Abstract: Deep learning-based image segmentation is a potent tool for medical image analysis, and the fetal echocardiogram segmentation is the frontier research topic. Due to the imbalance problem in the number of pixels in different heart parts and the blurred image, the existing algorithms have low segmentation accuracy for minority pixels, and the segmentation boundaries are usually inaccurate. To address these issues, this paper proposes a fetal echocardiogram segmentation algorithm that focuses on handling class imbalance problem. By embedding a cost-aware layer, which assigns higher loss weights to misclassifications of minority class pixels, thereby improving the classification accuracy of minority class pixels. Additionally, the global histogram equalization combined with dilated spatial pyramid pooling obtain clear image contours and multi-scale contextual information. In comparison with seven advanced segmentation algorithms on a real fetal heart dataset, the proposed algorithm demonstrates optimal results across multiple evaluation metrics.

Key words: fetal cardiac ultrasound, image segmentation, class imbalance, cost-aware, multi-scale

摘要： 基于深度学习的图像分割技术是处理医学图像的有力工具，胎儿心脏超声图像分割任务更是其中的热点研究内容。由于不同心脏部位像素点数量不平衡以及图像模糊等问题，导致现有算法对少数类像素点的分割准确度较低，分割边界通常也不够精确。为此，提出面向类别不平衡的胎儿心脏超声图像分割算法。该算法通过嵌入代价感知层，为少数类像素点的错误分类分配更高的损失权重，以提升少数类像素点的分类准确性。应用全局直方图均衡化，结合空洞空间金字塔池化，获取清晰的图像轮廓和多尺度上下文信息。在真实胎儿心脏数据集上与7种先进分割算法比较，该算法在多项评价指标上均获得了最优的分割结果。

关键词: 胎儿心脏超声, 图像分割, 类别不平衡, 代价感知, 多尺度