A 3D Dosimetric Data-Driven Multipath Network for Outcome Prediction in Early-Stage Non-Small Cell Lung Cancer Patients Treated with Stereotactic Body Radiotherapy
T Arsenault1*, A Amini1, B George2, S Bhat2, L Bailey2, P Vempati4, B Young4, C Towe1, P Linden1, Y Sun2, N Zaorsky1, D Spratt4, R Muzic3, T Biswas4, T Podder4, (1) University Hospitals Cleveland Medical Center, Cleveland, OH, (2) Case Western Reserve University, School Of Medicine, (3) Case Western Reserve University, Department of Biomedical Engineering, (4) Seidman Cancer Center /Uh Cleveland & CWRU, OH, Cleveland, OH,
Presentations
TU-D1000-IePD-F4-3 (Tuesday, 7/12/2022) 10:00 AM - 10:30 AM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 4
Purpose: To investigate the ability of a dose-integrated multi-path neural network (MPNN) in predicting the 2-year overall survival (2yr-OS) for non-small cell lung cancer (NSCLC) patients treated with stereotactic body radiotherapy (SBRT)
Methods: Data were acquired through an IRB-approved retrospective review for patients with node negative stage I-II NSCLC who underwent SBRT at our institution between 2012 and 2021. Primary endpoint for model training was 2yr-OS.The cohort included 371 patients among whom 2yr-OS was available for 136 patients. A MPNN was composed of two pathways: (1) An image-down pathway with 2-Channel inputs consisting of (a) three-dimensional (3D) CT and (b) 3D RTdose, each with a patch size of 124x124x124 voxels, and (2) A feature up pathway with four input values (age, adjusted Charlson comorbidity index, 18-fluoro-deoxy-D-glucose positron emission tomography maximum standard uptake value, and smoking pack-years). Training was implemented in Keras with Tensorflow using Adam gradient descent, a mini-batch size of eight, and 100 epochs. Data were randomly divided into training (n=104), and testing (n=32). Performance was measured by accuracy defined as the ratio of correct classifications to the sample number, and by the area under the curve (AUC) for receiving operator characteristics. Prediction based on NCCN standard staging via univariable analysis was used as benchmark for comparison.
Results: Median age was 75 years, and the 2yr-OS was 65%. On the testing cohort, MPNN performed better than standard age grouping and yielded an accuracy=0.69 and AUC=0.74, versus NCCN stage grouping with an accuracy=0.68 and AUC=0.66.
Conclusion: This work highlights the clinical predictive potential using the treatment dose distribution via an intelligent integration of 3D dosimetric data with clinical, demographic, and radiological data. To our knowledge, this is the first MPNN using 3D dosimetric data as input for outcome prediction in SBRT.
Keywords
Taxonomy
TH- External Beam- Photons: extracranial stereotactic/SBRT
Contact Email