EndoNet: A Deep Architecture for Recognition Tasks on Laparoscopic Videos

Andru Putra Twinanda, Sherif Shehata, Didier Mutter, Jacques Marescaux, Michel de Mathelin, Nicolas Padoy

Abstract

Surgical workflow recognition is essential for developing "context-aware" operating rooms. This paper introduces EndoNet, a multi-task convolutional neural network (CNN) designed to automatically recognize surgical phases and detect tool presence in laparoscopic videos. By leveraging visual information alone and jointly learning tool and phase cues, EndoNet achieves state-of-the-art performance on cholecystectomy datasets, outperforming traditional methods that rely on handcrafted features.

Introduction

In the quest for the "Operating Room of the Future," computer-assisted intervention (CAI) systems must understand the surgical workflow. Recognizing whether a surgery is in the "Preparation" or "Gallbladder Dissection" phase allows for optimized staff scheduling, automated video indexing, and real-time surgical assistance.

However, laparoscopic video is a challenging data source due to camera movement, lens staining, and smoke. Previous methods relied on handcrafted features or manual tool annotations, which are difficult to scale. EndoNet addresses these challenges by learning features automatically from raw pixels.

The EndoNet workflow involves tool presence detection, fine-tuning, and phase recognition.

Methodology

The core of the proposed approach is EndoNet, a deep learning architecture based on AlexNet, modified for multi-task surgical video analysis.

Multi-Task Learning

EndoNet performs two tasks simultaneously: tool presence detection and phase recognition. It uses a shared feature-extraction backbone. One branch of the network predicts the presence of seven surgical tools. These predictions are concatenated with high-level visual features to feed into a second branch for surgical phase prediction. This joint learning forces the network to develop visual features that are highly discriminative for the overall workflow, as tool usage is a strong indicator of the current phase.

The EndoNet architecture, based on AlexNet, features multi-task heads for tool and phase recognition.

Transfer Learning and Temporal Modeling

To overcome the scarcity of surgical data, the model is pre-trained on ImageNet and then fine-tuned on laparoscopic videos. To handle the temporal nature of surgery and prevent "jittery" phase transitions, a two-level Hierarchical Hidden Markov Model (HHMM) is applied to the CNN outputs, ensuring the predicted sequence follows a logical surgical timeline.

Datasets

The study utilizes Cholec80, a significant dataset of 80 cholecystectomy videos performed by 13 surgeons, labeled for 7 phases and 7 tools at 1 fps. The model was also evaluated on the public EndoVis dataset to test generalization across different hospitals and phase definitions.

Key Results

EndoNet significantly outperformed traditional methods:

  • Tool Presence: Achieved 81.0% Mean Average Precision (mAP), compared to 58.4% for Deformable Part Models (DPM).
  • Phase Recognition: Reached 92.0% accuracy in offline mode and 81.7% in real-time online mode on Cholec80.
  • Generalization: The learned features remained robust even when tested on datasets with different phase definitions, proving that visual information alone is sufficient for high-accuracy workflow recognition.

Conclusion and Future Work

EndoNet demonstrates that deep learning can extract sufficient information directly from video feeds to understand surgical workflows without external sensors. While the current pipeline uses a separate HHMM for temporal modeling, future work could integrate Recurrent Neural Networks (RNNs) or LSTMs for a fully end-to-end system. Additionally, expanding dataset diversity will further improve the model's robustness to non-linear transitions and surgical complications.

Summary updated

Abstract

Surgical workflow recognition has numerous potential medical applications, such as the automatic indexing of surgical video databases and the optimization of real-time operating room scheduling, among others. As a result, phase recognition has been studied in the context of several kinds of surgeries, such as cataract, neurological, and laparoscopic surgeries. In the literature, two types of features are typically used to perform this task: visual features and tool usage signals. However, the visual features used are mostly handcrafted. Furthermore, the tool usage signals are usually collected via a manual annotation process or by using additional equipment. In this paper, we propose a novel method for phase recognition that uses a convolutional neural network (CNN) to automatically learn features from cholecystectomy videos and that relies uniquely on visual information. In previous studies, it has been shown that the tool signals can provide valuable information in performing the phase recognition task. Thus, we present a novel CNN architecture, called EndoNet, that is designed to carry out the phase recognition and tool presence detection tasks in a multi-task manner. To the best of our knowledge, this is the first work proposing to use a CNN for multiple recognition tasks on laparoscopic videos. Extensive experimental comparisons to other methods show that EndoNet yields state-of-the-art results for both tasks.
BibTeX
@article{twinanda2017endonet,
  author = {Andru Putra Twinanda and Sherif Shehata and Didier Mutter and Jacques Marescaux and Michel de Mathelin and Nicolas Padoy},
  title = {{EndoNet: A Deep Architecture for Recognition Tasks on Laparoscopic Videos}},
  journal = {IEEE Trans. Medical Imaging},
  year = {2017},
  url = {https://openreview.net/forum?id=hADbtPKwcr},
}
Figure