基于多尺度残差注意力网络的重叠染色体分割模型

Overlapping Chromosome Segmentation Model Based on Multi-Scale Residual Attention Network

摘要: 在医学诊断中，染色体核型分析是检测由染色体数量和结构异常引起的遗传疾病的重要手段。然而由于染色体的非刚性特质，在不同图像中染色体经常发生随机卷曲和重叠，因此重叠染色体分割是染色体核型分析的关键环节。针对染色体图像重叠区域大小不一，多尺度特征无法有效自适应提取的问题，本文提出了一种基于多尺度残差注意力的重叠染色体分割模型，采用残差注意力模块代替U-Net中原有的标准卷积，自适应地进行通道间的特征重标定，提高模型的全局感知能力；同时提出了一个多尺度密集金字塔空洞卷积模块，扩大特征感受野，实现多尺度特征融合，进一步提高分割精度；并设计了一种残差路径式的跳跃连接方式，实现上下文对应尺度的特征复现，解决特征拼接时产生的语义差距问题，更精确地恢复高分辨率图像。实验结果表明，分割重叠区域的IoU指标达到了98.58%，证明了本文分割方法的有效性。

Abstract: In medical diagnosis, chromosome karyotype analysis is an important means to detect genetic dis-eases caused by chromosome number and structure abnormalities. However, due to the non-rigid nature of chromosomes, chromosomes often randomly curl and overlap in different images, so overlapping chromosome segmentation is the key link of chromosome karyotype analysis. In view of the problem that the overlapping regions of chromosome images are different in size and mul-ti-scale features cannot be effectively and adaptively extracted, this paper proposes an overlapping chromosome segmentation model based on multi-scale residual attention. The residual attention module is used to replace the original standard convolution in U-Net, and the feature recalibration between channels is carried out adaptively to improve the global perception ability of the model; At the same time, a multi-scale dense pyramid atrous convolution module is proposed to expand the feature receptive field, realize multi-scale feature fusion, and further improve the segmentation accuracy; A residual path type jump connection mode is designed to realize the feature reproduc-tion of the corresponding scale of the context, solve the semantic gap problem caused by feature concatenating, and more accurately restore the high-resolution image. The experimental results show that the IoU index of overlapping region segmentation reaches 98.58%, which proves the ef-fectiveness of the segmentation method in this paper.