# SR-Mamba: Effective Surgical Phase Recognition with State Space Model

**Authors:** Rui Cao, Jiangliu Wang, Yun-Hui Liu  
**arXiv:** [2407.08333](https://arxiv.org/abs/2407.08333)  
**Date:** 2024-07-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2407.08333v1/

## Key points

- Surgical phase recognition is crucial for improving efficiency and safety in computer-assisted interventions.
- A key challenge in this field is effectively modeling long-distance temporal relationships in surgical videos.
- SR-Mamba is a novel attention-free model designed to address the complexities of surgical phase recognition, building on the success of Mamba.
- SR-Mamba utilizes a bidirectional Mamba decoder to capture temporal context within lengthy sequences for more accurate phase recognition.
- The model allows for single-step neural network training, streamlining the workflow and enhancing accuracy without separate training steps.
- SR-Mamba achieves state-of-the-art results on benchmark datasets like Cholec80 and CATARACTS Challenge, setting a new standard in surgical video analysis.
- The code implementation for SR-Mamba is openly accessible at https://github.com/rcao-hk/SR-Mamba.

## Summary

In the field of computer-assisted interventions, surgical phase recognition plays a critical role in improving efficiency and safety. One of the key challenges faced in this area is effectively modeling the long-distance temporal relationships present in surgical videos. Building on the success of Mamba, a state space model known for its linear scalability in sequence length, this paper introduces SR-Mamba, a novel attention-free model specifically designed to address the complexities of surgical phase recognition. SR-Mamba utilizes a bidirectional Mamba decoder to capture the temporal context within lengthy sequences, enabling more accurate phase recognition. By optimizing the Mamba decoder efficiently, the model allows for single-step neural network training, eliminating the need for separate training steps seen in previous approaches. This streamlined training process not only simplifies workflow but also enhances accuracy, even when using a lighter spatial feature extractor. The performance of SR-Mamba sets a new standard in surgical video analysis by achieving state-of-the-art results on benchmark datasets such as Cholec80 and CATARACTS Challenge. The code implementation for SR-Mamba is openly accessible at https://github.com/rcao-hk/SR-Mamba. Authors Rui Cao, Jiangliu Wang, and Yun-Hui Liu have made significant contributions to advancing surgical phase recognition through their innovative approach with SR-Mamba.