Prediction of the number of lymph nodes with extranodal extension in patients with oropharyngeal carcinoma through a deep learning imaging platform for autosegmentation may help to guide pretreatment patient risk stratification and treatment decision-making, according to the results of a multisite, retrospective study published in the Journal of Clinical Oncology.
"Our tool may help identify which patients should receive multiple interventions or would be ideal candidates for clinical trials of intensive strategies such as immunotherapy or additional chemotherapy," said senior author Benjamin Kann, MD, Radiation Oncologist at Dana-Farber Cancer Institute and Brigham and Women's Hospital and Faculty Member of the Artificial Intelligence in Medicine (AIM) Program at Mass General Brigham. "Our tool can also help identify which patients should undergo de-intensification of treatment, such as surgery alone."
Study Methods
Researchers conducted a multisite, retrospective study across three institutions of 1,733 patients with oropharyngeal carcinoma who received pretreatment computed tomography (CT) scans and underwent definitive radiation therapy. They used a validated deep learning autosegmentation model to segment malignant lymph nodes from the CT scans and sequentially process them to calculate the number of nodes predicted to have extranodal extension per patient.
Researchers assessed the artificial intelligence (AI)-predicted extranodal extension and its association with disease outcomes, adjusting for human papillomavirus (HPV) status, smoking status, tumor and nodal stage, age, and sex. Risk stratification improvement was also considered for the AI-predicted extranodal extension calculations by Radiation Therapy Oncology Group (RTOG)-0129 risk groupings and American Joint Committee on Cancer (AJCC) 8th edition staging.
Key Findings
The median AI-predicted number of extranodal extension nodes was 1 (range, 0–6). AI-predicted extranodal extension node number was independently associated with both worse distant control (hazard ratio [HR] = 1.44; 95% confidence interval [CI] = 1.23–1.69; P < .001) and worse overall survival (HR = 1.30; 95% CI = 1.16–1.46; P < .001).
Increasing numbers of extranodal extension nodes were associated with worse outcomes incrementally, especially for distant control (P < .001). Concordance indices (C-index) showed improvement when including RTOG-0129 groupings into the AI predictions in terms of both overall survival (C-index = 0.70 with RTOG-0129 groupings vs 0.65 without) and distant control (C-index = 0.65 vs 0.57), and similar trends were observed when incorporating AJCC 8th edition stages for both overall survival (C-index = 0.75 with RTOG-0129 groupings vs 0.70 without) and distant control (C-index = 0.72 vs 0.67).
When adjusting for individual variables, the greatest improvement was seen in patients with HPV-negative disease, with the C-index improving by 15% for overall survival and by 14% for distant control.
"The AI tool enables the prediction of number of lymph nodes with [extranodal extension], which could not be done before, and shows that it is a powerful, novel prognostic biomarker for oropharyngeal cancer that could be used to improve the current staging scheme and treatment planning," Dr. Kann said.
Disclosure: For full disclosures of the study authors, visit ascopubs.org.
