Safe-ICE: AI for ICE Catheter Guidance

• Safe-ICE is an AI-driven framework that guides ICE catheter manipulation by predicting the device tip's position and approach angle in real time.
• It employs a hybrid dataset combining clinical ICE sequences with synthetic augmentation to enhance model robustness across diverse anatomical scenarios.
• The system integrates a pretrained vision transformer with a temporal prediction network to enable real-time, closed-loop robotic catheter control with built-in safety checks.

Safe-ICE is an AI-driven framework for guiding intra-cardiac echocardiography (ICE) catheter manipulation so as to maintain continuous visibility of the therapy device tip during electrophysiology (EP) and structural heart disease (SHD) interventions. In Huh et al., the framework is formulated as a sequential prediction problem in which an AI model estimates the device tip incident angle and passing point within the ICE imaging plane from short ICE frame windows, with the stated aim of ensuring continuous visibility and facilitating robotic ICE catheter control (Huh et al., 8 May 2025).

1. Clinical role and problem formulation

Intra-cardiac Echocardiography plays a critical role in EP and SHD interventions by providing real-time visualization of intracardiac structures. The operational difficulty addressed by Safe-ICE is that maintaining continuous visibility of the therapy device tip remains a challenge due to frequent adjustments required during manual ICE catheter manipulation. The framework therefore targets a specific guidance problem rather than generic ultrasound interpretation: it predicts both where the therapy device tip lies in the 2D ICE plane and how the tip approaches that plane in orientation space (Huh et al., 8 May 2025).

A recurrent misconception is to treat continuous tip visibility as a purely framewise localization task. Safe-ICE is explicitly defined more broadly. Each sample couples a bounding box with an incident-angle representation, and the downstream control objective is not only detection but also conversion of those estimates into ICE-probe pitch and yaw adjustments. This suggests that the method is designed as a perception component for closed-loop assistance rather than as an isolated image-analysis model.

The intended significance is procedural. By accurately predicting both where the tip lies in the 2D ICE plane and its approach angle, the model informs minimal ICE-probe pitch/yaw adjustments so that the tip never leaves the field of view. The supplied description further states that continuous visibility reduces the need for manual re-scanning, shortens procedure time, and lowers exposure to ultrasound energy.

2. Hybrid dataset generation

A defining feature of Safe-ICE is its hybrid dataset generation strategy, which combines clinical ICE sequences with synthetic data augmentation to enhance model robustness (Huh et al., 8 May 2025). Real ICE cine loops were acquired during electrophysiology ablation procedures. A clinical ultrasound scanner and standard ICE catheter imaged the right atrium and interatrial septum in a water-tank setup with black background. Frames were sampled at 25 Hz, cropped around the fan, and resized to $224 \times 224$ px.

The synthetic component was built from isolated imaging of the physical ablation catheter tip. The tip was moved at speeds of 10–20 mm/s against the same water-tank backdrop, and these sequences were automatically segmented to produce “tip patches.” Each patch was then composited onto clinical ICE frames in a temporally coherent way, with spatial jitter $< 2$ px/frame and intensity modulation $\pm 15$ HU, to simulate realistic device motion through anatomies. The supplied description emphasizes that this preserves motion continuity while simulating diverse anatomical scenarios.

The resulting dataset covers multiple anatomical and maneuver classes rather than a single static configuration.

Element	Specification
Total dataset	5,698 ICE-tip image pairs
Split	5,400 train / 48 validation / 250 test
Temporal input	$N = 5$ sequential frames
Targets	Bounding box $B=[x_{\min}, y_{\min}, x_{\max}, y_{\max}]$ and incident angle $A=[a_{\text{entry}}, a_{\text{rot}}]$
Angle coverage	Entry angles approximately $-90^\circ$ to $+90^\circ$; rotation angles $0^\circ$–$360^\circ$
Anatomical scenarios	Right atrium free wall, fossa ovalis, tricuspid annulus, variety of insertion-withdrawal maneuvers

Another common misunderstanding would be to classify the training set as purely synthetic. The source material is explicit that Safe-ICE uses clinical ICE sequences together with synthetic tip overlays, and that the final dataset is intended to ensure comprehensive training coverage.

3. Ground truth, coordinate definition, and target variables

Safe-ICE uses electromagnetic sensor instrumentation to establish precise ground-truth locations (Huh et al., 8 May 2025). A six-degree-of-freedom EM sensor was affixed to the ICE catheter’s shaft to define the ICE frame, and another was affixed to the tip of the therapy catheter (9-Fr). By initializing the ICE frame so that its $< 2$0-axis aligns with the catheter shaft and the $< 2$1-axis aligns with the ultrasound fan centerline, the relative transform

$< 2$2

yields the true 3D orientation of the tip in ICE coordinates.

The two angular targets are defined operationally. Entry angle $< 2$3 is taken from the tilt of the tip’s $< 2$4-axis. Rotational angle $< 2$5 is inferred both from that transform and from the diagonal orientation of the bounding-box overlay. This dual representation links the 3D EM-derived geometry to the 2D image-plane depiction of the therapy device tip.

