Cell Classification of FT-IR Spectroscopic Data for Histopathology using Neural Networks скачать в хорошем качестве

Cell Classification of FT-IR Spectroscopic Data for Histopathology using Neural Networks

FT-IR spectroscopy has shown potential as an imaging modality for cancer detection. This method measures the local absorbance of infrared light by a specimen, providing rich, quantitative, molecular and morphological data for analysis without the need of staining or altering the sample. Current practice in cancer detection using FT-IR data is to image a tissue sample and label each pixel according to cell type. The morphological information of these 2D masks or maps of cells informs pathologists about the presence of disease or cancer. In this presentation, I will demonstrate my work to automate the classification of cell types using the tools of Deep Neural Networks. I will also describe how Azure can be used to aid in similar tasks that deal with large datasets and parallelizable learning algorithms and explain how to set-up an appropriate workspace in Azure. For this problem, I deployed DNNs using the Pylearn2 Python library in the cloud on Azure VMs, which allowed for parallelized training of the DNNs through the tuning of hyper-parameters. So far, trained DNNs have reached labeling accuracies of 90%. I will discuss the classification results and their impact for clinical use of FT-IR, as well as my experience of distributing deep learning on Azure.