3753 - AI-Based Predict and Auto-Segment of PET-Avid Lymph Nodes from Contrast-Enhanced CT Simulation in Head and Neck Cancer Radiotherapy
Presenter(s)
W. C. You1,2, Y. T. Shao3, Y. Y. Hsu4, and Y. F. Lu5; 1Department of Post-Baccalaureate Medicine, National Chung Hsing University, Taichug City, Taiwan, 2Department of Radiation Oncology, Taichung Veterans General Hospital, Taichung, Taiwan, 3Department of Post-Baccalaureate Medicine, National Chung Hsing University, Taichung City, Taiwan, 4Department of Radiation Oncology, Taichung Veterans General Hospital, Taichung City, Taiwan, 5Department of Radiation Oncology, Taichung Veterans General Hospital, Taichung, Taiwan, Taichung City, Taiwan
Purpose/Objective(s): Accurate delineation of PET-avid lymph nodes (LNs) is crucial for radiotherapy planning in head and neck cancer (HNC), but it often necessitates resource-intensive PET imaging. We developed and clinically validated a nnUNet-based deep learning model to predict and auto-segment PET-avid LNs directly from contrast-enhanced CT simulation scans. This model aims to assist radiation oncologists in accurately identifying high-risk LN areas, thereby improving clinical workflows and reducing reliance on PET scans.
Materials/Methods: Our dataset included paired PET and contrast-enhanced CT simulation scans from 142 HNC patients. Expert consensus was used for ground truth labeling, identifying 266 PET-avid and 255 PET-negative lymph nodes (LNs). A 3D nnUNet deep learning network was trained to automatically segment PET-avid nodes based solely on the CT simulations. The model's performance was clinically validated on an independent cohort of 42 patients with 56 paired PET and contrast-enhanced CT simulations. We calculated diagnostic metrics to assess predictive accuracy, including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and overall accuracy. In addition, we compared the model's performance to radiologist reports derived from CT scans.
Results: In clinical validation, the nnUNet-based model demonstrated a sensitivity of 77.8%, specificity of 84.2%, PPV of 70.0%, NPV of 88.9%, and overall accuracy of 82.1% in predicting PET-avid LNs. Compared to radiologist interpretation (sensitivity: 61.5%, specificity: 86.7%, PPV: 80.0%, NPV: 72.2%, accuracy: 75.0%), the AI model showed greater sensitivity, NPV, and accuracy while decreasing potential interpretation variability. This CT-based predictive approach significantly lowers the need for routine PET imaging during radiotherapy planning, providing substantial clinical efficiency and resource management benefits.
Conclusion: Our nnUNet-based predictive model accurately identifies and auto-segments PET-avid lymphadenopathy from contrast-enhanced CT simulations in head and neck cancer patients, outperforming traditional radiologist reporting. This AI-driven approach is feasible and has the potential to facilitate radiotherapy planning by decreasing reliance on PET scans and optimizing clinical workflows by helping radiation oncologists identify high-risk lymph node regions.