Ophthalmology Personalized Healthcare Program
Daniela Ferrara MD PhD
Senior Medical Director
Personalized Health Care
Clinical Science Ophthalmology
Genentech, Roche
Presented at the OIS Ophthalmology Innovation Summit | July 25, 2019 | Chicago, IL
Deep Learning
Advanced Analytics
Clinical
Imaging
Meaningful Data at Scale Vision Loss Prevention
Treat intermediate
disease early…
Personalized Therapy
Faricimab
Port
Delivery
System…To prevent
irreversible vision loss
‘Omics
Ranibizumab ITV
Goal to Treat Vision Loss and Preserve Vision
Real
World
Ophthalmology Personalized Healthcare Program
To Predict and Prevent Vision Loss
Meaningful
data at scale
Advanced
analytics
Smarter
faster
efficient
R&D
Personalized
care
& Improved
access
Impact
Digital
PathologyGenomicsDigital HealthReal World Data
Electronic
Medical RecordsImaging
Ophthalmology Personalized Healthcare Program
From Data to Analytics to Impact: Begins and Ends with the Patient
Algorithm Development Strategy
Real-world Performance Monitoring & Retraining
Train & TuneAlgorithms
Prototype
algorithms
on legacy data
Refine & EvolveAlgorithms
Mature
algorithms
on larger external
or real-world data
ValidateAlgorithms
DeployAlgorithms
Algorithms on
prospective
clinical data
Potential Intended Uses
Incre
asin
g C
om
ple
xity
For an individual patient:
Disease Detection Tool
Detects disease
Risk Prediction Tool
Provides risk prediction of disease conversion
Treatment Recommendation Tool
Recommends best management option,
either monitoring or treatment
Prevention Tool
Predicts disease conversion early enough to
recommend prophylactic treatment
Ophthalmology Personalized Healthcare Program
Algorithm Development Pathway
Engaging with
regulators,
Health care
providers and
payers
1. Smith AF. Br J Ophthalmol. 2010;94:1116-7
2. International Diabetes Federation. IDF Diabetes Atlas. Eighth edition 2017. https://www.idf.org/e-library/welcome.html. Accessed January 9, 2019.
Global prevalence of diabetes is projected to increase by ~50% in 20452
2017: 425 million 2045: 629 million
Ophthalmology Personalized Healthcare Program
Addressing High Unmet Medical Needs
Global projected prevalence of neovascular age-related macular degeneration
2010: 23 million 2050: 80 million1
1. https://www.idf.org/e-library/welcome.html. Accessed January 9, 2019. 2. Bressler NM et al. JAMA Ophthalmol. 2014;132(2):168-173. 3. Willis JR et al. Ophthalmic Epidemiol. 2018;25(5-6):365-372. 4.
Fenner BJ et al. Ophthalmol Ther. 2018;7(2):333-346. 5. Tran K et al. Curr Opin Ophthalmol. 2018;29(6):566-575. 6. Zimmer-Galler IE et al. Curr Opin Ophthalmol. 2015;26(3):167-172.
DME, diabetic macular edema; DR, diabetic retinopathy.
Optical Coherence Tomography (OCT)4-6
✓ 3D volume at the central retina: requires specialized acquisition
✓ Current gold standard for DME diagnosis
But… Modality is not always available for tele-ophthalmology
✓ 2D image of the retina: easy to be acquired✓ DR screening is currently largely based o CFP But… DME screening is limited in this modality
Color Fundus Photography (CFP)4-6
Ophthalmology Personalized Healthcare Program
Challenges in Screening for Diabetic Macular Edema
Normal Diabetic Macular Edema
Many individuals with Diabetes are not getting screened or receiving care that can prevent
visual impairment and blindness1-3
Nat Biomed Eng. 2018;2(2):158-164.
JAMA. 2016;316(22):2402-2410.
JAMA Ophthalmol. 2017;135(11):1170-1176.
Invest Ophthalmol Vis Sci. 2018;59(8):3199-3208.
ARTIFICIAL INTELLIGENCE
A program that aims to simulate
human intelligence
MACHINE LEARNING
An AI process in which the
performance of algorithms
improves over time with more
data exposure
DEEP LEARNING
Subset of machine learning
that uses multiple layers of
neural networks to learn
from a large volume of data
1. Hogarty DT et al. Clin Exp Ophthalmol. 2018 Aug 28 [Epub ahead of print].
AI, artificial intelligence; AMD, age-related macular degeneration.
