Scientific Session
Wednesday, October 5, 2011
7:30 am – 8:55 am
Locations:
Grand Ballroom 1
Kings Garden North
Kings Garden South/LeBateau
1
Wednesday, October 5, 2011
Session #1 - IMAGING
quantitative analysis of morphological tissue changes
and cell infiltration.
Grand Ballroom 1
A New Method for Analyzing Cellular Protein
Patterns From Routine Histopathology Images
An Entirely Automatic Method to Score Crypt
Loss and Infiltration From Pathology Slides in a
Mouse DSS Model of Colitis
Gustavo K. Rohde, PhD (gustavo.rohde@gmail.com),
Wei Wang, Cheng Chen, Dejan Slepcev, John A.
Ozolek
Cleopatra Kozlowski, PhD (cleopatk@gene.com),
Jason deVoss, PhD, Joseph Beyer, PhD, DVM, Lauri
Diehl, PhD, DVM
Carnegie Mellon University, Department: Biomedical
Engineering, Pittsburgh, PA
Genentech, South San Francisco, CA
Content:
We describe a new technology for analyzing the
morphology of cells and subcellular structures from
digital pathology image data. The new approach is
substantially different from existing technology in that
it allows for visualization and discovery of meaningful
differences in subcellular protein patterns in normal
vs. cancerous cells.
Content:
The DSS (dextran sodium sulfate) model of colitis is a
mouse model of inflammatory bowel disease, a
condition that affects more than 1 million people in
the US alone. Microscopic symptoms include loss of
crypt cells from the gut lining, and infiltration of
inflammatory cells into the colon. An experienced
pathologist requires over 6h for 100 H&E slides to
manually score loss of crypts in selected regions of
the mouse gut. In order to increase the efficiency of
scoring, we devised an entirely automatic method to
quantify loss of crypt cells in the whole slide.
Technology:
Following automated image segmentation, the
differences between normal and cancerous cellular
patterns are computed by transporting (morphing)
one pattern into another. The morphing operation
allows for the definition of a distance as well as a
path from one structure to another (see figure), to
which all other structures can be compared. The
differences between two sets of patterns can be
found by computing the path that best separates
them according to the standard p value. In contrast
to standard numerical feature-based approaches, the
method we describe allows for easy and meaningful
visualization.
Technology:
We used Definiens Developer (Definiens AG, Munich,
Germany) to design an image analysis algorithm.
Design:
The method relies on a combination of morphological
and intensity based features of every tissue area, as
well as a neighbor-based classification method.
Further, we developed another entirely automatic
method to count infiltrating macrophages,
neutrophils, and T cells in IHC of serial sections to the
H&E slides.
Design:
We apply our methods to decoding differences
between protein patterns (nuclear chromatin, as well
as golgi protein patterns) in normal and cancerous
cells. Groups of cells are then compared to elucidate
the differences in structure that are present in the
data. Nuclear chromatin patterns imaged using
Feulgen staining from 10 thyroid cases (normal
thyroid, follicular adenoma, follicular carcinoma), 5
liver cases (normal, fetal-type hepatoblastoma) and
two proteins (giantin and gpp130) in the golgi
apparatus of HeLa cells imaged with
immunofluorescence. Patterns are first normalized for
translation, rotation, as well as size (optional).
Results:
We found an overall correlation of > 0.7 between
independently assigned automatic and manual scores,
in approximately 800 DSS treated mice. The
correlations were robust across 12 studies, in spite of
large variations in slide quality and staining intensity.
Though the analysis requires approximately 40min of
CPU time per slide, cluster computing allows us to
analyze up to 500 slides overnight without any
manual intervention.
Conclusion:
Overall, our methods greatly facilitate the scoring of
mouse colon in DSS induced colitis, and enable
Results:
Results show that nuclear chromatin patterns in
normal liver vs. hepatoblastoma differ, and that the
2
most significant difference (in the sense of p value) is
that normal cells tend to have chromatin more
densely packed while cancerous cells tend to have a
more disperse distribution (Figure 1). Results
describing differences between normal thyroid,
follicular adenoma and carcinoma, as well as two
golgi protein patterns in HeLa cells, will be shown.
classification. Efficiency of CV increases as the need
to edit classification (i.e. technologist time) decreases.
We assessed our experience with CV by evaluating
both percentage of correct pre-classified cells out of
CV initial classification (CPC/CVI) and percentage of
correct pre-classified cells out of total number of
verified cells for different cell classes (CPC/VC) in a
university hospital with a large hematology/oncology
patient population.
Conclusions:
The spatial distribution of cellular patterns can contain
important biological and diagnostic information. The
new method described is able to process the image
data automatically with the end of facilitating
visualization of important differences between normal
vs. cancerous cells.
Technology:
CellaVisionTM DM96 (see Content).
Methods:
Using CV’s Database Query Tool, Version 4.2
(CellaVision AB, Lund, Sweden), we analyzed
classification accuracy of CV for white blood cells,
erythroblasts, platelets and artefacts over six months.
Results:
The CV classified 211218 events. For all events,
CPC/CVI and CPC/VC reached 94% and 93%,
respectively. Values for segmented neutrophils,
eosinophils, lymphocytes, monocytes, giant platelets,
smudge cells and artefacts were > 80% for both
CPC/CVI and CPC/VC. CPC/CVI and CPC/VC were <
80% for immature granulocytes (band neutrophil,
promyelocyte, myelocyte and metamyelocytes)
(differences usually within one stage of maturation).
Table 1. Additional cell types with lower accuracy.
Figure 1: discrimination between normal and fetalhepatoblastoma nuclei based on chromatin structure.
