Artificial Intelligence in Surveillance of Barrett's Esophagus

Esophageal adenocarcinoma (EAC) has the fastest growing incidence of any cancer in the United States, with a 7-fold increase projected by 2030. Barrett's esophagus, the precursor lesion for EAC, affects 1% to 2% of the general population and has an annual progression rate to EAC of 0.1% to 0.5% per year. For this reason, patients with Barrett's esophagus undergo routine surveillance to identify those at risk for progression to EAC. Standard endoscopic surveillance includes use of high-resolution white light endoscopy with multiple biopsies. Surveillance programs are however recognized as less than ideal due to a number of factors including—relevant to this article—the endoscopic inability to identify some regions of Barrett's esophagus–associated neoplasia. This in part relates to the heterogeneous non-mass lesion pattern of growth of some Barrett's esophagus–associated neoplasms. Also, random surveillance biopsies are inherently imprecise due to the necessity of only being able to sample a fraction of the total surface area of Barrett's esophagus. It is recognized that a certain number of EACs occur during Barrett's esophagus surveillance—so called “interval cancers” (1). Given what is understood about esophageal carcinogenesis, it is likely that these cancers, or at least their immediate precursor dysplastic lesions, were present at the time of prior surveillance endoscopy.

Improved confidence in the endoscopic detection of Barrett's esophagus–associated dysplasia will go a long way in decreasing over treatment of Barrett's esophagus and decreasing endoscopy rates in patients with Barrett's esophagus overall. Multiple initiatives are in progress to integrate newly discovered biomarkers into screening and diagnostic algorithms (2). Importantly, as new nonendoscopic technologies become available to identify Barrett's esophagus in the general population (3, 4), highly accurate stratification of endoscopic Barrett's esophagus will be very important. Validation of emerging nonbiopsy mucosal sampling devices of Barrett's esophagus-risk assessment will crucially require improved “gold standard” technologies to facilitate their implementation into practice (5, 6). The approach presented by Waterhouse and colleagues is part of a rich decades-long tradition of innovative approaches to improve detection of Barrett's esophagus–associated neoplasia (7). Proposed advanced imaging techniques include software enhanced chromoendoscopy, autofluorescence, endoscopic optical coherence tomography, endoscopic confocal laser endomicroscopy, as well as narrow band imaging (NBI; refs. 8, 9). None of these promising technologies, including NBI, have been universally implemented for a variety of reasons including technical complexities, degree of user experience required, lack of validated clinical data, and cost.

In a previous issue of Cancer Research, Waterhouse and colleagues present a new technology, implemented in the context of an interventional clinical trial, that consists of a custom spectral endoscope that captures detailed attenuation spectra using diverse wavelength light channels (7). This method enhances contrast to identify changes in neovascularization in different categories of Barrett's esophagus including nondysplastic Barrett's esophagus, Barrett's esophagus with dysplasia, and intramucosal EAC. The instrument was easily integrated with a conventional endoscope used in routine clinical practice. On the basis of the spectral signatures obtained using this instrument, in a small number of prospectively imaged patients, the team was able to differentiate between healthy squamous epithelium versus nondysplastic Barrett's esophagus and neoplasia. In addition, the team also employed an artificial intelligence (AI)-based approach named convolutional neural network (CNN) to further aid in the diagnostic classification task, using the spectral signatures derived from the endoscope. The CNN is a specific type of machine learning algorithm that uses annotated training examples (in this case spectral data) corresponding to the categories of interest (e.g., Barrett's esophagus and Barrett's esophagus neoplasia) to identify representations of the spectral data that best distinguish the diagnostic categories. Unfortunately, the process of generating the CNN representations is unsupervised and hence the resulting representations are opaque and not visually interpretable. The CNN was trained using the spectral signatures obtained from regions of interest (ROI) sampled by the endoscope; 80% of the total sampled spectra were employed for training the CNN while the remaining 20% were employed for independent testing of the machine learning model. The CNN was employed to classify the ROIs into either nondysplastic Barrett's esophagus or neoplasia, yielding an accuracy of 84.4% on the test set.

AI-based approaches aiming to improve patient care have gained significant interest in clinical cancer research. AI-enabled technologies to enhance patient care have been identified in a broad array of specialties spanning from improved visual diagnosis of melanocytic skin lesions, pathology image recognition of cancer with abilities to define molecular phenotypes associated with disease prognosis and treatment response, and improved radiologic differentiation of malignant vs. benign masses.

In the context of endoscopic procedures, the FDA recently granted approval to Medtronic for an AI-driven endoscopic device for identifying suspicious polyps. A number of related AI studies have been conducted for interpretation of upper GI endoscopy (10), but the majority of these studies have been retrospective in nature. The study by Waterhouse and colleagues is unique in that it is one of the few studies to evaluate AI in a prospective cohort (7), even though the AI-based approach was not used to modulate management or intervention of the patients on the trial.

