CASE PRESENTATION: ACUTE RENAL FAILURE H

1
The Future of Diagnostic Medicine
Radiology-Pathology-Molecular Convergence
Michael Feldman, MD, PhD feldmanm_at_mail.med.upenn.
edu
2
Diagnostic Imaging Challenges in 2017

• Provide Diagnostically Informative Data from
  Non-Invasive or Minimally Invasive Techniques
• Provide Dense Multi-dimensional Data
• Provide Quantitative Data Which Integrates with
  Clinical and Molecular Information

3
Diagnostic Imaging today
MRI
In Vivo Optical
Pathology
Molecular
Proteomic
4
High Resolution Quantitative Imaging A Use
CaseFinding Significant Prostate Cancer
Non Invasive Imaging

• Machine Vision

Computer Assisted Diagnosis ( CAD)
Molecular Analysis of Tissues
5
High Resolution Magnetic Resonance Imaging ( MR
Microscopy) of Radical Prostatectomy Specimens
Prostatectomy specimen is placed on Endorectal
coil. Coil is then placed within 4T MRI
6
MRI of Prostatectomy 1 with 4T Magnet at 3 mm
Slice Thickness using 2D Fast spin echo
Specimen 1
Resolution 234mm x 234mm In plane 3 mm thick
slices 2D Fast spin echo Custom Bird Cage
Transmit receive coil TE 34 ms TR 3000 ms
BPH
8
9
Cancer
7
MRI of Prostatectomy 1 with 4T Magnet at 3 mm
MR Histopathology Correlation Specimen 1
10
BPH
Adenocarcinoma interrupts normal curvilinear
duct architecture
8
MRI of Prostatectomy 1 with 4T Magnet at 3 mm
MR Histopathology Correlation Specimen 1
V
American Journal of Surgical Pathology.
12(8)619-33, 1988
Adenocarcinoma interrupts normal curvilinear
gland architecture
Normal radial gland distribution
9
MRI of Prostatectomy 2 with 4T Magnet at 0.8 mm
Slice Thickness using 3D Fast spin echo
17
18
19
20
21
22
23
24
25
26
27
28
Resolution 234mm x 188 mm in plane 0.8 mm thick
slices 3D Fast spin echo Custom Bird Cage
Transmit receive coil TE 102 ms TR 3000 ms
10
MRI of Prostatectomy 2 with 4T Magnet at 0.8
mmMR Histopathology Correlation Specimen 2
17










Adenocarcinoma interrupts normal curvilinear
architecture

BPH with adenocarcinoma impinging on one edge

Fig. 8
BPH without carcinoma

11
MRI of Prostatectomy 2 with 4T Magnet at 0.8
mmMR Histopathology Correlation Specimen 2
27





Adenocarcinoma BPH
Capsular distortion by adenocarcinoma, Bulge
sign
12
Conclusions

• Magnetic resonance imaging provides signal
  contrast (MR stain) that allows for the
  identification of carcinoma and benign
  hyperplasia similar to a 2-4X optical lens.
• Features associated with carcinoma are similar to
  recognized low power histopathological features
  and include
• A. Interruption of normal curvilinear duct
  structure
• B. Intermediate T2 weighted signal with a smudged
  ground glass texture
• C. Interruption of capsular contour Bulge Sign

