A Semi-Supervised Approach with Error Reflection for Echocardiography Segmentation

TL;DR

This work tackles semi-supervised echocardiography segmentation under challenging image quality by introducing an error reflection strategy that improves pseudo-label reliability. It combines a reconstruction reflection step and a guidance correction step within a Mean Teacher framework and couples it with a multi-scale puzzle-patch augmentation to reduce labeled/unlabeled distribution shift. Key contributions include the two-step error reflection mechanism, dynamic error-guided region refinement, and the puzzle-patch data augmentation, all applied without changing the U-Net backbone. On CAMUS and a private clinical dataset, the method achieves superior performance at very low labeled data, demonstrating improved robustness and practicality for automatic echocardiography analysis.

Abstract

Segmenting internal structure from echocardiography is essential for the diagnosis and treatment of various heart diseases. Semi-supervised learning shows its ability in alleviating annotations scarcity. While existing semi-supervised methods have been successful in image segmentation across various medical imaging modalities, few have attempted to design methods specifically addressing the challenges posed by the poor contrast, blurred edge details and noise of echocardiography. These characteristics pose challenges to the generation of high-quality pseudo-labels in semi-supervised segmentation based on Mean Teacher. Inspired by human reflection on erroneous practices, we devise an error reflection strategy for echocardiography semi-supervised segmentation architecture. The process triggers the model to reflect on inaccuracies in unlabeled image segmentation, thereby enhancing the robustness of pseudo-label generation. Specifically, the strategy is divided into two steps. The first step is called reconstruction reflection. The network is tasked with reconstructing authentic proxy images from the semantic masks of unlabeled images and their auxiliary sketches, while maximizing the structural similarity between the original inputs and the proxies. The second step is called guidance correction. Reconstruction error maps decouple unreliable segmentation regions. Then, reliable data that are more likely to occur near high-density areas are leveraged to guide the optimization of unreliable data potentially located around decision boundaries. Additionally, we introduce an effective data augmentation strategy, termed as multi-scale mixing up strategy, to minimize the empirical distribution gap between labeled and unlabeled images and perceive diverse scales of cardiac anatomical structures. Extensive experiments demonstrate the competitiveness of the proposed method.

Figures (4)

• Figure 1: Comparison between echocardiography and cardiac MRI. (a) and (b) present a cardiac MR image and an echocardiographic image, respectively. Visually, echocardiography exhibits poor contrast, blurred edge details and more noise compared to cardiac MRI. (c) depicts their spectra, revealing fewer high-frequency components in echocardiography.
• Figure 2: The pipeline of the proposed approach. (a) Error reflection strategy, including the reconstruction reflection step and the guidance correction step. The lines in $\mathbf{X}^u_{ct}$, $\mathbf{X}^{pl}_{ct}$, and $\mathbf{X}^{mix}_{ct}$ in the figure are pseudo-colored, with red lines derived from the semantic masks of pseudo-labels and green lines from the auxiliary sketches of unlabeled images. (b) Multi-scale mixing up (data augmentation) strategy. The patches marked red in the figure are from unlabeled images, while those marked green are from labeled images. Note: (a) All student networks (green ones) are the same network with shared parameters. (b) The figure shows only the case where N=3.
• Figure 3: Visualization results of comparative experiments between our method and others on the CAMUS and private clinical dataset with 1% and 5% labeled ratios.
• Figure 4: Visualizing the quality of pseudo-label generation under varying conditions in error reflection strategy ablation studies. (Refer to Table \ref{['ablation_reflection']} for the explanation of All, -, S1 and S2)