Probabilistic Deep Learning Cervical Lymph-Node Auto-Segmentation For Imaging-enhanced Evaluation of Extracapsular Extension Risk (PDL-CLASIFIER) - PROJECT SUMMARY Head and neck cancer (HNC) is a growing public health concern in the US, with an increasing number of cases of Human Papillomavirus associated Oropharyngeal cancer (HPV+ OPC). Locoregional therapy in the form of surgery and/or radiation therapy (RT) is often successful in treating HPV+ OPC, but the resulting damage to normal tissue can lead to a decreased quality of life (QOL) for survivors. To improve treatment planning for HPV+ OPC patients, we aim to use artificial intelligence (AI) to accurately identify patients who can benefit from curative trans-oral robotic surgery (TORS) without adjuvant therapy. Currently, identifying HPV+ OPC patients that can benefit from TORS without adjuvant therapy is based on ad hoc estimation of risk factors for adjuvant therapy, including tumor resectability and the presence of extracapsular nodal extension (ENE). ENE is often used as an indicator for the requirement of adjuvant therapy and plays a critical role in patient risk stratification and decision-making for treatment planning. To address the issue of accurately identifying ENE in HPV+ OPC, we will develop AI methods with performance that meets or exceeds segmentation and detection accuracy of metastatic lymph nodes (MLNs) and ENE using pretreatment imaging/clinical data. This approach has the potential to enable clinicians to make better-informed decisions regarding treatment planning, which can ultimately lead to better outcomes and improve the QOL for HPV+ OPC survivors. Current approaches to identify ENE radiologically using diagnostic imaging such as computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography (PET) are unsatisfactory for robust clinical implementation. To address this issue, we will combine novel deep learning (DL) approaches for 3D multi-modality (PET/contrast-enhanced CT [CECT]) medical image segmentation and classification with predictive uncertainty approximation methods, i.e., probabilistic DL (PDL), and evaluate models against human performance for clinical acceptability. The central hypothesis of this research is that PDL methods can incorporate pretreatment multi-modal anatomical and functional imaging (PET/CECT) with clinical data to provide probabilistic segmentation of MLNs and ENE detection with high accuracy and quantified level of confidence that is sufficient for clinical decision making. Therefore, if successful, this approach would redefine the standard-of- care regarding TORS/definitive RT treatment planning for HPV+ OPC patients. The proposed AI-assisted approach aims to improve the accuracy of risk stratification for HPV+ OPC patients by identifying ENE more accurately and providing a more precise risk assessment. This approach has the potential to enable clinicians to make better-informed decisions regarding treatment planning, which can ultimately lead to better outcomes and improve the QOL for HPV+ OPC survivors.
