Parameter-efficient framework for surgical action triplet recognition

Surgical action triplet recognition is a clinically significant yet challenging task. It provides surgeons with detailed information about surgical scenarios, thereby facilitating clinical decision-making. However, the high similarity among action triplets presents a formidable obstacle to recognition. To enhance accuracy, prior methods necessitated the utilization of larger models, thereby incurring a considerable computational burden.

We propose a novel framework known as the Lite and Mega Models (LAM). It comprises a CNN-based fully fine-tuned model (LAM-Lite) and a parameter-efficient fine-tuned model based on the foundation model using Transformer architecture (LAM-Mega). Temporal multi-label data augmentation is introduced for extracting robust class-level features.

Our study demonstrates that LAM outperforms prior methods across various parameter scales on the CholecT50 dataset. Using fewer tunable parameters, LAM achieves a mean average precision (mAP) of 42.1%, a 3.6% improvement over the previous state of the art.

Leveraging effective structural design and robust capabilities of the foundational model, our proposed approach successfully strikes a balance between accuracy and computational efficiency. The source code is accessible at https://github.com/Lycus99/LAM.