In robotics and autonomous driving, while cameras and LiDAR are extensively employed as primary sensors, they exhibit substantial performance limitations under harsh weather conditions such as rain, fog, and snow. Millimeter-wave (mmWave) radar offers superior all-weather capabilities but faces challenges including lower resolution and susceptibility to noise, artifacts, and ghost points in raw data. Therefore, developing effective methods to enhance radar-based perception while maintaining its all-weather advantages is crucial for robust autonomous navigation. In this paper, we introduce a novel deep learning-based radar odometry method. This method leverages a decoupled conditional diffusion model to process radar point clouds, generating LiDAR-like point clouds for effective cross-modal denoising. Additionally, we design a novel matching enhancement network relying on coarse local descriptors, combined with a reliability map to filter high-confidence keypoint pairs for accurate motion estimation. The proposed method has been extensively evaluated on the publicly available Oxford Radar RobotCar Dataset. The results demonstrate our approach exhibits excellent localization accuracy and robustness across various conditions.