Abstract

Instrument-tissue interaction detection, a task aimed at understanding surgical scenes from videos, holds immense importance in constructing computer-assisted surgery systems. Existing methods consist of two stages: instance detection and interaction prediction. This sequential and separate model structure limits both effectiveness and efficiency, making it difficult to deploy on surgical robotic platforms. In this paper, we propose an end-to-end Action-Instance Progressive Learning Network (AIPNet) for the task. The model operates in three steps: action detection, instance detection, and action class refinement. Starting with coarse-scale proposals, the model progressively refines them into coarse-grained actions, which then serve as proposals for instance detection. The action prediction results are further refined using instance features through late fusion. These progressive learning processes improve the performance of the end-to-end model. Additionally, we introduce Dynamic Proposal Generators (DPG) to create dynamic adaptive learnable proposals for each video frame. To address the training challenges of this multi-task model, semantic supervised training is introduced to transfer prior language knowledge, and a training label strategy is proposed to generate unrelated instrument-tissue pair labels for enhanced supervision. Experimental results on PhacoQ and CholecQ datasets show that the proposed method achieves superior accuracy and faster processing speed than state-of-the-art models.