A deep learning framework for prostate localization in cone beam CT-guided radiotherapy.

Area of Science:

• Medical Imaging
• Radiation Oncology
• Artificial Intelligence

Background:

• Accurate patient setup is crucial for effective radiation therapy (RT).
• Cone-beam computed tomography (CBCT) is widely used for image-guided patient positioning.
• Current methods for prostate planning target volume (PTV) localization can be time-consuming and may require fiducial markers.

Purpose of the Study:

• To develop and validate a deep learning model for automated prostate PTV localization on CBCT.
• To enhance the efficiency and accuracy of CBCT-guided patient setup in RT.

Main Methods:

• A two-step task-based residual network (T²RN) was developed to identify prostate PTV landmarks directly from CBCT images.
• The T²RN model was trained using over a thousand synthetic CT images with varied anatomical configurations to simulate diverse clinical scenarios.
• The model's performance was evaluated on 240 CBCT datasets (120 original, 120 synthetic) from six patients.

Main Results:

• The T²RN model achieved high accuracy in PTV localization, with systematic/random setup errors below 0.25/2.46 mm and 0.14/1.41° for translation and rotation, respectively.
• Pearson's correlation coefficients exceeded 0.94, indicating strong agreement between model predictions and reference values.
• Bland-Altman analysis confirmed good agreement, demonstrating the reliability of the deep learning approach.

Conclusions:

• A novel deep learning technique (T²RN) effectively localizes the prostate PTV for radiation therapy patient setup.
• The study demonstrates the feasibility of highly accurate, marker-less prostate setup using advanced deep learning strategies.
• This approach has the potential to significantly improve the workflow of CBCT-guided patient positioning in radiation oncology.