Shared LowLatency Medical Imaging and High Bandwidth for the Masses:

1
Shared Low-Latency Medical Imaging and High
Bandwidth for the Masses

• Current and Future Collaborative Projects
• Nathan Stone
• Pittsburgh Supercomputing Center
• MetaComputing 2000 June 6-7, 2000

2
(No Transcript)
3
HUBS Project Goals

• to enable the infrastructure to achieve common
  goals through an IT environment that promotes
  effective collaboration
• to integrate data and services
• to develop and disseminate advanced IT
  applications that will enable regional
  organizations to effectively and efficiently
  collaborate

4
HUBS Task 1

• Next Generation Virtual Private Network

5
VPN Motivation

• The three HUBS applications we propose to
  develop require both NGI performance and VPN
  security. Today, these are mutually exclusive.
  We propose to combine both capabilities, without
  compromising either, to enable Next Generation
  VPNs (NG-VPN). By developing this technology in
  parallel with demanding HUBS applications ranging
  from medical research to missile defense, we will
  ensure the generality of the resulting
  implementation.

6
Security

• Host Security
• User-based trust model is risky
• 98 of systems compromised via the network
  utilize user-based security
• Local login ID is not adequate for authentication
  
• dedicated authentication infrastructure is
  necessary
• Transparency is desirable for ease of use
• Network Security
• NGI (etc.) has all the security problems of the
  current Internet
• LAN security may have to be considered as well
• (trust between users at a specific site)

7
VPN Project Goals

• Security
• Medical data has special legal restrictions
• HIPAA (http//www.hcfa.gov/medicaid/HIPAA/topics/m
  ore.asp)
• Network Auto-Tuning
• Modify the server kernel(s) to enable dynamic
  auto-tuning of TCP transport parameters
• Quality of Service (QoS)
• Utilizing QoS protocols the HUBS VPN will
  endeavor to enhance the performance of the
  inter-site network where it is needed

8
VPN Deployment Issues

• Firewall throughput (bypass it?)
• Node deployment and dedicated networks
• Database replication and data scrubbing

9
HUBS Task 2

• Intelligent Archiving for Medical Images

10
IA Motivation

• Intelligent Archiving is the storing of large
  quantities of information in a privacy-assured,
  fault-tolerant, geographically distributed manner
  while providing the ability to access it from
  anywhere in the region as efficiently as if it
  were on-line locally. Virtual Hospitals require
  it for patient care, research, continuing
  education, and rapid disaster recovery.

11
IA Design Goals

• Transparency. The users of the data shall not be
  aware of the physical location of the data.
• Confidentiality. Only the necessary data for the
  specific level of access will be returned by the
  archive.
• Authorization Access based on use (patient care,
  research, education) or site of user's origin.
• Consistency Assure users from all sites a common
  level of function and service across the region.

12
IA Design Goals (cont.)

• Flexibility Specify search criteria at any level
  of specificity and on any supported data fields.
• Replication To support disaster recovery, the
  archive shall mirror the relevant production
  database.
• Completeness (Integrity) For research, ensure
  100 of patients record fragments are retrieved.
• Appropriateness For education, assure that
  examples retrieved are appropriate to the users
  level.

13
IA High-Level View
Johns Hopkins Medical Inst.
UPMC Medical System
14
Schedule Service
WF Proxy
IMG SRC
CLIENT
15
(No Transcript)
16
(No Transcript)
17
(No Transcript)
18
HUBS Task 3

• Collaborative Telemicroscopy

19
CT Motivation

• Collaborative Telemicroscopy is a form of
  virtual microscopy that permits medical
  facilities to share and assimilate pathology
  images and data electronically throughout the
  Smart Region. It provides a high-bandwidth,
  low-latency interface that makes it practical to
  operate equipment, manipulate specimen samples,
  and react to the data as effectively as if
  physically present.

20
CT Project Goals

• to use the NGI VPN and standard metadata model
• to build a sustainable CT System
• to make it available across the four-state region
• Specifically, to provide
• decision support to pathologists
• real time image queries
• image feature matching
• intensive computing
• high resolution shared pathology images for
  clinical, research and academic purposes

21
(No Transcript)
22
CT Activities

• Establish Interconnectivity
• (1 NGI backbone, 3 applications, 5 sites)
• Virtual Microscope
• Johns Hopkins School of Medicine
• Content Based Image Retrieval
• Pittsburgh Supercomputing Center / University of
  Pittsburgh
• Distributed Telemicroscopy (remote robotic
  microscope) and Decision Support system
• University of Medicine and Dentistry of New Jersey

23
CT Activities

• Intelligent Image Archiving Component
• Develop a standard data model
• Incorporate the SNOMED vocabulary and
  DICOM-compliant imaging characteristics
• Utilize unique image spectral and spatial
  signatures to facilitate CBIR queries
• Automatically extract signatures (a vector
  a.k.a. tag) that will be used later

24
CT Activities

• Performance Evaluation Testing
• Initially launch on current Internet
• Measure performance (progressively)
• Baseline on vanilla Ethernet (100 Mbps)
• Upgrade some sites to vBNS, re-measure
• Upgrade all sites to VPN over vBNS, re-measure
• Tests will include
• multiple distributed clients simultaneously
  exploring several large data sets
• multiple interactive users acquiring, viewing and
  exchanging large images coordinated by the system

25
Related Applications

• Virtual Microscope
• Content Based Image Retrieval (CBIR)
• emphasis on prostate cancer
• Image-Guided Decision Support
• Visible Human Project

26
Virtual Microscope

• Digitally scan slides at high resolution
• Share them with remote institutions
• Provide a graphical interface for visual
  inspection analogous to typical analog viewing
  patterns, e.g. panning, zoom, etc.
• Done back in the early 90s (!)

