1. Deep Neural Networks for Machine Learning
Development of next generation computing processor for deep neural network-based applications is mainly ongoing. Network architecture compression and additional mathematical approximation are essential for fastening the inference time to meet real-time condition, targeting Convolution Neural Network (CNN)-based image classification, object detection, image super-resolution and Recurrent Neural Network (RNN) if time domain needs to be considered. In addition, for non-Euclidean (graph) data analytics, I am exploring variants of Graph Neural Network (GNN) architecture. 2. Medical Image Processing
Image Registration:
For a concise analysis of diseases by means of Imaging methodologies such as MR or CT, it is indispensable that two or more image sets with time sequences, with different modalities (Histology-MR) or with different dimensions (2D-3D) be preprocessed through the image registration (motion correction) in order to compensate for the artifacts caused from a patient's motion or respiration during the scanning. Image Segmentation: Segmentation is an essential pre-processing task for defining the volume of interest and localizing the analysis. Automation of this process is highly required because manual segmentation is time consuming and error prone. Feature Analysis: Over the registered and segmented region in medical image domain, highly related features to the target disease are extracted by hand-craft engineering or big data-driven deep learning. To the end, it aims to build an automated computer-aided diagnostic system. |
Purple: Manual reference, Green: Computed segmentation
|
3. Data Parallel ProcessingDespite the efficiency improvement in algorithm, its execution time for the registration of typical liver MRI scans is on the order of one hour when implemented in CPU. However, execution times on the order of one minute or less are needed for regular clinical use. Graphics processing units (GPU), which typically have 100+ processing elements, on-board memory of over 1 GB, and a high bandwidth (25+ time faster than the bandwidth of CPU main memory) for data transfer, have been used for computationally intensive processing tasks. In case of liver motion correction using demons algorithm, it is known to be able to reduce the run-time by a factor of 50 without sacrificing accuracy.
For more Information: http://www.nvidia.com/object/tesla_computing_solutions.html http://www.nvidia.com/object/gtc2010-presentation-archive.html#session2018 |
|