The Gift From Yesterday Matthew Lungren, MD MPH Assistant - PowerPoint PPT Presentation

The Gift From Yesterday Matthew Lungren, MD MPH Assistant Professor, Radiology, Stanford University School of Medicine #FDAPDSsymp | #AIinHealth

COMPUTER SCIENTIST TASKS ARE SIMILAR TO DOCTOR TASKS

“No acute intracranial “Saggital knee MRI with “No acute airspace abnormality” acute ACL tear” disease” “Heterogenously dense. “PET CT metabolically “Mild cerebellar BIRADS 2” active GE junction mass” atrophy”

OPEN SOURCE DATA SETS ARE THE PRIMARY CATALYST FOR MACHINE LEARNING ADVANCEMENTS

6 years 14 years 1991 1997 1983 Negascout Planning The Extended Book of Deep Blue Defeats Algorithm Developed Chess Games Dataset Kasparov Released

2 years 20 years 2009 2011 1991 8.6 Million documents Mixture of Experts Watson becomes world from Wikipedia and Algorithm Developed Jeopardy Champion Project Gutenberg Released to Public

6 years 26 years 2009 2015 1989 1.5 Million classified Microsoft Research Convolutional Neural images with 1,000 object surpasses human Network Created categories released to recognition public performance

MEDICAL DATA ARE A PUBLIC GOOD AND SHOULD BE OPEN SOURCED TO BENEFIT ALL

DATA are not Traditional Form of Property Not divided or consumed and easily replicated “Ownership” of data is an imprecise concept o Rights to control access o Rights to share of profit Value of data and information is relative o Exclusive access to data may give relative advantage o Advantage negated when others have access to the same data Credit: David Larson MD MBA

but… Medical Data are PEOPLE

Our position: Once clinical data have been used to provide care, the primary purpose for acquiring the data is fulfilled. For secondary use: Clinical data should be treated as a form of public good used for the benefit of all thorough open source research and education NOT for sale or exclusive partner access Credit: David Larson MD MBA

Rajpurkar& Irvin et al., PLOS Medicine, 2018 Rajpurkar& Irvin et al., MIDL,201Bien & Rajpurkaret al., PLOS Medicine,2018Park & Chute & Rajpurkaret al., JAMA NetworkOpen. 2019

Benchmark Performance with Consensus Expert Ground Truth 7,200+ teams worldwide competing in medical imaging challenges hosted at Stanford

Publications on AI in Radiology Pesapane F. et al. Eur Radiol Exp. 2018

Process for Open Source Medical Imaging Data Curate and label data Manual PHI review of all images Data user agreement and registration

ALTERNATIVE STRATEGIES TO OPEN SOURCE MEDICAL IMAGING DATA ARE COMPLIMENTARY NOT REPLACEMENTS

Detecting and Simulating Artifacts in GAN Fake Images (Extended Version) Xu Zhang, Svebor Karaman, and Shih-Fu Chang

Open source medical data Responsible ethically sound and benchmark competitions processes exist to make will accelerate innovation medical data available Alternative approaches to Publishing without sharing solve data availability are data is just marketing NOT complimentary science

Thank You Matthew P Lungren MD MPH Assistant Professor of Radiology Associate Director Center for Artificial Intelligence in Medicine & Imaging Stanford University School of Medicine mlungren@stanford.edu @mattlungrenMD #FDAPDSsymp | #AIinHealth