Intraoperative Optical Coherence Tomography (iOCT)

Updated 8 June 2026

iOCT is an advanced imaging technique that provides real-time, high-resolution cross-sectional views of tissue microstructure and instrument-tissue interactions during surgery.
It employs spectral-domain and full-field OCT systems with high-speed B-scan acquisition and GPU-accelerated processing for immediate image reconstruction.
iOCT is pivotal in ophthalmology, neurosurgery, and robotic interventions, aiding in margin assessment, instrument localization, and precise surgical guidance.

Intraoperative Optical Coherence Tomography (iOCT) is an advanced imaging modality that delivers real-time, high-resolution tomographic information during surgery, enabling precise visualization of tissue microstructure, margin status, and instrument-tissue interactions. Initially developed for ophthalmic microsurgery, iOCT now encompasses broad intraoperative applications in neurosurgery, oncology, and reconstructive procedures, offering micron-level axial and lateral resolution with acquisition rates that support real-time guidance, feedback control, and autonomous robotic interventions.

1. Physical Principles, System Architectures, and Imaging Modalities

OCT operates by measuring interferometric backscatter of low-coherence broadband or swept-source light. The two fundamental clinical implementations are spectral-domain OCT (SD-OCT), typically operating at central wavelengths 840–1350 nm with A-scan rates 20–200 kHz, and full-field OCT (FF-OCT), which employs spatially incoherent illumination and Linnik or Mirau interferometers.

Key system parameters include:

Axial resolution: Determined by source coherence length, typically 1–10 μm (e.g., δ_z ≈ 1–1.5 μm for FF-OCT with λ=800 nm, Δλ=100–150 nm, n=1.35) (Assayag et al., 2013, Assayag et al., 2012).
Lateral resolution: NA-limited, order 1–30 μm, e.g., 25 μm (wide-field) to 8 μm (high-res) in spinal tumor imaging (He et al., 2024), 1–1.5 μm in FF-OCT (Assayag et al., 2013, Assayag et al., 2012).
Imaging depth: 150–300 μm in dense tissue, 1–2.5 mm in transparent tissues.

Intraoperative adaptations include high-speed line-by-line B-scan acquisition, focus tracking to counter tissue motion (e.g., via electrically tunable lenses) (He et al., 2024), and parallelized acquisition/processing for real-time display or instrument guidance.

FF-OCT, used for ex vivo neurosurgical and oncological margin assessment, provides en face histology-like images over mm–cm fields of view with 1 μm³ isotropic voxels (Assayag et al., 2012, Assayag et al., 2013).

2. Data Processing Pipelines and Instrument Visualization

B-scans (cross-sectional OCT images) are processed via a sequence of signal digitization, k-space linearization, windowing, FFT, and intensity extraction. Real-time implementations integrate these pipelines using high-speed acquisition hardware and GPU-accelerated image reconstruction (He et al., 2024).

Metallic surgical instruments in B-scans exhibit hyper-reflective arcs/lines with pronounced shadowing. These optical signatures are exploited for:

Instrument localization: Multi-step ellipse fitting using candidate extraction, tissue-surface fitting (RANSAC/poly), morphological filtering, ellipse fitting with constrained DOF (fixed radius/center), and nonlinear boundary refinement (Weiss et al., 2018).
Pose estimation: Extended Kalman Filtering (EKF) fuses sequential B-scan-derived geometric features (ellipse parameters) into 5DOF instrument pose estimation at >180 Hz with sub-millimeter precision (Weiss et al., 2018).
Volumetric and tool-aligned slicing: Virtual B-scan extraction through 3D OCT volumes by aligning with the needle shaft direction detected via neural network, yielding sparse “intelligent” cross-sections that substantially reduce computational load (Dehghani et al., 2023).

Real-time throughput is central: e.g., native C++ can process 1k × 1k B-scans in ≤5.4 ms (∼186 FPS) (Weiss et al., 2018); newer pipelines running complex sonification and segmentation maintain <30 ms end-to-end latency per frame (Vargas et al., 14 May 2026, Mach et al., 2024).

3. Segmentation Algorithms and Statistical Modeling

Robust, real-time segmentation of tissues and instruments in iOCT is enabled by statistical modeling and machine/deep learning architectures:

Speckle statistics: Patch-wise Gamma distribution modeling (MLE) of intensity speckle yields device- and patient-invariant feature space for segmenting layer boundaries and surgical tools. Binary maps from learned class-specific parameter ranges feed into light-weight deep networks with residual U-Net topology (Mach et al., 2024).
Deep CNNs/U-Net architectures: Raw B-scans or engineered feature maps are used as input for U-Net–like networks with skip connections and multi-task outputs for layer, tool, and tissue probability (Kim et al., 2023, Mach et al., 2024, Xiong et al., 2 Feb 2026).
Topology-aware constraints: Losses penalizing anatomical inconsistencies (e.g., enforcing epithelium above Descemet’s membrane), smoothness priors, and star-shape constraints improve segmentation continuity especially under low SNR and shadowing (Xiong et al., 2 Feb 2026).
Real-time performance: Topology-aware UNeXt-based pipelines for corneal DALK guideline M-mode segmentation reach 80–87 Hz end-to-end rates while maintaining Dice >0.98 (Xiong et al., 2 Feb 2026). Patch-based models using only Gamma parameters as input maintain robust generalization across unseen ex vivo datasets (Dice 0.90) (Mach et al., 2024).

