Application of Image Segmentation Based on U-Net Model in Radiation Oncology Treatment

The exact delineation of the target region and organs at risk in radiation oncology therapy greatly depends on the experience of the physician. Traditional manual or semi-automatic segmentation methods have problems such as subjective bias, which can easily lead to inaccurate radiotherapy dose distribution, affect the treatment effect and increase the risk of normal tissue complications. To solve this problem, this study proposes an intelligent segmentation system based on an improved U-Net architecture: First, in view of the small sample size and anatomical complexity of medical images, a multi-scale U-Net++ (U-Net with a Nested Skip Pathways Architecture) model with integrated attention mechanism is designed to enhance the ability to extract tumor edge features through skip connection optimization. Then, an adaptive data enhancement strategy is applied, combined with CT (Computed Tomography) and MRI (Magnetic Resonance Imaging) multi-modal image registration and normalization processing, to construct a cross-device, multi-center robust training dataset. Finally, a hybrid loss function is used to balance the class imbalance problem. Experiments show that the improved U-Net++ model exhibits excellent segmentation accuracy on multiple datasets, especially in the segmentation of tumor target areas and complex anatomical structures. The Dice coefficients reach 0.91 and 0.88, respectively, and the Hausdorff distances are reduced to 3.7 mm and 4.1 mm, respectively. This model also shows significant advantages in the segmentation of small target tumors. It can effectively improve the boundary recognition ability, reduce the rate of mis-segmentation and missed segmentation, and the inference speed meets the needs of clinical real-time processing. Studies have shown that the improved method based on U-Net++ provides a more accurate and efficient solution for automated image segmentation in radiation oncology treatment, and has strong clinical application potential.