Optimization and Validation of a Cost-effective Image-Guided Automated Extracapsular Extension Detection Framework through Interpretable Machine Learning in Head and Neck Cancer - Project Summary Squamous cell carcinoma of the head and neck (SCCHN) is the 6th commonest cancer in the world, leading to >300,000 deaths annually worldwide. The extracapsular extension (ECE) of the tumor in the lymph nodes is a significantly high-risk feature in head and neck cancers. Several studies demonstrate that ECE results in worse survival outcomes. Pathologic confirmation, such as neck dissection, is currently required as the gold standard for clinical identification of ECE. However, if it is ECE positive, postoperative chemoradiotherapy has to be considered. As a result, neck dissection followed by chemoradiation increases the toxicity, especially late toxicity such as fibrosis, and lymphedema can get worse due to the late toxicity. If we can detect ECE during preoperative evaluation, we can select those patients for chemoradiation before surgery. Thus, predicting ECE becomes a piece of critical information for clinicians planning treatment. The biggest obstacles to adopting AI/ML algorithms in the clinic are concerns about security, reliability, and transparency. If an algorithm could be developed to not only provide an accurate prediction of ECE (and/or clinical staging), but also to provide transparent and clinically understandable communication of how the predictive conclusion was reached, then clinicians would more readily adopt a tool utilizing that algorithm to be an ally. The purpose of this proposal is to optimize and pathologically validate AI/ML approaches for prognoses and diagnoses of ECE from medical images on head and neck cancer patients, which could be further implemented as a tool for diagnostic assistance and precision medicine. This translational research project focuses on filling the gap of AI/ML transparency and interpretability in the automated detection of ECE in head and neck cancers. In addition, our interpretable ML algorithm will not require the pre-annotated lymph node areas, which is a quite cost-effective technique by eliminating the time-consuming lymph node annotation process. We propose the following aims: 1) Validate and interpret the image-based ECE diagnosis model and its association with head and neck cancer anatomic organ sites and HPV status. 2) Optimize the cost-effective image-based ECE diagnosis model considering clinical markers within pathology, PETCT, and MRI results for clinical implementation. We will validate our model based on the dataset collected from both our team and existing data collected from The Cancer Image Archive. This proposal aligns with the mission of Oral & Salivary Cancer Biology Program and the NIDCR Special Interest in Supporting Dental, Oral, and Craniofacial Research Using Bioinformatic, Computational, and Data Science Approaches (NOT-DE-20-006). This study will provide the preliminary results of our future research on the precision treatment of high-risk head and neck cancer patients. We plan to submit an NIDCR R01 proposal in Year 2 for the precision treatment of high-risk head and neck cancer.
