EndoAbS dataset: Endoscopic abdominal stereo image dataset for benchmarking 3D stereo reconstruction algorithms

Area of Science:

• Computer Vision
• Medical Imaging
• Surgical Technology

Background:

• 3D reconstruction algorithms are crucial for augmented reality (AR) in computer-assisted surgery.
• Limited availability of validated endoscopic stereo image datasets hinders algorithm development.
• The EndoAbS dataset addresses this gap by providing a rich resource for algorithm validation.

Purpose of the Study:

• To introduce and describe the EndoAbS dataset, a novel collection of endoscopic stereo images.
• To provide a valuable resource for the development and validation of 3D reconstruction algorithms for surgical AR.
• To enhance the accuracy and reliability of surgical navigation systems.

Main Methods:

• The dataset comprises endoscopic stereo images of phantom abdominal organs.
• A high-accuracy 3D organ surface reference (RF) was generated using a laser scanner.
• Camera calibration parameters and detailed generation methods are included.

Main Results:

• The dataset creation error, specifically camera-laser calibration, was estimated at 0.43 mm.
• A 3D reconstruction algorithm evaluated on EndoAbS achieved accuracy below 2 mm.
• Results align with state-of-the-art performance benchmarks.

Conclusions:

• The EndoAbS dataset expands the availability of open-access surgical stereo image datasets.
• It offers a highly accurate 3D reference and diverse surgical conditions.
• This resource facilitates advancements in surgical augmented reality and 3D reconstruction techniques.