Rethinking the value of dynamic and static feature planes in 4D reconstruction of deformable tissues

Reconstructing deformable tissues is crucial for medical image computing and robotic surgery, as it enhances the safety and efficacy of surgical procedures. However, current methods face significant challenges, including errors in tissue reconstruction at occluded regions and limitations in real-time accurate observation of complex structures.

In this paper, we present a novel method called Rethink Plane (RPlane), an efficient framework based on Neural Radiance Fields (NeRF), designed to reconstruct global high-fidelity deformable tissues from binocular endoscopic videos efficiently. Our main contribution lies in rethinking the value of dynamic and static features that existing methods often overlook, and developing a Depth Uncertainty Filter. Throughout this work, the dynamic filter is an extremely important foundational component. Based on this, a Dynamic Feature Enhancement module is proposed to address the depth distortion problem caused by the occlusion of surgical instruments. Additionally, a Color Recurrent Refinement strategy is proposed to reduce dynamic blurring caused by instrument contact or tissue self-motion. We validate the effectiveness of RPlane on two datasets (ENDONERF and StereoMIS).

In all cases, RPlane achieves state-of-the-art (SOTA) performance in terms of tissue reconstruction quality and detail clarity (with a PSNR of 40.527 in ENDONERF and 36.267 in StereoMIS). Furthermore, RPlane demonstrates a 53.3% improvement in robustness without increasing training time or computational resources.

RPlane addresses depth reconstruction errors caused by surgical instrument occlusion, which are common in existing algorithms. The Dynamic Feature Enhancement module is used to enhance geometric modeling in occluded areas, while the Dynamic Weight Generation & Fusion and Color Recurrent Refinement strategies improve the texture details of tissues. This significant performance improvement promises to be an innovative solution for intraoperative applications.