The histogram shows the proportion of cells in each
group (normal, cancer) that are most similar to the
image directly beneath the bar in the geodesic path.
p = 0.009.
Verified cell
% of total
events
(n=211218)
0.37%
CPC/CVI
48.85%
CPC/VC
80.26%
Lymphocyte
Variant
0.64%
61.60%
67.11%
Plasma
0.01%
6.31%
100.00%
Blast cell
0.31%
57.72%
64.74%
Erythroblast
1.29%
77.57%
98.38%
Cell type
Basophil
Experience With CellvisionTM DM96 For
Peripheral Blood Differentials In a University
Hematology/Oncology Center
Conclusion:

Reduced immature granulocyte classification
accuracy may be due in part to the
subjectivity in classification of these cells,
length of experience with the system and
individual expertise of the technologist.

Cells with low CPC/CVI and CPC/VC
comprised a minority of the cells and should
not significantly affect the technologist reclassification time.

CV serves as a clinically useful instrument in
performance of peripheral blood differentials.
M. Rollins-Raval MD, MPH
(rollinsravalma@upmc.edu), Lydia Contis MD
University of Pittsburgh Medical Center Division of
Hematopathology, Pittsburgh, PA
Content:
Rapid, accurate peripheral blood differentials are
essential to maintain standards of patient care.
CellaVisionTM DM96 (CellaVision AB, Lund, Sweden)
(CV) is an automated digital morphology and
informatics system used to locate, pre-classify, store
and transmit platelet, red and white blood cell images
to a trained technologist who confirms or edits CV cell
3
database. Finally, computer generated segmentations
were compared to human annotations using distancebased, region-based metrics and statistical agreement
indices.
An Accuracy Validation Framework for
Automated Pathology Image Segmentation and
Classification Algorithms
Jingjing Gao, PhD (jgao@emory.edu)1, Jun Kong,
PhD1, Fusheng Wang, PhD1, Tahsin Kurc, PhD1, Lance
Waller, PhD2, Joel Saltz, MD, PhD1
Results:
To assure the scalability and generalizability of the
proposed framework, four types of brain tumors,
namely, glioblastoma, oligodendroglioma,
oligoastrocytoma and astrocytomas with various
grades, were included in the scope of tests. Four subregions were randomly selected from clusters
stratified on imaging features for each slide. A total of
431 nuclei were segmented by a computationally
efficient algorithm and an experienced
neuropathologist. With the theory of survey sampling
and the measures from stratified randomly sampled
sub-regions, we estimated the means of three
measures (overlapping-ratio, centroid-distance, and
Hausdorff-distance) for each slide and provided a
95% confidence interval for each estimated mean.
Our tests suggest that 95% of the time the mean of
overlapping ratio for a selected slide is within the
range (0.791, 0.802). This implies concordance
between human and algorithm on nuclei
segmentation.
Emory University, 1Center for Comprehensive
Informatics and 2Department of Biostatistics and
Bioinformatics, Atlanta, GA
Content:
A major challenge in digital pathology is the large
inter-observer variability among pathologists. As
pathology has largely been a qualitative discipline,
significant disconcordance in human diagnosis is well
observed. Consequently, computer-based methods
have been developed to execute such tasks as nuclei
segmentation and classification. Due to the large
number of nuclei in pathology slides, the algorithm
accuracy and precision validation are often lack of
statistically theoretical support. In practice,
researchers usually make visual assessments by
overlaying the results on raw images. To address this
problem, we designed a generic human-guided
validation workflow both statistically defensible and
reproducible.
Conclusions:
We have proposed a design pattern for a
representative use case from the biomedical and
clinical research domain that aims at validating
segmentation and classification algorithms for
pathology images.
Technology:
To minimize the validation efforts required from
human experts, statistical stratified sampling
technique was adopted to select sub-regions from
whole-slide pathology images. Sub-region images
were clustered into strata by K-means analysis on
imaging features. Sample size was calculated based
on preliminary analysis. For human annotations,
original images were pre-loaded to the Aperio image
server to facilitate nuclei annotations by pathologists.
Human markups were then transferred into a presetup Pathology Analytical Imaging Standards (PAIS)
database with built-in query functions for spatial
based measurements. Numerous distance- and
region-based measurements along with the statistical
agreement test were performed at sub-image and
whole-slide level, making the algorithm validation
process definitive.
Design and Evaluation of a Virtual Reality
Microscope
Darren Treanor, FRCPath (darrentreanor@nhs.net),
Rebecca Randell, PhD, Rhys Thomas, PhD, Roy
Ruddle, PhD
University of Leeds, Pathology and Tumour Biology,
Leeds, United Kingdom
Content:
We have been developing a virtual reality (VR)
microscope that could be used for viewing digital
slides for diagnostic work. Using existing systems, it
can take up to 60% longer to perform diagnoses
using digital slides than on the light microscope. Our
goal is to develop a VR microscope that allows
diagnoses to be made quicker, but as accurately as, a
conventional microscope, an essential step in making
digital pathology suitable for routine use. We are
achieving this by combining ultra-high resolution
displays with VR technology and ‘intelligent
navigation’ techniques.
Design:
We design the workflow as follows. First, a set of subimages with pre-determined sizes were randomly
selected from strata identified by the clustering
analysis based on imaging features. Secondly, we
applied a variety of algorithms to selected sub-images
for nuclei segmentation and classification. Human
expert markups were also captured through an image
viewer (Aperio ImageScope Client), transformed into
PAIS document format and then loaded into PAIS
4
difficulty were digitized. WSI stored on a portable
hard drive were randomly reviewed on 4 different
monitors under standard conditions and also
examined with an OM by 6 pathologists with a
washout period of 2 weeks. Each pathologist
reviewed 30 slides based on their area of expertise.
