Robotic palpation for in situ tissue biomechanical evaluation is crucial for disease diagnosis, especially in luminal organs. However, acquiring real-time information about the tissue's interaction state and physical characteristics remains a substantial challenge. While commercial surgical robotic systems have integrated tactile feedback, the absence of tactile intelligence and autonomous decision-making limits the surgeon's ability to comprehensively assess tissue mechanics, hindering the efficient detection of abnormalities. Endoscopic optical coherence tomography has emerged as a promising technology for real-time, 3-dimensional visualization of tissue microstructures and subtle lesions in luminal organs. However, it does not address the tactile sensing required for lesion profiling and boundary identification. "To bridge this gap, we develop a hybrid sampling technique that utilizes OCT-based tactile sensing." said the author Wenchao Yue, a researcher at The Chinese University of Hong Kong, "This method combines dual-mode distal B-scan imaging measurement, including circumferential and sliding B-scan modes, to actively locate lesion centers and enable programmable boundary segmentation, enhancing the task-specific computational complexity reduction to 6,249-fold, lowering the center error to 0.032 mm, and improving the shape fitting accuracy to 0.983."
The authors develop a robotic tactile-sensing system that integrates a deployable continuum endoscope with an OCT-based elastographic probe (ElastoSight). Delivered through the endoscope's working channel, the probe supports motorized circumferential B-scans for active palpation and stiffness-center localization, as well as drag-based sliding B-scans for boundary following and shape extraction. The endoscope provides four degrees of freedom—axial rotation, axial translation, and bidirectional bending—covering most intraluminal surfaces. On the imaging side, a spectrometer-based SD-OCT unit is paired with custom rotary and linear modules, achieving ~2.7 µm axial resolution, ~1.05 mm imaging depth, an A-line rate up to 250 kHz, ~40 fps rotational scanning, and 100–400 µm/s linear drag. Once positioned at the target site, tumor assessment proceeds in two stages: (1) circumferential B-scan–based sequential palpation to rapidly estimate the stiffness field and lock onto its centroid; and (2) multi-direction sliding B-scans that pass through the centroid under low-friction continuous contact to detect paired boundary points and reconstruct the lesion's contour.
The core algorithm of the paper is divided into two aspects: active search for the tumor centroid based on Gaussian adaptive sampling and tumor shape segmentation based on the sliding B mode. For centroid localization, the author describes it as finding the global maximum stiffness distribution on the surface of the tissue. It first discretizes the accessible surface into a multi-resolution grid that can be progressively refined—for example, from a coarse 5×5 to a fine 50×50 grid, improving spatial resolution by roughly 10×—and then densifies sampling within suspicious regions to boost accuracy. The stiffness field f(x) is modeled with a Gaussian Process, initialized with a few random observations. Using the current posterior mean and uncertainty, the method iteratively selects the next palpation location and updates the GP with each new measurement, thereby improving exploration efficiency under a limited budget of circumferential B-scan palpations and accelerating convergence to the tumor center at the peak of the stiffness field. For tumor shape segmentation, the author adopted a sliding B-scan strategy based on the centroid. After the centroid is localized in the first stage, the sensor performs low-friction continuous sliding along multiple radial paths that pass through the centroid. When the probe crosses the tumor–soft tissue boundary, the lateral deformation on either side has opposite directions, producing a pair of sharp optical pulses with opposite polarity—a negative spike at the entry boundary point P1 and a positive spike at the exit boundary point P1'-which enables robust boundary detection. Unlike conventional endoscopic OCT, which requires ~6,250 A-lines per circumferential scan, the sliding B-scan records an intensity trace along a single A-line as a function of displacement, dramatically reducing data volume and improving real-time performance. Subsequently, the "single A-line pulse detection" is repeated along multiple directions from the centroid outward, yielding one boundary point per direction; aggregating these points reconstructs the lesion contour.
The article reports quantitative validation on two tracks: centroid localization and shape segmentation. On phantoms with three shapes, multiple active sampling strategies (EVR, EI, UCB, LSE, ILS-UCB, RASEC) were compared. F1 scores rose with iterations and converged to ~0.9 around the 10th iteration. Averaged over three active searches, EVR performed best, achieving F1 = 0.89 and the lowest centroid error of 0.20 mm; EI and UCB reached centroid errors of 0.35 mm and 0.38 mm, respectively, while all other methods exceeded 0.55 mm. For sliding B-scan segmentation, the authors reconstructed contours of circular, rectangular, and horseshoe phantoms using a progressive sector-density scheme (quartered → octant → 12-direction → 16-direction) and evaluated against ground-truth areas (circle 12.57 mm², rectangle 7.79 mm², horseshoe 7.88 mm²). Progressive densification substantially reduced area error and improved accuracy: for the circle, error fell from 1.99 to 0.08 mm² and accuracy rose from 0.842 to 0.994; for the rectangle, from 2.25 to 0.20 mm² and 0.712 to 0.974; and for the horseshoe, from 3.14 to 0.19 mm² and 0.602 to 0.976—supporting the effectiveness of centroid-anchored, multi-direction sliding segmentation. Additionally, experiments visualize the paired optical pulses and deformation contrasts observed as the probe traverses boundaries, illustrating both the feasibility and the intuitive interpretability of boundary detection and localization under the sliding B-scan mode.
This work introduces a novel robotic bimodal tactile tomography that integrates Bayesian optimization with OCT-based sensing to profile intracavitary targets. The proposed ElastoSight sensor offers dual-modal functionality, using circumferential B scans for precise localization of tumor centroids and sliding B scans for accurate boundary segmentation. This technology is promising to enhance surgical perception during minimally invasive procedures, particularly in early-stage tumor detection and margin assessment during oncological interventions. This hybrid approach effectively addresses critical limitations inherent in conventional RMIS by enabling the real-time stiffness mapping and visualization of subsurface microstructures. "Our current research primarily focuses on developing decision-making algorithms, with ongoing efforts to enhance engineering details. We plan to strengthen these components and overall system performance by integrating a surgical robotic system for real-time point cloud registration and alignment. Our future work includes experiments on autonomous searching in phantom models and dynamic in vivo environments with live animals. Further research will expand their framework to include multimodal signal alignment, depth-resolved elastography for layered tissue characterization, and integration with robotic control mechanisms for closed-loop palpation." said Wenchao Yue.
Authors of the paper include Wenchao Yue, Chao Xu, Tao Zhang, Jianing Qiu, Wu Yuan, and Hongliang Ren.
This work was supported in part by the Science, Technology, and Innovation Commission (STIC) of Shenzhen Municipality (SGDX20220530111005039), the Research Grants Council (RGC) of Hong Kong SAR (CRFC4026-21GF, RIF-R4020-22, GRF14203821, GRF14216222, GRF14216022, GRF14203323, GRF14201824, GRF 14204524, NSFC/RGC Joint Research Scheme N_CUHK420/22 GRF14213125, and GRF14203323), key project 2021B1515120035 (B.02.21.00101) of the Regional Joint Fund Project of the Basic and Applied Research Fund of Guangdong Province GDSTC, and the Innovation and Technology Commission (ITC) of Hong Kong SAR (ITS/252/23).
The paper, "Bimodal Tactile Tomography with Bayesian Sequential Palpation for Intracavitary Microstructure Profiling and Segmentation" was published in the journal Cyborg and Bionic Systems on Sept. 2, 2025, at DOI: 10.34133/cbsystems.0348.
