Accurate segmentation of prostate ultrasound and magnetic resonance (MR) images is crucial for robot-assisted diagnosis and treatment, particularly in guiding precise needle insertion. However, conventional methods often suffer from limited performance due to low image contrast, noise, and anatomical variability. We propose BiResUNet, a novel U-Net-based segmentation framework. It integrates multi-supervised learning and a customized attention mechanism for enhanced robustness and accuracy, even with poor image quality. We validate the method using data from 704 patients, including a private dataset from Shanghai East Hospital and public datasets from The Cancer Imaging Archive (TCIA) and the μ-ProReg Challenge. BiResUNet is benchmarked against eight state-of-the-art segmentation methods across multiple evaluation metrics. Experimental results demonstrate that BiResUNet achieves superior performance on both MR and ultrasound images (Dice: 0.93 & 0.89; IoU: 0.93 & 0.90, respectively), highlighting its potential for clinical deployment in multimodal prostate image segmentation. The code is publicly available at https://github.com/Prps7/BiResUnet .