Raman spectroscopy has tremendous potential for material analysis with its molecular fingerprinting capability in many branches of science and technology. It is also an emerging omics technique for metabolic profiling to shape precision medicine. However, precisely attributing vibration peaks coupled with specific environmental, instrumental, and specimen noise is problematic. Intelligent Raman spectral preprocessing to remove statistical bias noise and sample-related errors should provide a powerful tool for valuable information extraction. Here, we propose a novel Raman spectral preprocessing scheme based on self-supervised learning (RSPSSL) with high capacity and spectral fidelity. It can preprocess arbitrary Raman spectra without further training at a speed of ~1 900 spectra per second without human interference. The experimental data preprocessing trial demonstrated its excellent capacity and signal fidelity with an 88% reduction in root mean square error and a 60% reduction in infinite norm (\({L}_{{\infty }}\)) compared to established techniques. With this advantage, it remarkably enhanced various biomedical applications with a 400% accuracy elevation (ΔAUC) in cancer diagnosis, an average 38% (few-shot) and 242% accuracy improvement in paraquat concentration prediction, and unsealed the chemical resolution of biomedical hyperspectral images, especially in the spectral fingerprint region. It precisely preprocessed various Raman spectra from different spectroscopy devices, laboratories, and diverse applications. This scheme will enable biomedical mechanism screening with the label-free volumetric molecular imaging tool on organism and disease metabolomics profiling with a scenario of high throughput, cross-device, various analyte complexity, and diverse applications.

拉曼光谱在科学技术的许多分支中具有分子指纹识别能力，在材料分析方面具有巨大的潜力。它也是一种新兴的组学技术，用于代谢分析以塑造精准医学。然而，精确地归因与特定环境、仪器和样本噪声相结合的振动峰值是有问题的。用于消除统计偏差噪声和样本相关错误的智能拉曼光谱预处理应该为有价值的信息提取提供强大的工具。在这里，我们提出了一种基于自监督学习（RSPSSL）的新型拉曼光谱预处理方案，具有高容量和光谱保真度。它可以以每秒约 1 900 个光谱的速度预处理任意拉曼光谱，无需进一步训练，无需人工干预。实验数据预处理试验证明了其出色的容量和信号保真度，与现有技术相比，均方根误差降低了 88%，无限范数 ( \({L}_{{\infty }}\)降低了 60%） 。凭借这一优势，它显着增强了各种生物医学应用，癌症诊断准确率 (ΔAUC) 提高了 400%，百草枯浓度预测平均提高了 38%（少量）和 242%，并开启了生物医学高光谱的化学分辨率图像，特别是在光谱指纹区域。它对来自不同光谱设备、实验室和不同应用的各种拉曼光谱进行精确预处理。该方案将在高通量、跨设备、各种分析物复杂性和多样化应用的场景下，利用无标记体积分子成像工具对生物体和疾病代谢组学分析进行生物医学机制筛选。