Skip to Main Content

An official website of the United States government

Principal Investigator
Anant Madabhushi
Emory University
Position Title
Robert W Woodruff Professor
About this CDAS Project
PLCO (Learn more about this study)
Project ID
Initial CDAS Request Approval
Jan 30, 2024
Spatial Interplay of Tumor-Infiltrating Lymphocytes and Cancer Cells Identify Low-Risk Patients with Favorable Prognosis across Multiple Gastrointestinal Cancers
Gastrointestinal (GI) cancers (including esophagus, colon, stomach, liver, pancreas, and rectal) account for more than a quarter of all cancer diagnoses and 35% of cancer-related fatalities worldwide. Quantification of tumor infiltrating lymphocytes (TILs) is a well-known cancer prognostic marker. In this study, we used machine learning to evaluate the prognostic significance of the spatial interplay between tumor-infiltrating lymphocytes (TILs) and tumor nuclei across five different types of gastrointestinal (GI) adenocarcinomas, namely colon (COAD), stomach (STAD), liver (LIHC), pancreas (PAAD), and rectal (READ). Stage II and III colon adenocarcinoma whole slide images (WSIs) from 137 patients (pts) at University Hospital were collected to form Dataset 1 (D1). Additionally, WSIs from 661 patients with GI cancers across five organ sites (COAD: 152; STAD: 185; LIHC: 137; PAAD: 78; READ: 109) were collected from The Cancer Genome Atlas (TCGA) to form Dataset 2 (D2). All cases were microsatellite stable. Hand-crafted (HC) features relating to nuclear morphology as well as spatial interaction between TILs and tumor nuclei were extracted from the images. The top features, selected using the Least Absolute Shrinkage and Selection Operator, were used to train 2 Cox regression models that assigned a risk of death and recurrence to each patient, respectively. Pts with Stage III CRC in D2 defined as “high risk” had significantly worse disease-free survival than those identified as “low risk” with hazard ratio (HR) =3.98 (95% confidence interval (CI): 0.924-17.1, p = 0.003) Similarly, pts in D2 identified as “low risk” had better overall survival with HR of 3.36 (95% CI: 1.29-3.17, p<0.05) in COAD; HR=1.96 (95% CI:1.05-3.68, p<0.05) in PAAD; HR=5.85 (95% CI:2.25-15.2, p<0.005) in READ; HR of 2.6 (95% CI:1.51-4.49,p<0.005) in LIHC; HR of 1.9 (95% CI:1.12-3.21,p<0.05) in STAD. Multivariable survival analysis showed that our model was prognostic independent of T/N stages, age, race, and sex with HR=2.22 (95% CI: 1.49-3.28, p=0.0001). Our findings suggest that nuclear morphology as well as spatial interplay of TILs and cancer nuclei are prognostic of survival across multiple GI cancers. Future work will focus on prospective multi-site validation of these findings and to validate the signature in terms of benefit of therapy.

The goal of this project is to evaluate the prognostic value of TIL-Tumor Nuclei interactions we previously identified on PLCO cases. Our results show that interaction features are not only prognostic in D1_COAD but in all the studied GI cancers and achieved HR=2.31 (95% CI: 1.14-4.71, p=0.0158) in COAD; HR=2.1 (95% CI:1.11-3.97, p=0.0332) in PAAD; HR=5.85 (95% CI:2.25-15.2, p=0.00135) in READ; HR=2.34 (95% CI:1.33-4.12,p=0.0102) in LIHC; HR=1.86 (95% CI:1.1-3.16,p=0.0321) in STAD. Furthermore, Figure 6 shows that M2 is prognostic in Stage 2 (HR=3.09, 95% CI: 1.16-8.27, p=0.0181) and 3 COAD cases (HR=3.36, 95% CI: 1.05-10.7, p=0.0127) and Stage 1 LIHC (HR=2.66, 95% CI: 1.14-4.71, p=0.0158). We would like to also add other upper GI cases from PLCO to the analysis to evaluate our trained model.


Chuheng Chen, Reetoja Nag, Krunal Pandav, Tilak Pathak, Joseph Willis, Anant Madabhushi