Complex surgeries such as liver resections require robust AR tracking under deformation and occlusion. We propose an AI-enhanced depth framework using a CNN-LSTM hybrid to denoise depth maps and predict motion trajectories. GAN-based augmentation improved training on limited intraoperative datasets. On 500 annotated liver frames, mean error dropped from 2.5 mm (baseline) to 1.0 mm, robustness under occlusion increased by 40%, and latency stayed under 45 ms. The system maintained 24 fps, outperforming conventional ICP-based methods.