From Sparse to Precise: A Practical Editing Approach for Intracardiac Echocardiography Segmentation

Ahmed H. Shahin, Yan Zhuang, Noha El-ZehiryLecture Notes in Computer Science 2023 · DOI · OpenAlex

Abstract

Accurate and safe catheter ablation procedures for patients with atrial fibrillation require precise segmentation of cardiac structures in Intracardiac Echocardiography (ICE) imaging. Prior studies have suggested methods that employ 3D geometry information from the ICE transducer to create a sparse ICE volume by placing 2D frames in a 3D grid, enabling training of 3D segmentation models. However, the resulting 3D masks from these models can be inaccurate and may lead to serious clinical complications due to the sparse sampling in ICE data, frames misalignment, and cardiac motion. To address this issue, we propose an interactive editing framework that allows users to edit segmentation output by drawing scribbles on a 2D frame. The user interaction is mapped to the 3D grid and utilized to execute an editing step that modifies the segmentation in the vicinity of the interaction while preserving the previous segmentation away from the interaction. Furthermore, our framework accommodates multiple edits to the segmentation output in a sequential manner without compromising previous edits. This paper presents a novel loss function and a novel evaluation metric specifically designed for editing. Results from cross-validation and testing indicate that our proposed loss function outperforms standard losses and training strategies in terms of segmentation quality and following user input. Additionally, we show quantitatively and qualitatively that subsequent edits do not compromise previous edits when using our method, as opposed to standard segmentation losses. Overall, our approach enhances the accuracy of the segmentation while avoiding undesired changes away from user interactions and without compromising the quality of previously edited regions, leading to better patient outcomes.

Introduction

Atrial Fibrillation (AFib) affects millions of patients worldwide, often requiring catheter ablation to treat problematic heart tissue. For these procedures to be safe and effective, clinicians require a precise 3D map of the Left Atrium (LA). Intracardiac Echocardiography (ICE) is a preferred imaging modality as it involves inserting an ultrasound probe directly into the heart. However, generating dense 3D volumes from the sparse 2D slices produced by ICE is challenging due to cardiac motion and frame misalignment. While automated 3D segmentation models provide an initial estimate, they are rarely perfect and require manual refinement.

Existing interactive tools often struggle with "catastrophic forgetting" during the editing process; attempting to fix one area can inadvertently degrade accurate segments elsewhere. This paper introduces a novel interactive framework that treats editing as a localized refinement task, allowing clinicians to correct errors with simple 2D scribbles that propagate accurately into 3D space.

Methods

The proposed framework utilizes a 3D U-Net architecture designed for localized refinement. The model takes three primary inputs: the raw ICE volume, the initial (imperfect) segmentation mask, and the user's interaction, which is represented as a 3D Gaussian heatmap derived from 2D scribbles.

Editing-Specific Loss Function

The core innovation is a weighted loss function designed to balance correction and preservation:

  • $\mathcal{L}_{edit}$: This term forces the model to align with the ground truth in the immediate vicinity of the user's interaction.
  • $\mathcal{L}_{preserve}$: This term penalizes deviations from the initial segmentation in regions far from the interaction site.

By using 3D Gaussian heatmaps to weight these terms, the system ensures a smooth transition between edited and preserved regions, preventing jagged edges or disconnected components in the final 3D mask.

Overview of the interactive editing framework: (a) initial 2D contours, (b) initial 3D segmentation volume, (c) refined 3D segmentation after editing, and (d) the resulting 2D segmentation masks.

Results and Discussion

The framework was evaluated on a clinical dataset of 712 patients. The proposed "Editing Loss" was compared against standard Cross-Entropy (CE) and Dice losses, as well as state-of-the-art iterative methods like InterCNN.

Segmentation Accuracy

On the independent test set (107 patients), the proposed method achieved an overall error of 0.844 mm, significantly outperforming CE Loss (1.389 mm) and Dice Loss (1.039 mm). More importantly, the preservation error—measuring unintended changes in non-edited regions—was only 0.184 mm for the proposed method, compared to over 1.0 mm for standard losses.

Sequential Integrity

A critical requirement for clinical tools is the ability to handle multiple sequential corrections. Qualitative results demonstrate that while baseline methods begin to "drift" and lose the shape of earlier corrections after several edits, the proposed loss function maintains the integrity of the entire edit history. This allows for cumulative improvements in segmentation quality without the risk of undoing previous work.

Detailed view of the final 2D segmentation masks (green) overlaid on ICE frames, showing high precision following interactive refinement.

Conclusion

This research presents a practical and efficient approach to refining 3D ICE segmentations. By employing a localized, editing-specific loss function, the framework provides clinicians with a trustworthy tool that respects previous corrections while enabling rapid, precise updates. With an inference time of approximately 20 milliseconds, the system is well-suited for real-time clinical integration. Future work will investigate adaptive influence zones to handle larger errors and improve robustness to noisy user input.

Summary updated

BibTeX
@inproceedings{shahin2023from,
  author = {Ahmed H. Shahin and Yan Zhuang and Noha El-Zehiry},
  title = {{From Sparse to Precise: A Practical Editing Approach for Intracardiac Echocardiography Segmentation}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2023},
  doi = {10.1007/978-3-031-43901-8\_73},
}

Loading related content...

Figure