By Christina Henson, MD, University of Oklahoma Health Sciences Center
Annie Chan, MD, and Rahul Paul, PhD, from the Department of Radiation Oncology at Massachusetts General Hospital, together with their otolaryngology colleague Jeremy Richmon, MD, at the Massachusetts Eye and Ear Infirmary reported on utilizing machine learning algorithms that incorporate perinodal microenvironment to improve the performance of their AI model in predicting for extranodal extension (ENE) in patients with oropharyngeal cancer.
Between 2016 and 2022, 171 patients at these institutions were deemed to have resectable oropharyngeal cancer with no evidence of ENE on pre-operative imaging and underwent upfront TORS and neck dissections. Various radiomics and topological features from both the nodes and perinodal microenvironment were extracted to generate novel predictive models for ENE.
From these 171 patients, there were a total of 264 involved lymph nodes, of which 20% exhibited ENE on surgical pathology. When radiomics features of the nodal microenvironment were added to radiomics features of the nodal metastases themselves, it improved the performance of the model, as did incorporating topological features of the perinodal region. A multimodal predictive model incorporating both these radiomics features and topological features yielded an impressive AUC of 0.94, which improved to 0.95 when HPV status was incorporated. An AI model that will generate a correct prediction 95% of the time, if generalizable, represents a marked improvement over the performance of other currently existing prediction models. This has the potential to aid in better selection of patients with early-stage oropharyngeal cancer who are appropriate (or not) for deintensification of treatment and could help clinicians to better avoid delivering trimodality therapy with surgery, radiation and chemotherapy, which is known to negatively impact patient quality of life outcomes without any survival benefit.
Overall, this study illustrates the power that can be harnessed by machine learning algorithms. For putting this into practice on a wider scale, this model will need to be validated at other institutions and made broadly available. We look forward to the publication of the manuscript and learning more about the specific radiomics and topological features of the perinodal microenvironment. It will also be interesting to learn which – if any – system for diagnosing radiographic ENE was utilized by the head and neck radiologists at these institutions.
Abstract 133 - AI-ing Microenvironment Improves Prediction of Extracapsular Nodal Extension in Oropharyngeal Carcinoma was presented during the Poster Session I at the 2024 Multidisciplinary Head and Neck Cancers Symposium February 29 – March 2, 2024.
