Controlling Prediction Dynamics for Reliable Intraoperative Segmentation

Le résumé fourni par la source

Automatic detection and segmentation in intraoperative imaging sequences remains challenging because procedural events can change image appearance abruptly. Instrument motion, material injection, irrigation and suction, and acquisition changes introduce strong artifacts. As a result, models that process each image independently can achieve high overlap scores yet remain unreliable in practice, because their outputs fluctuate from moment to moment. These fluctuations appear as flicker and spatial drift and disrupt time-critical decision support, where clinicians need predictions that are not only accurate on average but also stable enough to trust from one moment to the next. A natural response is to smooth predictions over time, but naive smoothing introduces a second failure mode. It can hide real procedural changes by over-stabilizing, or it can propagate mistakes when artifacts dominate. We propose stability-gated temporal inference to address this stability-versus-change tension directly. Keeping the underlying single-image segmenter fixed, our method builds a reliable reference from recent stable outputs, uses it to prevent abrupt prediction swings caused by transient artifacts, and relaxes this constraint only where the scene is truly changing due to the procedure. This makes the model stable when it should be stable and responsive when it must be responsive. The same mechanism also supports short-horizon trend forecasting, allowing the system to issue prompts such as whether a material boundary is likely to keep expanding toward a safety-critical region. We evaluate on 300 intraoperative sequences and three public surgical benchmarks, comparing against per-frame, smoothing, propagation, open-source VOS, and promptable video foundation baselines. At comparable segmentation quality, our method improves Dice from 0.62 to 0.73, reduces drift from 5.8 to 2.6 pixels and flicker rate from 0.21 to 0.09, while keeping change lag low. On the internal dataset, trend prompts improve lead time from 1.0 to 2.2 seconds on average, enabling more reliable intraoperative decision support.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.