High-Throughput-Screening of Medical Image Data on Heterogeneous Clusters
Non-invasive medical imaging by means of computed tomography (CT) and fMRI helps clinicians to improve diagnostics and - hopefully - treatment of patients. Due to better image resolutions as well as ever increasing numbers of patients who undergo these procedures, the amount of data that have to be analyzed puts great strain on radiologists. In an ongoing development with SALK (Salzburger Landeskrankenhaus) we propose a system for automated screening of CT data for cysts in the patient’s kidney area. The proper detection of kidneys is non-trivial, due the high variance of possible size, location, levels of contrast and possible pathological anomalies a human kidney can expose in a CT slice. We employ large-scale, semi-automatically generated dictionaries (based on 107 training images) to be used in injunction with principal component analysis (PCA). Heterogeneous clusters of CPU-, GPGPU-, and Cell BE-processors are used for high-throughput-screening of CT data. For data-parallel programming CUDA, OpenCL and the IBM Cell SDK have been used. Task parallelism is based on OpenMPI and a dynamic load-balancing scheme, which demonstrates very low latencies by means of double-buffered, multi-threaded queues.
KeywordsCompute Tomography Data Dictionary Entry Compute Tomography Slice Work Item Heterogeneous Cluster
Unable to display preview. Download preview PDF.
- 9.Weber, R.: Similarity Search in High-Dimensional Data Spaces. Ph.D. thesis, Grundlagen von Datenbanken (1998)Google Scholar
- 10.Zinterhof, P.: Distributed Computation of Feature-Detectors for Medical Image Processing on GPGPU and Cell Processors. In: Guarracino, M.R., Vivien, F., Träff, J.L., Cannatoro, M., Danelutto, M., Hast, A., Perla, F., Knüpfer, A., Di Martino, B., Alexander, M. (eds.) Euro-Par-Workshop 2010. LNCS, vol. 6586, pp. 339–347. Springer, Heidelberg (2011)CrossRefGoogle Scholar