Neuroblastoma Stroma Classification on the Sony Playstation 3
Best Scientific Session - Telemedicine, High-Performance Computing and Imaging InformaticsMetin Gurcan PhD; Ohio State University; Olcay Sertel MSc; Ohio State University; Umit Catalyurek PhD; Ohio State University; Joel Saltz MD, PhD; Ohio State University; Timothy Daniel Raul Hartley MSc; Ohio State University; Mansoor Khan BS; Ohio State University;
Content:
Neuroblastoma is a nerve cell cancer which affects infants and children. The prognosis of the disease is based on classification of a tissue sample as stroma-rich or stroma-poor. The overall size of a single digital tissue sample image can be tens of gigabytes. The large size of these images requires unacceptably long processing times for computer analysis with traditional computer architectures. Further, the cost of computer systems which enable acceptable image analysis throughput can be prohibitive. Therefore, we present an application to analyze neuroblastoma images on the Sony Playstation 3 with the IBM Cell Broadband Engine microprocessor.
Technology:
The IBM Cell Broadband Engine represents a radical step forward in microprocessor design. It is a parallel-processor-on-a-chip, providing unprecedented levels of application performance. By decomposing the images into smaller tiles, we can leverage the inherent parallelism in the microprocessor to enable extremely fast analysis. The Sony Playstation 3 represents an inexpensive, powerful, and professional form factor for high-performance computing. Sony encourages the installation of Linux, and IBM makes many of their software tools available for the Cell Broadband Engine.
Design:
Our application leverages the parallelism of the Cell Broadband Engine by decomposing the input image into smaller image tiles and processing these tiles in parallel. Each tile undergoes a color transformation, statistical feature calculation, and ultimately classification as stroma-rich or stroma-poor.
Results:
The IBM Cell Broadband Engine in the Sony Playsation 3 gives a 6x performance improvement over state-of-the-art PCs for a small fraction of the cost of a high performance image analysis server. The total time required to process a complete slide of 100,000 x 100,000 pixels is under 15 minutes.
Conclusion:
We achieve better performance while analyzing neuroblastoma tissue sample images than a traditional computer design which costs several times that of the Sony Playstation 3. We believe the low cost and the high performance of our solution makes this favorable over others.
