1057 - AI Automated Measures of Tumor Burden and Correlation with Pre-Treatment Circulating Tumor DNA in Head and Neck Cancer
Presenter(s)
M. Bakhtiar1, Z. Ye2, J. D. Schoenfeld3, Q. Jiao4, J. P. Guenette5, E. M. Rettig5, G. J. Hanna5, and B. H. Kann6; 1Harvard Radiation Oncology Program, Boston, MA, 2Artificial Intelligence in Medicine (AIM) Program, Mass General Brigham, Harvard Medical School, Boston, MA, 3Brigham and Women’s Hospital/Dana-Farber Cancer Institute, Harvard Medical School, Boston, MA, 4Harvard Medical School, Boston, MA, 5Dana-Farber Cancer Institute, Boston, MA, 6Department of Radiation Oncology, Dana-Farber Cancer Institute/Brigham and Women’s Hospital, Harvard Medical School, Boston, MA
Purpose/Objective(s): Personalized circulating tumor DNA (ctDNA) assays are being increasingly used to support the diagnosis and surveillance of a variety of malignancies, including head and neck squamous cell carcinoma (HNSCC). CtDNA has been shown to have variable correlation to pre-treatment disease characteristics, including clinical stage. Artificial-intelligence (AI) automated analysis of tumor burden on pre-treatment imaging may provide insight into the dynamics of ctDNA in HNSCC.
Materials/Methods: This was a retrospective cohort study of patients treated definitively for HNSCC. All patients had pre-treatment measures of ctDNA using a personalized, tumor-informed, blood-based assay (Signatera, Natera). Clinical data was abstracted from electronic medical records. A previously validated, U-net based, AI auto-segmentation algorithm was applied to pre-treatment diagnostic or radiation planning CT scans to auto-segment primary tumor and lymph nodes and calculate 3D volumetrics. Univariable and multivariable regressions were performed to test associations between ctDNA level (mean tumor molecules [MTM]/mL) and automated volumetrics, AJCC 8th edition clinical tumor (T) and nodal (N) staging, smoking pack-years (> or = 10), primary tumor site, and human papillomavirus (HPV) status.
Results: Seventy-eight patients treated for HNSCC between November 2023 and May 2024 were included. Median age was 64.7 years (interquartile range [IQR]: 58.1-74.6), 68.8% male. Median pre-treatment ctDNA level was 0.57 MTM/mL (IQR: 0.06-6.8). The most common primary tumor site was oral cavity/lip (36 patients, 46%), followed by oropharynx (14 patients, 18%) and larynx (10 patients, 13%). Eight patients were positive for high-risk HPV by in-situ hybridization. On univariable testing, ctDNA level was associated with automated node volume (coeff=438.72, p=0.004), unknown primary tumor site (coeff=164.65, p=0.001), and high-risk HPV (coeff=64.60, p=0.003) but was not associated with automated primary tumor volume (p=0.54) clinical T (p=0.63), or clinical N stage (p=0.32). On multivariable testing, when adjusting for clinical T and N staging, smoking pack-years, primary tumor site, and HPV status, ctDNA remained significantly associated with automated nodal volume (coeff=752.98 p<0.001), but not with clinical N stage (p=0.15), and approached statistical significance for clinical T stage (p=0.06).
Conclusion: This is the first study of AI-based imaging correlates for a personalized ctDNA assay in HNSCC. AI-based analysis of pre-treatment imaging demonstrates an association between automated node volume and ctDNA in HNSCC; this association is stronger than clinical T and N staging and is independent of other disease factors. Automated volumetrics may provide a non-invasive complement to ctDNA level, and AI-based analysis may improve risk categorization in the future. Further investigation into how automated volumetrics and nodal disease aligns with ctDNA in HNSCC is warranted.