Real-Time and High-Accuracy Switchable Stereo Depth Estimation Method Utilizing Self-Supervised Online Learning Mechanism for MIS

In minimally invasive surgery (MIS), clinicians often rely on 2-D laparoscopic images to assess the size and distances between internal structures. However, this subjective estimation introduces significant uncertainty and can increase surgical risks while reducing efficiency. Modern 3-D laparoscopes offer improved stereoscopic perception and can incorporate stereo depth estimation methods for quantitative analysis. However, existing methods struggle with real-time and high-accuracy demands in diverse surgical scenarios. To address this issue, we propose a novel intraoperative stereo depth estimation framework termed metainitialized online learning (MIOL), aiming to assist surgeons in quantitatively controlling surgical targets during the procedure. This framework features two switchable modes and does not require annotated data. One mode enables rapid depth recovery through surgical videos, providing real-time 3-D reconstruction to help surgeons understand in vivo structures. The other mode achieves high-precision measurements of critical tissues in fixed frames, assisting in surgical decision-making. Our approach employs self-supervised adaptation to train a model specific to each stereo image, eliminating the need for generalization and achieving outstanding accuracy. The framework establishes a lightweight network that converges rapidly under self-supervised losses and incorporates meta-learning pretraining, sparse optical flow guidance, and effective region identification to ensure speed and accuracy. Extensive experiments on two public datasets demonstrate the superiority of our method over existing approaches. Furthermore, we implement an intraoperative measurement system and conduct clinical trials, confirming its practical utility. The code is available at https://github.com/Darcy-vision/MIOL.