重叠染色体分割的图像处理综合实验设计

Comprehensive experimental design of image processing for overlapping chromosome segmentation

医学图像的处理是图像处理的重要应用领域。该文针对核型分析中重叠染色体的分割问题,进行了综合性教学实验设计。引入多层空洞卷积MAC和同步长池化技术SSP,实现了对不同大小分割区域的检测及多尺度特征提取。MAC根据染色体重叠区域普遍偏小的特点组合具有连续较小空洞率的空洞卷积,以实现多尺度特征的有效提取;SSP组合池化尺寸不同但池化步长固定的池化操作,有利于减少上采样中语义信息的丢失;在编、解码器间引入残差块连接,缓解了语义信息差异。实验结果表明,该实验设计网络在染色体重叠部分的分割性能明显提升,且噪声鲁棒性更强。

Medical image processing is an important application field of image processing.Aiming at the segmentation of overlapping chromosomes in karyotype analysis,a comprehensive teaching experimental design is carried out in this paper.By introducing Multi-layer Atrous Convolution(MAC)and Same Stride Pooling Network(SSP),the detection and multi-scale feature extraction of segmented regions with different sizes are realized.MAC combines the atrous convolution with continuous small atrous rate according to generally small chromosome overlapping area,so as to realize the effective extraction of multi-scale features;SSP adopts pool operation with different pool size but fixed pool step,which is helpful to reduce the loss of semantic information in up-sampling;Residual block connection is introduced between codec and decoder to alleviate the difference of semantic information.The experimental results show that the segmentation performance of the experimental design network in chromosome overlap is significantly improved with better robustness and generalization.