Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MuST: Multi-Scale Transformers for Surgical Phase Recognition

Alejandra Pérez, Santiago Rodríguez, Nicolás Ayobi, Nicolás Aparicio, Eugénie Dessevres, Pablo Arbeláez

TL;DR

MuST tackles surgical phase recognition by enabling multi-scale temporal reasoning through a two-stage Transformer framework. It introduces MTFE to build a temporal pyramid of sequences around a keyframe, fused via a Multi-Temporal Attention Module into a multi-term embedding, followed by a Temporal Consistency Module that operates on long sequences to enforce coherence. The model relies on a shared MViT backbone and achieves state-of-the-art results on HeiChole, GraSP, and MISAW in both online and offline settings, outperforming prior methods that rely on fixed temporal windows. This work demonstrates the value of explicit multi-scale temporal modeling and long-range coherence for reliable surgical workflow understanding and provides open-source code to support reproducibility.

Abstract

Phase recognition in surgical videos is crucial for enhancing computer-aided surgical systems as it enables automated understanding of sequential procedural stages. Existing methods often rely on fixed temporal windows for video analysis to identify dynamic surgical phases. Thus, they struggle to simultaneously capture short-, mid-, and long-term information necessary to fully understand complex surgical procedures. To address these issues, we propose Multi-Scale Transformers for Surgical Phase Recognition (MuST), a novel Transformer-based approach that combines a Multi-Term Frame encoder with a Temporal Consistency Module to capture information across multiple temporal scales of a surgical video. Our Multi-Term Frame Encoder computes interdependencies across a hierarchy of temporal scales by sampling sequences at increasing strides around the frame of interest. Furthermore, we employ a long-term Transformer encoder over the frame embeddings to further enhance long-term reasoning. MuST achieves higher performance than previous state-of-the-art methods on three different public benchmarks.

MuST: Multi-Scale Transformers for Surgical Phase Recognition

TL;DR

MuST tackles surgical phase recognition by enabling multi-scale temporal reasoning through a two-stage Transformer framework. It introduces MTFE to build a temporal pyramid of sequences around a keyframe, fused via a Multi-Temporal Attention Module into a multi-term embedding, followed by a Temporal Consistency Module that operates on long sequences to enforce coherence. The model relies on a shared MViT backbone and achieves state-of-the-art results on HeiChole, GraSP, and MISAW in both online and offline settings, outperforming prior methods that rely on fixed temporal windows. This work demonstrates the value of explicit multi-scale temporal modeling and long-range coherence for reliable surgical workflow understanding and provides open-source code to support reproducibility.

Abstract

Phase recognition in surgical videos is crucial for enhancing computer-aided surgical systems as it enables automated understanding of sequential procedural stages. Existing methods often rely on fixed temporal windows for video analysis to identify dynamic surgical phases. Thus, they struggle to simultaneously capture short-, mid-, and long-term information necessary to fully understand complex surgical procedures. To address these issues, we propose Multi-Scale Transformers for Surgical Phase Recognition (MuST), a novel Transformer-based approach that combines a Multi-Term Frame encoder with a Temporal Consistency Module to capture information across multiple temporal scales of a surgical video. Our Multi-Term Frame Encoder computes interdependencies across a hierarchy of temporal scales by sampling sequences at increasing strides around the frame of interest. Furthermore, we employ a long-term Transformer encoder over the frame embeddings to further enhance long-term reasoning. MuST achieves higher performance than previous state-of-the-art methods on three different public benchmarks.
Paper Structure (9 sections, 6 equations, 3 figures, 2 tables)

This paper contains 9 sections, 6 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Experiments and Results
  4. Experimental Setup
  5. Experimental Validation
  6. Ablation Experiments
  7. Conclusions
  8. Acknowledgments.
  9. Disclosure of Interests.

Figures (3)

  • Figure 1: MuST employs a Multi-Term Frame Encoder to generate rich embeddings containing short- and mid-term dependencies for a long-term sequence of $F'$ frames. The Temporal Consistency Module introduces long-term analysis by processing relationships among frame embeddings to enable coherence in the predictions.
  • Figure 2: MuST's Multi-Term Frame Encoder utilizes MViT to capture embeddings of a temporal sequence sampled at multiple rates $(x_{i})$ with the keyframe placed at the middle position for offline setups and at the last position for online setups. Each sequence embedding $(l_{i})$, along with their corresponding class token $(cls_{i})$, goes through the Multi-Temporal Attention Module with the rest of the sequence embeddings of the pyramid as temporal context. A final rich multi-term embedding of the keyframe is generated by concatenating class tokens from each scale and processing them through a multi-layer perceptron (MLP).
  • Figure 3: Qualitative results. Qualitative comparisons with state-of-the-art methods on one representative video from each dataset.