Date of Publication : July 2026

Abstract: Real-time surgical video understanding is a pre-requisite for context-aware, AI-assisted operative intervention, yet the computational demands of deep inference remain a barrier to practical deployment. Existing methods apply uniform-rate inference across all frames regardless of clinical content, subjecting frames without instrument-tissue interaction to the same computational cost as those containing critical operative events. In laparoscopic cholecystectomy, over 40% of frames are clinically redundant under this criterion, yet existing pipelines do not systematically exploit this structure to reduce inference cost.

This paper presents a multi-stage adaptive inference framework with learned importance-based routing. A lightweight MobileNetV2 controller generates per-frame importance scores, stabilized via sliding-window dynamic normalization and moving-average temporal smoothing to suppress illumination drift and kinematic noise.

A learned threshold function routes each frame to one of three paths: a ResNet18 dual-head network for instrument and verb recognition, a DenseNet169 with Efficient Channel Attention for full triplet prediction (instrument, verb, target), or a discard path for clinically irrelevant frames.

Evaluated on the CholecT50 benchmark, the framework achieves 31.8% triplet mAP, retaining 97.8% of the full-frame baseline, while reducing amortized GFLOPs per frame by 67.9% and increasing throughput from 18.5 to 46.2 FPS, surpassing the 30 FPS threshold for intra-operative deployment.

Ablation studies confirm that both stabilization components are independently necessary, with removal of either reducing triplet mAP by 4–8 points.

The system is released with a FastAPI/Docker inference backend and a React 18 TypeScript clinical dashboard, establishing a reproducible reference implementation for the translation of adaptive surgical AI from research prototype to intra-operative deployment.

Reference :