Deep learning-based detection of hepatobiliary disorders in ophthalmic imaging

Abstract

Artificial intelligence has emerged as a powerful tool that can synthesize and analyse complex multimodal data across all domains of health care. Some artificial intelligence technologies enhance clinical workflow by automatically performing repetitive and time-consuming tasks faster and more efficiently than humans, while others lead to innovations through development of novel biomarkers or therapeutic tools.1Bohr A Memarzadeh K The rise of artificial intelligence in healthcare applications.Artif Intell Health. 2020; (published online June 26.)https://doi.org/10.1016/B978-0-12-818438-7.00002-2Crossref Scopus (71) Google Scholar The discovery of a new role for an inexpensive, standardised, and widely available test can sometimes be even more impactful than a new invention because it might be much easier to implement universally to resource-limited environments. In The Lancet Digital Health, Wei Xiao, Xi Huang, Jinghui Wang and colleagues present an interesting study of applying deep learning models to detect seven categories of hepatobiliary disorders on two common types of ocular images: slit-lamp and retinal fundus images.2Xiao W Huang X Wang J et al.Screening and identifying hepatobiliary diseases through deep learning using ocular images: a prospective, multicentre study.Lancet Digit Health. 2021; 3: e88-e97Summary Full Text Full Text PDF PubMed Scopus (12) Google Scholar The authors used single-modality images from the ocular imaging datasets to train a deep neural network, ResNet-101, to develop a screening model and several independent identifying models (seven slit-lamp models and seven retinal fundus models in total). The authors used data collected from 1252 participants for the development dataset and data from 537 additional participants for the test dataset. The area under the receiver operating characteristic curve for screening for hepatobiliary diseases was 0·74 (95% CI 0·71–0·76) for the slit-lamp model and 0·68 (0·65–0·71) for the fundus model. The slit-lamp model showed high performance for detection of liver cancer (0·93, 95% CI 0·91–0·94) and cirrhosis (0·90, 0·88–0·91), whereas the slit-lamp and fundus models did not perform as well for detection of other conditions such as chronic viral hepatitis, non-alcoholic fatty liver disease, cholelithiasis, and hepatic cyst. The true novelty of this study is that it ties together two seemingly distant disciplines of medicine: hepatology and ophthalmology. The application of artificial intelligence has seen an exponential growth in all fields of medicine, with hepatology and ophthalmology being no exceptions. In hepatology, artificial intelligence is being used to develop prognostic models, chart review tools using natural language processing, and deep learning-based software for interpretation of cross-sectional imaging and histopathological images.3Ahn JC Connell A Simonetto DA Hughes C Shah VH The application of artificial intelligence for the diagnosis and treatment of liver diseases.Hepatology. 2020; (published online Oct 23.)https://doi.org/10.1002/hep.31603Crossref Scopus (28) Google Scholar In ophthalmology, there have been successful applications of deep learning algorithms for diagnosis of ocular conditions such as glaucoma, diabetic retinopathy, and macular degeneration.4Ting DSW Pasquale LR Peng L et al.Artificial intelligence and deep learning in ophthalmology.Br J Ophthalmol. 2019; 103: 167Crossref PubMed Scopus (390) Google Scholar Nevertheless, this is the first study to propose an entirely new role for ophthalmological imaging to serve as a screening tool for early detection of hepatobiliary disorders. Cirrhosis develops insidiously over many years of injury and many patients first get diagnosed when they present with advanced liver cancer or an acute decompensation event such as ascites, variceal haemorrhage, or hepatic encephalopathy.5Asrani SK Kamath PS Natural history of cirrhosis.Curr Gastroenterol Rep. 2013; 15: 308Crossref PubMed Scopus (55) Google Scholar Therefore, there is a clear unmet need for a simple, non-invasive, and reliable screening tool for cirrhosis, which makes this study timely and important. Another useful step that Wei Xiao and colleagues took in this study was the incorporation of elements of explainable artificial intelligence into their model. Deep learning networks typically consist of more than ten layers and millions of artificial neurons and rely on extremely complex interconnected representations of the data to make their predictions.6Esteva A Robicquet A Ramsundar B et al.A guide to deep learning in healthcare.Nat Med. 2019; 25: 24-29Crossref PubMed Scopus (951) Google Scholar This leads to the so-called black-box problem where the deep neural networks are capable of producing highly accurate predictions, but humans cannot understand how the decision was made.7Castelvecchi D Can we open the black box of AI?.Nature. 2016; 538: 20-23Crossref PubMed Scopus (512) Google Scholar There is a growing emphasis on explainable artificial intelligence, with efforts to enhance interpretability of deep learning models and find new ways to visualise the features that artificial intelligence models use to make predictions.8Montavon G Samek W Müller K-R Methods for interpreting and understanding deep neural networks.Digit Signal Process. 2018; 73: 1-15Crossref Scopus (866) Google Scholar Wei Xiao and colleagues generated visualisation heatmaps and did occlusion tests to evaluate the relative contributions to the predictions from different regions in the eye and obtained consistent results showing that the model's performance was dependent on information obtained from the iris and conjunctiva and sclera regions. In future research projects using deep learning networks, including similar steps to explain the artificial intelligence model's decision making will be crucially important not only for the readers of the project manuscripts, but also for the health-care providers who will actually use the models for patient care. Despite the novelty and promises of this study, more work needs to be done before these ophthalmological imaging-based deep learning models can be recommended as a reliable screening tool for hepatobiliary disorders. Further studies focusing specifically on cirrhosis or liver cancer should be done on a larger cohort of better characterised patients with clearly delineated severity and extent of disease. Overall, this study is a nice example of how artificial intelligence can lead to innovation in health care by taking an existing and widely used test in one field of medicine and finding a completely new purpose for the test in another field by applying state-of-the-art artificial intelligence technology. VHS reports personal fees from Akaza Bioscience, Ambys Medicines, Durect Corporation, and Surrozen; and other support from Evive Biotech (formerly Generon Biomed), outside the submitted work. JCA declares no competing interests. Screening and identifying hepatobiliary diseases through deep learning using ocular images: a prospective, multicentre studyOur study established qualitative associations between ocular features and major hepatobiliary diseases, providing a non-invasive, convenient, and complementary method for hepatobiliary disease screening and identification, which could be applied as an opportunistic screening tool. Full-Text PDF Open Access

Ähnliche Arbeiten