Imaging Beyond RadioShack Autonomous Robots, Digital Stains and the Pathology PACS

1
Imaging Beyond RadioShackAutonomous Robots,
Digital Stains and the Pathology PACS

• John Gilbertson MD
• Director of Research Development
• The Center for Pathology and Oncology Informatics
• UPMC Health System
• Pittsburgh PA
• GilbertsonJR_at_msx.upmc.edu

2
The Promise of Computational Morphology

• Morphology matters
• If you can image, you can apply computational
  power and network connectivity to the study of
  morphology and the practice of anatomic pathology
• The ability to easily digitalize morphology and
  integrate it with traditional medical data is
  consistent with our clinical, educational and
  research mission
• What if we could effortlessly image every slide
  at high resolution and associate those images
  with with our LIS text, gene and protein
  expression data?

• Todays talk will center around
• The problems with single frame imaging in
  pathology
• How to achieve fast, automated whole slide image
  capture
• The implications of ubiquitous whole slide
  imaging for pathology
• The goal is to convince you that is possible
  and reasonable - to image all of our slides at
  high resolution.

3
The Positive Legacy of the Single Field Era

• Digital imaging can be diagnostic
• Telepathology works
• Many pathologists feel imaging is beneficial
• The LIS industry has become interested
• Imaging has been a driver for Pathology on the
  Web.
• There is a great deal of interest amongst
  Pathologists

4
What is wrong with Microscopic Imaging in
Pathology?

• In 1994 UPMC purchased 10 microscopic capture
  stations, upgraded its network, installed a
  centralized imaging server, upgraded all the
  pathologists PCs and Monitors
• UPMC also purchased 2 digital macro cameras for
  the cutting rooms
• What happened is instructive

5
What is wrong with Microscopic Imaging in
Pathology?

• Within two years, every large gross specimen at
  UPMC (all our hospitals) was getting imaged,
  archived and made available with the report.
• Less than one percent of slides get an image.
• If we tried to remove gross imaging there would
  be a rebellion, if we turned off microscopic
  imaging there might be quiet satisfaction.

6
Gross Imaging

• Documents the entire specimen
• The best documentation of the gross specimen
• Useful at signout
• Clinicians (and Pathologists) can relate to the
  image
• It is a fairly easy process (and much easier than
  film)
• Can be operationalized - There can be rules, so
  PAs can do the imaging
• Pathologists need not be involved

7
Microscopic Imaging

• Does not completely document the entire slide (or
  specimen)
• It is not only documentation of the specimen
• Limited Usefulness - without text and context
• Subsampling forces the pathologist to be involved
  and makes it hard to operationalize
• Microscopic digital imaging is a complex
  system/process that is hard to do well

8
Can Microscopic Imaging Be Improved?

• Can the process better represent the slide and
  the microscope
• Can it be operationalized
• Can it be made horribly simple - (can we
  automate it)
• Can we get the pathologist out of the loop -
  (can we make it like radiology)
• Can it be better integrated into the LIS data
  environment
• Can it be more maintainable and scalable. (can we
  centralize it)
• Three existing technologies Virtual Microscope,
  HL7 feed and the WWW
• One new technology the High Speed, Autonomous
  Imaging Robot

9
The Virtual Microscope

• A virtual microscope images the entire tissue
  section, field by field, and then knits the
  field together to form a seamless montage
• One can either capture at multiple magnifications
  or capture at high resolution and bin pixels to
  generate lower resolution
• Software then allows the user to pan and zoom
  across is multi-resolution montage just like one
  would examine a slide under a microscope

• To my knowledge, initial work done by Joel Saltz
  (Hopkins) and Bacas Instrument Co.
• http//www.cs.umd.edu/projects/hpsl/ResearchAreas/
  vm.htm
• There are many others
• http//neuroinformatica.com/mbfvs/index.html
• http//demo.interscopetech.net/gallery/Gallery.asp
  

