ERCP: 3D Planning & Robotic Innovations

• ERCP is a hybrid endoscopic and radiologic procedure that diagnoses and treats pancreatobiliary diseases using both conventional 2D fluoroscopy and advanced 3D reconstruction.
• The procedure overcomes complex ductal anatomies with techniques like contrast-enhanced cholangio-MRI and structured segmentation, enhancing clinical decision making.
• Innovations such as deep learning segmentation and modular soft robotic catheters are improving procedural accuracy, reducing radiation exposure, and shortening intervention times.

Endoscopic retrograde cholangiopancreatography (ERCP) is a hybrid endoscopic and radiologic intervention for the diagnosis and treatment of biliary stone disease, biliary and pancreatic neoplasia, and strictures involving the pancreatobiliary ducts. ERCP’s technical complexity arises from highly variable and often intricate patient-specific ductal anatomy, particularly in cases of hilar cholangiocarcinoma. Initial access involves cannulation of the major duodenal papilla using a side-viewing duodenoscope and guide-wire-loaded catheter under fluoroscopic guidance. Standard imaging modalities (abdominal CT, 2D MRCP) and intraoperative fluoroscopy provide sectional or projectional duct views, whereas three-dimensional (3D) reconstruction with contrast-enhanced cholangio-MRI offers an explicit roadmap for optimized ERCP planning and navigation (Essamlali et al., 2023, Calmé et al., 21 Jan 2026).

1. Clinical Workflow and Procedural Complexity

ERCP commences with endoscopic positioning at the papilla, followed by fluoroscopy-guided guidewire cannulation of the ductal system. Procedural success rates depend heavily on anatomy visualization and manipulation, with adverse event and failure rates reported at 2–9%, particularly in complex anatomy such as malignant hilar strictures or cholangiocarcinoma (Essamlali et al., 2023). Current preoperative planning relies on sectional imaging, which does not directly resolve intricate 3D spatial relationships within the biliary tree. Intraoperative guidance is traditionally limited to 2D fluoroscopic projection, resulting in significant operator dependence, prolonged procedure time, and radiation exposure (Essamlali et al., 2023, Calmé et al., 21 Jan 2026).

2. Three-Dimensional Biliary Tract Reconstruction

The integration of 3D reconstruction from cholangio-MRI data using platforms such as 3D Slicer represents a substantial advancement over conventional 2D imaging. Segmentation and reconstruction workflows involve extracting biliary tree masks using techniques such as thresholding, flood filling, and region growing, each with domain-specific algorithmic and practical trade-offs (Essamlali et al., 2023).

Segmentation Method	Degree of Automation	Mean DSC	Mean HD (mm)	RVD	Key Limitation
Thresholding	Manual	$0.819\pm0.057$	$0.816\pm0.353$	$0.188\pm0.118$	Labor-intensive, poor reproducibility
Flood Filling	Semi-automatic	$0.772\pm0.059$	$1.291\pm0.656$	$0.235\pm0.172$	Time-consuming, not reproducible
Region Growing	Automatic	$0.657\pm0.110$	$1.963\pm1.091$	$0.539\pm0.349$	Anatomical errors, suboptimal fine structure capture

Manual or threshold-based methods provide the highest Dice similarity coefficient (DSC), but at the cost of 20–30 hours per case, extending to 200–300 hours for complex hilar disease. Fully automatic region growing (employing a Sauvola local threshold) substantially reduces manual time but fails to achieve satisfactory anatomical fidelity or overlap, with increased risk of clinical misguidance (Essamlali et al., 2023).

3. Limitations of Conventional Segmentation and the Drive for Deep Learning

The 3D segmentation of bile ducts in ERCP planning is fundamentally constrained by a manual/automation trade-off: high-quality, anatomically faithful segmentations remain labor-intensive and poorly reproducible, while automated pipelines expedite processing yet lack sufficient spatial accuracy. Qualitative evaluations by expert endoscopists indicate failure modes including over-segmentation (false positives, incorporation of non-biliary axial structures), under-segmentation (ductal discontinuities), and false interductal communications (Essamlali et al., 2023). No evaluated method achieved DSC $>$0.82 or HD $<$0.8 mm across all cases. These findings underscore the necessity for alternative approaches, notably deep learning–based frameworks (e.g., U-Net with attention modules, data augmentation), trained on large annotated cholangio-MRI datasets and capable of generalizing to diverse hepatic and biliary pathologies (Essamlali et al., 2023).

