A Biomedical Imaging Acceleration Testbed: NSF Award Number 0946463
InvestigatorsDavid Kaeli, William Karl, Homer Pien, Badrinath Roysam, Nayda Santiago, Miriam Leeser
DescriptionThis project will develop the methodology for rapid parallelization of biomedical imaging applications which will result in a testbed that will provide the capability to "right-size" a multi-GPU system to best meet the goals of any biomedical imaging application. This testbed will utilize a web-based framework that will allow a larger community to leverage the technology available and it will also contain a rich library of parallelized biomedical imaging codes.
- Rodrigo Dominguez - Graduate Student - NEU
- Byunghyun Jang - Graduate Student - NEU
- Perhaad Mistry - Graduate Student - NEU
- Richard Moore - Director of Breast Imaging Research, Massachusetts General Hospital, Boston
- Dana Schaa - Graduate Student - NEU
- Matthew Sellito - Graduate Student - NEU
- Justin White - Undergraduate Student - NEU
SystemsThe Medusa Cluster:
- 4 Server Nodes with 4 quad-core AMD Opteron CPUs in each node.
- Two NVIDIA Tesla S1070 Units
- Four ATI 5800 GPUs.
- Two storage nodes with Infiniband connectivity and quad core CPUs in each node
The Tesla system:
- 2 Server Nodes with dual-core Intel Xeon CPUs
- NVIDIA Tesla S870 Units
- 2 GeForce GTX 285
- 1 GeForce GTX 9800
- 4 GeForce GTX 8800
- 1 ATI Radeon HD 5800
- 1 AMD FireStream 9270
Plans the Microscopy domain have have been formulated as follows.
Step 1: Multiscale Multi-dimensional Image Pre-processor (MMiPP): Enable near-interactive exploration of scale space characteristics of images, rapid identification of optimal scale space settings & orientation diversity at different scales, and image reconstruction methods for denoising, generic smoothing, texture analysis, enhancement, and specialized enhancement (with a focus on stick, plate, and blob segmentation).
Step 2: Basic Differential geometry package: First and second order derivates, Jacobian, Hessian, and Weingarten matrices, curvature, ridges, distance maps, plateness, ballness, and stickness fields.
Step 3: Flexible tensor voting package: (i) voxel based (ii) object based. Both extended to handle ball, plate, and tube classes, leveraging Stage 2.
Step 4: GPU implementation of time-critical parts of existing FARSIGHT segmentation codes for blobs, network of tubes (thin/thick/hollow), leveraging codes from previous 3 stages. Focus on rapid parameter selection methods.
Cardiac CT IRT
Recent Talks and Tutorials
Tutorial slides can be found here.