Purpose: To implement an attention-gated deep convolutional neural network (cnn) to predict the optimal 3D dose distribution for Total marrow irradiation (TMI) plans. TMI using intensity modulated radiation therapy has many potential benefits compared to total body irradiation for patients receiving hematopoietic cell transplant. The iterative optimization process in TMI is challenging due to dose constraints on organ-at-risks (OARs) and multiple target volumes (PTVs).
Methods: Twenty-Seven Tomotherapy TMI plans with 20Gy/10fx prescription doses were collected. We focused on the most difficult to be optimized: the thoracic region where lung limits must be met. In this area, there are 6 PTV structures and 3 OARs. We incorporated an attention mechanism in an U-Net network to develop an Attention-Gated U-Net model for predicting 3D dose distributions. The model was trained on twenty randomly selected plans and was validated on the other 7 datasets.
Results: The Attentional-gated U-Net model predicted the PTV and OAR doses accurately. The training time on was around 150 minutes (2500 epochs) on NVIDIA Quadro P6000 GPU. The average absolute difference (|D_predict-D_true |/D_true) of mean dose, max dose and D95 for (1) PTV_Ribs were 2.79%±1.86%, 3.24%±2.75%, 6.95%±5.18%. (2) PTV_Bone was 0.86%±0.71%, 5.09%±2.79%, 7.79%±5.05%. The average absolute difference of mean and max dose waw 3.65%±2.64%, and 5.97%±4.01% for lungs, 6.67%±5.1%, and 12.59%±11.45% for heart.
Conclusion: The Attention-Gated U-Net network models predicted the optimal dose distribution for complex multiple targets TMI plans. Our future work includes training and testing the model on a larger patient cohort with different prescription levels as input into plan optimization. We will also implement the model on other body regions and build a whole-body model for TMI planning.
Not Applicable / None Entered.
TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation