Memory-based Ensemble Learning in CMR Semantic Segmentation

2025-02-13

AI Summary

In cardiac cine MRI ventricular segmentation, existing methods suffer from significantly degraded accuracy at end slices, compromising the reliability of clinical functional metrics. To address this, we propose a novel streaming ensemble framework integrating spatial continuity modeling and uncertainty-driven adaptation. We introduce the End-slice Coefficient (EC) to quantitatively assess segmentation quality specifically at end slices. Leveraging voxel-wise segmentation variance as a global uncertainty map, we formulate it as a memory signal to dynamically weight ensemble classifiers for patient-specific optimization. The framework further incorporates Dice loss with joint 3D/2D training. Evaluated on the ACDC and M&Ms benchmarks, our method achieves near-state-of-the-art overall Dice scores while substantially outperforming all prior approaches in end-slice segmentation accuracy. Consequently, key clinical indices—particularly ejection fraction—exhibit markedly improved robustness and clinical trustworthiness.

Technology Category

• Computer Vision: Segmentation
• Machine Learning: Ensemble Methods
• Reasoning under Uncertainty: Uncertainty Representations

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• User Modeling, Personalization and Recommendation: Metrics for user behavior and evaluating success

Abstract

Existing models typically segment either the entire 3D frame or 2D slices independently to derive clinical functional metrics from ventricular segmentation in cardiac cine sequences. While performing well overall, they struggle at the end slices. To address this, we leverage spatial continuity to extract global uncertainty from segmentation variance and use it as memory in our ensemble learning method, Streaming, for classifier weighting, balancing overall and end-slice performance. Additionally, we introduce the End Coefficient (EC) to quantify end-slice accuracy. Experiments on ACDC and M&Ms datasets show that our framework achieves near-state-of-the-art Dice Similarity Coefficient (DSC) and outperforms all models on end-slice performance, improving patient-specific segmentation accuracy.

Problem

• Improves end-slice segmentation accuracy
• Leverages spatial continuity for uncertainty extraction
• Introduces End Coefficient for quantification

Innovation

• Leverages spatial continuity for segmentation
• Uses global uncertainty as memory in ensemble learning
• Introduces End Coefficient to quantify accuracy