Data was captured on a graded scale (high, medium
and low) for 8 parameters (see table).
Methods:
An iterative approach to development and evaluation
is being taken. A first version of our VR microscope
went through a phase of formative evaluation and the
feedback from this was used to refine the software.
We then evaluated the VR microscope using a mixed
factorial experimental design with technology (2
levels, conventional microscope and VR microscope)
and task (2 levels) as within-participant variables and
grade of pathologist (2 levels) as a betweenparticipant variable. We had 16 participants in the
evaluation.
Results:
Results from 900 encounters (720 electronic and 180
OM over 3 months) are shown in the table.
Parameter
Analyzed
Adequacy of
Tissue
Excellent
Image
Quality
High
Diagnostic
Confidence
Speed
(faster than
glass)
Unacceptable
Pixilation
Unacceptable
Refresh Lag
Obscured
Nuclear
Features
Results:
No significant difference in time taken to come to a
diagnosis was found between the conventional
microscope and VR microscope. We will also present
findings regarding diagnostic confidence, diagnostic
accuracy, pathologist preference, and navigation
patterns in the two conditions.
Evaluation of Different Display Modalities for
Whole Slide Images in Pathology
Gaurav Sharma MD (sharmag@med.umich.edu),
Gautam Sharma MD, Abid Shah PhD, Anil Parwani MD
PhD, Walid E. Khalbuss MD PhD, Sara E. Monaco MD,
Alka Palekar MD, Rajendra Singh MD, Liron
Pantanowitz, MD
LPTP
NMGM
MGM1
MGM2
OM
83.3%
94.4%
92.2%
90.0%
93.3%
32.2%
37.2%
83.3%
95.5%
93.3%
6.1%
15.0%
27.8%
58.9%
52.2%
0.0%
0.0%
34.4%
13.3%
NA
5.5%
2.8%
2.2%
47.8%
NA
9.4%
8.9%
3.9%
35.5%
NA
74.4%
26.6%
22.2%
17.8%
34.4%
Across SP and CP, the difference in image quality and
diagnostic confidence was statistically significant
(P<0.001) between medical grade modalities
(MGM1,2 and OM) compared to non-medical grade
modalities (LPTP and NMGM). Monitor brightness was
the most frequent (13.19%) screen adjustment.
University of Pittsburgh Medical Center, Division of
Pathology Informatics, Department of Pathology,
Pittsburgh, PA
Content:
Diagnostic accuracy and user confidence at
interpreting whole slide images (WSI) of digitized
surgical pathology (SP) and cytopathology (CP) slides
depends in part on the presentation and perception of
WSI. The aim of this study was to compare the user
experience with different display (computer monitor)
modalities.
Conclusions:
Non-medical grade displays had inferior user
experience compared to medical grade monitors and
the light microscope. Although comparable to the
microscope, limiting issues with MGM include
pixilation and prolonged refresh lag. With advances in
display technology, MGM would be the optimum
modality to electronically display WSI.
Technology:
Laptop monitor - IBM Thinkpad 14.1 inch (LPTP),
non-medical grade monitor - HP ZR24w 24 inch
(NMGM), medical grade monitor (small) - Eizo
FlexScan SX2462W 24 inch (MGM1), medical grade
monitor (large) - Barco Coronis Fusion 6MP 30 inch
(MGM2), optical microscope - Leica DM2000 (OM),
WSI scanner (Aperio Scanscope XT), WSI browser
(Aperio Imagescope 10.2), data storage - portable
hard drive (Seagate Free Agent 1 TB).
Design:
Glass slides from 60 cases (30 SP, 30 CP) of varying
5
Wednesday, October 5, 2011
staff via email. In addition, a flag has been
established within CoPath to alert the pathologist
at the time of signout if the patient has been
recently merged or has a potential merge pending
processing.
Session #2 - HOSPITAL / LAB OPS
Kings Garden North
Duplicate Patients: Considering Clinical
Priority in Resolving Patient Registration
Errors
Results:
Resolving duplicates is now done in the context of
clinical priorities, increasing the likelihood that
pathologists will know about prior material before
finalizing the diagnosis or ordering additional tests.
The improved algorithm allows better control over
sensitivity and specificity for the selection of
potential duplicates.
Peter Gershkovich MD
(peter.gershkovich@yale.edu), John Sinard MD,
PhD
Yale Medical School, Department of Pathology
Yale University School of Medicine, New Haven CT
Conclusions:
More duplicate cases are resolved prior to
rendering of the Pathological Diagnosis, providing
better continuity of care and reducing the number
of unnecessary tests.
Content:
Over the past ten years the number of duplicate
patient records in our AP-LIS has significantly
increased. The increase is caused by electronic
integration between multiple Hospitals and Clinics
in the absence of a Unique/Universal Patient
Identifier. We discovered that the existing built in
mechanism for identifying potential duplicate
patients does not take into account clinical
timeline and priorities. In addition it is inadequate
in discovering all duplicates. Finally, processing
duplicate patients is time consuming and needs to
be done by skilled staff members. As a result, last
year in up to 70 percent of merged patients in a
given month, pathologists were not aware of prior
material at the time of making a diagnosis.
