RESUMO
Better knowledge of the breast tumor microenvironment is required for surgical resection and understanding the processes of tumor development. Raman spectroscopy is a promising tool that can assist in uncovering the molecular basis of disease and provide quantifiable molecular information for diagnosis and treatment evaluation. In this work, eighty-eight frozen breast tissue sections, including forty-four normal and forty-four tumor sections, were mapped in their entirety using a 250-µm-square measurement grid. Two or more smaller regions of interest within each tissue were additionally mapped using a 25 µm-square step size. A deep learning algorithm, convolutional neural network (CNN), was developed to distinguish histopathologic features with-in individual and across multiple tissue sections. Cancerous breast tissue were discriminated from normal breast tissue with 90 % accuracy, 88.8 % sensitivity and 90.8 % specificity with an excellent Area Under the Receiver Operator Curve (AUROC) of 0.96. Features that contributed significantly to the model were identified and used to generate RGB images of the tissue sections. For each grid point (pixel) on a Raman map, color was assigned to intensities at frequencies of 1002â¯cm-1 (Phenylalanine), 869â¯cm-1 (Proline, CC stretching of hydroxyproline-collagen assignment, single bond stretching vibrations for the amino acids proline, valine and polysaccharides) and 1309â¯cm-1 (CH3/CH2 twisting or bending mode of lipids). The Raman images clearly associate with hematoxylin and eosin stained tissue sections and allow clear visualization of boundaries between normal adipose, connective tissue and tumor. We demonstrated that this simple imaging technique allows high-resolution, straightforward molecular interpretation of Raman images. Raman spectroscopy provides rapid, label-free imaging of microscopic features with high accuracy. This method has application as laboratory tool and can assist with intraoperative tissue assessment during Breast Conserving surgery.