Background: Intensity-Modulated Radiation Therapy (IMRT) is one of the main treatments for brain tumors; however, treatment planning is complex and highly dependent on the physicist's experience. This study aimed to develop and evaluate a Hybrid Convolutional Neural Network (Hybrid CNN) model for accurate automated IMRT treatment planning prediction.
Methods: Data from 120 patients with glioblastoma treated between 2021 and 2024 were used. Contoured CT and MRI images along with Overlap Volume Histogram (OVH) values for organs-at-risk and target volumes were provided as model inputs. Three approaches were evaluated: using OVH data, contoured structures, and a combined input of both. The model was trained using 1D and 3D-CNN layers with a fully connected network, and its performance was assessed using Mean Absolute Error (MAE) and Mean Squared Error (MSE).
Results: The hybrid model outperformed individual models. The mean MAE and MSE in the test set were 2.85 Gy and 7.55 Gy, respectively. Predicted doses for organs-at-risk showed high agreement with actual values (p>0.05). Combining OVH and contoured images improved model accuracy and captured complex relationships between anatomical structures and dose–volume objectives.
Conclusion: The Hybrid CNN can perform automated IMRT treatment planning with high accuracy, improve plan quality, reduce dependence on physicist experience, and integrate with clinical treatment planning systems. This approach paves the way for safer and personalized radiotherapy in patients with brain tumors.