Automatic Pathology Software for Diagnosis of Non-Alcoholic Fatty Liver Disease (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: Jessica Silvaggi Senior Licensing Manager 1440 East North Ave. Milwaukee, WI 53202 Tel: 414-906-4654 Jessica@uwmrf.org UWMRF 2017 1
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 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, costfocused 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 UWMRF 2017 3
Liver Anatomy-Lobule Multiple lobules A single lobule 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 UWMRF 2017 4
Liver Biopsies for Pathological Assessment Biopsy sections are cut into different planes for staining H&E Stain Trichrome Stain White regions are identified from the stained biopsy sections from a patient UWMRF 2017 5
Pathologists use the NAFLD Activity Score (NAS) 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 UWMRF 2017 6
Examples of Histological Legions Normal Steatosis Lobular Inflammation Hepatocyte Ballooning UWMRF 2017 7
Approach for Machine Learning Study Patients Patient P H&E Stained Images Pathologis t Annotatio ns TC Stained Images Fibrosis Classifier Lobular Inflammatio n Classifier Image HE Image TC Portal Triad Identification Heuristic White Region Classifier Ballooning Classifier Ballooning Calculator Steatosis Calculator Lobular Inflammatio n Calculator Fibrosis Calculator B P S P L P F P UWMRF 2017 8
Classification of White Regions 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 UWMRF 2017 9
Method for Classification of White Regions 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 UWMRF 2017 10
Results of White Region Classification Accuracy Model Overall Accuracy P-Value Morphology 74.8% Advanced feature set 92.1% 0.0007 Supervised Machine Learning Results The results show that the algorithm is predicting the white regions with very high accuracy. UWMRF 2017 11
Results of White Region Classification-Cont d 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 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) UWMRF 2017 12
Model Percent Steatosis FLE038 FLE050 FLE013 FLE020 FLE041 FLE031 FLE032 FLE012 FLE046 FLE001 FLE045 FLE018 FLE037 FLE025 FLE016 FLE015 FLE047 FLE026 FLE034 FLE044 FLE022 FLE027 FLE036 FLE003 FLE004 FLE005 FLE051 FLE033 FLE035 FLE042 FLE028 FLE040 FLE009 FLE024 FLE017 FLE011 FLE048 FLE021 FLE007 FLE029 FLE049 FLE043 FLE023 FLE014 FLE008 FLE039 FLE030 Steatosis Grade Machine Learning Pathologist Grade Model Percent Steatosis 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 Correlation of Computed Percentage Steatosis with Pathologist Grade Relationship of Computed Percentage Steatosis with Pathologist Grade 26% 24% 22% 20% 18% 16% 14% 12% 10% 8% 6% 4% 2% 0% Avg. Pathologist Grade % Steatosis 3 2 1 0 30% 25% 20% 15% 10% 5% 0% y = 0.0252x 2 + 0.0067x + 0.003 R² = 0.9392 0 0.5 1 1.5 2 2.5 3 Pathologist Grade Patient Sample UWMRF 2017 13
Lobular Inflammation Classification 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 UWMRF 2017 14
Precision True Positive Rate Approach for Lobular Inflammation Classification: 95.6% accuracy (vs. 94.0% baseline) 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 Precision vs. Recall Curve 0 0.2 0.4 0.6 0.8 1 Recall 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 Receiver Operating Characteristics 0 0.2 0.4 0.6 0.8 1 False Positive Rate Feature Precision Recall ROC Area Lobular Inflammation 0.696 0.489 0.946 Not-Lobular Inflammation 0.968 0.986 0.946 OVERALL 0.952 0.956 0.946 UWMRF 2017 15
Precision True Positive Rate Hepatocyte Ballooning: Classification Results 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 Precision vs. Recall Curve Receiver Operating Characteristics 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.2 0.4 0.6 0.8 1 Recall 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.2 0.4 0.6 0.8 1 False Positive Rate UWMRF 2017 16
Computed Model NAS Overall Results for NAFLD Activity Score (NAS) 73.8% Correlation with Pathologists 5.0 Comparison of Computed NAS with Average Pathologist NAS 4.5 4.0 3.5 y = 0.7624x + 0.4561 R² = 0.6532 3.0 2.5 2.0 Computed Model NAS Linear (Computed Model NAS) 1.5 1.0 0.5 0.0 0 1 2 3 4 5 6 Average Pathologist NAS Circled region accounts for diagnosis by software of a worse case than that predicted by pathologist UWMRF 2017 17
Summary 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 UWMRF 2017 18
Next Steps: Product Development Company must convert Matlab algorithms into a production software environment for commercial use Estimated 500 hours of development to move this technology into production UWMRF 2017 19
31B Treats Ovarian Cancer in Mice Automatic Pathology Software for Diagnosis of Non-Alcoholic Fatty Liver Disease (OTT ID 1236) For more information contact: Jessica Silvaggi Senior Licensing Manager 1440 East North Avenue Milwaukee, WI 53202 Tel: 414-906-4654 Jessica@uwmrf.org UWMRF 2017 20