An Automated System for Managing the
Storage and Retrieval of Organisms Isolated in
the Microbiology Laboratory
Richard Hill, MBA (hillr4@ccf.org), Kavous Roumina,
PhD, Walter H. Henricks, MD
Cleveland Clinic, Center for Pathology Informatics,
Cleveland, OH
Content:
A microbiology isolate is an organism that is “isolated”
from culture of a clinical specimen. Microbiology
laboratories have a need to identify and archive
isolates that represent organisms that may require
additional antimicrobial testing, are unusual, are from
critical sites, or are used in validation of new
equipment and assays. The Micro Isolate Storage
(MIS) system was developed to replace manual,
error-prone, time-consuming manual data
maintenance for stored microbiology isolates.
Technology:
The new software to identify duplicates is written
in Java and is encapsulated as a “Job” in RTSE, an
acronym for Repetitive Task Scheduling Engine,
developed in-house. RTSE uses a Quartz
framework for scheduling and running jobs; it is
integrated into the operation of our CoPathPlus
(Cerner DHT, Waltham MA) information system
through a direct access to the underlying
database.
Design:
The encapsulated “Job” is activated based on a
configurable schedule. It selects active cases from
CoPath and passes them to an algorithm with an
intuitive scoring system for evaluating potential
duplicate patients. Selected patients are listed with
corresponding demographic information, case
priority, number of prior specimens associated
with the potential duplicate, and the time since
working draft for the case has been printed. The
Duplicate Patient Report is sent to appropriate
Technology:
Application in Visual Basic 6.0 (Microsoft); relational
database (Access 2003, Microsoft); laboratory
information system (Sunquest).
Design:
In the laboratory information system, laboratory
technicians enter isolate data including organism
name, result codes, and patient demographics, and
also set a “flag” that the isolate is to be archived.
Data elements from flagged specimens are extracted
6
daily from the laboratory information system. The
system imports extracted isolate data into a database,
and presents data to users to review for
completeness and accuracy. The system then
automatically assigns storage (box) locations and
generates a report containing the isolate type,
accession number, specimen identification number
and storage location, in the exact sequence to be
archived. The system allows for specification of
storage box types and sizes, and tracks isolates
removed from or returned to storage. The system
can retrieve data based on patient demographics,
organism type, date range, and/or specimen
identification number.
Design:
A survey of medical examiner offices was conducted
in 2007 to evaluate their usage of forensic pathology
laboratory information systems. Information regarding
the systems that were in use and the user's opinions
of the systems were obtained and analyzed. The
same survey is being conducted in 2011 with the goal
of observing changes and trends in the market.
Results:
In 2007, a total of 78 medical examiner offices in 36
states responded to the survey. Among the
responding offices, 22% did not have a forensic
pathology laboratory information system, 33% were
using systems developed in house and 46% were
using a commercially available product. Only half of
the responders would recommend their system to
another office, with 46% of in-house and 57% of
vendor purchased offices making such a
recommendation. Only 63% of systems had some
access to images. The 2011 survey is currently being
conducted. The 2011 results will be evaluated,
compared to 2007 survey results and presented at
the conference.
Results:
Over 16 months, 9,859 isolates have been processed,
averaging 616/month. An efficient electronic
procedure replaced unnecessary manual work and
logs, increasing data accuracy and saving 4-8
hours/month labor in managing isolates. The system
also enables quick retrieval of isolates, saving an
additional 2-4 hours/month.
Conclusion:
The Micro Isolate Storage system has improved the
efficiency and reliability of managing microbiology
isolates. Manual archiving processes have been
eliminated, and the accurate retrieval of isolaterelated data and status using a variety of search
criteria is now possible.
Conclusions:
Preliminary results indicate that the market for
forensic pathology laboratory information systems
appears to remain immature. Most commercially
available systems continue to be represented by
adaptations of in-house systems, and many offices
continue to use and develop their own systems. The
small size of this market will continue to keep major
developers uninterested and restrict development of
more advanced systems despite user’s interest in
improved systems.
Implementation and User Satisfaction with LIS
in Forensic Pathology
Bruce P. Levy, MD (bplevy@partners.org)
Massachusetts General Hospital, Department of
Pathology, Boston, MA
Development and Deployment of PathologistDriven Electronic Billing Module: Overcoming a
Dysfunctional Workflow
Content:
Medical examiner offices have unique requirements
for the collection, preservation and reporting of data
and conclusions related to death investigations.
Forensic pathology laboratory information systems
need to appropriately address these issues. The
market for these systems appears to remains
immature, with the few available commercial products
being adaptations of home-grown, in-house products.
Philip J. Boyer, MD, PhD
(Philip.boyer@ucdenver.edu), Mark L. Gallen, Adam L
Kanallakan, MHA, Heidi M. Gullord Wendt, MSM, CPC,
Kathleen Zeleski, MS, Miriam D. Post, MD
University of Colorado Denver, Department of
Pathology and University Physicians, Inc., Aurora, CO
Content:
The anatomic pathology billing process at the
University of Colorado Hospital (UCH) had been
entirely paper-based with an inefficient workflow that
lead to delays of over 38 days on average from "date
of service" (DOS) to the initial "date of posting"
(DOP) into the UCH physician practice billing division
(University Physicians Incorporated, UPI) billing
Technology:
The significant commercially available products
include CME-VertiQ, Justice Trax, Quincy and BEAST.
In-house systems have been created using mostly
utilizing Microsoft Access or Microsoft Excel.
7
system. Originally, departmental billing experts did all
coding manually on paper, with paperwork returned
to the sign-out pathologist for review and manual
signature, with transfer of finalized paperwork to UPI
for data entry. We sought to design an electronic
billing (E-billing) system that would stream-line the
billing workflow.
