Practical deep learning solutions for prostate cancer segmentation and implementation in radiotherapy planning

Accurate segmentation of the prostate and surrounding structures is crucial in radiotherapy planning to ensure effective treatment delivery while minimising radiation exposure to nearby tissues. However, existing deep learning (DL) segmentation studies often exclude clinically important structures such as the penile bulb (PB) and seminal vesicles (SV), rely heavily on large proprietary datasets that are difficult to acquire in clinical settings, and lack direct comparisons with traditional atlas-based methods. To address these gaps, this study provides practical insights for hospitals considering the implementation of DL systems by evaluating 2D U-Net models across six structured experiments, including one experiment comparing a 3D U-Net model. Notably, this is the first study to utilises publicly available prostate CT data in DL training and integrates public and private datasets to improve model robustness for prostate and adjacent anatomical structures. The inclusion of underrepresented structures, such as PB and SV, enhances the clinical applicability of the models. Our results demonstrated that moderate-sized datasets, ranging from 60 to approximately 300 in our study, can achieve sub-millimetre Mean Distance Agreement (MDA) for key organs at risk like the bladder (0.60 mm) and rectum (0.93 mm), reducing reliance on large-scale annotated data. Models trained on routine clinical contours performed well without extensive manual refinement, demonstrating that existing clinical data can be leveraged for model training. In comparisons with four commercial atlas-based tools, the 2D U-Net models achieved superior accuracy for key anatomical structures, including the prostate, bladder, and SV, with our deep learning model achieving an MDA of 1.28 mm for the prostate compared to the 1.4-2.5 mm range of the atlas tools. Although 3D models provided improved spatial context for small structures such as SV, 2D models proved to be a practical alternative due to lower computational demands, making them suitable for resource-limited clinical settings.