• Department of Biomedical Engineering, Case Western Reserve University, Cleveland, USA.
• Department of Pathology and Laboratory Medicine, Emory University, Atlanta, USA; Radiology and Imaging Sciences, Biomedical Informatics (BMI) and Pathology, Georgia Institute of Technology and Emory University, Atlanta, USA.
• Wallace H. Coulter Department of Biomedical Engineering, Georgia Institute of Technology and Emory University, Atlanta, USA.
• Department of Pathology and Laboratory Medicine, Emory University, Atlanta, USA.
• Department of Population and Quantitative Health Sciences, Case Western Reserve University, Cleveland, USA.
• Department of Biomedical Engineering, Case Western Reserve University, Cleveland, USA; Department of Pathology, University Hospitals Cleveland Medical Center and Case Western Reserve University, Cleveland, USA.
• Radiology and Imaging Sciences, Biomedical Informatics (BMI) and Pathology, Georgia Institute of Technology and Emory University, Atlanta, USA; Wallace H. Coulter Department of Biomedical Engineering, Georgia Institute of Technology and Emory University, Atlanta, USA; Atlanta Veterans Administration Medical Center, Decatur, USA. Electronic address: anantm@emory.edu.

BACKGROUND: Gastrointestinal (GI) cancers (including esophagus, stomach, colon, rectum, pancreas and liver) account for more than one-quarter of all cancer diagnoses and 35% of cancer-related fatalities worldwide. Quantification of tumor-infiltrating lymphocytes (TILs) is a known cancer prognostic marker. In this study, we used computer vision and machine learning [artificial intelligence (AI)] approaches to evaluate the prognostic significance of computational pathology features relating to spatial arrangement and diversity in the appearance of TILs and cancer nuclei across five different types of GI cancers: colon, stomach, pancreas, and rectum adenocarcinoma and liver hepatocellular carcinoma.

PATIENTS AND METHODS: The study comprises >1700 patients from four different sites. Pathomic features (2236) were extracted from hematoxylin-eosin stained whole slide images and the top 9 features were selected by Least Absolute Shrinkage and Selection Operator (LASSO) Cox model. The top prognostic features identified were related to the spatial relationships between TILs and the closest cancer nuclei and tumor nuclei shape and texture features captured within local cellular clusters.

RESULTS: Our trained model identified that 'low-risk' patients have significantly better overall survival than those identified as 'high risk' with a hazard ratio (HR) of 2.28 [95% confidence interval (CI) 1.32-3.93, P = 0.0032] in liver hepatocellular carcinoma; an HR of 2.79 (95% CI 1.66-4.68, P = 0.0001) in pancreatic adenocarcinoma; an HR of 5.85 (95% CI 2.53-15.5, P = 0.0002) in rectal adenocarcinoma; an HR of 1.81 (95% CI 1.07-3.07, P = 0.0268) in gastric adenocarcinoma. Across three different external validation sets of colorectal cancer (CRC) patients, our model yielded an HR of 2.32 (95% CI 1.67-3.23, P < 0.0001) in The Cancer Genome Atlas-Colon Adenocarcinoma (TCGA-COAD), an HR of 2.32 (95% CI 1.67-3.23, P < 0.0001) in Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial (PLCO)-COAD, and an HR of 3.38 (95% CI 1.99-5.71, P < 0.0001) in the Emory dataset. Multivariable survival analysis showed that our trained model was prognostic independent of stage, age, race, and sex.

CONCLUSIONS: Our findings suggest that the spatial relationships of TILs and cancer nuclei are prognostic of survival across multiple GI cancer types.