Development of an artificial intelligence-driven, imaging-based platform for pretreatment identification of extranodal extension in head and neck cancer

Project Summary. The goal of this project is to develop, optimize, and evaluate an artificial intelligence (AI)- driven, medical imaging platform that utilizes computed tomography (CT) imaging to identify the presence of extranodal extension (ENE) in head and neck squamous cell carcinoma (HNSCC). HNSCC is a debilitating

disease with significant patient-related morbidity related to the disease itself and its management, which is complex and consists of a combination of surgery, radiation, and chemotherapy. A key factor in determining proper HNSCC management is the presence of ENE, which occurs when tumor infiltrates through the capsule

of an involved lymph node into the surrounding tissue. ENE is both an important prognostic factor and an indication for adjuvant treatment escalation with the addition of chemotherapy to radiation following surgery. This “trimodality therapy” is problematic, as it is associated with increased treatment-related morbidity and

healthcare costs, but no improvement in disease control compared to upfront chemoradiation alone. The challenge is that ENE can only be definitively diagnosed pathologically after surgery, and pretreatment radiographic ENE identification has proven unreliable for even expert diagnosticians, leading to high rates of

trimodality therapy and suboptimal treatment outcomes. In HNSCC management there is a critical need for improved pretreatment ENE identification to 1) select appropriate patients for surgery to avoid the excess morbidity and costs of trimodality therapy, 2) risk-stratify patients optimally, and 3) select appropriate patients

for treatment de-escalation or intensification clinical trials. In recent years, Deep learning, a subtype of machine learning, under the umbrella of AI, has generated breakthroughs in computerized medical image analysis, at times outperforming human experts and discovering patterns hidden to the naked eye. While AI is poised to

transform the fields of cancer imaging and personalized cancer care, there remain significant barriers to clinical implementation. The hypothesis of this project is that AI can be used to successfully identify HNSCC ENE on pretreatment imaging in retrospective and prospective patient cohorts and to develop a platform for lymph

node auto-segmentation that will promote clinical utility of the platform. This hypothesis will be tested by rigorous optimization and evaluation of a deep learning ENE identification platform. Specifically, the platform will be validated for accuracy, sensitivity, specificity, and discriminatory performance on two heterogeneous retrospective datasets and two prospective cohorts derived from

institutional and national Phase II clinical trials for HNSCC patients. The platform will then be directly compared with head and neck radiologists to determine if radiologist performance can be augmented with AI. In parallel, AI will be utilized to develop an auto-segmentation platform for tumor and lymph nodes, which will 1) improve

the platform's clinical impact and 2) provide a valuable tool for treatment planning and future imaging-based research for HNSCC patients. 1