Room: AAPM ePoster Library
Purpose: Total marrow irradiation (TMI) using intensity modulated radiation therapy has been used in place of conventional total body irradiation as the conditioning regimen in hematopoietic cell transplantation. It’s a challenging iterative process to optimize the plan and balance between the planning-target-volume(PTV) coverage and organ-at-risk (OAR) dose-sparing. This feasibility study aims to implement and test convolutional neural network (cnn) deep learning model to accurately predict the optimal TMI plan dose distribution and to help plan optimization.
Methods: Fourteen Helical Tomotherapy TMI plans with 20Gy/10fx prescription doses were collected. The study focused on predicting thoracic region dose with 6 PTV structures and 4 OARs, In this area, navigating the trade-off between PTV_Ribs, PTV_Lymph nodes, PTV_Bone coverage and sparing the Lungs, esophagus is most difficult. A convolution neural network Unet model is applied for predicting dose distributions. The model was trained on 2D single slices from ten random selected plans (training set: 330 slices) on NVIDIA RTX2070 GPU card. The other four plans were used as testing dataset.
Results: The Unet network model predicted the PTV and OAR doses accurately. The training time was around 150 minutes (2500 epochs). For the test data result, the average absolute difference (|D_predict-D_true |/D_true) of mean dose, max dose and D95 for (1) PTV_Ribs were 2.10%±0.81%, 1.54%±0.85%, 7.33%±2.96%. (2) PTV_Bone were 1.18%±0.70%, 4.30%±1.08%, 2.36%±0.96%. The average absolute difference of mean and max dose were 3.29%±2.66%, and 2.22%±1.47% for lungs, 5.96%±3.24%, and 2.79%±1.47% for heart, 8.61%±3.82%, and 6.15%±3.48% for esophagus.
Conclusion: Convolutional neural network models are promising to accurately predict the optimal dose distribution for complex multiple target/multiple dose-level TMI plans. This can help derive optimization objectives and reduce unnecessary treatment planning time. Further model development and test on a larger patient cohort is warranted.