Fast contour propagation for MR-guided prostate radiotherapy using convolutional neural networks.

Area of Science:

• Medical Physics
• Radiotherapy
• Medical Imaging

Background:

• Accurate organ contouring is crucial for effective radiotherapy planning and delivery.
• Manual contour propagation in MR-guided prostate radiotherapy is time-consuming and prone to inter-observer variability.
• Automated methods are needed to improve efficiency and consistency in contour propagation.

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN) for rapid and automatic propagation of organ contours.
• To assess the CNN's performance in transferring pretreatment contours to fraction images in MR-guided prostate radiotherapy.
• To compare the CNN's accuracy and speed against traditional deformable registration software.

Main Methods:

• A CNN was trained for combined image registration and contour propagation using T2-weighted 3D MR imaging from five prostate cancer patients.
• The CNN estimated propagated contours and deformation fields, trained on synthetically generated data.
• Performance was evaluated using leave-one-out cross-validation and compared to Elastix software.

Main Results:

• CNN variants optimized for segmentation overlap or a combined objective significantly outperformed Elastix in Hausdorff distance.
• The CNN achieved a registration speed of 0.5 seconds, substantially faster than conventional methods.
• CNNs trained to maximize prostate overlap and minimize registration errors yielded the best propagation results.

Conclusions:

• A CNN-based approach offers a fast and accurate solution for deformable contour propagation in MR-guided prostate radiotherapy.
• Optimizing for segmentation overlap and registration accuracy is key to achieving superior performance.
• This automated method has the potential to enhance clinical workflow and treatment precision.