CalcSeg: Confidence-aware 3D Latent Context Curriculum Learning For Myocardial Scar Segmentation From Single-Stack LGE-CMRs
CalcSeg employs confidence-aware semi-supervised curriculum learning with 3D latent context to improve myocardial scar segmentation from single-stack LGE-CMR, achieving a Dice of 0.677.
Key Findings
Methodology
CalcSeg integrates a transformer-based encoder with a multi-scale residual decoder, capturing inter-slice dependencies via slice-wise self-attention. It introduces a dynamic confidence scoring function based on Dice similarity, scar burden error, and Monte Carlo Dropout uncertainty, enabling adaptive curriculum training. The model progressively expands from high-confidence to challenging samples, optimizing stability and robustness. This approach effectively models 3D spatial relationships from 2D slices, significantly enhancing segmentation accuracy in low-contrast and diffuse scar cases.
Key Results
- On multi-center clinical datasets, CalcSeg achieved an average Dice coefficient of 0.677, outperforming five state-of-the-art models such as TransUNet and ScarNet by 4-6%. The model showed particular strength in difficult cases, reducing scar error to 38.88%. Performance on low-confidence samples improved by approximately 10%, demonstrating robustness.
- In ablation studies, removing either slice-wise self-attention or curriculum learning decreased Dice scores by 5-8%, confirming their importance. The combined approach yielded the best results, especially in challenging scenarios involving diffuse or small scars.
- Uncertainty metrics decreased across curriculum stages, confirming effective sample difficulty regulation. The model maintained stable performance with high generalization across diverse datasets.
Significance
This work advances automated myocardial scar segmentation by effectively leveraging 3D spatial context from 2D images and adaptively focusing on difficult cases. It addresses longstanding challenges in low-contrast, diffuse, and small scar detection, providing a robust tool for clinical diagnosis and prognosis. The integration of confidence-aware curriculum learning offers a new paradigm for handling heterogeneous and uncertain medical data, with potential to improve diagnostic accuracy and reduce manual workload.
Technical Contribution
The paper introduces a novel semi-supervised curriculum learning framework guided by a confidence score derived from clinically relevant metrics. It innovatively combines transformer-based 3D context modeling with slice-wise self-attention, enabling effective spatial dependency capture from sparse 2D slices. The adaptive sample expansion mechanism improves training stability and robustness, setting a new benchmark in myocardial scar segmentation. The approach generalizes well across multi-center datasets, demonstrating scalability and clinical relevance.
Novelty
This is the first work to incorporate confidence-aware curriculum learning explicitly into myocardial scar segmentation from single-stack LGE-CMR. It uniquely models 3D spatial relationships using slice-wise self-attention within a semi-supervised framework, addressing the limitations of prior 2D-only or multi-sequence methods. The adaptive difficulty assessment based on prediction error and uncertainty is a key innovation, enabling targeted learning in challenging scenarios.
Limitations
- The reliance on a pre-trained transformer encoder increases computational complexity, limiting real-time deployment on resource-constrained devices.
- Parameter tuning for confidence thresholds and weighting factors remains manual and may affect robustness across different datasets.
- Extreme cases with very poor image quality or minimal scar tissue still pose challenges, requiring further robustness enhancements.
Future Work
Future directions include integrating multi-sequence CMR data for richer spatial and tissue information, developing self-adaptive confidence thresholds, and exploring boundary uncertainty modeling. Additionally, efforts will focus on reducing computational costs and deploying the model in real-world clinical workflows, aiming for real-time performance and broader applicability.
AI Executive Summary
Myocardial scar detection in cardiac MRI is crucial for diagnosing and predicting outcomes in heart disease patients. However, existing deep learning models often struggle with low contrast, diffuse scars, and small lesion regions, especially when limited to single-stack LGE-CMR images. Traditional 2D models lack the capacity to fully exploit the 3D spatial structure of the heart, leading to inconsistent and inaccurate segmentation results.
