Article Text

Download PDFPDF

829 Spatial arrangement and density of tumor-infiltrating lymphocytes (TILs) predicts response to immunotherapy in head and neck squamous cell carcinoma patients
  1. Reetoja Nag1,
  2. Germán Corredor1,
  3. Vidya Viswanathan1,
  4. Pingfu Fu1,
  5. James Lewis2,
  6. Jay Wasman1,
  7. Theodoros Teknos3,
  8. Monaliben Patel1,
  9. Quintin Pan3 and
  10. Anant Madabhushi1
  1. 1Case Western Reserve University, Cleveland, OH, United States
  2. 2Vanderbilt University Medical Center, Nashville, TN, United States
  3. 3University Hospitals, Cleveland, OH, United States


Background Anti-PD-1 immunotherapies (IOs) have robust clinical benefit in a subset of head and neck squamous cell carcinoma (HNSCC) patients and is considered as a standard of care option in the recurrent/metastatic setting.1 Previous studies showed association of higher number of lymphocytes and density of tumor-infiltrating lymphocytes (TILs) with better survival probability and prognosis in HNSCC patients.2 3 In this study, we evaluate whether the spatial interplay between TILs with surrounding nuclei and TIL density from digitized H&E-stained slides was associated with better immunotherapy response in HNSCC patients.

Methods Whole slide images (WSIs) from 43 HNSCC patients treated with IO at University Hospitals, Cleveland were selected. Response to immunotherapy was defined as per RECIST v1.1.4 Computerized algorithms identified and built clusters for nuclei of TILs and TIL density and spatial arrangement of the clusters was quantified using network graph-metrics. To assess the predictive ability of the combination of spatial arrangement and density features of TILs, a cross-validation scheme was used as follows: at each iteration, the dataset was randomly split into training (60%) and validation sets (40%). The Wilcoxon method selected top two features in the training set, then used to train a Naïve Bayes classifier to differentiate between responders and non-responders. Next, the classifier was applied to the validation set and its performance was evaluated by computing the area under the curve (AUC) for the receiver operating characteristic (ROC) curve. This process was repeated 250 times. For comparison, the same cross-validation procedure was used on (1) TIL arrangement and (2) TIL density features.

Results Figure 1(A) and 1(B) show the violin plots corresponding to the two selected top features across the cross-validation iterations for responders and non-responders. Figure 1(C) illustrates the average ROC curves for the three assessed models and corresponding AUC. The average AUCs for the model based on combination (TIL arrangement + TIL density), TIL arrangement and TIL density only yielded an average AUCs of 0.84+0.14, 0.81+0.02 and 0.74+0.06 respectively. After DeLong assessment of the ROC curves, significant difference was found for TIL combination vs. TIL density (p=0.03) whereas no significant difference was found for TIL combination vs. TIL arrangement (p=0.55).

Conclusions We present a predictive model based on image biomarkers i.e. spatial interplay between TILs and surrounding nuclei along with TIL density in order to distinguish HNSCC patients responding and not responding to immunotherapy. The model based on TIL combination performed better in comparison to only TIL density.


  1. Patel RR, Ludmir EB, Augustyn A, et al. De-intensification of therapy in human papillomavirus associated oropharyngeal cancer: a systematic review of prospective trials. Oral Oncol 2020;103:104608.

  2. Lilja-Fischer JK, Eriksen JG, Georgsen JB, et al. Prognostic impact of PD-L1 in oropharyngeal cancer after primary curative radiotherapy and relation to HPV and tobacco smoking. Acta Oncol 2020;59(6):666–672.

  3. Peled M, Onn A, Herbst RS. Tumor-infiltrating lymphocytes—location for prognostic evaluation. Clin Cancer Res 2019;25(5):1449–1451.

  4. Therasse P, Arbuck SG, Eisenhauer EA, et al. New guidelines to evaluate the response to treatment in solid tumors. J Natl Cancer Inst 2000;92(3):205–16.

Abstract 829 Figure 1

(A) Violin plot for topmost feature 1 selected by Wilcoxon test (Feature selected from number of clusters surrounding a specific cluster type) (Responder vs. Non-responder p=0.0006). (B) Violin plot for topmost feature 2 selected by Wilcoxon test (Feature selected from the intersected area from the cell clusters formed from the centroids of the cell clusters) (Responder vs. Non-responder p=0.00016) (C) Average receiver operating characteristic curves for the 3 assessed models i.e. TIL combination (AUC=0.84+0.14), TIL arrangement (AUC=0.81+0.02) and TIL density (AUC=0.74+0.06). Significant difference between the ROC curves i.e. TIL combination vs. TIL density (p=0.03), TIL combination vs. TIL arrangement (p=0.55), TIL arrangement vs. TIL density (p=0.32)

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.