Investigation and benchmarking of U-Nets on prostate segmentation tasks.

Area of Science:

• Medical imaging
• Radiotherapy
• Artificial intelligence

Background:

• Personalized radiotherapy for prostate cancer requires accurate delineation of target structures.
• Biomedical image segmentation is challenging due to time constraints, expertise requirements, and observer variability.
• Deep learning models, particularly U-Net architectures, have advanced medical image segmentation capabilities.

Purpose of the Study:

• To provide a reliable framework for assessing deep learning models in medical image segmentation.
• To objectively compare automatic prostate segmentation algorithms.
• To evaluate the strengths and weaknesses of different 3D prostate segmentation models.

Main Methods:

• Comprehensive review of state-of-the-art convolutional neural networks for 3D prostate segmentation.
• Development of a framework for objective comparison using public and in-house CT and MR datasets.
• Rigorous evaluation of automatic prostate segmentation algorithms on multi-modal images.

Main Results:

• Deep learning models can achieve clinician-level accuracy in segmenting anatomical structures.
• The developed framework allows for objective comparison of segmentation algorithms.
• Evaluation highlighted the performance variations and limitations of current deep learning models for prostate segmentation.

Conclusions:

• A robust framework is established for evaluating deep learning-based prostate segmentation.
• Objective comparison is crucial for advancing the clinical application of AI in radiotherapy.
• Further research is needed to address data heterogeneity and improve model generalizability.