13
Computer Assisted Diagnostic (CAD) Analysis Can
Machine Vision See More ?
14
CAD Analysis Texture Features
Madabhushi A, Feldman M, Metaxas D, Tomaszewski
JE, Chute D Automated detection of prostatic
adenocarcinoma from high resolution Ex vivo MRI.
IEEE Transactions on Medical Imaging 241611, 2005
15
CAD Analysis Gradient Features
Madabhushi A, Feldman M, Metaxas D, Tomaszewski
JE, Chute D Automated detection of prostatic
adenocarcinoma from high resolution Ex vivo MRI.
IEEE Transactions on Medical Imaging 241611, 2005
16
CAD Analysis Gabor Filters
Madabhushi A, Feldman M, Metaxas D, Tomaszewski
JE, Chute D Automated detection of prostatic
adenocarcinoma from high resolution Ex vivo MRI.
IEEE Transactions on Medical Imaging 241611, 2005
17
CAD Identification of Prostate Carcinoma from 4T
MRI Images Using Multiple Classifier Ensembles
Madabhushi A, Shi J, Rosen M, Tomaszewski JE,
Feldman M. IEEE 2006
18
Machine Learning for Finding CAP in High
Resolution MR Images
4T MR Exam of Ex-Vivo Prostates with CAD Analysis
19
High Resolution MR-Histology Data Convergence
20
Mapping and Data Co-Registration
21
Diagnostic Imaging today
MRI
In Vivo Optical
Pathology
Molecular
Proteomic
New Signature for CAP in MRI
22
Gleason Grading IdentifiesFive Primary
Hisopathological ( Micron Resolution)
Patternsof Gland Growth
Identifying Significant Prostate Cancer
23
Pattern 3 vs. Pattern 4/5
24
Gleason Sum is Strong Predictor of Clinical
Progression
Gleason Sum 6
Gleason Sum gt7
Pinto et al. Urol Int 7202-208, 2006
25
Can CAD be Used to Find Gleason Pattern 4 CAP ?
Doyle S, Hwang M, Shah K, Madabhushi A, Feldman
M, Tomaszewski J. IEEE, 2007
26
An Image Analysis Approach to CAP Grading
Texture Features Examining Pixels
Graph Features Interrrogating Nuclei or Glands
27
(No Transcript)
28
Non-Linear dimension reduction followed by SVM
G3 vs. BE 85.43 G4 vs. BE 92.60 G3 vs. G4
95.80
Blue CAP3 Green CAP4 Red - Benign
29
Diagnostic Imaging today
MRI
Histologic signature Gr4 vs Gr 3
In Vivo Optical
Pathology
Molecular
Proteomic
CAP MRI signature
30
How to Meet Diagnostic Challenges of the Future
An Opinion

• In Finance , the Axiom is
• Follow the Money
• In Medicine
  Follow the Data

31
Characteristics of Data Used in Diagnostics
in the Future

• Extremely Large and Quantitative Data Sets
• Scaled Data
• Require Dimensionality Reduction
• Machine Learning
• Different Types of Data Converge

Path Image Data 2 GB
Microarray Data 30K 7.5 MB
32
Data is Scaled

• There is Informative Data at Every Resolution of
  Examination
• Informative Data at Each Level of Resolution can
  be Mapped to the Other Levels.
• The Data at Each Level of Resolution has Unique
  Attributes

33
OPTICAL 1x MAGNIFICATION RESOLUTION
10-3m TEXTURE COLOR VARIGATION SUSPICIOUS for
CAP
34
1.5 T MR T2 IN-VIVO 1x MAGNIFICATION RESOLUTION10
-3m HYPODENSE AREA SUSPICIOUS for CAP
35
4T MR T2 EX-VIVO 1x MAGNIFICATION RESOLUTION10-4m
HYPODENSE AREA INTERRUPTING NORMAL
CURVILINEAR GLAND ARCHITECTURE DIAGNOSTIC of CAP
36
OPTICAL 200X MAGNIFICATION RESOLUTION10-7m HAPHA
ZZARD PATTERN of INFILTRATING MICROACINI
DIAGNOSTIC of CAP GLEASON PATTERN 3
37
CGH DATA GENOMIC ANALYSIS RESOLUTION 10-7 to
10-9
38
MS2 Spectra from Paraffin Tissue of CAP
Fatty Acid Synthase in CAP
Mass Spec Protein Analysis Resolution 10-10
Caldesmon-1 in Benign Stromal Hyperplasia
39
The Curse of Multidimensionality

• In Large Multidimensional Space One Can Always
  Find Multiple Solutions to a Two Class Problem (
  CAP or Not CAP)
• In Order to Avoid This Multiplicity of Solutions
  One Must Reduce the Dimensions of the Data Used
  for Classification
• Manifold Learning Methods Reduce the
  Dimensionality of a Data Set from N Dimensions to
  M Dimensions where NgtgtgtM

40
Machine Learning
41
Diagnostic Imaging in 2017
101
Convergence of Large, Diverse, Scaled Data
Streams
10-10
42
Future Challenges

• Tools to manage and manipulate data types with
  following characteristics (business opportunity
  ???)
• Multidimensional
• Scaled
• Fused
• Registered
• Clinical
• Architecture must allow new methodologies to plug
  in using standards
• Molecular imaging
• Optical imaging
• Spectral imaging

43
The Team
Image Science Anant Madabhushi Scott Doyle M.
Hwang Shivang Naik Kinsuk Shah Jianbo Shi John
Chappelow James Monaco
Pathology Mike Feldman Deb Chute John
Tomaszewski Li Ping Wang
Radiology Mitch Schnall Mark Rosen