Efficient Processing of Pathological Images Using the Grid: ComputerAided Prognosis of Neuroblastoma

1
Efficient Processing of Pathological Images Using
the GridComputer-Aided Prognosis of
Neuroblastoma

• B. Barla Cambazoglu
• Ohio State University
• Department of Biomedical Informatics

2
Overview

• Neuroblastoma classification problem
• Grid overview
• Grid-enabled parallel computing solution
• Experimental results
• On-going work

3
Neuroblastoma Classification Problem

• Neuroblastoma is a childhood cancer
• Peripheral neuroblastic tumors are a group of
  embryonal tumors of the sympathetic nervous
  system
• International Neuroblastoma Prognosis
  Classification System developed by Shimada et
  al., classifies the disease into various
  prognostic groups in terms of different
  pathologic features
• In clinical practice, two typical criteria for
  classification of the neuroblastic tumors are
• Grade of neuroblastic differentiation
  (undifferentiated, poorly-differentiated, and
  differentiating)
• The presence of Schwannian stromal development
  (stroma-poor and stroma-rich)

4
Sample Neuroblastoma Images

• In the current clinical practice, prognosis of
  neuroblastoma is largely dependent on the
  examination of haematoxylin- and eosin-stained
  tissue images by expert pathologists under the
  microscope
• considerably time consuming
• subject to inter- and intra-reader variations

5
Sample Segmentation
Background
Original image
Nuclei
Segmented
Neuropil
Cytoplasm
6
Challenges in Neuroblastoma Classification

• The size of a single neuroblastoma image is in
  the order of a few Gigabytes when compressed
• A typical image repository contains data whose
  size is in the order of Terabytes
• Complicated, time-consuming image classification
  algorithms are required
• Sequential systems are not practical due to the
  massive size of the image data and hence the
  processing requirements, justifying the need for
  parallel large-scale data processing

7
Grid for Biomedical Applications

• The collaborative nature of the grids
• Lets scientists share distributed resources and
  applications
• Eliminates the need for replication and waste of
  resources
• Fosters the collaboration among developers
• Large computational resources offered by the grid
• Large memory and storage capacities
• Distributed computational resources
• The grid comes with built-in security mechanisms
• Authentication
• Authorization
• Encryption

8
Grid-Enabled Neuroblastoma Classification

• Service-based infrastructure
• Multiple, geographically distributed scientists
  and developers access a common image data
  repository
• Share a common code repository allowing
  reusability of the developed codes
• Remote job execution
• A multi-processor backend
• Fast parallel processing of images
• Specifically designed for very large-scale image
  processing
• Pipelined processing capabilities

9
General System Architecture
10
Neuroblastoma Grid Service

• The service is developed
• Based on the caGrid 1.0 middleware
• Using Introduce service development toolkit
• Strongly-typed interfaces
• Provided operations on images/algorithms
• Query
• CQL (caGrid Query Language)
• Retrieve/Upload
• Bulk data transfer
• GridFTP
• Execute

11
Grid Service Client
12
Parallel Backend
13
Execution Times
14
Speedups (Single Reader)
15
Speedups (Multi-Reader)
16
On-going/Future Work

• Integration of the demand-driven code with the
  multi-reader code
• Dynamic service deployment
• Security infrastructure
• Adaptation from In Vivo Imaging Middleware