多特征融合高通量dPCR荧光图像识别

Multi-feature fusion high-throughput dPCR fluorescence image recognition

传统高通量dPCR荧光图像分析结果易因假阳性点与非特异性扩增而导致阳性点识别率较低,因而本文提出一种多特征融合高通量dPCR荧光图像识别方法(HDFINet),以提高阳性点识别准确性。首先,在特征融合部分引入自上而下结构,使得下层特征在顶层被更有效地利用。在自上而下结构中,使用通道注意力来分配荧光图像通道权重,并使用空间注意力来分配特征图中荧光图像像素相应权重,使得特征映射能够更好地响应荧光图像特征。然后,在RPN中使用自适应交并比IOU计算阳性点包围框置信度,减少阳性点信息丢失可能性。最后,ROI Align将荧光图像候选区域中阳性点特征重新固定尺寸后,输入至全连接层和全卷积层,进行类别与回归框回归,输出阳性点识别结果。本文提出的HDFINet网络具有较高识别率,可以有效地实现荧光图像阳性点识别,与YOLOv4、VF-Net、GROIE相比,本文方法综合指标F1最高,相比于经典的深度学习网络Mask R-CNN网络,本方法对阳性点识别真阳性率提高了1.13%,阳性预测值提高了0.36%,综合指标F1的值提高了0.75%。本文提出的HDFINet网络具有良好的识别性能,能够有效识别荧光图像阳性点,对其他荧光图像分析研究具有参考价值。

The results of traditional high-throughput dPCR fluorescence image analysis are prone to low positive spot recognition rate due to false positive points and non-specific amplification.Therefore,in this paper,a multi-feature fusion high-throughput dPCR fluorescence image recognition method(HDFINet)is proposed to improve the accuracy of high-throughput dPCR fluorescence image recognition.Firstly,a upbottom structure is introduced in the feature fusion part so that the lower layer features can be used more effectively in the top layer.In the up-bottom structure,channel attention is used to assign channel weight of fluorescent image,and spatial attention is used to assign corresponding weight of fluorescent image pixels in the feature map,so that the feature map can better respond to the feature of fluorescent image positive points.Then,the confidence of the bounding box of positive points was calculated by using the adaptive Intersection-over-Union(IOU)in RPN to reduce the possibility of loss of positive points information.Finally,ROI Align re-fixed the size of the features in the candidate areas of fluorescent images,and then input them to the full connection layer and fully convolution layer to perform category and regression box regression and output positive point recognition results.The experimental results show that the HDFINet network proposed in this paper has a high recognition rate and can effectively realize the recognition of positive points in fluorescent images.Compared with YOLOv4,VF-Net,and GROIE,the comprehensive index F1 of the method in this paper is the highest,compared with the classic deep learning Network Mask R-CNN network,this method increases the true positive rate of positive points by 1.13%,the positive predictive value by 0.36%,and the value of the comprehensive index F1 by 0.75%.The HDFINET network proposed in this paper has good recognition performance and can effectively identify positive spots in fluorescence images,which has reference value for other fluorescence image analysis and research.