Deep learning as a staging tool in gastric cancer

Accurate staging of gastric cancer, following the American Joint Committee on Cancer (AJCC)/Union for International Cancer Control (UICC) classification system,1In H. Solsky I. Palis B. et al.Validation of the 8th edition of the AJCC TNM staging system for gastric cancer using the National Cancer Database.Ann Surg Oncol. 2017; 24: 3683-3691Crossref PubMed Scopus (161) Google Scholar is crucial to determine appropriate treatment. Although most patients in Western countries will present with advanced disease, those with locoregional disease are candidates for curative surgical intervention, usually combined with (neo-)adjuvant chemotherapy and chemoradiotherapy where optimal surgery is guided by tumour size (T-stage) and number of malignant nodes (N-stage). Imaging is the cornerstone of gastric cancer staging and routinely consists of endoscopic ultrasound and cross-sectional imaging. A Cochrane systematic review showed that endoscopic ultrasound could accurately discriminate T1/T2 tumour from T3/T4, but performed less well regarding lymph node characterisation.2Mocellin S. Pasquali S. Group CUG and PDDiagnostic accuracy of endoscopic ultrasonography (EUS) for the preoperative locoregional staging of primary gastric cancer.Cochrane Database Syst Rev. 2015; 2015: CD009944Google Scholar Computed tomography (CT), and increasingly positron emission tomography (PET)/CT, allows for the detection of metastatic disease; however, diagnostic performance of local staging and particularly lymph node involvement are variable,3Choi J.-I. Joo I. Lee J.M. State-of-the-art preoperative staging of gastric cancer by MDCT and magnetic resonance imaging.World J Gastroenterol. 2014; 20: 4546-4557Crossref PubMed Scopus (40) Google Scholar posing an ongoing diagnostic problem and driving the search for new imaging approaches. The application of artificial intelligence (AI) using deep learning models and radiomics is revolutionising the field of cancer imaging. Imaging is frequently carried out for cancer diagnosis and treatment response assessment. As such, there is a wealth of data available from which now new quantitative features, invisible to the human eye, can be extracted which may aid in disease detection, treatment decisions and prognostication of clinical outcomes. The potential of AI in imaging has been demonstrated in several recent publications. For instance, a deep learning algorithm was able to accurately detect breast cancer on screening mammograms, attaining lower false-positive and false-negative rates, outperforming radiologists.4McKinney S.M. Sieniek M. Godbole V. et al.International evaluation of an AI system for breast cancer screening.Nature. 2020; 577: 89-94Crossref PubMed Scopus (701) Google Scholar In another study by Lu et al.,5Lu H. Arshad M. Thornton A. et al.A mathematical-descriptor of tumor-mesoscopic-structure from computed-tomography images annotates prognostic- and molecular-phenotypes of epithelial ovarian cancer.Nat Commun. 2019; 10: 764Crossref PubMed Scopus (79) Google Scholar a radiomic signature improved prognostication in patients with epithelial ovarian cancer, identifying those with a median overall survival (OS) of less than 2 years. In gastric cancer, deep learning software has been employed for histopathological characterisation6Iizuka O. Kanavati F. Kato K. et al.Deep Learning models for histopathological classification of gastric and colonic epithelial tumours.Sci Rep. 2020; 10: 1504Crossref PubMed Scopus (114) Google Scholar or risk stratification,7Nakahira H. Ishihara R. Aoyama K. et al.Stratification of gastric cancer risk using a deep neural network.JGH Open. 2019; https://doi.org/10.1002/jgh3.12281Crossref PubMed Scopus (12) Google Scholar but is still in an early stage in imaging. The largest imaging study to date used a real-time AI tool to analyse endoscopic images in 84 424 subjects for the detection of gastric cancer, improving the performance of non-expert endoscopists while being noninferior to expert assessment. Only few and generally small-sized studies investigated the use of CT-derived deep learning or radiomic algorithms in gastric cancer and found that these tools may aid in discriminating histopathological features or predicting response and survival. A larger-scale study by Jiang et al.8Jiang Y. Chen C. Xie J. et al.Radiomics signature of computed tomography imaging for prediction of survival and chemotherapeutic benefits in gastric cancer.EBioMedicine. 2018; 36: 171Abstract Full Text Full Text PDF PubMed Scopus (86) Google Scholar interrogated the prognostic role of radiomics based on contrast-enhanced preoperative CT in 1591 patients with gastric cancer. Their radiomic signature was able to predict disease-free survival (DFS) and OS better than clinicopathological features or TNM stage. Moreover, the radiomics score allowed for the identification of patients with stage II and III disease likely to respond to adjuvant chemotherapy. In the current issue of Annals of Oncology, Dong et al.9Dong D. Fang M.