Author: Ibtisam Almajnooni, Victor Cobilean, Milos Manic, Harindra Sandun Mavikumbure, Elisabeth Weiss, Lulin Yuan 👨🔬
Affiliation: Virginia Commonwealth University 🌍
Purpose: This study aims to optimize the 3D U-Net architecture for dose prediction in lung cancer radiation therapy (RT) plans, particularly in scenarios with limited clinical data, as well as to quantify the uncertainty caused by variations in plan quality and physician clinical preferences.
Methods: This study included 115 clinical RT plans for patients with locally advanced lung cancer who were treated at our institution. Using CT scans and organ segmentation masks as input, we tested various 3D U-Net architectures with three attention mechanisms: Patch Attention in residual convolutional blocks, Pixel Attention in the bottleneck, and Soft Attention in the encoder-decoder residual connections. We used the L1 and Smooth L1 loss functions, as well as their weighted versions, based on the region of interest masks generated by pixel intensity differences. Because of the limited data, we used augmentations such as flips, rotations, blurring, noise, and contrast variations. Furthermore, we used Monte Carlo dropout during training and investigated the prediction variance of the trained models.
Results: The lowest Mean absolute error (MAE) of the predicted 3D dose distribution among the test cases is 2.034 Gy. The lowest MAE for the Dose-Volume Histogram (DVH) metric is 6.115 %. Uncertainty prediction results showed consistent variance across all models (approx. 0.4 Gy), indicating that the majority of uncertainty is because of training data noise (physician preferences, plan quality variation).
Conclusion: In conclusion, L1 and Weighted L1 loss functions outperform Smooth L1 losses for the Dose metric, while no differences were found for the DVH metric. Models trained with the same loss function showed similar performance, regardless of architecture. Monte Carlo dropout analysis revealed consistent variance across all models, indicating that data, not the model architecture or loss function, is the most significant factor contributing to the prediction uncertainty.