Abstract
Triple-negative breast cancer (TNBC) and human epidermal growth factor receptor 2 (HER2)-positive breast cancer require distinct systemic treatment strategies, whereas definitive subtype assignment depends on immunohistochemistry and, when indicated, in situ hybridization. Microscopic hyperspectral imaging (HSI) may provide wavelength-resolved information directly from routine histological sections. We evaluated hematoxylin-eosin-stained surgical sections from 120 patients with invasive breast carcinoma (60 TNBC and 60 HER2-positive cases). Microscopic spectral imaging was performed in the visible-near-infrared region: the calibrated system covered 367.55-974.92 nm, and 450-900 nm was retained for analysis. Pathologist-annotated tumor regions underwent white-dark correction, Savitzky-Golay smoothing (window 11, polynomial order 2), and standard normal variate normalization. A lightweight one-dimensional convolutional neural network was compared with support vector machine, random forest, and red-green-blue image classifiers using stratified five-fold patient-level cross-validation. The calibrated spectra showed the clearest subtype separation across approximately 470-640 nm, with smaller differences in the 760-860 nm region. Data-derived local spectral-contrast features were observed near 497, 616, 812, and 851 nm. Exploratory principal component analysis of patient-level mean spectra showed partial separation between TNBC and HER2-positive cases, with PC1 and PC2 explaining 61.3% and 36.8% of the variance, respectively. The hyperspectral classifier yielded a patient-level area under the receiver operating characteristic curve of 0.886, accuracy of 83.3%, sensitivity of 85.0%, specificity of 81.7%, and F1-score of 83.6%. Microscopic HSI of routine hematoxylin-eosin sections captured subtype-associated optical differences between TNBC and HER2-positive breast cancer. The approach is best positioned as an adjunctive optical triage method rather than a replacement for immunohistochemistry or in situ hybridization, and external multicenter validation is required before clinical implementation.