Fig. 3
From: Intra-operative spectroscopic assessment of surgical margins during breast conserving surgery
Raman spectral acquisition and annotation. Tumor regions (clusters of blue dots in the H&E image in (c)) appear darker in the auto-fluorescence (AF) image (a). The region in the green box was measured by a Raman raster scan. K-means cluster analysis of these spectra identifies similar spectra to create a hyperspectral image (b). Single spectra from locations marked in b are shown in d. Based on the information in a-d, pre-processed spectra from green areas (horizontal triangles) are marked as tumor, blue (square/circle) as inflamed stroma, and red (vertical triangles) as fat. Other clusters (cyan, yellow, and magenta) were background or noise and were withheld from the training set. Mean and standard deviation of all spectra in the training set show that the annotated tissue types (e) could be simplified to three classes used by the spectral classifier (f). Spectral features used for classification are marked as shaded areas (peak areas) and magenta lines (peak intensity differences). These peak areas are shaded blue for lipid-associated bands, green for protein-associated bands, and magenta for nucleic acid-associated bands. These features are consistent across all tumor types (g). Classes: IC, invasive carcinoma; OT, other tumor types (includes ductal carcinoma in situ (DCIS), lobular carcinoma in situ (LCIS), malignant phyllodes (MP)); BG, benign growths (includes fibroadenoma, sclerosing adenosis, hyperplasia); IN, inflammation; P, parenchyma; S, healthy stroma; F, fat; F + S, mixture of fat and stroma