Three-Dimensional Dose-Function Data-Based Deep Convolutional Neural Network for Predicting Pulmonary Toxicity After Radiotherapy for Lung Cancer
Y Fujita1*, Y Nakajima1, D Kawahara2, T Kimura3, T Yamamoto4, (1) Komazawa University, Tokyo, JP, (2) Hiroshima University, Hiroshima, JP, (3) Kochi University, Kochi, JP, (4) UC Davis School of Medicine, Sacramento, CA
Presentations
PO-GePV-M-351 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: To quantify the predictive power of a 3-dimensional (3D) deep convolutional neural network (CNN) model based on dose distribution and pulmonary functional image data for predicting grade ≥2 radiation pneumonitis (RP) after radiotherapy for lung cancer.
Methods: We studied 60 patients with lung cancer who underwent radiotherapy. All patients underwent single-photon emission computed tomography perfusion scans before radiotherapy. A pre-trained deep 3D residual network architecture with 200 layers on a large-scale video dataset was fine-tuned on the combined images of 3D dose and perfusion image data. The predictive power was quantified by the area under the receiver operating characteristic curve (AUC-ROC) and area under the precision-recall curve (AUC-PR) using a 5-fold cross-validation method. The predictive power was compared between the 3D CNN model and different logistic regression (LR) models based on a range of dose-function metrics (e.g., fV20, the percentage of total lung perfusion receiving ≥20 Gy). Moreover, the guided gradient-weighted class activation maps (Grad-CAM) were generated to identify salient regions that received the highest attention during classification decisions. In different salient regions, we compared the perfusion-weighted mean dose between patients with grade ≥2 RP and those with grade ≤1 RP using a Wilcoxon rank-sum test.
Results: The 3D CNN model had significantly higher AUCs compared with the LR models (AUC-ROC 0.992 vs. 0.700-0.785; AUC-PR 0.992 vs. 0.811-0.905). Salient regions identified by the guided Grad-CAM were mainly distributed in higher perfusion-weighted dose regions. In the top 20% salient region, the perfusion-weighted mean dose was significantly higher in patients with grade ≥2 RP than in those with grade ≤1 RP (median, 27.4 Gy vs. 14.7 Gy, p < 0.01).
Conclusion: The dose-function data-based 3D CNN model demonstrated greater predictive power for RP compared with the LR models. Dose to well-perfused lung regions were associated with RP.
Keywords
Radiation Therapy, Dose Response, Lung
Taxonomy
TH- Response Assessment: Modeling: Machine Learning
Contact Email