Explainable prediction model for the human papillomavirus status in patients with oropharyngeal squamous cell carcinoma using CNN on CT images

Abstract

Several studies have emphasised how positive and negative human papillomavirus (HPV+ and HPV-, respectively) oropharyngeal squamous cell carcinoma (OPSCC) has distinct molecular profiles, tumor characteristics, and disease outcomes. Different radiomics-based prediction models have been proposed, by also using innovative techniques such as Convolutional Neural Networks (CNNs). Although some of these models reached encouraging predictive performances, there evidence explaining the role of radiomic features in achieving a specific outcome is scarce. In this paper, we propose some preliminary results related to an explainable CNN-based model to predict HPV status in OPSCC patients. We extracted the Gross Tumor Volume (GTV) of pre-treatment CT images related to 499 patients (356 HPV+ and 143 HPV-) included into the OPC-Radiomics public dataset to train an end-to-end Inception-V3 CNN architecture. We also collected a multicentric dataset consisting of 92 patients (43 HPV+ , 49 HPV-), which was employed as an independent test set. Finally, we applied Gradient-weighted Class Activation Mapping (Grad-CAM) technique to highlight the most informative areas with respect to the predicted outcome. The proposed model reached an AUC value of 73.50% on the independent test. As a result of the Grad-CAM algorithm, the most informative areas related to the correctly classified HPV+ patients were located into the intratumoral area. Conversely, the most important areas referred to the tumor edges. Finally, since the proposed model provided additional information with respect to the accuracy of the classification given by the visualization of the areas of greatest interest for predictive purposes for each case examined, it could contribute to increase confidence in using computer-based predictive models in the actual clinical practice.

Keywords: Convolutional neural network; Explainable artificial intelligence; Grad-CAM; Human papillomavirus; Oropharyngeal squamous cell carcinoma.

Figure 1

Workflow of the proposed model.

Figure 2

Performance classification on training and test sets.

Figure 3

Examples of activation maps referring to correctly classified patients generated by the Grad-CAM algorithm. The first two images refer to two HPV+  cases, while the last two images refer to two HPV− cases. The red areas are the regions that have most influenced the process of assigning the positive or negative class.

Figure 4

Examples of activation maps for misclassified patients generated by the Grad-CAM algorithm. The first image refers to a real HPV+  case, while the second to a real HPV− case. The red zones are the regions that have most influenced the positive or negative class assignment process.