目的 染色体核型分析从细胞分裂中期图像中分离和分类染色体，是遗传疾病诊断广泛采用的方法，其中形态多样的重叠染色体簇的分割，依赖于准确的边界等细节特征。为此，本文融合目标的上下文信息，构建了一种两阶段的重叠染色体分割模型SC-Net。方法 首先，在语义分割基线模型U-Net++中增加混合池化模块捕获重叠染色体的局部上下文信息，在解码器网络中并联上下文融合模块和上下文先验辅助分支，增强通道和空间上的全局上下文信息。其次，利用已标注样本的类别先验信息生成真实亲和矩阵，加入训练过程以有效区分重叠染色体图像中易混淆的空间信息。最后，通过染色体实例重建算法对重叠与非重叠区域的元素迭代进行配对，拼接形成单条染色体。结果 在公开的ChromSeg数据集上进行实验，结果表明SC-Net分割出的重叠染色体区域交并比值为83.5%，比现有最佳方法提升2.7%。结论 本文构建的重叠染色体分割模型通过融合上下文信息能够更有效地解决形态多样的重叠染色体簇的分割问题，相比现有的方法得到更精细和准确的结果。

Objective Chromosome karyotype analysis, which separates and categorizes chromosomes in mid-cell division images, is a widely used method for the diagnosis of genetic diseases, in which overlapping chromosome segmentation is one of the key steps in chromosome karyotype analysis. Based on the image analysis of overlapping chromosomes, the morphologically diverse chromosome clusters rely on more detailed features such as accurate boundaries in the segmentation process, in addition to obtaining the basic contour, texture and semantic information. For this reason, in this paper, a two-stage overlapping chromosome segmentation model SC-Net is constructed by fusing the contextual information of the target to improve the segmentation performance of the network. Method First, the model SC-UNet++ adds Hybrid Pooling Module (HPM) to the baseline model U-Net++ for semantic segmentation to capture the local context information of overlapping chromosomes, and complements the detailed features of chromosomes such as color, thickness, and stripes based on the superposition operation of empty space pyramid pooling (ASPP) and stripe pooling . The Context Fusion Module (CFM) is connected in parallel in the decoder network, i.e., the channel correlation of the input features is extracted by the Efficient Channel Attention Module (ECAM), and the features obtained by multiplying the output with the input are subsequently fed to the HPM and the SAM which explore the correlation of the region around the pixel to obtain the local context as well as extract the global context through global pooling operation, respectively. And Context Prior Auxiliary Branch (CPAB) is introduced after CFM to enhance the global context information on channel and space. Second, the category a priori information of labeled training samples is used to generate the true affinity matrix, which serves as additional supervisory information during training and effectively distinguishes confusing spatial features in overlapping chromosome images. Finally, the elements of overlapping and non-overlapping regions are iteratively paired by the chromosome instance reconstruction algorithm to splice and form a single chromosome. In this paper, experiments are analyzed based on the ChromSeg dataset, and the hardware resources used are a desktop server with 32G of RAM, a 3.3GHz Intel Xeon CPU, and an NVIDIA RTX 3070 GPU. The model was implemented based on the semantic segmentation toolkit MMSegmentation with version 0.30.0, and was run under the operating system Ubuntu 18.04, with the deep learning framework PyTorch 1.10.0. The relevant hyperparameter settings and initialization methods for network training as well as the loss function selection strategy are described in the following sections. Result SC-Net fully extracts and utilizes the contextual and category prior information of overlapping chromosome images and shows good performance in segmentation scenarios with different numbers of overlapping chromosomes. The effect of each improvement on the performance of the algorithm is investigated through ablation experiments, where different combinations of CFM, HPM, CPAB and segmentation loss are designed on the baseline model U-Net++, and the results prove that SC-UNet++ outperforms the compared models in all evaluation metrics. This proves the effectiveness of the method proposed in this paper, i.e., SC-UNet++ obtains better performance in segmenting overlapping chromosomes. Through comparative experiments, the performance of SC-Net proposed in this paper on the ChromSeg dataset outperforms several models compared in all metrics, and the model achieves an overlapping chromosome region intersection and merger ratio of 83.5%, and the overall accuracy obtained after chromosome instances are reconstructed is 92.3%, which is higher than the best same two-stage ChromSeg segmentation methods by 2.7% and 1.8%. SC-Net outperforms these models mainly due to its ability to extract contextual information and category relevance of the target, which enables the model to have a better understanding of overlapping regions. Conclusion The overlapping chromosome segmentation model constructed in this paper can solve the problem of segmenting morphologically diverse overlapping chromosome clusters more effectively by fusing the contextual information, and can obtain finer and more accurate results compared with the existing methods.