부산대학교 도서관

Big Data presents considerable challenges to deep learning for transforming complex, high-dimensional, and heterogeneous biomedical data into health care data. Various kinds of data are analyzed in recent biomedical research that includes e-health records, medical imaging, text, and IoT sensor data, which are complex, badly labeled, heterogeneous, and usually unstructured. Conventional statistical learning and data mining methods usually require first to extract features to acquire more robust and effective variables from those data. These features help build clustering or prediction models. New useful paradigms are provided by the latest advancements based on deep learning technologies for obtaining end-to-end learning techniques from complex data. The abstractions of data are represented using the multiple layers of deep learning for building computational models. Clinician performance is augmented by the prospective of deep learning models in medical imaging interpretation, and automated segmentation is used to reduce the time for the diagnosis. This work presents a convolution neural network-based deep learning infrastructure that performs medical imaging data analysis in various pipeline stages, including data-loading, data-augmentation, network architectures, loss functions, and evaluation metrics. Our proposed deep learning approach supports both 2D as well as 3D medical image analysis. We evaluate the proposed system's performance using metrics like sensitivity, specificity, accuracy, and precision over the clinical data with and without augmentation.