Automatic delineation of organs at risk for prostate cancer on computed tomography images using convolutional neural networks

Abstract

Background: Prostate cancer is the fourth most common cancer worldwide and the second most common in men. Accurate identification of organs at risk (OARs) is crucial for improving treatment planning, making the process faster and more efficient. Artificial intelligence has become a vital tool in medicine, reducing delays and improving treatment workflow by enabling precise, consistent automatic organ delineation. Objectives: This study investigates and compares the performance of deep learning models for automatic OAR segmentation in CT images of prostate cancer patients. Materials and methods: Using 2,011 CT images from the Prostate-Anatomical-Edge-Cases dataset, the U-Net, Dense V-Net, and DeepMedic models were evaluated under the same parameters for segmenting the bladder, rectum, and left and right femoral heads. Performance was assessed using the Dice Similarity Coefficient (DSC), 95th percentile Hausdorff Distance (HD95), and Mean Surface Distance (MSD). Results: The Dense V-Net outperformed the other models, achieving the highest DSC (0.99) and the lowest HD95 (0.00) and MSD (0.00) for the bladder. U-Net performed well for well-defined structures, such as the femoral heads, while DeepMedic demonstrated the most stability. Conclusion: Dense V-Net’s superior performance highlights its effectiveness in medical image segmentation, particularly in prostate cancer treatment.