10
Digitizing a Slide - File Size and Capture Time

• Given base resolution of 0.33 um/pixel how many
  pixels are required to cover one square cm of
  tissue?
• 9 pixels / square um
• 9 million pixels / square mm
• 900 million pixels / square cm
• each pixel has 24 bits
• 2700 million bytes / square cm
• 2.7 GB
• 2.7 x 1.33 3.5 GB / square cm
• 101 compression 350 MB

• Given at Primary Magnification of 20x and a 2/3
  inch CCD how many fields are in a one square cm
  of tissue?
• FOV 0.44 x 0.33 mm
• FOW 0.145 sq. mm
• 700 fields per square cm
• 12 minutes at one field/second
• Most systems require significantly more time

11
What if we could build a machine that would image
an entire slide at very high resolution,
automatically and in one minute What if this
machine had a data link to the LIS, could read
slide number labels, and associated slides into
cases and link them to LIS text and make th
information available to pathologist securely on
a intranet The automated and in one minute
were tricky (others are fairly straight forward)
12
Disclosure
Interscope Technologies builds Robotic Imagers I
am a Founder, CTO, and Principle in
Interscope I am not a disinterested
party! Pathology needs more than one Robotic
Imaging Company
13
An Autonomous Imaging Robot

• An mechanism for Automated, High Throughput, high
  resolution imaging of whole slides and their
  integration into the clinical record
• Automated Give it a slide and walk away
• Fast 10 minutes/slide at 333 nm/pixel and
  getting faster
• High Volume 1 to 200 slide/batch
• Integrated into any LIS system through bar codes
  and a HL7 interface
• Images are standard TIFF (with optional JPG
  compression) in a multi-resolution format

• The system is an autonomous robot that handles,
  identified, focuses and images entire slides very
  rapidly using a montage image capture technique.
• The robot identifies the slide from a bar code,
  finds tissue is on the slide, focuses, flat
  fields and takes high resolution images, field
  after field, until all the tissue has been
  imaged. At the same time it creates a large,
  multi-resolution data set, stores it and
  associates it with LIS data from a HL7 feed

14
Speed

• Limitations on image capture speed
• The size of an individual frame (FOV)
• Size of the CCD Optical Magnification of System
• The Camera Frame Rate
• CCD Integration Time CCD Readout Time
• Image Processing Time
• Flat fielding, Bayer Patterns, Compression, etc
• Time required to write the image to disk
• How fast one can move the slide
• Stage Precision, CCD Integration Time
• Focusing Time
• Required Resolution (N.A.)
• With faster Camera Frame Rates and lighter,
  stiffer stages I expect image time to approach
  one minute

15
Pathology PACS

• Clinical Applications
• Telepathology
• Internal Consult
• Q/A Q/C
• Distributed Department
• Conferencing / Mentoring
• Training/Standardization
• Sign out
• Frozen?
• In Vivo Imaging?
• Goal is to have widespread capture of clinical
  slides up at UPMC in one year?

• One can implement a slide handling robot to
  feed slides to the imaging robot
• The system could run in batch mode
• Take slide from the stack
• Put it on the imager
• Read a bar code
• Image slide
• Send image to server
• Associate images into cases
• Pull LIS data from a HL7 feed...

16
Digital Staining and Computational Morphology

• A digital image can be the input for a computer
  program
• One could consider image analysis as a type of
  special stain
• Sharpening
• Contrast and Brightness
• Finding Fibrosis
• Gleason Grading
• Image Segmentation
• As images become ubiquitous, we will see an
  explosion of image analysis programs
• This will give us new ways of describing and
  studying morphology

17
Conclusions

• Current imaging technology is limited by
  subsampling, complexity, lack of integration and
  and the need for direct pathologist input.
• Over the next several years we will begin to see
  a series of machines that will solve these
  problems and have the potential to truly make
  imaging useful and ubiquitous.
• The onset of large scale imaging should set the
  stage for a revolution in computational
  morphology.

18
People

• Many people are involved in this effort
• Patty Feineigle
• Jeff Beckstead
• Yukako Yagi
• Art Wetzel
• Eric Schubert