Author: Smith Apisarnthanarax, Stephen R. Bowen, Sunan Cui, Jie Fu, Clemens Grassberger, Yulun He, Yejin Kim, Matthew J. Nyflot, Sharon Pai 👨🔬
Affiliation: Department of Radiation Oncology, University of Washington and Fred Hutchinson Cancer Center, Department of Radiation Oncology, University of Washington, University of Washington, Department of Radiation Oncology, Fred Hutchinson Cancer Center, University of Washington, Department of Physics, University of Washington, University of Washington and Fred Hutchinson Cancer Center 🌍
Purpose: 99mTc-sulfur colloid SPECT imaging enables quantitative assessment of voxel-wise liver function in patients with hepatocellular carcinoma (HCC). Accurately predicting post-radiotherapy (RT) liver SPECT may enable early dose-response modeling, facilitating more informed treatment planning and timely interventions. In this study, we developed a 3D multi-path DenseNet to predict post-RT SPECT based on dose distributions and pre-RT SPECT.
Methods: We included 41 HCC patients who underwent SBRT or proton RT. Each patient had SPECT/CT scans before and after RT, which were deformably aligned to the planning CT. RT doses were converted to EQD2a/b=3 and GyRBE. SPECT images were normalized to the normal liver (liver-GTV) region that received less than 5GyEQD2. We constructed a multi-path DenseNet that separately encoded the EQD2 dose and pre-RT SPECT and predicted the post-RT SPECT. A log-linear model was fitted as the baseline method for comparison. Four-fold cross-validation was conducted. Prediction accuracy was assessed using mean square error (MSE) within normal liver between the predicted and ground truth post-RT SPECT. Additionally, regional dose-response curves were drawn by binning normal liver in 5 GyEQD2 increments and calculating mean dose and percentage uptake changes between pre-RT and post-RT SPECT in each binned region. Mean absolute error (MAE) between the ground truth and predicted dose-response curves was computed.
Results: The multi-path DenseNet achieved an average MSE of 0.099 across all test patients, superior to the baseline method with an average MSE of 0.145. Average MAEs between dose-response curves were lower for DenseNet (13.4%) compared to the baseline method (19.6%). All Wilcoxon Signed-Rank tests yielded p-values of less than 0.01.
Conclusion: DenseNet significantly enhances the accuracy of predicted post-RT SPECT compared with the baseline method. Future work will focus on utilizing the prediction to assess RT outcomes and toxicity, aiming to further refine treatment approaches and patient management.