Content:
Computerized provider order entry (CPOE) systems
offer clinicians the ability to order laboratory tests
electronically. Many CPOE systems support the
display of test-specific ordering alert messages and
institutions can leverage these alerts to implement
changes in test ordering guidelines or policies.
However, alerts may vary considerably in efficacy; a
better understanding of the characteristics and
mechanisms contributing to an alert’s success is
needed.
Technology:
Billing fields were established in a Microsoft (MS) SQL
2008 database used by Cortex, the UCD anatomic
pathology laboratory information system. A clientserver billing module was developed using MS VB.net,
with separate pathologist and biller clients. The
module is opened when a pathologist signs out a case
based on a call using MS VBScript from within an MS
Word .dot document. Pathologists access a "denial"
queue within the module by a desktop icon where
billing changes recommended by billers are accepted
or rejected.
Technology:
Our institutionally-developed CPOE system allows
clinicians to electronically search for potential
inpatient laboratory tests and order desired tests from
among those retrieved. The system captures highly
detailed data regarding user interactions.
Design:
We utilize our detailed user interaction data to
understand the mechanism by which CPOE alerts
influence test ordering. In this report, we evaluate a
CPOE alert created to support an institutional policy
restricting creatine kinase MB (CKMB) testing to
limited indications. This alert, displayed during CKMB
ordering, noted the new policy and required clinicians
to either cancel the order or enter an indication.
Design:
The original and post-E-billing implementation
workflows for surgical pathology cases were
compared with respect to the number of steps,
reasons for delays, staffing needs, and time-course
using LEAN process improvement methodology.
Primary goals included (1) reduction in the DOS to
DOP interval for one pathologist (MP) involved with
the billing process, (2) implementation of coding by
pathologists, with review by billers, (3) meeting of
Centers for Medicare and Medicaid Services (CMS)
guidelines, (4) electronic transfer of data to UPI, and
(5) elimination of paper- and manual-based
processes.
Results:
As shown in Figure 1, orders for CKMB decreased
rapidly following the implementation of the alert.
Searches for CKMB-associated search terms also
decreased significantly following the intervention.
The percentage of successful CKMB searches not
resulting in orders increased dramatically. These
results demonstrate the effectiveness of our alerting
strategy and provide insight into the mechanisms of
alert efficacy.
Results:
Implementation of an advanced version of the
physician-driven e-billing system reduced the DOS DOP interval by an average of 16 calender days (from
37.3 to 21.8 days) with electronic sign-off that meets
CMS requirements and a striking reduction in manual
processes and paper-based components.
Figure 1.
Conclusions:
The implementation of an e-billing system has
streamlined the billing workflow, nearly eliminating
paper from the process and significantly reducing the
delay between DOS and DOP.
A Novel Strategy for Evaluating the Effects of
an Electronic Test Ordering Alert Message
Jason M. Baron, MD (jmbaron@partners.org), Anand
S. Dighe, MD, PhD
Massachusetts General Hospital, Department of
Pathology Boston, MA
8
Conclusions:
The data we capture in the context of CPOE user
interactions helps distinguish the value of an alert in
providing just-in-time advice from its value in
providing longer-term education. For example, since
clinicians searching for CKMB but reconsidering in
response to an alert log a search without an
associated order, the percentage of searches not
associated with orders is suggestive of the immediate
advisory effects. Likewise, trends in total CKMB
search volume provide an indication of the level of
clinician education. Application of this analysis to a
larger cohort of alerts may provide an understanding
of which alert characteristics are most beneficial for
different effects and may lead to the development of
improved alert strategies.
wrapper, which combined data in these locations to
generate images with patient data, as required by
DICOM standards. The image and data were then
“wrapped” according to DICOM standards, transferred
to the PACS servers, and made accessible on an
institution wide basis.
Results:
In total, 26,966 gross images from 9,733 cases were
transmitted from the laboratory information system to
the EIS. The average process time for cases with
successful automatic uploads (n=9,688) to the EIS
was 98 seconds. Only 45 cases (0.5%) failed
requiring manual intervention. Uploaded images were
immediately available to institution-wide PACS users
(image 1). Since inception, user feedback has been
positive.
Integration of Digital Gross Pathology Images
for Enterprise-Wide Access
Image 1
Milon Amin, MD (aninm2@upmc.edu) Gaurav Sharma,
MD, Ralph Anderson, MS, Brian J Kolowitz, MS, MBA,
Anil V Parwani, MD, PhD, Rasu B. Shrestha, MD, MBA,
Liron Pantanowitz, MD.
University of Pittsburgh Medical Center, Department
of Pathology, Pittsburgh, PA
Content:
Sharing digital pathology images for enterprise-wide
use into a picture archiving and communication
system (PACS) is not yet widely adopted. Benefits of
sharing PACS images in institutional multimedia
repositories include amalgamation of clinical findings
for patient care and education. We share our solution
and three-year experience of transmitting such
images to an enterprise image server (EIS).
Conclusions:
Enterprise-wide PACS-based sharing of pathology
images is feasible, provides useful services to clinical
staff, and utilizes existing information system and
telecommunications infrastructure. However, shared
images require a proper “DICOM wrapper” for multisystem compatibility. Development of this process
requires significant programming skill and manpower.
Although our methods were developed in-house,
commercially available methods are emerging. Plans
are underway to begin sharing digital microscopic
pathology images with electronic health records and
EIS at our institution.
Technology:
Laboratory Information System (Cerner CoPath v.3.2
with PicsPlus image management system), Digital
Cameras (Nikon DS-U2), Pathology Image Server
(PicsPlus Server), EIS (Philips iSite v.3.5), Interface
Engine (HL7-Message-Routing), Enterprise Digital
Imaging and Communications in Medicine (DICOM)
Wrapper v.1.5.
