Comparing multi-image and image augmentation strategies for deep learning-based prostate segmentation.

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence

Background:

• MR-Linac systems acquire multiple images per patient during adaptive radiotherapy.
• Deep learning (DL) models can potentially reduce manual annotation efforts in radiotherapy segmentation.
• Prostate cancer treatment benefits from accurate organ segmentation for dose delivery.

Purpose of the Study:

• To evaluate the benefit of using multiple MR-Linac images versus single images for training deep learning models for prostate treatment segmentation.
• To assess the impact of data quantity (images per patient and number of patients) on segmentation performance.

Main Methods:

• A 2D U-net deep learning model was trained using single and multiple MR-Linac images from prostate cancer patients.
• Segmentation performance was evaluated using DICE and Hausdorff 95% metrics.
• The study compared training strategies involving varying numbers of images per patient and varying numbers of patients.

Main Results:

• Minimal improvement in DICE and Hausdorff 95% metrics was observed when using multiple images per patient compared to single images.
• The most significant difference was noted for the rectum segmentation in a low-data scenario (training with images from five patients).
• A 2D U-net achieved DICE values of 0.80 with one image/patient and 0.83 with five images/patient.
• Increasing the number of patients in the training set reduced the performance gap between single and multiple image strategies.
• Standard data augmentation methods demonstrated greater effectiveness than using multiple images per patient.

Conclusions:

• Utilizing multiple MR-Linac images per patient offers limited advantages for prostate treatment segmentation with deep learning models.
• Standard data augmentation techniques are more impactful for improving segmentation accuracy than increasing the number of images per patient.
• Future research should focus on optimizing data augmentation and exploring other deep learning architectures or training strategies.