VISION-ICE: Video-based Interpretation and Spatial Identification of Arrhythmia Origins via Neural Networks in Intracardiac Echocardiography

This paper proposes VISION-ICE, an AI framework utilizing 3D Convolutional Neural Networks to analyze intracardiac echocardiography videos for automated, real-time localization of arrhythmia origins, achieving 66.2% accuracy and demonstrating potential to streamline electrophysiological procedures.

The Gist

Imagine your heart is a bustling city with a complex electrical grid. Sometimes, this grid gets a glitch, causing the city to panic and beat chaotically. This is called an arrhythmia. To fix it, doctors need to find exactly where the glitch started (the "origin") and zap it with a tiny electrical shock (ablation) to reset the system.

Currently, finding that glitch is like trying to locate a specific faulty wire in a dark, crowded room using only a flashlight and a lot of guesswork. It takes a long time, requires a highly skilled expert, and uses extra tools (like CT scans) that are slow and expensive.

This paper introduces a new "smart assistant" called VISION-ICE that helps doctors find the glitch much faster and more accurately. Here's how it works, broken down into simple concepts:

1. The "Inside-Out" Camera (ICE)

Doctors usually look at the heart from the outside (like looking at a house from the street). But for this procedure, they use a special camera on a thin tube that goes inside the heart. This is called Intracardiac Echocardiography (ICE).

• The Analogy: Think of it like a drone flying inside a factory to inspect the machinery, rather than just looking at the factory from the outside. It gives a super-clear, real-time video of the heart's inner walls.

2. The Problem: Too Much Data, Not Enough Time

The drone (the ICE camera) records hours of video. A human doctor has to watch this video while also monitoring other machines, trying to spot tiny movements that indicate where the electrical glitch is coming from. It's like trying to find a specific needle in a haystack while the haystack is moving.

3. The Solution: The "Super-Observer" AI

The researchers built a computer brain (an AI) that acts as a super-observer. They taught this AI to watch the ICE videos and answer a simple question:

• Is the heart beating normally?
• Is the glitch coming from the Left side?
• Is the glitch coming from the Right side?

They didn't just show the AI one picture; they showed it movies (videos). This is crucial because the heart is a moving machine. The AI uses a special type of neural network (a 3D Convolutional Neural Network) that understands both space (where things are) and time (how they move).

• The Analogy: If a regular photo AI is like a security guard looking at a still photo of a room, this AI is like a security guard watching a live security feed, noticing how a person walks, not just what they look like.

4. How They Trained the AI

To teach the AI, they used data from 39 patients.

• The Setup: They took videos of the heart in four different "angles" (like taking photos of a statue from the front, back, left, and right).
• The Game: They played a game where they would "fake" an arrhythmia by sending a small electrical pulse from a specific spot (Left or Right). The AI had to watch the video and guess: "Did the pulse come from the Left or the Right?"
• The Practice: They let the AI practice thousands of times, making mistakes and learning from them, just like a student studying for a test.

5. The Results: Better Than Random Guessing

When they tested the AI on patients it had never seen before:

• Random Guessing: If you just guessed "Left," "Right," or "Normal" with your eyes closed, you'd be right 33% of the time.
• The AI: The AI got it right 66% of the time.
• Why this matters: While 66% isn't perfect yet, it's double the success rate of a random guess. In medicine, that's a huge leap. It means the AI can act as a reliable "second opinion" to help the doctor make a decision faster.

6. Making it Trustworthy (The "Why" Factor)

Doctors are skeptical of "black box" computers that give answers without explaining why. To fix this, the researchers added a feature called Grad-CAM.

• The Analogy: Imagine the AI is a detective. Instead of just saying "The criminal is in the kitchen," it highlights the kitchen on the video with a glowing red circle, showing the doctor exactly what it saw (e.g., a specific muscle twitching) that led to its conclusion. This builds trust.

The Bottom Line

This paper is like the blueprint for a GPS for heart doctors.
Right now, finding an arrhythmia is like driving a car with a foggy windshield and no map. This new AI system clears the fog and puts a "You are here" arrow on the screen, pointing directly to the problem area.

Why is this exciting?

• Faster Procedures: Doctors won't have to spend hours searching; the AI points them in the right direction immediately.
• Less Stress: It reduces the mental load on the doctor.
• Better Care: Faster, more accurate fixes mean patients recover quicker and have fewer complications.

The researchers admit the system isn't perfect yet (it needs more data to get even smarter), but they have proven that AI can "see" heart problems in video footage that humans might miss, paving the way for a future where heart surgery is faster, safer, and more precise.

Technical Summary

1. Problem Statement

Current methods for localizing arrhythmia origins (the source of abnormal heart rhythms) rely heavily on high-density mapping and pre-operative CT/MRI scans, which are time-consuming, resource-intensive, and often require significant operator expertise. While Intracardiac Echocardiography (ICE) is a routine, high-resolution imaging modality used during electrophysiology (EP) procedures to guide interventions, it has not been effectively leveraged by Artificial Intelligence (AI) for automated arrhythmia localization.

