This study introduces RISE-MAR (Radiologist-Integrated Self-Evolution for Metal Artifact Reduction), a novel framework that addresses the persistent challenge of metal artifacts in computed tomography (CT) imaging. Unlike previous approaches that rely solely on mathematical optimization or deep learning, RISE-MAR explicitly integrates radiologist expertise into a self-evolving system through an innovative feedback loop. Our dual-domain architecture combines a transformer-based image branch with a specialized sinogram network, enforcing physical consistency while capturing long-range dependencies essential for complex artifact patterns. The key innovation lies in our confidence-guided pseudo-labeling mechanism that selectively identifies high-confidence regions for self-training, refined by radiologist feedback that ensures clinical relevance. Experimental results demonstrate RISE-MAR's superior performance across synthetic and clinical datasets, with significant improvements in both quantitative metrics (PSNR: 36.2 dB, SSIM: 0.923) and clinical relevance scores (4.2/5) compared to state-of-the-art methods. Most notably, RISE-MAR shows remarkable generalization capability across different implant types and anatomical regions, effectively bridging the domain gap between training environments and clinical applications. Our work establishes a new paradigm for developing medical image enhancement techniques that leverage the complementary strengths of computational methods and clinical expertise.