917P - Automatic characterization of spatial arrangement of tumor-infiltrating lymphocytes identifies oral cavity squamous cell carcinoma patients with poorer prognosis

Background

Patients with oral cavity squamous cell carcinoma (OCSCC) are at high risk of death within 5 years of treatment. Previous works have demonstrated that spatial arrangement of tumor-infiltrating lymphocytes (TILs) and surrounding nuclei (e.g, cancer nuclei, macrophages, fibroblasts) is prognostic of overall survival (OS) in different types of cancer, including oropharyngeal and laryngeal. In this work, we evaluated the prognostic ability of a computerized strategy that characterizes the spatial arrangement of TILs & non-TILs on digitized H&E-stained slides of patients with OCSCC.

Methods

Whole slide images (WSIs) from a cohort of 283 patients with OCSCC were retrospectively collected from the Cancer Genome Atlas (TCGA, D1). Additionally, WSIs from 136 patients with OCSCC were obtained from Vanderbilt University Medical Center (VUMC, D2). D1 was used to train a prognostic model while D2 was used for independent validation. Computer algorithms automatically identified 2 types of nuclei (TILs & non-TILs) and built clusters for each nucleus type based on proximity. Metrics related to density, intersection, and neighborhood were computed from these clusters. A proportional hazard Cox regression model, regularized via the least absolute shrinkage and selection operator, was trained to predict risk of death. The percentile 66 risk score on D1 was used as a threshold for stratifying patients on D2 as either low or high risk. Survival analysis was then used to evaluate the association with OS.

Results

Patients in D2 defined as “low risk” (57%) based on spatial arrangement of TILs has significantly better OS than those identified as “high risk” (43%) with hazard ratio=3.84 (95% confidence interval: 1.39-10.6, p<0.01). Multivariable survival analysis showed that this model was prognostic independent of age and overall, T, and N stages with HR=2.88 (95% CI: 1.10-7.56, p=0.03).

Conclusions

A computerized image analysis model based on measurements of spatial arrangement of TILs & non-TILs was found to be prognostic in patients with OCSCC. With more modeling and development, this model may provide a risk classifier for use in routine clinical practice.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

A. Madabhushi.

Funding

US National Institutes of Health, US Department of Defense.

Disclosure

A. Madabhushi: Financial Interests, Personal, Advisory Board, Serve on SAB and consult: SimbioSys; Financial Interests, Personal, Advisory Board: Aiforia, Picture Health; Financial Interests, Personal, Full or part-time Employment: Picture Health; Financial Interests, Personal, Ownership Interest: Picture Health, Elucid Bioimaging, Inspirata Inc; Financial Interests, Personal, Royalties: Picture Health, Elucid Bioimaging; Financial Interests, Institutional, Funding: AstraZeneca, Bristol Myers Squibb, Boehringer Ingelheim, Eli Lilly. All other authors have declared no conflicts of interest.