Task 1 Surgical workflow recognition

Workflow analysis is fundamental to the scene understanding in surgery and will help to recognize current surgical phase and also provide high-level information to the other two tasks. In the dataset, the hysterectomy procedure is divided into 7 phases and each frame is annotated with a phase label:

Table 1. The total duration and average duration (hh:mm:ss) of the full-length videos in the dataset, and proportion of each phase in each video (P1-P7, %)

Figure 1.​ The ​order information of laparoscopic hysterectomy and sample frames of each phase (P1-P7)

Task 2 Laparoscope motion prediction

To promote learning-based automatic FoV control, this sub-dataset is built with selected clips that corresponds the laparoscope visual feedback to its motion mode.  In the dataset, 300 clips are carefully selected from Phase 2-4 of the aforementioned 21 videos and each clip lasts for 10 seconds. These clips contain typical laparoscope motions and the motion time T is set as the fifth second. Seven types of motion modes are defined, including one Static mode and six non-static mode: Up, Down, Left, Right, Zoom-in, and Zoom-out.

Table 2. Type and number of the clips with laparoscope motion annotation

Figure 2. From top to bottom: video examples of typical laparoscopic motions Up, Left, and Zoom-in.

Figure 2. From left to right: video examples of typical laparoscopic motions Up, Left, and Zoom-in.

Task 3 Instrument and key anatomy segmentation

For each clip in Task 2, six frames are sampled at 1fps, i.e., frames at time t=0,1,2,3,4,5, and annotated with pixel-wise segmentation, which provides ground truth for Task 3 and also support Task 1 and 2 as additional modality for advanced scene understanding.  In the dataset, four types of instruments and one key anatomy is annotated in the dataset:

Table 3. Total number of presences in clips and fram​es, and total number of annotations of each class in the segmentation dataset. m and s denote the manipulator and shaft of the instrument, respectively

The annotations are performed with an open annotation tool called LabelMe available online, following the settings of instrument part segmentation, i.e., the shaft and manipulator of each instrument are annotated separately. 

Figure 3. Pixel-wise annoatation for instrument and key anatomy segmentation.

Download and citation

[Data download ] 
- Please kindly fill the request form if you want to gain access to the dataset. We will send you a link to download the dataset. 
- IMPORTANT : The released dataset is provided only for academic research, NOT for commercial use. 
 
[Citation] 
- If you use this dataset, please cite: 

@InProceedings{wang2022autolaparo,
    title = {AutoLaparo: A New Dataset of Integrated Multi-tasks for Image-guided Surgical Automation in Laparoscopic Hysterectomy},
    author = {Wang, Ziyi and Lu, Bo and Long, Yonghao and Zhong, Fangxun and Cheung, Tak-Hong and Dou, Qi and Liu, Yunhui},
    booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
    pages = {486--496},
    year = {2022},
    organization = {Springer}
}

For further question about the dataset, please contact ziyiwangx@gmail.com.