CurConMix+: A Unified Spatio-Temporal Framework for Hierarchical Surgical Workflow Understanding

Surgical action triplet recognition models fine-grained surgical interactions as structured triplets of ⟨instrument, verb, target⟩, but remains challenging because of severe class imbalance, subtle visual ambiguity, and compositional dependencies among triplet components. Existing approaches typically address only a subset of these challenges. We present CurConMix+, a unified spatio-temporal framework for surgical action triplet recognition. Its spatial component, CurConMix, learns structured and discriminative action representations through curriculum-guided contrastive learning, hard-pair sampling, and feature-level mixup. Its temporal component, the Multi-Resolution Temporal Transformer (MRTT), learnably combines multi-resolution temporal features and balances spatial and temporal evidence. We also introduce LLS48, extending surgical triplet recognition beyond cholecystectomy through aligned step, task, and action annotations sampled across 48 complete laparoscopic left lateral sectionectomy videos. Standalone CurConMix+ models achieve up to 42.2% AP_IVT on CholecT45 and 58.4% on LLS48, while CurConMix+-Ens further reaches 43.4% and 59.0%, respectively. Across three matched spatial backbones, MRTT consistently improves action-level segmental coherence on both datasets. The action-triplet-supervised CurConMix representations also improve phase, step, and task recognition over matched high-level supervised baselines, demonstrating effective transfer from fine-grained actions to higher-level workflow understanding. The code and LLS48 dataset will be released to support reproducible research in hierarchical surgical workflow understanding.