AI-Predicted Extranodal Extension Improves Risk Stratification in Oropharyngeal Carcinoma

Key Clinical Summary:

• The number of lymph nodes with AI-predicted extranodal extension was independently associated with poorer distant control and overall survival in patients with oropharyngeal carcinoma treated with definitive radiation therapy.
• Incorporation of AI extranodal extension node count significantly improved prognostic performance of established risk models, including RTOG-0129 risk groupings and AJCC 8th edition staging, particularly for distant control outcomes.
• Prognostic gains were most pronounced in HPV-negative disease, highlighting the potential clinical value of AI-derived extranodal extension assessment in higher-risk patient subsets.

Benjamin Kann, MD, Mass General Brigham, Boston, Massachusetts, discusses results from a large retrospective study evaluating an automated artificial intelligence (AI)-based imaging platform to predict extranodal extension in patients with oropharyngeal carcinoma using pretreatment computed tomography (CT) scans.

Results demonstrate that the number of AI-predicted extranodal extension-positive lymph nodes was independently associated with worse distant control and overall survival and significantly improved risk stratification beyond existing staging systems. 

These findings demonstrate the potential of AI-driven imaging biomarkers to enhance pretreatment prognostication and clinical decision-making in oropharyngeal carcinoma. 

Transcript:

Extranodal extension is a very important risk factor to identify for patients with oropharyngeal carcinoma because we know that patients who go to surgery and are found to have malignant cancerous tissue invading outside of the lymph node capsule are at a higher risk for distant metastasis and for death due to their cancer. Not only is it a poor prognostic factor, but extranodal extension, also called ENE, is also an indication that these patients need to have escalated systemic therapy after surgery, or what’s called trimodality therapy. 

One of the issues that we face when trying to figure out if a patient has ENE or not is that we can really only definitively diagnose it on surgical resection, and trying to determine if ENE is present upfront from radiographic imaging or from clinical exam is very difficult. We would love to be able to determine or predict with some degree of confidence whether or not a patient is going to have ENE, but it is very difficult to do that with traditional means. In prior work, we developed an AI-based algorithm that uses deep learning, analyzing contrast-enhanced or unenhanced CT scans to give a prediction on a node-by-node basis of whether or not ENE is present. 

In this work, we took that algorithm and tied it together with an autosegmentation algorithm, which means that first the algorithm identifies where the lymph nodes are in the neck in an automated way. Then, for each lymph node, we run it through the algorithm to give us a prediction of extranodal extension or not—does that patient have ENE or not. We could then take a composite measurement of the entire neck and give an output of an ENE node number, how many lymph nodes in this patient’s neck are predicted to have ENE? Why this is really important and novel is that, in the past, even from pathologic exam, we generally don’t know how many lymph nodes have ENE. It’s generally thought of as a binary phenomenon—the patient has it or doesn’t—but we know from a biological standpoint that it may mean a lot more if a patient has, say, 3 or 4 lymph nodes with ENE versus just a single lymph node with ENE, as that may tell us something about the cancer aggressiveness and this patient’s future risk. What this AI algorithm now allows us to do is calculate the number of lymph nodes that are predicted to have ENE, and then we can use that to study different patient populations and potentially use it as a way to triage the appropriate patients for intensified therapy. 

In this study, we investigated the prognostic value of the ENE node number. We looked at this in 3 large multi-institutional cohorts of about 1,700 patients and what we found was that, consistently, if you added ENE node number based on this AI prediction into traditional risk schemas or risk frameworks, such as AJCC 8th edition staging or other standard-of-care risk assessment tools, it improved the ability to predict distant metastasis and overall survival. We saw this in both HPV-positive oropharyngeal cancer and HPV-negative oropharyngeal cancer, with an especially potent effect in HPV-negative oropharyngeal cancer, with the ability to predict distant metastasis and overall survival showing about a 15% gain over current staging frameworks. 

Taken together, we think this AI tool could potentially be used to help with both surgical and radiation planning because it provides a three-dimensional overlay of lymph node status, showing which nodes are likely to have ENE. This may have implications for radiation dose volumes or surgical approach. We also think it is a novel prognostic tool that we can use to better risk stratify our patients. Lastly, one thing that we found was that when we looked more closely at the number of ENE nodes—the number of lymph nodes with ENE—and how that predicted risk of distant metastasis and survival, we did indeed find a relationship between the number of nodes and risk. Patients who had just one lymph node with ENE did not have a substantially increased risk compared with those with none. However, once you started to see two or three lymph nodes predicted to have ENE, the gaps widened significantly in terms of risk of distant metastasis and survival. 

This again, shows that it is more important to know the number of lymph nodes with ENE and the overall burden of ENE than simply whether a patient has it or not. This AI tool allows us to predict that node number, which we really could not do before.

Source:

Ye Z, Mojahed-Yazdi R, Zapaishchykova A, et al. Automated lymph node and extranodal extension assessment improves risk stratification in oropharyngeal carcinoma. J Clin Oncol. Published online: December 23, 2025. doi: 10.1200/JCO-24-02679