A Dataset for Depth in Robotic Endoscopy with Dynamic Scenarios (DRENDS)

Area of Science:

• Robotics
• Computer Vision
• Medical Imaging

Background:

• Depth perception is crucial for robotic minimally invasive surgery.
• Existing datasets lack ground-truth depth data in dynamic surgical scenarios.
• Accurate depth estimation is vital for surgical task automation and safety.

Purpose of the Study:

• To introduce DRENDS (Depth in Robotic Endoscopy with Dynamic Scenarios), a novel dataset for robotic surgery depth estimation.
• To provide high-resolution stereo images with ground-truth point clouds for dynamic surgical environments.
• To facilitate research in metric, temporally consistent depth estimation for medical applications.

Main Methods:

• Collected high-resolution stereo image sequences during robotic laparoscopic manipulation of human phantom and ex vivo porcine tissue.
• Acquired ground-truth point clouds and calibration data for each frame.
• Recorded data under varying illumination and anatomical conditions, including tissue manipulation and deformation.

Main Results:

• The DRENDS dataset includes dynamic scenarios with ground-truth depth information.
• Baseline evaluations using state-of-the-art depth estimation models were performed.
• Results highlight challenges and potential for metric, temporally consistent depth estimation in robotic surgery.

Conclusions:

• DRENDS addresses the critical need for dynamic ground-truth data in robotic surgery depth estimation.
• The dataset and open-source code encourage advancements in tissue deformation prediction.
• Public release of DRENDS aims to foster innovation and collaboration in surgical robotics.