Random splicing assisted deep learning for breast cancer cell line classification via Raman spectroscopy.

Raman spectroscopy extracts rich biochemical information on a single cell, demonstrating significant potential for precise cancer identification. While machine learning enhances spectral analysis efficiency, conventional models remain constrained by data volume. Here, we developed Random Splicing-Convolutional Neural Network (RS-CNN), a deep learning framework that addresses data scarcity through spectral concatenation. By randomly splicing Raman spectra from the same cell line, RS-CNN enhances distinctive spectral features while simultaneously expanding dataset size and improving signal quality. Validation across six breast cancer cell lines demonstrated RS-CNN's superiority over five benchmark models (SVM, LDA, PCA-SVM, PCA-LDA, CNN). With 450 spectra per cell line, RS-CNN achieved 98.63 % classification accuracy compared to conventional models' accuracies of around 85 %. Under data-limited conditions (100 spectra/line), RS-CNN maintained 91.47 % accuracy, outperforming CNN's 70.83 %. The RS-CNN's generalizability was further validated by an independently acquired dataset, achieving at least 94 % classification accuracy. SHAP analysis suggested the spectral region around 980 cm⁻¹ was significant for cancer diagnosis, while the 1158-1160 cm⁻¹and 1603-1607 cm⁻¹ regions were particularly valuable for distinguishing between cancer subtypes. These findings establish RS-CNN as a robust analytical model for clinical Raman diagnostics, particularly valuable in applications requiring high accuracy with limited samples.