Generative Adversarial Network Augmentation of Whole-Blood Raman Spectra for Breast-Cancer Classification: A PCA–LDA and Patient-Aware Validation Study
Brian O. Mairura *
Department of Physical Sciences, Tharaka University, Tharaka Nithi, Kenya.
John I. Githaiga
Department of Physics, University of Nairobi, Nairobi, Kenya.
Ruth W. Ndung'u
Department of Physical Sciences, Tharaka University, Tharaka Nithi, Kenya.
*Author to whom correspondence should be addressed.
Abstract
Background: Whole-blood Raman spectroscopy provides a label-free molecular fingerprint with potential for breast-cancer classification, but limited sample sizes and substantial biological variability can restrict model generalization. This study evaluated whether generative adversarial network (GAN) augmentation could improve PCA–LDA classification while preserving patient-independent evaluation.
Methods: Whole-blood Raman spectra from breast-cancer patients and healthy controls across Stages I–III were analysed using PCA–LDA under four data-composition scenarios: RealOnly, Real25, Real50 and Real100, corresponding to 0%, 25%, 50% and 100% synthetic augmentation. Vanilla GAN and conditional GAN (cGAN) data were evaluated using SpectrumKFold, SubjectKFold and leave-one-patient-out cross-validation (LOPOCV). Classification performance was assessed using accuracy, sensitivity, specificity, balanced accuracy, F1-score, MCC and ROC-AUC, with spectral correspondence and prediction confidence additionally examined.
Results: Spectrum-level validation frequently produced near-perfect discrimination, whereas patient-aware validation produced more conservative performance, demonstrating the influence of inter-patient variability. GAN augmentation generally improved sensitivity, balanced accuracy, F1-score and MCC, with the largest benefits observed for the more challenging patient-aware settings. cGAN generally produced greater improvements than vanilla GAN and frequently achieved strong performance at lower augmentation levels. Under LOPOCV, for example, Stage II sensitivity increased from 79.09% with real data alone to 87.32% with 25% cGAN augmentation, while Stage III sensitivity increased from 75.23% to 82.98%. Stage III remained the most challenging setting, with RealOnly LOPOCV AUC of 0.9521 compared with 0.9918 for Stage II.
Conclusion: GAN-based augmentation can strengthen PCA–LDA classification of whole-blood Raman spectra, particularly for patient-aware detection of breast-cancer spectra. cGAN showed greater augmentation efficiency than vanilla GAN in several settings. However, the small number of patients within individual stages and absence of independent external validation limit generalizability. Larger multicentre datasets and external validation are required before clinical applicability can be established.
Keywords: Breast cancer, Whole-blood Raman spectroscopy, Generative adversarial network (GAN), Conditional generative adversarial network (cGAN), Principal component analysis–linear discriminant analysis (PCA–LDA), Synthetic data augmentation, Patient-aware cross-validation, Cancer classification