PO.CL01.19 · 临床研究
整合AI与SERS技术的创新性无标记、非侵入性尿液代谢物分析用于癌症早期检测：一项涉及五种癌症类型的回顾性临床研究
Innovative label-free and non-invasive urinary metabolite analysis integrating AI and SERS technology for early cancer detection: A retrospective clinical study involving five cancer types
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
背景：由于传统诊断方法成本高、敏感性和特异性有限，准确的癌症早期检测仍是一项关键的临床挑战。事实上，迫切需要能够对生物体液进行分子水平表征的非侵入性方法，以提高早期诊断准确性并改善患者预后。人工智能辅助表面增强拉曼散射（AI-SERS）提供了一个强大的平台，能够以卓越的敏感性检测尿液中复杂的代谢特征。通过将高分辨率SERS光谱与基于深度学习的分类相结合，AI-SERS平台超越了传统检测方法的局限性，实现了高度精确的癌症鉴别，并为发现用于癌症早期检测的基于代谢物的生物标志物开辟了新途径。
方法：本研究共纳入287份临床尿液样本，分为六组：前列腺癌（PRC，n = 49）、胰腺癌（PAC，n = 16）、卵巢癌（OC，n = 64）、肺癌（LC，n = 30）、乳腺癌（BC，n = 29）和正常对照（NOR，n = 99）。将极小体积（10 μL）的临床尿液样本施加于专利SERS传感器上以增强代谢物信号，并使用拉曼光谱仪进行测量。所得SERS光谱采用基于卷积神经网络（CNN）的深度学习方法进行分析。
结果：6个不同组别尿液样本的SERS光谱呈现出尖锐的拉曼光谱峰，为分析提供了重要信息。使用所开发的CNN模型，将所有癌症类型合并可与正常对照相区分，准确率为96.6%，敏感性为99.3%，特异性为94.0%，表明癌症与非癌症样本之间总体分类性能优异。对各单一癌症类型与正常对照的进一步分析显示出稳健的预测性能，其准确率、敏感性和特异性分别为：PRC 98.0%、98.6%、97.3%；OC 97.9%、97.9%、97.9%；LC 97.8%、97.8%、97.8%；BC 98.9%、97.7%、100%。这些结果证明了AI-SERS平台在非侵入性、癌症特异性检测方面的强大预测能力。
结论：使用AI-SERS平台进行的非侵入性、无标记尿液分析在区分癌症与正常对照方面表现出卓越的检测性能，提供了一种快速癌症筛查方法，并凸显了其在癌症早期诊断中的潜力。正在进行的临床研究旨在鉴定癌症类型特异性代谢生物标志物，以进一步提高诊断特异性并有助于改善患者预后。
查看英文原文 English abstract
Background: Accurate early cancer detection remains a critical clinical challenge due to the high cost and limited sensitivity and specificity of conventional diagnostic methods. Indeed, non-invasive approaches capable of molecular-level characterization of biofluids are urgently needed to improve early diagnostic accuracy and patient outcomes. Artificial intelligence-assisted surface-enhanced Raman scattering (AI-SERS) offers a powerful platform to test complex metabolic signatures in urine with exceptional sensitivity. By integrating high-resolution SERS spectra with deep learning-based classification, AI-SERS platform transcends the limitations of conventional assays, enabling highly precise cancer differentiation and unlocking new avenues for the discovery of metabolite-based biomarkers for early cancer detection.
Methods: This study enrolled a total of 287 clinical urine samples across six groups: prostate cancer (PRC, n = 49), pancreatic cancer (PAC, n = 16), ovarian cancer (OC, n = 64), lung cancer (LC, n = 30), breast cancer (BC, n = 29), and normal controls (NOR, n = 99). A minimal volume (10 μL) of clinical urine samples was applied to a patented SERS sensor to enhance metabolite signals, measured using a Raman spectrometer. The resulting SERS spectra were analyzed using a convolutional neural network (CNN)-based deep learning approach.
Results: The SERS spectra of urine samples from 6 different groups exhibited sharp Raman spectral peaks, providing significant information for analysis. Using the developed CNN model, all cancer types combined could be distinguished from normal controls with an accuracy of 96.6%, sensitivity of 99.3%, and specificity of 94.0%, demonstrating excellent overall classification performance between cancer and non-cancer samples. Further analysis of individual cancer types versus normal controls showed robust predictive performance, with accuracy, sensitivity, and specificity of 98.0%, 98.6%, and 97.3% for PRC; 97.9%, 97.9%, and 97.9% for OC; 97.8%, 97.8%, and 97.8% for LC; and 98.9%, 97.7%, and 100% for BC, respectively. These results demonstrate the strong predictive capability of the AI-SERS platform for noninvasive, cancer-specific detection.
Conclusions: The noninvasive, label-free urine analysis using the AI-SERS platform revealed remarkable test performance in classifying cancers from normal controls, offering a rapid cancer screening approach and highlighting its potential for early cancer diagnosis. Ongoing clinical studies aim to identify cancer-type-specific metabolic biomarkers to further improve diagnostic specificity and contribute to enhanced patient outcomes.
利益披露 Disclosure
J. Kim, None..
H. Choi, None..
E. Koh, None..
T. Vo, None..
G. Ryu, None..
E. Lee, None..
D. Kwon, None..
S. Lew, None..
S. Song, None.