Feasibility study of multispectral imaging guided Raman spectroscopy for breast cancer diagnosis
Abstract
Accurate breast cancer diagnosis on fresh surgical specimens requires an optical workflow that can first localize diagnostically relevant tissue regions and then verify these regions using molecular information. In this feasibility study, multispectral imaging (MSI) was used to delineate regions of interest (ROIs) on freshly excised breast specimens, followed by high-wavenumber (HW) Raman spectroscopy from the MSI-defined ROIs. A total of 28 specimens (fibro-adipose, n = 18; tumor, n = 10) and 48,390 pixel-wise multispectral reflectance spectra (fibro-adipose, n = 27,934; tumor, n = 20,456) were included. The MSI profiles showed reproducible optical differences between fibro-adipose and tumor regions. For MSI-based classification, partial least-squares discriminant analysis (PLS-DA) generated probability-based tissue maps with a specimen-wise mean intersection over union (mIoU) of 85.39%, a tumor-class sensitivity of 89.9%, a tumor-class specificity of 90.7%, an accuracy of 90.4%, and an area under the receiver operating characteristic curve (AUC) of 0.96. A one-dimensional convolutional neural network (1D-CNN) applied to the same MSI inputs achieved a sensitivity of 92.7%, a specificity of 92.46%, an accuracy of 92.53%, and an AUC of 0.98, with posterior probabilities showing stronger class separation than PLS-DA. Raman spectra acquired from the MSI-defined ROIs provided complementary biochemical evidence: fibro-adipose tissue showed stronger lipid-associated carbon-hydrogen stretching features, whereas tumor tissue exhibited relatively increased protein-related spectral contributions. For Raman-based classification, PLS-DA achieved a sensitivity of 90.4%, a specificity of 92.8%, an accuracy of 91.69%, and an AUC of 0.96. The 1D-CNN model further achieved a sensitivity of 99.8%, a specificity of 90.31%, an accuracy of 94.66%, and an AUC of 0.97. These findings indicate that MSI can identify diagnostically suspicious breast tissue regions and guide targeted Raman confirmation, supporting the feasibility of an MSI-guided Raman spectroscopy workflow for breast cancer diagnosis.