Confidence-aware Monocular Depth Estimation for Minimally Invasive Surgery

A collaborative research team has introduced a novel confidence-aware framework for monocular depth estimation (MDE) specifically designed for the chaotic visual environment of minimally invasive surgery (MIS). Published on arXiv, the work addresses a critical limitation: current MDE models for endoscopy provide depth maps but offer no indication of their own reliability, which is problematic when video is contaminated by smoke, blood, blur, or surgical tool occlusions. The proposed system not only aims to be more accurate but, for the first time, generates a simultaneous confidence map telling surgeons and robotic systems which parts of the estimated 3D scene can be trusted.

The technical approach features three key innovations: using an ensemble of stereo-matching models to create calibrated confidence targets for training, a new confidence-aware loss function that prioritizes reliable pixels, and a lightweight convolutional head that predicts per-pixel confidence during inference. On the internal 'StereoKP' clinical dataset, the framework improved dense depth accuracy by approximately 8% over baseline models. The ability to output a confidence metric is a paradigm shift, moving surgical AI from providing potentially flawed answers to offering answers with quantified uncertainty. This paves the way for more reliable augmented reality overlays, safer robotic assistance, and ultimately greater surgeon trust in AI-guided procedures.