-J. Tang L. et al.Deep learning radiomic nomogram can predict the number of lymph node metastasis in locally advanced gastric cancer: an international multicenter study.Ann Oncol. 2020; 31: 912-920Abstract Full Text Full Text PDF PubMed Scopus (89) Google Scholar developed a deep learning radiomic model derived from preoperative contrast-enhanced CT, which is able to predict lymph node metastasis in locally advanced gastric cancer more accurately than clinical N stage and other clinical characteristics, and could be useful to determine the needed extent of lymphadenectomy. The results of these studies are promising, yet also raise the question whether these AI algorithms can be used in different clinical settings and different tumour types. To prove algorithm robustness, the software was first developed in a training set and thereafter validated in multiple external training cohorts according to ruling methodological best practice. Furthermore, Dong et al. released their nomogram via online open access, encouraging other groups to test their algorithm on their own data and thereby supporting further validation. Nevertheless, the nodal discriminatory performance of the nomogram was considerably lower in a nongastric cancer patient cohort, as was its ability to predict OS compared with the radiomic signature of Jiang et al.8Jiang Y. Chen C. Xie J. et al.Radiomics signature of computed tomography imaging for prediction of survival and chemotherapeutic benefits in gastric cancer.EBioMedicine. 2018; 36: 171Abstract Full Text Full Text PDF PubMed Scopus (86) Google Scholar These observations indicate that these AI tools are not a ‘one-size-fits-all’ solution and very much tailored to a specific clinical question and tumour type, which considerably limits the implementation of AI in practice. Interpretability of imaging-derived deep learning and radiomic models remains complicated, as these essentially represent mathematical features without an evident biological substrate. Previous research has shown that deep neural networks are easy to fool, with minimal changes in the images resulting in substantive classification changes.10Szegedy C. Zaremba W. Sutskever I. et al.Intriguing properties of neural networks. 2nd Int. Conf. Learn. Represent. ICLR 2014 - Conf.Track Proc. 2014; Google Scholar Therefore, the integration of imaging AI features with clinical, genetic and histopathological data is of great importance to ensure biologically meaningful features have been calculated. This has been achieved in several studies, including a previously-mentioned radiomics study in ovarian cancer patients,5Lu H. Arshad M. Thornton A. et al.A mathematical-descriptor of tumor-mesoscopic-structure from computed-tomography images annotates prognostic- and molecular-phenotypes of epithelial ovarian cancer.Nat Commun. 2019; 10: 764Crossref PubMed Scopus (79) Google Scholar in which the radiomic signature was shown to correlate with specific oncogenetic pathways and a stromal phenotype. Similarly, Sun et al.11Sun R. Limkin E.J. Vakalopoulou M. et al.A radiomics approach to assess tumour-infiltrating CD8 cells and response to anti-PD-1 or anti-PD-L1 immunotherapy: an imaging biomarker, retrospective multicohort study.Lancet Oncol. 2018; 19: 1180-1191Abstract Full Text Full Text PDF PubMed Scopus (486) Google Scholar validated a radiomic signature estimating CD8 tumour-infiltrate in patients treated with programmed death-ligand 1 (PD-L1) and programmed death-1 (PD-1) targeted monotherapy against a genomic score predicting CD8 cell abundance and demonstrated a significant correlation between the radiomic signature and the underlying immunologic phenotype. Currently, no such studies have been carried out in gastric cancer. Imaging-derived AI research requires large datasets and obtaining sufficient data to properly train and validate the AI algorithm is a significant hurdle. To counter this problem, the National Cancer Institute has created a large image repository, the National Biomedical Imaging Archive, analogous to The Cancer Genome Atlas Program. The main purpose of this database is to set up a large data warehouse enabling advances in the development in deep learning and radiomic algorithms. However, this is still work in progress and presently mainly contains data of more prevalent malignancies such as lung and breast cancer. As a result, due to their lower occurrence, defining a large enough patient cohort in upper GI cancers continues to be challenging. AI is rapidly changing the precision medicine landscape. Although imaging-derived AI is still in its infancy, the first promising studies are now being published assessing its role in gastric cancer. Understanding the strengths and weaknesses of AI tools based on imaging and integrating imaging features with clinical findings, genetics and histopathological data is essential to interpret published data correctly and to set out the right direction in this new exciting field. There are no funders to report for this submission.