Design:
Gross pathology images acquired by prosectors were
integrated with clinical cases into the laboratory
information system, and stored in JPEG2000 format
on a local image server. Automated daily searches for
cases with gross images were used to compile an
ASCII text file that was forwarded to the institutional
Enterprise DICOM wrapper server. Concurrently, an
HL7-based image order for these cases was
generated, containing the locations of images and
patient data, and forwarded to the Enterprise DICOM
9
Wednesday, October 5, 2011
Session #3 – IVD / Molecular / Training
Kings Garden South / LeBateau
Investigation of KOC Expression as a Potential
Diagnostic Biomarker for Pancreatic Ductal
Adenocarcinoma
Results:
Immunohistochemical analysis of KOC expression
showed heterogeneity of staining in PDAC but not in
normal ductal tissues. Overall average percentage of
KOC expressing cells across all cores in 3 TMAs for
PDAC samples was 77.9% (range, 0%-100%),
however in normal pancreatic ducts it was 0.32%
(range, 0%-10%) [p <.0001, Independent sample t
test]. Similarly, the average histoscore for PDAC
samples was 153 (range, 0-300), while for normal
ducts it was 0.52 (range, 0-18) [p <.0001,
Independent sample t test]. Finally, with a cut-off of
12.5% positivity of cells and at least weak staining, a
sensitivity of 83.9% and specificity of 100% was
achieved.
Asif Ali, MBBS(a.ali@beatson.gla.ac.uk)1, Victoria
Brown2, NB, Jamieson NB3,4, SM Denley3, RC Carter3,
CJ McKay3, Karin A. Oien KA1,5
1Institute
of Cancer Sciences, College of MVLS,
University of Glasgow, Glasgow
2Department of Pathology, Forth Valley Royal
Hospital, Larbert
3West of Scotland Pancreatic Unit
4University of Glasgow, Department of Surgery
5University Department of Pathology, Glasgow Royal
Infirmary, Alexandra Parade, Glasgow,UK
Introduction:
Pancreatic ductal adenocarcinoma (PDAC) is common
and aggressive with a 5-year survival of 2-3%. Initial
assessment of patients involves imaging and
diagnostic cytology at endoscopy. Pre-treatment
diagnosis on cytology samples can be difficult to
establish, and may be improved with diagnostic
biomarkers. KOC (k homology domain protein) is
highly expressed in PDAC but less in normal ducts.
Therefore, KOC could help to increase the diagnostic
accuracy of pancreatic ductal adenocarcinoma (PDAC)
versus benign or inflammatory tissue. Our aim was to
validate the expression of KOC in a local surgical
cohort.
Conclusion:
The results of this study with 100% specificity and
almost 84% sensitivity demonstrate that KOC
expression is essentially restricted to PDAC in tissue
samples. Thus it has the potential to increase the
diagnostic accuracy of cytological samples of PDAC,
but further studies on cytologic samples of PDAC and
normal ducts are required.
Selected Hematologic Malignancies Show
Interspersed Repeat Elements Clustering Near
Chromosomal Translocations Breakpoints
Nemanja Rodic, MD, PhD (nrodic1@jhmi.edu)1,
Kathleen Burns, MD, PhD2
Material and Methods:
Optimised antibodies for KOC were applied to 3
Tissue Microarray (TMA) slides containing a total of
256 cores (149 PDAC, 107 Normal Ducts) from 89
patients. Afterwards, we analysed the
immunohistochemical expression of KOC with the
intension of differentiating PDAC from normal
pancreatic ducts. We selected cores for comparison
purposes as negative, weak, moderate and strong for
staining. The expression was visualised as a
cytoplasmic stain of varying intensity and proportion
in distiller 2.1. Results were quantitatively measured
as a “weighted histoscore”, taking into account both
the intensity of staining and the proportion of cells
being positive [Histoscore= (negative=0×%cells +
weak=1×%cells + moderate=2×%cells +
strong=3×%cells)].
1Johns
Hopkins Hospital, Department of Pathology,
Baltimore, MD
2McKusick-Nathans Institute of Genetic Medicine,
Sidney Kimmel Comprehensive Cancer Center
Throughput Biology Center, Baltimore, MD
Content:
Chromosomal translocations are important to the
development of hematologic malignancies, though
their mechanisms are only partially understood.
Some chromosomal translocations involve V(D)J-type
recombination or selected sequences, such as CpG
dinucleotides. However, most chromosomal
translocations are thought to be random, presumed
to be due to ionizing radiation of reactive oxygen
species. A publically available database of selective
recurrent chromosomal translocations breakpoints is
10
maintained, though no bioinformatics pipeline exists
for its evaluation with respect to interspersed repeats.
MD PhD1,2,4,5, Carlos S. Moreno, PhD2,4,5, Daniel J.
Brat, MD PhD2,4,5
Technology:
We are evaluating associations between
recombination-promoting motifs, such as repeat
elements, and chromosomal breakpoints by
developing and applying a computational technique to
a large publically available breakpoint database.
1Departments
of Biomedical Informatics, 2Pathology
and Laboratory Medicine, 3Neurosurgery, 4Center for
Comprehensive Informatics, 5Winship Cancer
Institute, Emory University School of Medicine,
Atlanta, GA
Content:
Gene expression profiling has made possible the
further subclassification of human malignancies,
providing data that identifies new potential
therapeutic targets. The Cancer Genome Atlas (TCGA)
project, through cluster analysis based on gene
expression, has demonstrated four distinct molecular
subclasses of glioblastoma (GBM), designated
proneural, neural, classical and mesenchymal. The
mesenchymal class, associated with the worst
prognosis, is regulated by six transcription factors,
with C/EBP-β/δ and STAT3 identified as master
transcriptional regulators. Our study investigates
whether the tumor microenvironment influences GBM
molecular subclass and master transcriptional
regulator expression.
