融合上下文增强与结构感知损失的二维心脏磁共振图像分割

A Method for 2D Cardiac MRI Segmentation Integrating Context Enhancement and Structure-Aware Loss

摘要: 针对心脏磁共振图像中结构形态差异大、组织对比度低引发的右心室拓扑碎裂与心肌边界模糊问题，本文提出一种融合深层选择性上下文增强与结构感知损失的二维心脏分割方法。该方法以nnU-Net为基础框架，在编码器阶段采用分层优化策略：深层嵌入轻量级上下文增强模块，利用多尺度空洞卷积和通道注意力机制来增强深层语义建模；浅层保留原高分辨率特征提取模块，以兼顾全局结构感知与精细边界的表达。在损失函数设计中，引入Laplacian算子的类别感知窄带边界曲率损失，以增强模型对器官边界结构的学习。此外，总损失函数采用基于Sigmoid函数的课程学习调度策略来渐进式增强边界结构约束力，从而平衡基础区域分割学习与局部边界结构优化。实验表明，本文方法在ACDC数据集上对RV、MYO、LV的Dice分别达到0.9078、0.9064、0.9485，平均Dice达0.9209。相比于基线，本文在边界敏感的RV和MYO表现更优。

Abstract: To address the challenges of right ventricular topological fragmentation and obscure myocardial boundaries—which are primarily induced by pronounced structural morphological variations and low tissue contrast in cardiac magnetic resonance images—this paper proposes a novel two-dimensional cardiac segmentation method that synergistically integrates deep selective context enhancement with a structure-aware loss function. Building upon the foundational architecture of the nnU-Net, the proposed framework adopts a hierarchical optimization strategy within the encoder phase. Specifically, a lightweight context enhancement module is embedded into the deeper layers, leveraging multi-scale dilated convolutions and channel attention mechanisms to fortify deep semantic modeling. Concurrently, the original high-resolution feature extraction modules are preserved in the shallower layers, thereby achieving an optimal balance between global structural perception and the precise delineation of fine-grained boundaries. Regarding the loss function design, we introduce a class-aware narrow-band boundary curvature loss predicated on the Laplacian operator to amplify the model’s capacity for learning complex organ boundary structures. Furthermore, the overall loss objective employs a curriculum learning scheduling strategy governed by a Sigmoid function. This strategy progressively intensifies the structural constraints on boundaries, effectively orchestrating a balance between fundamental regional segmentation learning and localized boundary optimization. Comprehensive experiments demonstrate that the proposed method achieves Dice Similarity Coefficients of 0.9078, 0.9064, and 0.9485 for the right ventricle (RV), myocardium (MYO), and left ventricle (LV), respectively, on the ACDC dataset, culminating in an average DSC of 0.9209. Compared to the baseline model, our approach exhibits superior performance, particularly in segmenting the boundary-sensitive RV and MYO regions.