automatic pathology software for diagnosis of non
TRANSCRIPT
5/18/17©UWMRF 2017 1
(OTT ID 1236)Inventors: Joseph Bockhorst and Scott Vanderbeck, Department of Computer
Science and Electrical Engineering , University of Wisconsin-Milwaukee; Samer Gawrieh, Department of Gastroenterology and Hepatology and
Director of Liver Transplantation, Medical College of Wisconsin
For further information please contact:
Automatic Pathology Software for Diagnosis of Non-Alcoholic Fatty Liver Disease
Jessica Silvaggi Licensing Manager
1440 East North Ave. Milwaukee, WI 53202
Tel: [email protected]
Shortfalls of current liver pathology assessment
Problems/Unmet Needs:
• Non-Alcoholic Fatty Liver Disease (NAFLD) is the most common liver disease in the U.S.
• Accurate distinction of mild vs. severe phenotypes is essential and liver biopsy is the only way to accurately diagnose and stage NAFLD
• Current NAFLD scoring method relies on manual pathologist assessment and grading of histological features
• Studies demonstrate substantial pathologist disagreement and inter-observer error (Gawrieh, S.,
Knoedler, 2011. 15: 19-24.)
• Poor reproducibility; intra-observer error
• Does not reflect the true continuous nature of “scoring” lesions
Technological Solution:
• The inventors have utilized a supervised machine learning technique for identification of white regions in liver biopsies with 92% accuracy (based on results from 47 patients)
• This is the first work to identify white regions using machine learning and automatically quantify lobular inflammation and hepatocyte ballooning
• This technology is automatic, faster, and more accurate than human pathologists
• Applications include diagnosis of NAFLD, assessment of candidate liver donors for transplants, and biopsy index database searching
5/18/17©UWMRF 2017 2
Road to Commercialization: Market, Intellectual Property, and Partnering
Market
• The global market for liver disease treatments was $400 million in 2009 and expected to increase to more than $700 million by 2014
• Computer aided detection and diagnosis (CAD), originally used in image analysis and the development of algorithms for image recognition, has evolved to offer full workflow management packages for a range of healthcare conditions
• Use of CAD can provide a more economically viable proposition for busy, cost-focused healthcare providers; CAD allows physicians to deal with enormous data sets more efficiently, making their job easier and in turn making physicians more accurate
Intellectual Property
• Copyrighted Software
Partnering
• We are looking for a partner to license the algorithm and software and to develop the technology into a final product for use by pathologists or other end users
5/18/17©UWMRF 2017 3
5/18/17©UWMRF 2017 4
Liver Anatomy-Lobule
A single lobuleMultiple lobules
• The bile duct, portal artery and vein, and central vein are some of the white regions that must be distinguished from the fatty white regions in a liver lobule
Liver Biopsies for Pathological Assessment
5/18/17©UWMRF 2017 5
H&E Stain Trichrome Stain
• Biopsy sections are cut into different planes for staining
• White regions are identified from the stained biopsy sections from a patient
Pathologists use the NAFLD Activity Score (NAS)
5/18/17©UWMRF 2017 6
• State-of-the-art scoring system for NAFLD (Kleiner et al., 2005)
• Based on 3 key histological features:
– Steatosis [0-3]
– Lobular Inflammation [0-3]
– Hepatocyte Ballooning [0-2]
• NAS = Steatosis + Lobular Inflammation + Hepatocyte Ballooning
Not
Steatohepatitits Borderline NASH
0 1 2 3 4 5 6 7 8
Examples of Histological Legions
5/18/17©UWMRF 2017 7
Steatosis Lobular
Inflammation
Hepatocyte
Ballooning
Normal
Approach for Machine Learning
5/18/17©UWMRF 2017 8
Pathologist
Annotations
Study Patients
Patient P
H&E Stained Images
White Region
Classifier
Lobular Inflammation Calculator
Lobular Inflammation Classifier
Ballooning Classifier
TC Stained Images
ImageTCImageHE
Portal Triad Identification
Heuristic
Ballooning Calculator
Steatosis Calculator
Fibrosis Calculator
FibrosisClassifier
SP FPBP LP
Classification of White Regions
5/18/17©UWMRF 2017 9
• Motivation
– Classifying all white regions provides a direct methods for quantifying steatosis
– Many anatomical land markers manifest as white or with a white center. This is useful for other tasks
Method for Classification of White Regions
5/18/17©UWMRF 2017 10
Hypothesis:
• Comparing previous methods that use hand crafted rules based purely on morphology, to a supervised machine learning approach that uses a richer feature vector representation
• Richer feature vector representations lead to more accurate classifications than ones based purely on morphology
– We compare 2 representations:
1. Morphology Only
2. Advanced feature set
Approach
• Represent each white region by a fixed length feature vector
• 10 fold cross validation
Results of White Region Classification
5/18/17©UWMRF 2017 11
Model Overall Accuracy P-Value
Morphology 74.8%
Advanced feature set 92.1% 0.0007
Supervised Machine Learning Results
Accuracy
The results show that the algorithm is predicting the white regions with very high accuracy.