To address these challenges, this study introduces CalcSeg, a novel framework that combines confidence-aware semi-supervised curriculum learning with a 3D latent context modeling approach. The core innovation lies in dynamically assessing sample difficulty based on clinically meaningful metrics—Dice similarity, scar burden error, and model uncertainty—allowing the training process to progressively focus on more challenging cases. This adaptive curriculum strategy enhances the model’s robustness and generalization, particularly in complex scenarios.
Simultaneously, CalcSeg employs a transformer-based encoder with slice-wise self-attention, capturing inter-slice dependencies and enforcing anatomical consistency. This design effectively models the 3D structure from sparse 2D slices, overcoming the limitations of traditional methods that process slices independently. The integration of these techniques results in a highly accurate and robust segmentation model.
Experimental results on multi-center datasets demonstrate that CalcSeg achieves an average Dice coefficient of 0.677, surpassing existing state-of-the-art models by 4-6%. It performs especially well on difficult cases, reducing scar volume error to 38.88%. Ablation studies confirm that both the curriculum learning and self-attention mechanisms contribute significantly to performance gains.
This work represents a meaningful step forward in automated cardiac MRI analysis, offering a practical tool for clinical use. It enhances the reliability of myocardial scar detection, which is vital for risk stratification and treatment planning. Future research will focus on multi-sequence integration, boundary uncertainty modeling, and deployment in real-time clinical environments, aiming to further improve accuracy and usability.
Deep Analysis
Background
心肌瘢痕检测在心血管疾病诊断中具有重要意义，早期识别有助于预后评估。近年来，深度学习模型如U-Net、AttentionU-Net、TransUNet等在心脏影像分割中取得突破，但多为二维切片级方法，难以充分利用三维空间信息。多序列、多模态融合技术逐步兴起，但在低对比、弥散和小瘢痕区域仍存在性能瓶颈。近年来，课程学习、生成模型和自注意力机制被引入，旨在提升模型鲁棒性和空间表达能力，但多数学者未能充分结合样本难度自适应调控和3D空间建模。
Core Problem
单堆叠LGE-CMR在临床中广泛应用，但其空间信息有限，导致心肌瘢痕自动分割面临低对比、弥散和微小区域难以准确识别的问题。现有方法多依赖多序列或多模态数据，增加采集成本，且在复杂病例中表现不佳。如何在单堆叠2D图像中有效捕获3D空间关系，提升模型鲁棒性，成为亟待解决的难题。此外，缺乏针对不同病例难度的自适应训练策略，限制了模型在临床复杂场景中的应用。
Innovation
本研究提出结合置信感知的半监督课程学习和潜在空间的跨切片自注意力机制，创新性地解决了上述难题。具体包括：
1）引入动态置信评分函数，基于预测误差和不确定性自动调节样本难度，逐步扩展训练集，增强模型对复杂病例的适应性；
2）设计Transformer编码器与切片自注意力机制，融合多尺度特征，捕获跨切片空间关系，提升空间一致性；
3）采用多中心临床数据验证模型的泛化能力，显著优于传统2D模型和多序列融合方法。
Methodology
- �� 输入：单堆叠LGE-CMR图像，经过预训练Transformer编码器提取2D特征。
- �� 3D潜在空间：利用自注意力机制在编码特征中建模跨切片关系，形成空间一致的潜在表示。
- �� 置信评分：结合Dice系数、瘢痕负荷误差和Monte Carlo Dropout不确定性，动态评估样本难度。
- �� 课程学习：根据置信评分逐阶段扩展训练样本，从易到难，优化模型稳定性。
- �� 损失函数：采用多尺度残差解码器，结合焦点损失和Dice损失，增强微小瘢痕的检测能力。
