Low-dose CT (LDCT) has been widely adopted in clinical practice to reduce radiation exposure, especially for brain CT imaging in the rapid diagnosis of neurological disorders. However, LDCT images commonly contain spatially non-uniform and nonlinear noise, which significantly degrades diagnostic quality and can compromise clinical decisions. Although CNN-based denoising methods have shown the outstanding performance among deep learning approaches, their use of spatially invariant convolutional kernels restricts their capacity to handle spatially varying noise and anatomical structures. To model the nonlinear and spatially variant noise in LDCT images, we propose the Feature-space Kernel Prediction Network (Feat-KPN), which dynamically generates adaptive kernels in the feature domain. Furthermore, we introduce an Anatomical Perception Loss that incorporates anatomical context to guide the recovery of fine structural details. Experimental results demonstrate that our method consistently outperforms existing approaches across multiple quantitative metrics, particularly in restoring critical brain structures.