The per-sample formulation combines temporal and geometric supervision. Each sample consists of a temporal window of five sequential frames plus ground-truth bounding box and incident angle. In that sense, Safe-ICE is neither a purely geometric tracker nor a purely image-only regressor; it is a temporally conditioned estimator with explicit coordinate grounding. This suggests that the framework was designed to stabilize predictions under the frequent viewpoint changes associated with ICE catheter manipulation.

4. Model architecture and optimization

The model architecture integrates a pretrained ultrasound foundation model for feature extraction with a transformer-based temporal prediction network (Huh et al., 8 May 2025). The backbone is a 12-layer Vision Transformer style encoder with patch size 16 px and feature dimension $< 2$6. Its pretraining corpus comprises 37.4 million unlabeled echocardiography frames from multiple views, including parasternal, apical, and subcostal, using a self-supervised contrastive objective identified as EchoApex¹. The backbone outputs per-frame feature vectors $< 2$7.

The main network receives three classes of inputs: a sliding window of image features $< 2$8, the previous timepoint’s predicted bounding box $< 2$9, and the previous timepoint’s predicted angle $\pm 15$0. The latter two are each projected via learned linear layers into $\pm 15$1, and a learnable [CLS] token is prepended. The transformer comprises 8 encoder layers, each with multi-head self-attention with 6 heads, model dimension 768, and feed-forward dimension 3,072. The final [CLS] embedding is linearly mapped to box coordinates $\pm 15$2 and angles $\pm 15$3.

Training uses a sum of mean-squared losses over spatial and angular outputs:

$\pm 15$4

The reporting metrics for angle error are

$\pm 15$5

The training regime was 117 epochs with batch size 6, AdamW optimizer, and initial learning rate $\pm 15$6 with cosine decay. Implementation used PyTorch and training on a single NVIDIA A100 GPU. Real-time inference is reported at approximately 25 Hz, or approximately 40 ms/frame.

5. Quantitative results, robustness, and failure modes

The supplied source material contains two numerical summaries of performance (Huh et al., 8 May 2025). The abstract reports that the method achieves 3.32 degree entry angle error and 12.76 degree rotation angle error. The detailed experimental section reports performance on held-out 250 test cases as mean entry angle error $\pm 15$7, mean rotation angle error $\pm 15$8, and bounding-box IoU $\pm 15$9. Both summaries are part of the provided record and indicate that the method is evaluated in terms of both orientation accuracy and image-plane localization.

Robustness was assessed by varying insertion/withdrawal speed profiles and adding up to 30 dB of simulated speckle noise. Performance degraded gracefully, with entry and rotation errors increasing by 5–8 degrees. Worst-case errors occur at extreme entry angles, specifically $N = 5$0, or when deep occlusions such as trabeculations overlap the tip.

These failure modes are clinically relevant because they correspond to precisely the geometries in which visibility maintenance is most difficult. A plausible implication is that the framework’s value is greatest when used with supervisory checks rather than as a fully unsupervised autonomy module. The source material supports that interpretation through its explicit specification of confidence thresholds, EM redundancy, and watchdog conditions in the deployment workflow.

6. Robotic integration, safety logic, and future scope

Safe-ICE is described as laying the foundation for real-time robotic ICE catheter adjustments, minimizing operator workload while ensuring consistent therapy device visibility (Huh et al., 8 May 2025). The real-time integration pathway is specified as a closed-loop tip tracking and catheter adjustment sequence:

1. Acquire raw ICE frame and preprocess to $N = 5$1.
2. Infer $N = 5$2 with the Safe-ICE network.
3. Convert $N = 5$3 into desired ICE-probe pitch/yaw increment $N = 5$4 and yaw increment $N = 5$5.
4. Send motion commands to robotic manipulator under PID supervision.

The latency budget is stated as $N = 5$6 ms per frame to sustain 20–25 Hz closed-loop control, and the target hardware is any GPU with at least 16 GB memory, such as NVIDIA A100 or RTX 3080Ti. These deployment constraints place the model in a real-time guidance setting rather than an offline analysis workflow.

The safety checks and fallbacks are explicit. A confidence threshold is placed on $N = 5$7; if predicted angle change exceeds $N = 5$8 in one step, automation is suspended and the operator is alerted. Dual monitoring compares EM-sensor-derived $N = 5$9 against the Safe-ICE estimate; if the discrepancy exceeds $B=[x_{\min}, y_{\min}, x_{\max}, y_{\max}]$0, the system reverts to manual ICE manipulation. A watchdog timer holds robot motion if no valid prediction is available within 100 ms. This makes clear that Safe-ICE is framed as a guarded assistive-control component with redundancy and fallback logic.

Future work is described in dataset and anatomical terms. It will focus on expanding clinical datasets to further enhance model generalization, including incorporation of in vivo clinical ICE cases across left-heart structural interventions and additional anatomy classes such as mitral-clip steering, occluder deployment, and transseptal puncture. This suggests a trajectory from the current right-heart and interatrial-septal scenarios toward broader structural-heart intervention coverage.