms[65,66]. Some studies have developed survival prediction models with the help of radiomics analysis of computed tomography images. Following image segmentation, the analysis was accomplished and risk scores stratified, showing the superiority of radiomics over normal nomograms in stomach cancer patients[67,68].

In colorectal liver metastases using computed tomography and magnetic resonance imaging, radiomics have depicted splendid accuracy in finding early liver spread, which might aid in effective management and better prognosis. It is also helpful in predicting responses to chemotherapy[69]. Studies have tried to predict peritoneal spread with the help of convolutional neural networks and diagnosing metastatic lymph node spread in carcinoma of the stomach[39,70]. Three-dimensional imaging delineation of anatomy and three-dimensional reconstructions improve planning in pancreatic, liver, or biliary surgery[71-73]. Navigation systems technology helps in pre and postoperative surgeries to depict difficult anatomy, and is currently used in pancreatic, esophageal, and splenic surgeries[74-76]. Intuitive Surgical has also developed IRIS 1.0, providing segmented three-dimensional images for better analyses of anatomy and planning surgery[77]. Recently, some studies have shown use of optical coherence tomography in finding metastatic lymph nodes using artificial neural networks, which may help intraoperatively[78].

SHORTCOMINGS OF ARTIFICIAL INTELLIGENCE

As with any new technology, artificial intelligence also has limitations and drawbacks, as it depends heavily on the data provided. The quality, amount, and variation in the data used to train the model is an important aspect. Outcomes depend on the quality of the dataset provided for input, and any shortcomings will affect the results. Biases while accumulating clinical data have an impact on the kind of patterns artificial intelligence perceives and the predictions it makes[79,80]. In gastrointestinal malignancy, IBM Watson Oncology—a question–answer-based computer device—was supposed to help physicians in decision making and to keep updated with recent evidence; however, it has not lived up to the standard[81,82]. In the presence of well-devised models, depending entirely on artificial intelligence is not recommended, as a skilled endoscopist is still needed to accurately capture the data[83].

Radiological models are deficient in automatic segmentation in gastrointestinal malignancies due to large discrepancies in imaging of the gastrointestinal tract. Radiomics appears to be efficient in finding malignant probability with analyses of intratumor heterogeneity; however, it also has limitations in the form of interobserver variability during segmentation, scanning equipment, image acquisition protocols, and algorithm reconstruction[84-86]. In endoscopic imaging, most studies are retrospective in nature and do not represent real clinical scenarios, and it is not clear that artificial intelligence will be cost effective and beneficial in improving clinical results[87]. Similarly, in pathology, there is interobserver discord commonly noticed in various diagnoses among pathologists, which affects the learning or input data and the cost effectiveness of artificial intelligence[88,89].

Traditionally, the responsibility and legal implications onus was on the surgeon; however, with autonomous artificial intelligence, these need to be sorted out in respect of accountability, liability, and culpability. Only accountability can be assessed in artificial intelligence with the help of recording actions, but liability and culpability require further introspection[90]. One of the limitations is lack of sixth sense, which a machine cannot possess. Another drawback is sympathy or sensitivity toward the patient, which cannot be replaced by artificial intelligence. Ethical and legal doubts are another issue which must be dealt with. Moreover, in surgical practice, operative efficacy comes by observing and performing cases repeatedly[77]. One more issue is trust, the basis of the patient–doctor relationship, which is lacking in artificial intelligence models[91].

CONCLUSION

Artificial intelligence is advancing at a rampant pace, and is playing a major role in shaping the way patients will be managed in coming years. The implementation of artificial intelligence in diagnosing gastrointestinal disease seems to be at par with the experts; however, it is important to realize that artificial intelligence cannot totally replace humans, and will need assistance for its use, further development, and enhancement. The aim of artificial intelligence is to enrich the skills of a surgeon and not to replace him. Due to its various limitations, more must be done to use artificial intelligence to improve healthcare.