Abstract
Common dense stereo Simultaneous Localization and Mapping (SLAM) approaches in Minimally Invasive Surgery (MIS) often require high-end parallel computational resources for real-time implementation. However, in clinical settings, these resources are frequently allocated to other critical tasks like segmentation and tracking. This paper presents BDIS-SLAM, a lightweight dense stereo SLAM system designed to operate on a single-core CPU while achieving real-time performance (over 30 Hz). By integrating a novel Bayesian Dense Inverse Searching (BDIS) module with the ORB-SLAM2 framework, the system provides a low-cost solution for dense mapping in MIS. Experiments on in-vivo and synthetic datasets demonstrate that BDIS-SLAM maintains high accuracy with minimal computational overhead, making it suitable for surgical robots and AR systems.
Introduction
In Minimally Invasive Surgery (MIS), real-time 3D reconstruction of the surgical site is vital for both surgeons and robotic assistance systems. While SLAM is the standard approach for localization and mapping, "dense" systems—which map every pixel to create a detailed surface—typically rely on powerful GPUs.
In a clinical environment, GPU resources are a luxury. They are often fully utilized by AI-driven tasks such as tissue segmentation or instrument tracking. Furthermore, integrating high-performance GPUs into surgical hardware is challenging due to strict thermal and physical size constraints. BDIS-SLAM addresses this gap by providing a "lightweight" alternative that delivers usable 3D maps using only standard CPU power.
System Architecture and Methodology
BDIS-SLAM is built upon the robust ORB-SLAM2 framework, augmented with a specialized dense mapping module. The system architecture is designed to handle tracking and mapping in parallel threads to ensure smooth performance.
Bayesian Dense Inverse Searching (BDIS)
The core of the dense mapping module is the BDIS algorithm. Unlike standard stereo matching, which can struggle with the dark, specular, or textureless surfaces common in internal anatomy, BDIS employs a Bayesian probability model. It evaluates the confidence of overlapping image patches to filter out unreliable depth predictions. This makes the system significantly faster and more robust than traditional CPU-level algorithms like Lucas-Kanade (LK) tracking, which often fails due to reflections.
Rigid Mosaic Strategy
To maintain a low computational footprint, BDIS-SLAM utilizes a "map-on-keyframe" approach rather than expensive non-rigid fusion. Instead of mathematically warping and fusing every frame into a seamless mesh—a process that is computationally taxing—the system generates a local 3D point cloud for each new keyframe. These clouds are then "mosaicked" into the global map, substituting old data with new data in overlapping regions. This strategy allows the dense mapping thread to run with only approximately 12% additional time compared to the sparse backbone system.
Experimental Evaluation
The system was validated using two primary data sources:
- Hamlyn In-Vivo Datasets: Real-world stereo endoscopy and laparoscopy videos of porcine subjects, covering various resolutions (e.g., 640x480 and 720x288).
- Synthetic Colon Dataset: A 12,000-frame sequence generated in Unity3D, providing perfect ground truth for camera trajectories and 3D shapes to allow for precise error measurement.
Key Results and Discussion
- Real-Time Performance: The system consistently achieved over 30 Hz on a modern single-core CPU for standard 640x480 surgical images, meeting the requirements for real-time surgical use.
- Accuracy: On synthetic data, BDIS-SLAM achieved an average mean error of 0.296 mm and a median error of 0.215 mm. While slightly less precise than GPU-intensive state-of-the-art systems, these results are highly usable for surgical navigation and AR overlays.
- Resource Efficiency: The system maintained a memory footprint of less than 5GB, ensuring compatibility with standard medical workstations without requiring specialized hardware upgrades.
Conclusion and Future Work
BDIS-SLAM demonstrates that dense mapping in surgery does not have to be a resource-heavy task. By prioritizing efficiency through Bayesian filtering and a rigid mosaic strategy, the system provides a viable path for dense SLAM on computationally constrained devices.
Future developments will focus on addressing current limitations, such as the "gaps" or seams visible in the mosaic and the lack of non-rigid deformation modeling. By utilizing the depth variance predicted by the BDIS algorithm, the researchers aim to develop a more sophisticated, yet still lightweight, fusion process that can account for the movement of soft tissues during surgery.
Summary updated