Results of White Region Classification-Cont’d
5/18/17©UWMRF 2017 12
• Results again show that the algorithm is classifying white regions with very high accuracy. Of particular interest is how well it performs for macro and micro steatosis (fat)
Feature Precision Recall ROC Area
Bile Duct 0.900 0.849 0.981
Central Vein 0.732 0.788 0.938
Macro Steatosis 0.954 0.974 0.980
Micro Steatosis 0.930 0.927 0.967
Other 1.000 0.857 0.994
Portal Artery 0.786 0.767 0.973
Portal Vein 0.901 0.880 0.976
Sinusoid 0.924 0.899 0.982
OVERALL 0.921 0.921 0.974
Steatosis Grade Machine Learning
5/18/17©UWMRF 2017 13
• Percent Steatosis = total steatosis area / total tissue area
• Learn a different model for each patient using a leave one out approach
• The model is able to correlate best with the average grade
0
1
2
3
0%2%4%6%8%
10%12%14%16%18%20%22%24%26%
FLE0
38
FLE0
50
FLE0
13
FLE0
20
FLE0
41
FLE0
31
FLE0
32
FLE0
12
FLE0
46
FLE0
01
FLE0
45
FLE0
18
FLE0
37
FLE0
25
FLE0
16
FLE0
15
FLE0
47
FLE0
26
FLE0
34
FLE0
44
FLE0
22
FLE0
27
FLE0
36
FLE0
03
FLE0
04
FLE0
05
FLE0
51
FLE0
33
FLE0
35
FLE0
42
FLE0
28
FLE0
40
FLE0
09
FLE0
24
FLE0
17
FLE0
11
FLE0
48
FLE0
21
FLE0
07
FLE0
29
FLE0
49
FLE0
43
FLE0
23
FLE0
14
FLE0
08
FLE0
39
FLE0
30
Pat
ho
logi
st G
rad
e
Mo
de
l Pe
rce
nt
Ste
ato
sis
Correlation of Computed Percentage Steatosis with Pathologist Grade
Avg. Pathologist Grade
% Steatosis
y = 0.0252x2 + 0.0067x + 0.003R² = 0.9392
0%
5%
10%
15%
20%
25%
30%
0 0.5 1 1.5 2 2.5 3
Mo
del
Per
cen
t St
eato
sis
Pathologist Grade
Relationship of Computed Percentage Steatosis with Pathologist Grade
Patient Sample
Lobular Inflammation Classification
5/18/17©UWMRF 2017 14
Hypothesis/Approach:
• Lobular inflammation classification can be performed by supervised machine learning
• Provide a representative quantification for actual lobular inflammation regions
• Use a feature vector very similar to the one used for white regions
• 10 fold cross validation
Approach for Lobular Inflammation Classification: 95.6% accuracy (vs. 94.0% baseline)
5/18/17©UWMRF 2017 15
Feature Precision Recall ROC AreaLobular Inflammation 0.696 0.489 0.946
Not-Lobular Inflammation 0.968 0.986 0.946OVERALL 0.952 0.956 0.946
00.10.20.30.40.50.60.70.80.9
1
0 0.2 0.4 0.6 0.8 1
Pre
cis
ion
Recall
Precision vs. Recall Curve
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.2 0.4 0.6 0.8 1
Tru
e P
osit
ive R
ate
False Positive Rate
Receiver Operating Characteristics
Hepatocyte Ballooning: Classification Results
5/18/17©UWMRF 2017 16
• 98.9% accuracy (vs. 97.9% baseline)
Feature Precision Recall ROC Area
Hepatocyte Ballooning 0.912 0.542 0.983
Not-Hepatocyte Ballooning 0.990 0.999 0.983
OVERALL 0.989 0.989 0.983
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.2 0.4 0.6 0.8 1
Pre
cis
ion
Recall
Precision vs. Recall Curve
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.2 0.4 0.6 0.8 1
Tru
e P
os
itiv
e R
ate
False Positive Rate
Receiver Operating Characteristics
Overall Results for NAFLD Activity Score (NAS)
5/18/17©UWMRF 2017 17
• 73.8% Correlation with Pathologists
y = 0.7624x + 0.4561R² = 0.6532
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
4.5
5.0
0 1 2 3 4 5 6
Co
mp
ute
d M
od
el N
AS
Average Pathologist NAS
Comparison of Computed NAS with Average Pathologist NAS
Computed Model NAS
Linear (Computed Model NAS)
• Circled region accounts for diagnosis by software of a worse case
than that predicted by pathologist
Summary
5/18/17©UWMRF 2017 18
• Computer imaging techniques can effectively be used as a diagnostic aid for NAFLD / NASH
• Supervised learning techniques are superior to hand crafted computational rules
• Steatosis grading accurate enough for clinical use
• Next Steps
– Additional data to study lobular inflammation and ballooning
– Model refinement, tile size, etc.
– Features in other domains (FFT, Wavelets, etc)
– Impact of magnification of scan
Next Steps: Product Development
5/18/17©UWMRF 2017 19
• Company must convert Matlab algorithms into a production software environment for commercial use
• Estimated 500 hours of development to move this technology into production
31B Treats Ovarian Cancer in Mice
5/18/17©UWMRF 2017 20
Automatic Pathology Software for Diagnosis of Non-Alcoholic Fatty Liver Disease
(OTT ID 1236)
For more information contact:
Jessica SilvaggiLicensing Manager
1440 East North AvenueMilwaukee, WI 53202
Tel: 414-906-4654