Deep learning-based prediction of interfractional anatomic variations in prostate cancer radiotherapy

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• Interfractional prostate displacement, driven by bladder volume variations, impacts radiotherapy accuracy.
• Accurate patient-specific anatomy estimation is crucial for effective prostate radiotherapy.

Purpose of the Study:

• To develop a deep learning (DL) model for predicting daily megavoltage computed tomography (MVCT) images.
• To enable patient-specific anatomic estimation in prostate radiotherapy using bladder volume and treatment fraction.

Main Methods:

• Retrospective analysis of 700 MVCT scans from prostate cancer patients treated with tomotherapy.
• Training a 3D U-Net deep learning model to generate synthetic MVCT images from kilovoltage computed tomography (KVCT), bladder volume, and fraction number.
• Evaluating model performance using image similarity metrics (SSIM, NCC, Dice) and contour accuracy (Dice, MSD).

Main Results:

• The DL model achieved accurate MVCT predictions with high SSIM (0.76-0.80) and NCC (0.84-0.89).
• Predicted anatomy allowed for accurate bladder (Dice 0.83) and prostate (Dice 0.92) contouring.
• Bladder volume changes showed moderate correlation with bladder contour accuracy metrics.

Conclusions:

• The deep learning framework shows promise for predicting anatomy and generating contours using noninvasive inputs.
• This patient-specific method estimates daily treatment anatomy based on planning KVCT and bladder volume.
• The approach is suitable for helical radiotherapy systems where daily MVCT is standard practice.