Predicting HPV association using deep learning and regular H&E stains allows granular stratification of oropharyngeal cancer patients

Abstract

Human Papilloma Virus (HPV)-associated oropharyngeal squamous cell cancer (OPSCC) represents an OPSCC subgroup with an overall good prognosis with a rising incidence in Western countries. Multiple lines of evidence suggest that HPV-associated tumors are not a homogeneous tumor entity, underlining the need for accurate prognostic biomarkers. In this retrospective, multi-institutional study involving 906 patients from four centers and one database, we developed a deep learning algorithm (OPSCCnet), to analyze standard H&E stains for the calculation of a patient-level score associated with prognosis, comparing it to combined HPV-DNA and p16-status. When comparing OPSCCnet to HPV-status, the algorithm showed a good overall performance with a mean area under the receiver operator curve (AUROC) = 0.83 (95% CI = 0.77-0.9) for the test cohort (n = 639), which could be increased to AUROC = 0.88 by filtering cases using a fixed threshold on the variance of the probability of the HPV-positive class - a potential surrogate marker of HPV-heterogeneity. OPSCCnet could be used as a screening tool, outperforming gold standard HPV testing (OPSCCnet: five-year survival rate: 96% [95% CI = 90-100%]; HPV testing: five-year survival rate: 80% [95% CI = 71-90%]). This could be confirmed using a multivariate analysis of a three-tier threshold (OPSCCnet: high HR = 0.15 [95% CI = 0.05-0.44], intermediate HR = 0.58 [95% CI = 0.34-0.98] p = 0.043, Cox proportional hazards model, n = 211; HPV testing: HR = 0.29 [95% CI = 0.15-0.54] p < 0.001, Cox proportional hazards model, n = 211). Collectively, our findings indicate that by analyzing standard gigapixel hematoxylin and eosin (H&E) histological whole-slide images, OPSCCnet demonstrated superior performance over p16/HPV-DNA testing in various clinical scenarios, particularly in accurately stratifying these patients.

Fig. 1. Study design and overall concept.

a Patients from Cologne (C, GER), Giessen (G, GER), Maastricht (M, NL), Heidelberg (H, GER) and TCGA (T, USA; database) were included in the study (n = 906). HPV association was defined as either dichotomous HPV-DNA and p16 IHC status if both markers were available or by p16/detection of high-risk HPV-DNA. All patients received standard treatment of care. b CONsolidated Standards Of Reporting Trials (CONSORT)-like flow chart representing the study population of the training/validation cohort of 267 patients. Cases that could not be identified as OPSCC or with missing information on HPV-status were excluded. There were four cases where follow-up data were not available in the training cohort, which were used for training the model for HPV association, but not for survival analysis. (c) CONSORT-like flow chart representing the study population of the external test set of 639 patients. . Both primary tumors and lymph node metastases were included in the test set.

Fig. 2. Performance of the model to predict HPV association for six different patient populations.

a Pie charts for patient characteristics for the training cohort (left panel) and test cohort (right panel). The number besides the panel indicates the number of patients originating from each center. b Area under the receiver operator curve (AUROC) for six different patient populations. Cologne (n = 177), Giessen (n = 240), Maastricht (n = 88), Heidelberg (n = 31), and lymph node metastasis (n = 103) which originated from the Giessen site and were independent to the training cohort. c Overview of the study population after applying a threshold of cases with a variance below $7x{10}^{-2}$. d Area under the receiver operator curve (AUROC) for cases that were filtered by the threshold of variance. e Visualization of the Cox proportional hazards model for tile class prevalence of the HPV-positive class. The red line indicates the smoothened function, the vertical lines at the horizontal axis indicates individual patients with a given risk. The green bar indicates the error of the function. f Kaplan–Meier curve of n = 258 patients stratified for tile class prevalence of the HPV-positive class. A Cox proportional hazards model was used to compare survival curves.

Fig. 3. Prognostic relevance of predicted HPV association.

a Hazard ratio plot of the predicted HPV association and overall survival for patients from the test cohort: Cologne (C; n = 177), Giessen (G; n = 240), Maastricht (M; n = 88) and Heidelberg (H; n = 31). b Corresponding Kaplan–Meier curve for the same population as depicted in A (n = 531). The cohort is divided into three separate groups by the combined score. A Cox proportional hazards model was used to compare survival curves. c Hazard ratio plot of the predicted HPV association and overall survival for patients from the training cohort: Cologne (C; n = 233), Giessen (G; n = 13), and TCGA (T; n = 21), with a total of n = 263 patients. d Corresponding Kaplan Meier curve for the same population as depicted in C (n = 263). The cohort is divided into three separate groups by the combined score. A Cox proportional hazards model was used to compare survival curves. e Hazard ratio plot of the predicted HPV association and overall survival for patients from the test cohort, exclusively containing lymph node metastatic samples: Giessen (G; n = 102). f Corresponding Kaplan–Meier curve for the same population as depicted in E (n = 102). The cohort is divided into three separate groups by the combined score. A Cox proportional hazards model was used to compare survival curves. g Histological image of three cases with corresponding classes of the combined score. The scale bars have a length of 0.052 mm.

Fig. 4. Prognostic significance of the predicted HPV association for stage I/II patients compared to gold-standard.

a Patients with early-stage disease I/II (Union for International Cancer Control; UICC 8th) from different centers of the test cohort are included for the subsequent analysis. b Hazard ratio plot of the selected patients with stage I/II disease. c Kaplan–Meier curve for the survival benefit of stage I/II patients according to the combined score. A Cox proportional hazards model was used to compare survival curves. d Schematic, illustrating that for comparison HPV testing, with both HPV-DNA assessment and p16 status (dichotomous testing) was used. e Kaplan–Meier curve for regular HPV testing. A Cox proportional hazards model was used to compare survival curves. f Schematic illustration of the filtering criteria for the subsequent analysis. Cases from Cologne, Giessen, Maastricht, and Heidelberg that were tested as HPV-positive (n = 232). g Hazard ratio plot of the selected patients with stage a positive HPV-status (n = 232). h Kaplan–Meier curve for the combined score of patients tested as HPV-positive (n = 232). A Cox proportional hazards model was used to compare survival curves.