4. Autonomous Robotic Interventions and Real-Time Guidance

iOCT is a core enabler of autonomous and semi-autonomous robotic surgery. Landmark implementations include:

Instrument/needle tracking: 5DOF geometric modeling supports direct injection guidance, with pose drift <0.2 mm even with failed detections (Weiss et al., 2018).
Robot–OCT registration: Direct mapping between segmented OCT instrument axes and robot frames allows real-time hand–eye calibration and trajectory planning without iterative optimization (Dehghani et al., 2023, Kim et al., 2023).
Trajectory generation: Model predictive control (MPC) under remote center of motion (RCM) constraints fuses real-time B-scan depth feedback with kinematic safety constraints (e.g., avoiding subretinal layer breach) to achieve sub-30 μm accuracy in porcine subretinal injections (Kim et al., 2023).
Deformation-aware feedback: Real-time B⁵-scans (5 parallel B-scans at ∼9 Hz update) enable dynamic virtual target layer tracking, with robotic controllers adjusting needle insertion depth in response to tissue deformation, achieving a 90% success rate in bleb generation compared to 35% for point targeting (Arikan et al., 2024).
Multimodal integration: Fusing iOCT with microscope RGB data via cross-attention and temporal recurrent models enhances instrument localization (mAP50 95.79%) and reduces close-range distance MAE to 33 μm from 284 μm (OPMI only) (Rohrmoser et al., 26 Mar 2026).

5. Clinical Applications and Validation Studies

iOCT clinical integration extends across domains:

Ophthalmic microsurgery: Enables subretinal injection, membrane peeling, lamellar keratoplasty, and depth-critical maneuvers with real-time layer and instrument feedback (Weiss et al., 2018, Kim et al., 2023, Arikan et al., 2024).
Neurosurgery and oncology: FF-OCT provides en face digital histology (1 μm³ voxels, 1 cm² in 5–7 min) for intraoperative brain tumor/epileptogenic margin assessment, and breast cancer margin analysis with 97%/90% sensitivity and 74–77% specificity (Assayag et al., 2013, Assayag et al., 2012).
Spinal tumor resection: FACT-ROCT achieves artifact-free, high-resolution in situ imaging of spinal cord tumors, enabling tumor boundary delineation, microvascular mapping (OCTA), and grading based on attenuation coefficient heterogeneity (σ(μ_t) threshold yields >90% accuracy) (He et al., 2024).
Margin assessment: Deep neural network-based margin assessment in breast lumpectomy reduces EER from ~12% (prior art) to 5% (function-norm–regularized CNNs), with inference <2 s per B-scan (Triki et al., 2017).
Perceptual augmentation: Physics-based sonification translates B-scan–derived segmentations and deformation estimates into auditory cues, substantially improving temporal event identification (e.g., bleb onset, ILM/RPE contact) and reducing surgeon cognitive load (Vargas et al., 14 May 2026).

6. Limitations, Performance, and Future Directions

Several limitations persist:

Depth penetration: Imaging depth is limited (<200–300 μm in dense tissues), restricting assessment to tissue surfaces or superficial margins (Assayag et al., 2013, Assayag et al., 2012).
Shadowing/artifact handling: Instrument-induced shadowing and motion artifacts require inpainting, topology-regularization, and robust frame rejection strategies (Weiss et al., 2018, Xiong et al., 2 Feb 2026).
Resolution/throughput: Direct in situ volumetric imaging in vivo is challenging at depths/resolutions >1 mm³; high-speed focus tracking and MHz-line-rate sources are active areas (He et al., 2024, Arikan et al., 2024).
Registration and robustness: Out-of-plane instrument movements and physiological tissue dynamics necessitate continual algorithmic refinement for 6DOF tracking and real-world clinical translation (Arikan et al., 2024, Dehghani et al., 2023).

Future directions include GPU-accelerated real-time volume rendering, probe miniaturization for truly in situ imaging, self-supervised pretraining for improved scene understanding, adaptive multi-modal fusion, and clinical studies to quantify the impact on surgeon workload and patient outcomes (He et al., 2024, Rohrmoser et al., 26 Mar 2026).

7. Summary Table: Representative iOCT System Performance

Modality/Domain	Resolution (Axial/Lateral)	Real-Time Rate	Application	Key Performance
FF-OCT (Brain/Breast)	1–1.5 μm / 1–1.5 μm	5–7 min/cm²	Margin assessment, neuro/onco	≥90% sensitivity(Assayag et al., 2012, Assayag et al., 2013)
SD-OCT Ophthalmology	5–10 μm / 20–30 μm	≥180 Hz	Needle tracking, retina surgery	Sub-mm, <0.5° ang. var(Weiss et al., 2018)
FACT-ROCT (Spine)	10 μm / 8 μm (high-res)	280 Hz/B-scan	In situ spinal tumor imaging	Tumor grading acc. >90%(He et al., 2024)
Robotic autopilot	5–15 μm / 15 μm	9 Hz (B⁵-scan)	Deformation-aware injection	Bleb success 90%(Arikan et al., 2024)
OPMI + iOCT Fusion	≈10 μm-depth, OPMI lateral	44 FPS	Tool/tissue distance estimation	MAE 33 μm (<1 mm)(Rohrmoser et al., 26 Mar 2026)