159 - Predicting Dysphagia and Identifying Anatomical Regions Responsible for Dysphagia Following Intensity-Modulated Radiation Therapy (IMRT) for Oropharyngeal Carcinoma
Presenter(s)
S. K. Das1, F. Dong2, R. Morse3, B. M. Anderson4, B. S. Chera5, C. Shen4, M. C. Repka4, W. Yarbrough6, T. Hackman7, J. Blumberg8, C. Lumley6, R. Ferris9, S. Trivedi1, J. Weiss10, S. Patel11, S. Sheth12, and X. S. Chen13; 1University of North Carolina, Chapel Hill, NC, 2University of North Carolina Department of Radiation Oncology, Chapel Hill, NC, United States, 3Department of Radiation Oncology, University of North Carolina School of Medicine, Chapel Hill, NC, 4Department of Radiation Oncology, University of North Carolina, Chapel Hill, NC, 5Department of Radiation Oncology, University of North Carolina at Chapel Hill, charleston, SC, 6UNC School of Medicine, Chapel Hill, NC, 7Department of Otolaryngology, University of North Carolina School of Medicine, Chapel Hill, NC, 8UNC, Chapel Hill, NC, 9Department of Otolaryngology/Head and Surgery, UNC Lineberger Comprehensive Cancer Center, The University of North Carolina at Chapel Hill, Chapel Hill, NC, 10University of North Carolina Lineberger Comprehensive Cancer Center, Chapel Hill, NC, 11University of North Carolina Hospitals, Chapel Hill, NC, 12Division of Oncology, Department of Medicine, University of North Carolina, Chapel Hill, NC, 13University of North Carolina at Chapel Hill, Chapel Hill, NC
Purpose/Objective(s): Dysphagia is common following radiation therapy (RT) for oropharyngeal carcinoma (OPC) despite dose de-intensification and use of standard dose constraints for pharyngeal constrictors in IMRT planning. This study investigates: (a) whether dysphagia can be predicted at an individual-patient level using artificial intelligence (machine learning, ML) with CT and dose as inputs; (b) which anatomic regions are most responsible for dysphagia; and (c) whether these regions can be identified using CT alone to inform dose sparing during treatment planning.
Materials/Methods: This study retrospectively analyzed 155 patients with human papilloma virus (HPV)–related OPC treated with definitive IMRT (60 Gy [58%] or 70 Gy [42%]) from 2017–2021 at a single academic center. Dysphagia at 6 months post-RT was graded using CTCAE v5 and categorized as none/mild (Grade 0–1, N=82) versus significant (Grade 2–3, N=73). Patients were randomly divided into 125 training and 30 testing cases. An ML classification model using contrast-enhanced CT and 3D dose inputs employed a U-Net followed by convolutional and dense layers with sigmoid activation and ensemble bagging of three sub-models. Performance was compared against traditional NTCP models using pharyngeal constrictor fractional-volume dose metrics and LASSO logistic regression. The ML classification model was employed to generate gradient heatmaps for all patients with significant dysphagia. These heatmaps were used to train a second ML model to reproduce regions of interest using CT alone.
Results: The ML model achieved an AUC of 0.96 on the training set, with predicted scores significantly different between dysphagia groups (t-test p < 0.001). Test set performance yielded AUC=0.81 (p=0.002). Traditional NTCP models performed worse: the best dosimetric metric (superior pharyngeal constrictor V70Gy) achieved training/test AUCs of 0.68/0.62, while the best logistic regression model (selected metrics: superior pharyngeal constrictor V50Gy and V70Gy; middle pharyngeal constrictor V46Gy and V53Gy) yielded 0.73/0.63. Heatmaps consistently identified a region including some portion of the inferior pharyngeal constrictor (IFC) and cricopharyngeus muscles as highly associated with dysphagia. The ML heatmap model using CT-input alone was able to accurately reproduce dysphagia heatmaps normalized to their highest values.
Conclusion: The ML classification model with CT and dose distribution as inputs demonstrated superior discrimination for predicting dysphagia compared with traditional NTCP models (AUC=0.81 vs. 0.63). The ML model can guide treatment planning by identifying and minimizing dose to regions most responsible for post-RT dysphagia.