TL;DR
This paper introduces a novel robotic palpation system that combines optical coherence tomography (OCT) with tactile sensing to create a "bimodal tactile tomography." This system uses a Bayesian approach to efficiently locate and segment abnormalities within luminal organs, offering a significant improvement over current methods that lack sophisticated tactile intelligence.
What Problem Does This Paper Address?
Diagnosing diseases, especially in luminal organs, often relies on understanding the biomechanical properties of tissues. While robotic surgery offers precision, current systems lack the advanced tactile feedback needed for surgeons to accurately assess tissue mechanics and detect subtle abnormalities like early-stage tumors. Existing technologies like OCT provide great visualization but don't offer tactile information, and current tactile sensors often operate as single modalities, limiting their effectiveness. This paper aims to bridge this gap by developing a system that integrates tactile sensing with OCT for more comprehensive and intelligent tissue assessment during minimally invasive surgery.
Methods
The core of this work is a novel robotic bimodal palpation technique utilizing an OCT-based tactile sensor called ElastoSight. This sensor employs two distinct scanning modes:
- Circumferential Rotary Scanning (Volumetric Mode): This mode is used for actively locating the center of potential lesions. It leverages Bayesian sequential palpation, specifically a Gaussian adaptive sampling strategy, to efficiently search for regions of high stiffness. This method intelligently selects sampling points to reduce the number of probes needed, making the process more efficient.
- Sliding B-scan Mode (Drag-and-Slide Scanning): Once a lesion's center is identified, this mode is used for precise boundary segmentation. By sliding the sensor across the tissue surface, it captures subtle changes in optical signals that correspond to tissue stiffness gradients, allowing for accurate contouring of abnormalities.
The system integrates these modes to achieve both localization and segmentation, aiming to overcome the limitations of purely visual or single-modal tactile sensing.
Datasets
The paper primarily uses phantom models for evaluation. These phantoms are designed to mimic tissue properties, featuring hard inclusions (simulating tumors or lumps) embedded within softer materials. The paper also mentions experiments with various fruits (onion, cherry tomatoes, grape) to test the system's ability to recognize subcutaneous anatomy and estimate contact forces in a more complex, real-world-like scenario. No specific named datasets were used for training; rather, the system's performance was evaluated through these controlled experimental setups.
Key Insights and Ideas
- Bimodal Sensing is Key: The combination of OCT's visualization capabilities with tactile sensing for biomechanical properties is crucial for comprehensive tissue assessment. Relying on just one modality is insufficient for accurate diagnosis and lesion identification.
- Bayesian Optimization for Efficiency: Using Bayesian sequential palpation significantly reduces the number of probes required to locate a lesion's center, making the process much faster and more practical for surgical applications.
- Sliding B-scan for Precision Segmentation: The sliding B-scan mode offers a non-abrasive and effective way to map the boundaries of hard inclusions, overcoming limitations of previous sliding tactile sensing methods.
- OCT-based Tactile Sensing: The ElastoSight sensor's ability to derive tactile information (stiffness) from OCT signals is a significant advancement, enabling simultaneous imaging and force sensing.
Key Results
- Lesion Localization: The system achieved tumor localization within 30 iterations, with high F₁ scores over 0.976 and a centroid error below 0.032 mm in tumor phantom models.
- Segmentation Accuracy: Using the sliding B-scan mode, the system achieved accurate segmentation of hard tissue inclusions with a precision rate of 0.983 and an area error below 0.25 mm².
- Computational Efficiency: The technique demonstrated a significant computational complexity reduction (6,249-fold) and improved shape fitting accuracy to 0.983.
- Fruit Palpation: Experiments with fruit phantoms showed the system's ability to simultaneously estimate contact force and visualize internal structures, with force readings showing a 5% error compared to a reference sensor.
Limitations and Future Work
- Limitations:
- The current work primarily relies on phantom models and fruit phantoms. While promising, real-world clinical validation in human patients is still needed.
- The paper acknowledges that the calibration process for tactile sensing has inherent limitations in acquiring ground truth stiffness.
- While the sliding process is designed to be non-abrasive, the risk of tissue abrasion in previous sliding tactile sensing methods is mentioned as a concern that this work aims to mitigate.
- Future Work:
- The authors plan to integrate the system with a surgical robotic system for real-time point cloud registration and alignment.
- Experiments are planned for autonomous searching in dynamic in vivo environments with live animals.
- Future research will also focus on multimodal signal alignment, depth-resolved elastography for layered tissue characterization, and integration with robotic control mechanisms for closed-loop palpation.
Summary updated