Multi-modal Deep Learning for Time-to-Event analysis in Head and Neck Squamous Cellular Carcinoma patients

Head and Neck Squamous Cellular Carcinoma (HNSCC) constitutes the seventh most common cancer diagnosis worldwide, and the long term survival of the affected patients is highly influenced by possible development of distant metastasis and tumor relapse. Hence, a prognostic model able to predict such occurrences would significantly benefit these patients and could be employed for treatment recommendation in order to optimize the handling of the individual subject. In the depicted context, this Thesis aims at describing the development of a multi-modal Deep Learning model for Time to Event analysis of HNSCC patients; our approach involves the employment of clinical and imaging data for the prediction of distant metastasis, loco-regional failure and overall survival of the single patient. In the context of automated models for clinical outcome prediction, our work involves the innovation given by the use of a combination of clinical and imaging (i.e., CT and 18FDG-PET) data: while the latter are universally known to hold a significant amount of information in the investigation of tumors, in this study clinical data are thought to provide further knowledge to the conditions of the patient, and their prognostic power is therefore assessed and employed for the predictions. Therefore, a significant weight is given to pre-treatment clinical variables and the investigation of their influence on the future evolution of the disease. Moreover, this study aims at investigating the predictive power of CT and PET volumes without primary and lymph node Gross Tumor Volume segmentation; indeed, the goal is the development of an automated model that allows to disengage from a manual segmentation while granting an equally satisfactory performance. This research is based on Deep Learning techniques and involves the comparison of single- and multi-modality inputs to investigate the predictive performance of clinical data in different combinations with imaging data. To this purpose, clinical and imaging data were employed to train three-dimensional Convolutional Neural Networks. Specifically, this Thesis focuses on the retrospective data analysis phase of this research: our models were trained on the retrospective dataset and used to predict individual clinical outcomes; the resulting model will then be validated during the prospective validation. The ultimate goal is to achieve a model able to accurately predict the timing of HNSCC-related events for the single patient, in order to contribute to the goal of personalized medicine by allowing, with such a technology, a step closer to the development of individualized therapies. The findings presented in this Thesis provide valuable information concerning the prognostic power of clinical and imaging data in the progression HNSCC, together with notable insights resulting from the comparison of different DL-based prediction models. The present work was carried out in collaboration with Berlin's Charitè Lab of Artificial Intelligence in Medicine.