This paper presents O-DAT (Optimized DA-TransUNet), a deep learning framework for medical image segmentation. Accurate segmentation is essential for clinical diagnosis but challenging due to complex anatomical structures. O-DAT enhances U-Net with optimized dual attention (ODA) modules for better spatial and channel feature extraction, and uses patch expansion layers in the decoder for efficient upsampling. This approach combines the strengths of U-Net and Transformers, improving both accuracy and efficiency. Experiments show O-DAT outperforms existing methods, achieving a DSC of 80.30% and an HD of 25.88 mm on the Synapse dataset. Ablation studies confirm the effectiveness of ODA blocks and patch expansion layers. O-DAT sets a new benchmark for medical image segmentation, with potential to enhance clinical diagnosis and guide future research.