Enhancing Diagnostic Reliability of Ex Vivo Adenocarcinoma Identification Using Minimally Processed, Heterogeneous Raman Spectra and Deep Learning

ABSTRACT Reliable Raman Spectroscopy (RS)–based identification of lung adenocarcinoma remains challenging due to autofluorescence, low signal‐to‐noise ratios (SNR), and substantial spectral heterogeneity in clinical tissue samples. To evaluate whether robust diagnostic performance can be achieved under realistic acquisition conditions, this study investigated the use of minimally processed Raman spectra from ex vivo lung tissues to differentiate adenocarcinoma from matched normal samples. Spectra from 56 paired specimens underwent minimal preprocessing, limited to baseline subtraction and vector normalization. All spectra, including those with low‐SNR, which are typically excluded, were retained to reflect real‐world heterogeneity. Unsupervised K‐means clustering captured intra‐class variability, guiding a grouped, stratified train–validation–test split to ensure robust evaluation. The optimized ResNet‐34 model achieved a classification accuracy of 91.4%, sensitivity of 92.5%, specificity of 90.3%, and an area under the receiver operating characteristic curve (AUROC) of 0.956 on an independent test set. Integrated gradients (IG) analysis revealed that the model prioritized spectral regions associated with nucleic acid synthesis, protein turnover, and membrane remodeling, aligning with published studies. These findings demonstrate that deep learning applied to minimally processed, heterogeneous Raman spectra can yield reliable diagnostic performance without extensive preprocessing, supporting the feasibility of clinically realistic Raman‐based tissue assessment. Future work will extend this approach to additional lung cancer subtypes and in vivo settings.