27
Virtual Microscope
28
Content-Based Image Retrieval
29
CBIR Project Goals

• To provide
• a repository of high quality images with attached
  diagnostic notations
• tools for matching unknown images to the
  repository
• rank ordered diagnostic possibilities inferred
  from matching cases
• To discover diagnostic methods using HPC
• which are accessible via typical workstations /
  PCs
• which will become part of common practice

30
(No Transcript)
31
Emphasis on Prostate Cancer
32
Discover useful discriminators
33
Image-Guided Decision Support

• The system allows physicians to interactively
  review diagnostic images and to delineate regions
  containing structures which are either
  unidentifiable or are known to be key to the
  diagnosis
• Numerically, automatically, in real-time

34
(No Transcript)
35
Visible Human Project Goals

• Provide high performance visualization of the VH
  data to classrooms of anatomy students
• Provide data structures and compression
  algorithms to facilitate high speed retrieval of
  arbitrary VH slice views
• Provide network tuning for high speed delivery of
  visual data to distributed client stations

36
Arbitrary viewing alignments
37
EdgeWarp Prototype Navigator
38
Common Features

• Storage and rapid access of large quantities of
  medical image data
• Network delivery of custom views to each user
• Image feature extraction and pattern matching
• Remote utilization of HPC processing
• Network tuning for high speed interaction
• Use of collaboratory tools to facilitate
  efficient cooperation between remote partners
• See also http//telepathology.upmc.edu/hubs/

39

• High Bandwidth for the Masses

40
Web100 Motivation

• Even users of the highest speed networks dont
  always get gt100Mbps
• The limitations often do not lie in the network
  itself, but in the OS networking implementation
  and the applications that rely upon them.
• Network engineers can improve this, but their
  services can be expensive or unavailable.

41
Web100 Project Components

• Software Package
• Kernel modifications (per-session record
  keeping), AutoTuning, Tuned FTP (works w/o kernel
  mods), TCP gauges (users), Diagnostic Tools
  (developers)
• Support
• Early adopters (test sites) will have on-site
  support, training, and a feedback path to the
  developers.
• Vendor Liaison
• Goal is to get modifications adopted and
  supported by vendors.

42
AutoTuning

• Implement per-session TCP MIB in kernel
• Similar to UNIX netstat information, but
  per-TCP-session and more useful information
• Also publish and shepherd new TCP MIB through
  the MIB standards process
• and then
• Develop user-level algorithms that dynamically
  optimize the maximum TCP buffer size based on TCP
  congestion-feedback variables

43
Implementation Information

• 1,200 diff lines against Linux 2.2.14
• API is through /proc
• About 2 dozen variables right now
• All counters are cumulative
• Counters updated continuously in kernel /proc
  updates each time accessed
• One instance of data structure for each TCP
  session in /proc
• curses demo interface

44
Current Issues

• The US Presidents Information Technology
  Advisory Council (PITAC) has identified this
  project as one which clearly addresses the needs
  of the community.
• Vendor discussions
• Working with the Linux community
• Pursuing projects with Cisco, IBM, etc.
• Inter-Agency interest
• Ongoing discussions with the US Dept. of Energy
  and other agencies.

45
Sample MIB Database
128.182.61.238.22 lt-gt 128.182.61.156.1022
ESTABLISHED -----------------------
------------------------------------------------
------- PktsIn 1974 PktsOut
1951 Enabled DataPktsIn 972
DataPktsOut 1002 SACK
N AckPktsIn 1975 AckPktsOut
949 ECN N DataBytesIn 19823
DataBytesOut 74651 Timestamps
N DupAcksIn 0 PktsRetran
0 BytesRetran
0 ---------------------------------------
--------------------------------------- loss
episodes 0 cwnd 1453792
winscale rcvd 0 timeouts 0
max cwnd 1453792 rwin rcvd
986816 TO after FR 0 ssthresh
0 max rwin rcvd 986880
min ssthresh 0 winscale
sent 0 max
ssthresh 0 rwin sent 32120

max rwin sent 32120 ----------------------
----------------------------------------------
------- rto (ms) 20 rtt (ms)
1 mss 1448 Rate min rto (ms) 20
min rtt (ms) 0 min mss 1448 Out
(kbps) 0.1 max rto (ms) 20 max rtt (ms)
1 max mss 1448 In (kbps)
0.0 --------------------------------------------
--------------------------------- Overall
rate-controlling effects (only valid if we are
the sender) aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa
aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa
Receiver(S)topped,(A)pp,(B)ufsize /
Path(C)ongestion / Sender(b)ufsize,(a)pp
46
Work with Vendors

• Persuade commercial vendors to integrate this
  code base as quickly as possible
• else who would it benefit ?!
• Working with the Linux community to have Web100
  code included in standard Linux kernel releases

47
MetaComputing at the PSC

• It is
• BioMedicine
• Clinical Data and Analysis
• High-Performance Computing
• High-Performance Storage Delivery
• Advanced Networking Research
• For the benefit of
• the Scientific Community
• users everywhere