A short chromosome classification method based on spatial attention and texture enhancement
Chromosome classification is a crucial task in karyotype analysis.Despite the significant achievements made by residual neural networks in chromosome classification,the classification still presents challenges due to the short length,difficult-to-identify classification features,and high morphological similarity of certain chromosomes.To address this issue,a spatial information attention and texture enhancement network(SIATE-Net)model was proposed for chromosome classification.The SIATE-Net model utilized the Inception_ResNetV2 model as its backbone network to extract deep features of chromosomes.By introducing self-attention mechanisms and depth-wise separable convolution,the model could better focus on and retain the spatial information of short chromosomes.The short length of certain chromosomes often leads to confusion in banding information.To mitigate this issue,the model integrated a texture enhancement mechanism to amplify the differences between chromosomes,providing the model with more discriminative features for classification.The SIATE-Net model was validated on both private and public datasets,demonstrating superior classification performance compared to other methods,especially in classifying short chromosomes.On the private dataset,the SIATE-Net model achieved the best overall classification accuracy of 98.05%,with a high accuracy of 97.42%for short chromosomes.On the public dataset,the overall classification accuracy of the SIATE-Net model was 98.95%,with short chromosomes reaching an accuracy of 98.51%.Experimental results demonstrated that the targeted self-attention module,depth-wise separable convolution,and texture enhancement module effectively addressed the classification of short chromosomes without compromising overall classification accuracy.
medical image processingshort chromosome classificationInception_ResNetV2 modelself-attention mechanismdepth-wise separable convolutiontexture enhancement
NETL
NSTL
万方数据