Design:
We have analyzed a database of over 470 breakpoints
from recurrent interchromosomal rearrangements in
human hematopoietic tumors. We started by
sequence alignment of each breakpoint to a unique
genomic coordinate. Next we used pattern recognition
software to visualize repeat element motifs
distributions near chromosomal breakpoints. We then
developed a computational approach to map and
calculate distances between chromosomal breakpoints
and nearest repeat element motifs. Finally, we
applied formal statistical method to identify
breakpoints that are closer to the motifs than they
should be given random chance.
Results:
Of nine breakpoint loci analyzed, we show that
human translocation breakpoints at Transcription
factor 3, C-abl oncogene 1receptor tyrosine kinase,
and Myeloid/lymphoid or mixed-lineage leukemia
variant gene loci show distinctive patterns of
clustering in proximity to repeat elements. Organizing
the breakpoints by the stage of development during
which the rearrangement occurs reveals that the
repeat element-type translocations occurred in the
Pro-B/Pre-B Cell and Lymphoid-Myeloid Hematopoietic
Stem Cells. Interestingly, we do not observe repeat
element sequence hotspots in lymphoid progenitor,
mature B, or T cells.
Technology:
Digitized TCGA GBM frozen section slides were
downloaded and areas of necrosis and angiogenesis
were selected and quantified using a humancomputer interface. Significance Analysis of
Microarrays (SAM) and Cox regression were used to
identify genes significantly correlated with
microenvironment. Ingenuity Pathway Analysis was
used to identify pathways enriched with necrosiscorrelated genes.
Design:
177 slides from 91 patients were analyzed. Extent of
necrosis and angiogenesis were determined as
percentage of total tissue. SAM and Cox regression
analysis were used to identify genes correlated at 5%
significance (with multiple hypothesis correction).
Immunohistochemistry was performed on a separate
set of 10 GBM samples for the transcription factors
C/EBP-β/δ, and STAT3.
Conclusion:
We show that chromosomal breakpoints cluster near
repeat elements in translocations incurred at the ProB/Pre-B Cell and Lymphoid-Myeloid hematopoietic
stem cell stage. The stage specificity and repeat
element targeting may be related to locus specific
epigenetic modifications within selected repeat
element sequence domains.
Results:
GBM samples with a high degree of necrosis were
more likely to be of the mesenchymal class.
Furthermore, the mesenchymal class had a higher
mean percent necrosis than the other classes. Among
non-mesenchymal classes, as extent of necrosis
increased, the transcriptome more closely resembled
that of the mesenchymal class. Angiogenesis did not
show a significant correlation with transcriptional
class. Cox regression analysis identified 2,422 genes
correlated with extent of necrosis. The transcription
factors C/EBP-β/δ, STAT3, FOSL2, bHLHE40 and
RUNX1 (regulators of the mesenchymal class) were
Tumor Microenvironment Influences
Glioblastoma Molecular Subclass and
Transcription Factor Expression
Christina Appin, MD (cappin@emory.edu) 2 , Lee AD
Cooper, PhD1,4, David A.Gutman, MD, PhD1,4,
Candace Chisolm, BA2, Yuan Rong, MD, PhD2, Tahsin
Kurc, PhD1,4, Erwin G. Van Meir, PhD3,5, Joel H. Saltz,
11
among those showing the tightest correlation with
necrosis. Integrated Pathway Analysis of genes
associated with necrosis identified enrichment of
canonical pathways including hypoxia signaling, Rac,
Rho, PI3K/AKT, NFκB, IL-6, ERK/MAPK, and
JAK/STAT. Immunohistochemistry showed strong
expression of C/EBP-β/δ, specific for the hypoxic,
peri-necrotic “pseudopalisading” tumor cells.
hypothesis. If residents are relying on the ‘forward
reasoning’ strategy, models (i) and (iii) will be better
predictors than models (ii) and (iv), respectively. The
reverse will be true if they are using the ‘hypotheticaldeductive’ model.
Results:
Preliminary analyses suggest no statistically significant
differences in prediction between models (i) and (ii).
However, significant differences seem to occur
between models (iii) and (iv), with model (iii) (using
identified diagnostic criteria to predict correctness of
newly reported diagnostic hypothesis) being a
stronger predictor (average area under the ROC = Az
= 0.820) than model (iv) (average area under the
ROC = Az = 0.776). Statistical analyses were carried
out using the Mann-Whitney U-test, P<0.05.
Conclusion:
This study demonstrates that necrosis is strongly
associated with the GBM mesenchymal subclass and
influences the expression of master transcriptional
regulators C/EBP-β/δ and STAT3.
How Do Pathology Residents Acquire Expertise
in the Reading of Dermatopathology Slides?
Conclusion:
Connectionist modeling has been used extensively to
explain cognitive processes in many domains. Here
we apply it for the first time to identify decision
making patterns among Pathology residents.
Claudia Mello-Thoms, MS, PhD
(mellothomsc@upmc.edu)
University of Pittsburgh, Department of Biomedical
Informatics, Pittsburgh, PA
Content:
There are many theories regarding how medical
residents acquire expertise in Pathology. We will focus
on two theories, the ‘forward reasoning’ strategy,
which advocates that identification of diagnostic
criteria is paramount to the formation of correct
diagnostic hypotheses, and the ‘hypotheticaldeductive’ model, which poses that residents first
generate a number of possible diagnostic hypotheses
and then extract diagnostic criteria in order to confirm
or to dismiss the generated hypotheses.