The specific challenge addressed is the lack of AI frameworks capable of analyzing ICE video data in real-time to distinguish between:

1. Normal Sinus Rhythm (NSR)
2. Left-sided arrhythmia origins (simulated via distal coronary sinus pacing)
3. Right-sided arrhythmia origins (simulated via proximal coronary sinus pacing)

2. Methodology

A. Dataset Construction

• Source: Data was collected from 39 patients undergoing elective EP studies at Baylor St. Luke's Hospital.
• Protocol: Patients were imaged using a Soundstar ultrasound catheter and a GE Vivid S70 machine.
• Views: Four standard anatomical views were captured: Tricuspid Valve (TV), Mitral Valve (MV), Left Pulmonary Vein (LPV), and Crista Terminalis (CT).
• Stimulation: Baseline recordings were taken during sinus rhythm. Subsequently, pacing was induced at 600 ms cycle lengths from the distal (electrodes 1-2) and proximal (electrodes 9-10) ends of the coronary sinus catheter to simulate left and right-sided arrhythmia origins.
• Volume: The dataset contains thousands of heartbeats across 39 unique patients.

B. Preprocessing Pipeline

The pipeline ensures data integrity and standardization:

1. Masking: Irrelevant background regions are masked out.
2. Normalization: Pixel intensities are normalized to $[0, 1]$.
3. Temporal Segmentation: Recordings are decomposed into individual cardiac cycles (heartbeats) guided by expert annotations of ECG signals (marking PR segments and heartbeat ends).
4. Spatial Cropping & Resizing: Non-informative regions are removed, and frames are resized to $553 \times 756$.
5. Standardization: Sequences are standardized to 32 frames. Longer sequences are truncated; shorter ones are padded by replicating the last frame.
6. Augmentation: Applied only to the training set to combat data scarcity. Techniques include random brightness/contrast adjustments, random frame dropping, and adding Gaussian noise.

C. Model Architecture

• Task: Formulated as a 3-class video classification problem (NSR, Distal Pacing, Proximal Pacing).
• Backbone: A pre-trained 3D ResNet-18 (torchvision r3d_18).
• Customization:
  • The original input stem was replaced with a custom single-channel adapter (grayscale input) using a $9 \times 7 \times 7$ kernel.
  • Followed by batch normalization and 3D dropout ($p=0.1$).
  • The final layers include dropout ($p=0.2$) and a fully connected layer for 3-class logits.
• Training:
  • Optimizer: AdamW (LR $10^{-5}$, weight decay $10^{-3}$).
  • Loss: Class-weighted cross-entropy.
  • Strategy: Mixed precision (AMP) on GPU, gradient clipping, and early stopping based on validation accuracy.

D. Evaluation Strategy

To prevent data leakage and ensure robust generalization, the study employed a strict 10-fold patient-level cross-validation:

• Splitting: In each fold, 4 patients were assigned to the test set, 4 to validation, and 31 to training. This ensures no patient data appears in both training and testing sets within the same fold.
• Hierarchical Prediction:
  1. Sample-level: Prediction on individual heartbeats.
  2. Clip-level: Aggregation of heartbeat predictions within a video clip via majority voting.
  3. Patient-level (Cross-View): Aggregation of predictions across all four anatomical views (TV, MV, LPV, CT) via majority voting to produce a final patient diagnosis.

3. Key Results

• Performance: The model achieved a mean accuracy of 66.2% on the test set (patient-level, cross-view majority voting) across the 10 folds.
• Baseline Comparison: This significantly outperforms the random baseline of 33.3% (1/3 chance for 3 classes).
• View-Specific Performance: The Mitral Valve (MV) view showed the highest individual accuracy (77.09% validation, 65.05% test), while the Crista Terminalis (CT) view performed lowest.
• Interpretability: 3D Grad-CAM visualizations confirmed that the model focuses on physiologically relevant structures (e.g., interatrial septum, mitral annulus, crista terminalis) rather than noise, validating that the model learns meaningful spatiotemporal features.

4. Key Contributions

1. First AI Framework for ICE-based Arrhythmia Localization: Introduces a novel deep learning approach specifically tailored to identifying arrhythmia origins using ICE video data, a previously unexplored application domain.
2. Curated Dataset: Provides a carefully annotated dataset of 39 patients with multi-view ICE videos and corresponding pacing labels, serving as a resource for future EP research.
3. Adapted Architecture: Demonstrates how 3D CNNs can be effectively adapted to the unique spatiotemporal characteristics of ICE video data to achieve high-precision classification.
4. Clinical Feasibility: Proves that AI can assist in reducing procedural time and improving the targeting of cardiac ablation by providing real-time, objective guidance on arrhythmia sources.

5. Significance and Future Work

• Clinical Impact: This work has the potential to transform electrophysiology by enabling faster, more targeted interventions. It could be particularly valuable in centers with limited EP expertise, offering expert-level decision support.
• Limitations: The study is limited by a small cohort size (39 patients) and high inter-patient variability, which caused performance fluctuations across different cross-validation folds.
• Future Directions:
  • Expanding the dataset to improve model robustness and generalizability.
  • Exploring advanced architectures like Swin3D Transformers.
  • Integrating domain adaptation techniques to handle different echocardiographic modalities.
  • Further refining interpretability to build greater clinical trust.

In conclusion, VISION-ICE establishes a foundational step toward AI-guided electrophysiology, demonstrating that deep learning can extract actionable diagnostic insights from routine ICE videos to localize arrhythmia origins automatically.