Real-time validation of Surface-Enhanced Raman Scattering substrates via convolutional neural network algorithm
Real-time validation of Surface-Enhanced Raman Scattering substrates via convolutional neural network algorithm
复制标题
DOI：
10.1016/j.imu.2022.101076
复制
发表时间：
2022-09
影响因子：
--
通讯作者：
Paulo de Carvalho Gomes;Alexander Crossman;Emily A Massey;Jonathan James Stanley Rickard;P. Oppenheimer
中科院分区：
文献类型：
--
作者：
Paulo de Carvalho Gomes;Alexander Crossman;Emily A Massey;Jonathan James Stanley Rickard;P. Oppenheimer
下一代传感器技术的发展可以通过利用表面增强拉曼光谱（SERS）衬底来推进，这些衬底能够大大增强对感兴趣的分析物的信号检测，直至单分子水平。然而，SERS的广泛实施需要克服许多现有的和新出现的挑战，这些挑战与基材不同部分的增强不一致有关，从而导致读数不一致。在这里，开发了一种定制设计的卷积神经网络（CNN）算法，用于实时分析和验证衬底，提供快速输出和识别每个衬底上的SERS有源结构，从而实现一致的高信号增强。已经开发了一种计算算法，用于从获得的SERS结构的光学显微镜图像中实时识别高且一致的SERS活性区域。在输入的光学图像上对CNN模型进行测试，并输出叠加在输入图像上的预测一致的sers活性结构或区域。使用的优化CNN模型是一个预训练的神经网络（VGG16），最后一层进行了微调。该模型在从光学图像中对我们独特开发的电流体动力学（EHD）制造，高度增强的SERS活性结构进行分类时达到了90%的精度，从而实现了SERS基板的实时预测工具。CNN输出生成热图，热图提供对高SERS活性区域的反馈，覆盖在光学图像上，以快速促进SERS光谱采集测量。除了EHD的应用之外，所开发方法的多功能性使其易于扩展，以实现SERS基板的进一步，当前和未来的制造技术。完整的脚本运行产生了87.5±12.0%的现场测量采集精度，成功地展示了通过光学图像的实时神经网络处理快速识别一致SERS结构的概念验证。
Development of next-generation sensor technologies could be advanced by exploiting surface enhanced Raman spectroscopy (SERS) substrates capable of considerably enhancing signal detection of analytes of interest down to single molecule levels. The widespread implementation of SERS, however, requires overcoming many of the existing and emerging challenges related to the enhancement inconsistency from the different parts of substrates, resulting in inconsistent readings. Here, a custom-designed convolutional neural networks (CNN) algorithm is developed to specifically analyse and validate substrates inreal-time, providing a rapid output and identification of SERS active structures on each substrate enabling a consistently high signal enhancement. A computational algorithm has been developed to identify regions of high and consistent SERS activity from acquired optical microscopy images of SERS structures inreal-time. The CNN model is tested on the inputted optical images and outputs the predicted consistent SERS-active structures or regions overlaid on top of the input image. The optimised CNN model used is a pre-trained neural network (VGG16) with the last layers fine-tuned. This model achieves an accuracy of 90% in classifying our uniquely developed electrohydrodynamically (EHD) fabricated, highly enhancing, SERS-active structures from optical images, enabling areal-timepredictive tool for SERS substrates. The CNN output generates a heatmap, which provides feedback on the regions of high SERS activity, overlaid on the optical image to rapidly facilitate the SERS spectral acquisition measurements. Beyond its application for EHD, the versatility of the developed method renders it easily extendable for implementation with further, current and future, fabrication techniques of SERS substrates. The full script run yields an accuracy of 87.5 ± 12.0% forin-situmeasurement acquisition, successfully demonstrating a proof-of-concept of rapidly identifying consistent SERS structuresviathereal-timeCNN processing of optical images.