The findings from this study are also important in that they provide real-world evidence of the potential for use of AI with spectral endoscopic procedures to better manage interventions. Improved accuracy of surveillance endoscopy with targeted biopsies would have multiple advantages over current practice, including decreasing the number of biopsies and restricting them to patients with a high probability of abnormality. This AI-guided approach would likely improve pathologist interpretation of dysplastic lesions—especially, low-grade dysplasia. Enhancing the confidence of a “negative” endoscopy, due to improved identification of suspicious lesions, would allow for better risk stratification, with more intensive surveillance/intervention in patients with high-risk Barrett's esophagus and prolonging surveillance intervals in patients with low-risk Barrett's esophagus. Among other benefits, this would significantly reduce the high cost burden of routine surveillance in patients with Barrett's esophagus overall.

While the results reported by the study are promising, the road to clinical adoption will have to overcome at least a few road-bumps. First, the study focused on classification of the individual spectra as opposed to classification at a patient level. Strictly speaking for independent assessment of AI-based classification approaches, truly independent training and test sets are needed; the study from Waterhouse and colleagues was a preliminary study and therefore did not have a sufficient sample size to achieve this (7). In addition, while the results compare favorably to other studies previously conducted, it is unclear whether the results are a clinically significant improvement on current practice to allow for clinical adoption or prospective interventional trials. Importantly, one of the limitations of the choice of machine learning approach, in this case CNNs, is that they tend to lack transparency. In other words, it is unclear which attributes of the spectral data were critical for classification, a consideration that could potentially impede clinical adoption.

In spite of the limitations, this study represents an important inflection point in the use of AI for diagnosis of Barrett's esophagus in a prospective setting. In addition, the fact that the study was able to achieve a greater than 84% classification accuracy based off spectral classification alone suggests that the accuracy could be even further improved by combining the approach with 2D endoscopic image classification. Perhaps most critically the study has provided an initial roadmap for how a prospective interventional clinical trial on the use of AI for diagnosis of Barrett's esophagus might be conducted. This will likely improve a larger retrospective validation of the algorithms, enabling locking down of the AI approach on a much larger powered set of cases. This will subsequently set the stage for a prospective randomized clinical trial with an AI-based arm and a control arm (representing the current status quo) for determining need for a biopsy for Barrett's esophagus diagnosis.

A. Madabhushi reports personal fees from Aiforia Inc, grants from Bristol Myers-Squibb, Astrazeneca, Boehringer-Ingelheim, other support from Elucid Bioimaging, and other support from Inspirata Inc. outside the submitted work; in addition, A. Madabhushi has a patent 10,528,848 issued, a patent 10,769,783 issued, a patent 10,783,627 issued, and a patent 10,902,256 issued. J.E. Willis reports grants from NIH and personal fees from Lucid Diagnostics during the conduct of the study, and grants from NIH outside the submitted work; in addition, J.E. Willis has a patent for molecular detection of Barrett's Esophagus issued and licensed to Lucid Diagnostics, a patent for development of a collection device for sampling the esophagus licensed to Lucid Diagnostics, and a patent for identification of esophagus cancer using aneuploidy analysis pending. No disclosures were reported by the other author.

The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH, the U.S. Department of Veterans Affairs, the Department of Defense, or the United States Government.

Research reported in this publication was supported by the NCI/NIH (under award numbers 1U24CA199374–01, R01CA202752–01A1, R01CA208236–01A1, R01CA216579–01A1, R01CA220581–01A1, 1U01 CA239055–01, R01CA249992–01A1, R01CA257612–01A1, 1U01CA239055–01, 1U01CA248226–01, 1U54CA254566–01, 2P50CA150964–06A1, 1UH2/UH3CA205105–01, 2U01 CA152756–06); the National Heart, Lung and Blood Institute 1R01HL15127701A1; the National Institute for Biomedical Imaging and Bioengineering 1R43EB028736–01; the National Center for Research Resources under award number 1 C06 RR12463–01; the VA Merit Review Award IBX004121A from the United States Department of Veterans Affairs Biomedical Laboratory Research and Development Service; the DOD Breast Cancer Research Program Breakthrough Level 1 Award W81XWH-19–1-0668; the DOD Prostate Cancer Idea Development Award (W81XWH-15–1-0558); the DOD Lung Cancer Investigator-Initiated Translational Research Award (W81XWH-18–1-0440); the DOD Peer Reviewed Cancer Research Program (W81XWH-16–1-0329); the Ohio Third Frontier Technology Validation Fund; the Wallace H. Coulter Foundation Program in the Department of Biomedical Engineering; the Clinical and Translational Science Award Program (CTSA) at Case Western Reserve University.