Differentiating Benign and Malignant Phenotypes in Breast Histology Sections with Quantum Cascade Laser Microscopy
StatisticsView Usage Statistics
The proposed research has two complementary goals: increasing productivity of pathologists and explore the utility of chemical imaging for diagnostic applications. The first goal is driven by an emerging shortage of pathologists is predicted for the near future; hence, increased productivity will be a necessity. The second goal is motivated by simplifying sample preparation, avoid fixation and staining, and increase throughput in a clinical setting. The efficacy of the stain-free protocol is evaluated through a pilot project of clinical samples that are scanned by a Quantum Cascade Laser InfRared (QCL-IR) microscope.QCL-IR microscopy has the potential to emerge as a unique modality for the diagnostics of histology sections. The pilot experiment is designed to evaluate whether benign and malignant breast histology sections can be differentiated using chemical profiling and without the use of spatial information. The experiment consists of fifteen independent samples that were randomly selected from paraffin-embedded blocks with 8 benign and 7 malignant labels. Spectral data are normalized and then used for both visualization and classification. Visualization is based on consensus clustering of the spectral signature within each group of the benign or malignant phenotype. Classification has been evaluated with four methods of tree-based, convolutional neural networks, Bayesian model, and an encoder module for compression followed by softmax classification. The latter has the best performance with using only 20\% of the data for training to arrive at the classification accuracy of 100\%. Direct analysis of the spectral signatures indicates that both malignant and benign samples express the similar spectral peaks; however, peaks corresponding to several nucleic acid sequences and protein groups are over expressed in malignant tissues.