Intrafractional rectum anatomy shape prediction based on 3D point cloud representation in online adaptive radiation...

Area of Science:

• Medical Physics
• Artificial Intelligence in Medicine
• Radiotherapy Technology

Background:

• Online adaptive radiotherapy (OART) for prostate cancer requires accurate intrafractional anatomical monitoring.
• Predicting rectal shape variations during treatment is crucial for maintaining radiotherapy accuracy and minimizing side effects.

Purpose of the Study:

• To develop a generative artificial intelligence (AI) model for predicting intrafractional rectal shape changes in prostate cancer OART.
• To evaluate the performance of the developed model against conventional deep learning approaches.

Main Methods:

• Retrospective analysis of MRI data from 42 prostate cancer patients undergoing OART.
• Development of the SA-UNet generative AI model for 3D rectal shape prediction from point cloud data.
• Benchmarking SA-UNet against Baseline-MLP and Baseline-PointCNN models using Dice Coefficient (CD), Earth Mover's Distance (EMD), and Jaccard Index (JAC).

Main Results:

• SA-UNet demonstrated superior performance with the lowest average CD (29.06 mm) and EMD (4.82 mm), and the highest average JAC (0.69).
• SA-UNet showed significantly greater consistency across treatment fractions compared to Baseline-MLP (p < 0.025).
• SA-UNet significantly outperformed Baseline-PointCNN in all evaluated metrics (p < 0.01).

Conclusions:

• The SA-UNet model shows preliminary feasibility for real-time intrafractional rectal shape prediction in OART.
• This AI-driven approach has the potential to enable early warnings for anatomical changes and improve the precision of radiotherapy.