Ophthalmology Personalized Healthcare Program
Deep Learning: Machine Learning for Knowledge Discovery
Deep learning offers a novel methodology to address scientific questions,
but its interpretability represents an open challenge1
1. Ting DSW et al. Br J Ophthalmol. 2019;103(2):167-175.
Input Processing Output
Answer to a
scientific
question
Raw Image Convolutional Neural Network
Ophthalmology Personalized Healthcare Program
Deep Learning: Potential for Macular Thickness Assessment
Transfer learning cascade is a powerful strategy to make deep learning feasible in a real-world setting
1. Russakovsky O et al. Int J Comput Vis. 2015;115(3):211-252. 2. Kaggle. Diabetic retinopathy detection. 2017. https://www.kaggle.com/c/diabetic-retinopathy-detection. Accessed January 31, 2019.
CST, central subfield thickness; DR, diabetic retinopathy; OCT, optical coherence tomography (*time-domain).
Network learns to recognize
natural images
Network specializes in
understanding fundus photos
Transfer Learning from ImageNet1
Transfer of “Knowledge”
KaggleDR Color
Fundus Photo Dataset2
KaggleDR challenge:
is it severe DR?
Inception-V3
YES
NO
CST 400 µm (Yes vs No)?
CST 250 µm (Yes vs No)?
Actual CST Value?
RIDE/RISE Color Fundus Photo Dataset
Deep learning (Inception-V3 architecture)
Convolution
Pooling
Softmax
Other
Convolution
Pooling
Softmax
Other
Ophthalmology Personalized Healthcare Program
Transfer Learning Cascade Strategy in Deep Learning
Deep learning applied on 17,997 fundus
photos from RIDE & RISE phase 3
DME trials to predict OCT-equivalent*
measures of macular thickening
1. Arcadu F, Benmansour F, Maunz A, et al. Deep learning predicts OCT measures of diabetic macular thickening from color fundus photographs. Invest Ophthalmol Vis Sci. 2019;60:852-857.
CST, central subfield thickness; AUC, area under the curve; n, number of cases; OCT, optical coherence tomography (time-domain); OP, operating point.
CST 250 µm
AUC 95% CI n
All validation samples 0.86 0.81–0.90 307
High-quality color fundus
photos without laser scars0.97 0.89–1.0 24
CST 400 µm
All validation samples 0.84 0.79–0.88 342
High-quality color fundus
photos without laser scars0.94 0.82–1.0 28
AUC: 0.86
Sensitivity OP: 80.3
Specificity OP: 78.9
Central Subfield Thickness (CST) 250 µm
(All Validation Samples)
The performance of the model increased with algorithm training
on fundus photos of high quality and without laser scars
Deep learning model successfully identified color fundus photos with Diabetic Macular Edema
and can support broader screening efforts
Ophthalmology Personalized Healthcare Program
Pilot Project on Diabetic Retinopathy
Central Subfield Thickness (CST)
R2 95% CI n
All validation samples 0.57 0.48–0.64 307
High-quality color fundus
photos without laser scars0.74 0.49–0.91 17
CST 250 µm CST 400 µm
1. Arcadu F, Benmansour F, Maunz A, et al. Deep learning predicts OCT measures of diabetic macular thickening from color fundus photographs. Invest Ophthalmol Vis Sci. 2019;60:852-857.
Ophthalmology Personalized Healthcare Program
Pilot Project on Diabetic Retinopathy
Attribution maps highlighting “hot spots” on fundus photos providing insight into:
• What the algorithm “sees”
• How the deep learning model develops its output
• Biological plausibility of features identified by the model
Ophthalmology Personalized Healthcare Program
Pilot Project on Neovascular Age-Related Macular Degeneration
Machine learning algorithm predicted response to anti-VEGF treatment and visual outcomes
and can support physician’s decision for clinical management
Ophthalmology Personalized Healthcare Program
Pilot Project on Geographic Atrophy
Assessment of potential prognostic variables in GA lesion occurrence and progression
and can enable research and development with smarter clinical trials
Genentech/Roche robust ophthalmology pipeline, rich longitudinal datasets,
innovative partners and assets, and strong analytics capabilities supports our program and
invites strong collaboration with the retina community and other expert leaders in the field
Ranibizumab ITV
Faricimab Port Delivery System
Robust ophthalmology pipeline
with new drug targets
and delivery systems
~10k patients and ~3M
fundus images
from legacy trialsUnique assets and
collaborations
Ophthalmology Personalized Healthcare Program
Collaborations to Achieve Meaningful Results for Patients