4. Advances in Instrumentation: Modular Soft Robotic Catheters

Emerging instrumentation platforms address core deficiencies in conventional ERCP tools. The 1.47 mm outer-diameter modular soft magnetic continuum robot (mSCR) combines compliant navigation with active sensing, therapeutic delivery, and high-resolution onboard imaging (Calmé et al., 21 Jan 2026). The modular architecture permits functionalization with four independently controlled tip modules:

• Anchoring balloon: hard-magnetic silicone structure, pressurization range 0–57 kPa, mean anchoring force $F_{\rm anchor}=3.3\pm0.2$ N
• Compliant gripper: flexure-based L-arm mechanism, coupled to a 250×250 pixel “NanEye” camera
• Onboard imaging and shape sensing: monocular camera and three-core fiber Bragg grating (FBG) for 27 Hz curvature feedback; tip orientation error $<$2°
• Ultrasound-triggered drug release: PLA micro-patterned shell ($20~\mu$m edge, $1$ MHz resonance) with confirmed T${}_{80}$ (80% payload release) of $104\pm22$ s

Integration is achieved through low-pressure injection molding, sequential assembly, and core-clearing. The result is a catheter compatible with a range of endoscopes (duodenoscope, gastroscope, bronchoscope) and capable of advanced interventions within previously inaccessible pancreatic regions (Calmé et al., 21 Jan 2026).

5. Modeling, Control, and Autonomous Navigation

The mSCR navigation and control suite is governed by a combined elastic (Neo-Hookean) and magnetic materials model. Key mathematical constructs include:

• Magnetic torque: $\tau(x) = m \times B(EPM, x)$ with $m$ as remanent dipole moment per unit length
• Magnetic field: $$B(p) = \frac{\mu_0}{4\pi}[3(\vec{p}\cdot \hat{m})\hat{p} - \hat{m}]/|\vec{p}|^3$$
• Maximum tip deflection: $$\theta_{\mathrm{max}} = \frac{3\mu_0 mB r^2 L^3}{EI_x}$$

Shape is reconstructed by integrating FBG strain data into the Frenet–Serret equations. Closed-loop semi-autonomous navigation is executed by a shared-control system running at $\sim27$ Hz. A YOLOv8-based detector identifies colored markers and the papilla landmark in real-time ($0.81$ precision, $0.76$ recall, F1 $0.78$), yielding vector-based angle error $e_\theta$ that is mapped to the desired magnetic deflection and implemented via damped least-squares field control. PID-like corrections and safety interlocks (triggered by FBG orientation deviation) further ensure robust cannulation and tissue protection. Manual operator override is available at all times (Calmé et al., 21 Jan 2026).

6. In Vivo Performance and Clinical Implications

Porcine model experiments document a navigation depth of 75 mm within the pancreatic duct, threefold increase in peak tip gripping force with balloon anchoring, and 11% reduction in insertion time using autonomous/shared control relative to fully operator-led insertion. X-ray fluoroscopy usage is reduced by up to 60% owing to algorithmically stabilized alignment. No papilla trauma or bleeding was observed. The catheter’s fully modular tip permits on-demand drug delivery, stable anchoring, and adaptive manipulation during high-precision navigation; successful cannulation rate was 100% across six trials (Calmé et al., 21 Jan 2026).

7. Prospective Directions

Conventional biliary tract segmentation for ERCP planning is constrained by speed, reproducibility, and anatomical fidelity limits inherent to thresholding, flood filling, and region growing. The integration of deep learning segmentation and semi-autonomous soft robotic catheters represents a disruptive paradigm for endoluminal intervention, enabling real-time, anatomically accurate, and reproducible 3D planning and execution (Essamlali et al., 2023, Calmé et al., 21 Jan 2026). Future research will focus on deep learning frameworks trained on multi-institutional annotated datasets and further clinical translation of multifunctional robotic architectures, with cross-domain implications for other delicate, tortuous, or otherwise inaccessible anatomical regions.