Facilitating Feedback and Education on the Hot
Seat Rotation
John H. Sinard, MD, PhD (john.sinard@yale.edu);
Neil Mutnick; Peter Gershkovich, MD, MHA
Yale University School of Medicine Department of
Pathology, New Haven CT
Content:
The “Hot Seat” rotation in anatomic pathology
residency training gives senior residents the
opportunity to review a large volume of slides from a
variety of services and to field clinician inquiries about
the cases. Traditionally, obtaining feedback on those
cases that they have seen has been difficult to
ensure, and residents have often resorted to making
lists of interesting cases and looking up these cases
days later in the AP-LIS to see how the attending
pathologist ultimately signed the case out.
Technology:
A custom-designed interface was used for this study.
Observers had the ability to zoom in on the slides (up
to 20x), pan, report diagnostic criteria, diagnostic
hypotheses and final diagnoses (or set of differential).
All interactions with the interface were time-stamped
and recorded.
Design:
Eleven Pathology residents read a set of 20 virtual
slides depicting inflammatory skin dermatitides.
Artificial Neural Networks were trained to represent
the decision making strategy of the residents. Under
this paradigm, four different types of models were
built: (i) using identified diagnostic criteria to predict
correctness of recently identified diagnostic criteria;
(ii) using reported diagnostic hypotheses to predict
correctness of identified diagnostic criteria; (iii) using
reported diagnostic criteria to predict correctness of
recently identified diagnostic hypothesis; and (iv)
using reported diagnostic hypotheses to predict
correctness of recently identified diagnostic
Technology:
We developed in-house a custom software solution
that is semi-integrated into the operation of our
CoPath (Cerner DHT, Waltham MA) information
system. The software is written in Java and deployed
as a web application using the Google Web Tool Kit.
Design:
The Hot Seat application is designed to provide: 1)
rapid access to all of the pertinent clinical and
ancillary information about a case (text, histology
status, events, clinician contact information,
photographs, prior specimens, etc.), thus, replacing
12
the need for the working draft; 2) an opportunity for
the resident to enter their impression of the case; 3)
the ability for the resident to compare (later and
efficiently) their impression with the signed out final
diagnosis, and to self-score the level of agreement;
and 4) summary statistics of the resident performance
by specialty area for self evaluation.
not have the hardware in place to accommodate such
conferencing. Departmental faculty and staff were
not in control of the process.
Technology:
Evaluations on the UCD campus and, when possible,
between teaching sites were undertaken with (1)
combined hardware and software solutions including
Polycom, Haivision, and LifeSize options and (2)
online software / conference call solutions including
GoToMeeting, Adobe Connect, and WebEx.
Results:
The application has been extremely well received by
the residents on the rotation. Resident evaluation of
the rotation has improved, and the residents feel they
can look at more cases more efficiently than the prior
method of flipping through pages in a logbook. Also,
because the interface allows rapid access to complete
information about a case, it has begun to be used
outside of the hot-seat rotation.
Design:
The various options were assessed for simplicity of
use including working within the VA system's lockeddown constraints, reliability, quality of transmitted
images and sound, flexibility including option to easily
switch between presenters at different sites during a
conference, ability to quickly create an unscheduled
conference, ability to incorporate live transmission of
whole slide images and microscopic slides, conference
archiving, and cost.
Conclusions:
Informatics tools can be used to improve the
educational experience of residents on a hot seat
rotation by tracking cases, providing feedback on
diagnosis, and enabling self-assessment of
performance.
Results:
The aging Polycom equipment at two sites would not
accommodate the transmission of high definition
signal, budget would not allow for replacement, and
three sites did not have the equipment to receive the
Polycom conferences. The budget for a replacement
multi-featured hardware-software system was not
available. The online software / conference call
solutions were similar, each with positives and
negatives.
Optimizing Pathology Teleconferencing in the
Disbursed Academic Medical Setting: Image
Quality, Flexibility, and Cost Considerations
Differentiate Options
Philip J. Boyer, MD, PhD1
(philip.boyer@ucdenver.edu), Lisa Litzenberger1,
Brandon Davis, BS1, Andrew B. Sholl, MD1, Frank
Moore, MD1, Andrew J. Rohrer, DO1, Mark L. Gallen1,
Fernando J. Castro-Silva, MD1, Francisco G. LaRosa,
MD1, Geza S. Bodor, MD2
Conclusions:
A combination of a GoToMeeting online session and a
concurrent conference call was found to be the least
complex option that allowed for outstanding image
quality, easy alternation of presenters between sites,
and minimal cost. Connections at each of the sites
were easily accomplished by residents and
departmental staff, for scheduled and ad hoc
meetings, and all six sites could easily join in on
conferences.
1University
of Colorado Denver, Department of
Pathology, Aurora, CO
2Denver Veterans Affairs Hospital, Denver, CO
Content:
The University of Colorado Denver (UCD) Department
of Pathology faculty, residents, and fellows are
dispersed among six training sites, necessitating a
teleconferencing solution for inter-site conferences
and meetings. A Polycom Distributed Media
Application system was in use, provided by UCD
Educational Support Services, requiring the use of a
complex, expensive hardware solution in primary
conference rooms with meetings established and
maintained by UCD staff, operated through a
"bridge," with staff juggling multiple concurrent
conferences. Virtually every Polycom session was
plagued by unacceptably poor image quality and
delays in or failure of transmission of images, with
frequent dropped connections. Three of our sites do
13