- �� 训练：多中心数据集上进行端到端优化，动态调整样本难度阈值，确保模型逐步适应复杂场景。
Experiments
使用四个临床中心的LGE-CMR数据集，总计976例，划分训练、验证、测试集。对比TransUNet、AttentionUNet、UNETR、ScarNet等模型，采用Dice系数和瘢痕误差作为评估指标。参数包括Transformer编码器预训练权重、置信评分阈值、学习率1e-4、Dropout 0.3。通过消融实验验证课程学习和自注意力机制的贡献，特别关注难例的性能提升。
Results
在多中心数据上，CalcSeg平均Dice达0.677，优于所有对比模型，提升幅度达4-6%。在低置信样本中，Dice达0.644，较未用课程学习模型提升10%。难例误差降低至38.88%，模型对弥散和微小瘢痕表现出更强鲁棒性。消融实验显示，两项创新技术结合效果最佳，整体性能提升明显。
Applications
该模型可作为临床心肌瘢痕自动检测的辅助工具，适用于心血管疾病风险评估、手术规划等场景。只需单堆叠LGE-CMR图像，无需多序列配准，便于临床快速部署。未来可结合多模态信息，进一步提升诊断准确性。
Limitations & Outlook
模型依赖预训练Transformer，计算资源需求高，难以在设备有限环境中部署。置信评分参数需调节，可能影响模型稳定性。极端低质量图像和极少标注样本仍会影响性能，未来需优化鲁棒性和自适应能力。
Plain Language Accessible to non-experts
想象你在一家工厂里，工人们要把不同的零件拼装成一台机器。每个工人只看到一部分零件（单个切片），但他们需要知道这些零件是怎么组合成完整机器的（3D结构）。传统的方法就像只让工人看一块零件，难以理解整体结构，导致拼装不准确。CalcSeg就像给工人们提供了一种聪明的工具，不仅能帮他们理解每个零件的细节，还能根据零件的复杂程度，逐步指导他们拼装，从简单到复杂。它还用一种特殊的“眼睛”——自注意力机制，帮助工人们看到不同零件之间的关系，确保拼装的机器既完整又合理。这种方法让工厂的生产效率大大提高，也能拼出更精确、更可靠的机器。
ELI14 Explained like you're 14
想象你在玩拼图游戏，但有时候拼图上的图片很模糊或者碎片很小，难以拼出完整的图案。传统的拼图助手只能帮你拼一块块的碎片，但不能帮你看整体。CalcSeg就像一个聪明的拼图助手，它不仅能帮你拼出每一块碎片，还能根据碎片的难度，告诉你什么时候需要更仔细地看。它还会用一种特别的“眼睛”观察所有碎片之间的关系，确保拼出来的图案既完整又漂亮。这样，即使拼图很难，也能帮你拼出一幅清晰的画面。这个助手在医学里用来帮医生找到心脏上的疤痕，既快又准，帮助病人得到更好的治疗。
Abstract
Myocardial scar segmentation from single-stack late gadolinium-enhanced cardiac magnetic resonance (LGE-CMR) imaging has been a longstanding and clinically important challenge, particularly in the presence of low tissue contrast, diffuse, and small scar regions. These challenges are further intensified by the limited availability of 3D spatial context. This paper presents CalcSeg, a Confidence-aware latent context curriculum learning framework that leverages fused 3D feature representations from single-stack 2D LGE-CMR images for robust scar segmentation. Specifically, we introduce a dynamic semi-supervised curriculum learning strategy that progressively expands training from easier to more challenging scar cases using a learned confidence-aware scoring function. Such a function integrates errors in the predicted scar maps with quantified epistemic uncertainty and scar burden estimation to automatically assess sample difficulty without requiring manual labels. To compensate for the limited spatial context in single-stack acquisitions, we then develop a latent slice-wise self-attention to capture inter-slice dependencies and infer 3D spatial representations from sparse 2D inputs. We evaluate CalcSeg on multi-center clinical LGE-CMR datasets and benchmark against existing scar segmentation networks. Experimental results show that CalcSeg consistently outperforms all competing methods, particularly with substantial improvements on clinically challenging cases. Our code is released on Github.