a novel ai-based algorithm for quan8fying volumes of ......introduction • there is a growing...
TRANSCRIPT
ANovelAI-BasedAlgorithmforQuan8fyingVolumesofRe8nalPathologiesinOCTScans
ePoster503AAO2019
SanFrancisco,CA
Michaella Goldstein MD, Moshe Havilio PhD, Omer Rafaeli MD, Anat Loewenstein MD
Financial Relationships
• Michaella Goldstein, MD – [E] Tel Aviv Medical Center – [C,L] Allergan, Bayer, Novartis – [L] Roche
• Moshe Havilio, PhD – [E] Notal Vision
• Omer Rafaeli, MD – [E] Notal Vision
• Anat Loewenstein, MD – [E] Tel Aviv Medical Center – [C,O] Notal Vision – [C,L] Allergan, Bayer, BeyeOnics, ForSight Labs, Novartis, Roche
Introduction • There is a growing clinical interest in automatic, Optical Coherence Tomography (OCT)
based, retinal pathologies detection[1] and fluid quantification [2] for retinal disease management.
• The Notal OCT Analyzer (NOA), a novel Artificial Intelligence based algorithm, was developed to detect and quantify the volumes of Age-related Macular Degeneration (AMD)-related pathologies, including retinal fluid and RPE elevations, and to follow these volume quantities over time.
Objectives • To evaluate the performance of the Notal OCT Analyzer (NOA) algorithm in detecting and
quantifying the volume of retinal fluid in individual macular OCT images
• To evaluate the performance of the Notal OCT Analyzer (NOA) algorithm in detecting and quantifying the volume of RPE elevations in individual macular OCT images
• To analyze sequential macular OCT images in individual eyes and populations of eyes to follow volume quantities over time
Methods Subjects: 172 eyes of 89 patients with exudative AMD Time Frame: Sequential OCT image cubes were collected for up to 10 years in each patient OCT Device: Heidelberg SPECTRALIS Retinal Fluid Detection:
For presence/absence of retinal fluid, Intra-retinal fluid (IRF), and Sub-retinal fluid (SRF) • 314 cubes were manually evaluated by a single human grader • 271/314(86%) of cubes were evaluated by NOA • 43/314 (14%) were excluded by NOA due to ineligibility
Retinal Fluid Quantification: 135 cubes containing retinal fluid, either IRF or SRF, were identified • The fluid in 135 cubes was manually delineated by a single human grader and its volume quantified • 155 cubes (135 cubes with fluid + 20 randomly selected cubes without fluid) were evaluated by NOA for
Quantification-eligibility analysis • 146 eligible cubes were fluid-volume quantified by NOA
RPE Elevation Quantification: • 146 cubes evaluated by NOA for eligibility in Fluid Quantification analysis were evaluated for RPE Elevations • In 146 cubes the largest RPE Elevation in a cube was manually delineated by a single human grader and its volume
was quantified • In 146 cubes the the largest RPE Elevation in a cube was identified by NOA
Retinal Volume Quantification in Sequential Images over time (Time Series Analyses): • Analyses were performed on up to 10 years of images per eye
Results – Fluid Detection: NOA compared to Human Grader High Sensitivity and Specificity
Detec8onofRe8nalFluid:IRFand/orSRF
Posi8ve Nega8ve Total
Posi8ve 127 10 137
Nega8ve 8 126 134
Total 135 136 271
0.94 0.93
Detec8onofIRF
Posi8ve Nega8ve Total
Posi8ve 86 20 106
Nega8ve 9 156 165
Total 95 176 271
0.91 0.89
Detection sensitivity of 94%±4% (CI of 95%)
Detection specificity of 93%±5% (CI of 95%) Detection sensitivity of 91%±6% (CI of 95%)
Detection specificity of 89%±5% (CI of 95%)
Detec8onofSRF
Posi8ve Nega8ve Total
Posi8ve 83 15 98
Nega8ve 11 162 173
Total 94 177 271
0.88 0.92
Detection sensitivity of 88%±7% (CI of 95%)
Detection specificity of 92%±5% (CI of 95%)
Results – Fluid Volume Quantification: NOA Compared to Human Grader High Correlation Coefficient
Fluid measured by average height (=fluid volume / cube area) in microns
Total Retinal Fluid
Correlation coefficient = 0.95
IRF
Correlation coefficient = 0.96
SRF
Correlation coefficient = 0.96
NOA IRF average height [µ] NOA SRF average height [µ] NOA Total Fluid average height [µ]
Gra
der I
RF
aver
age
heig
ht [µ
]
Gra
der S
RF
aver
age
heig
ht [µ
]
Correlation coefficient = 0.95 Correlation coefficient = 0.96 Correlation coefficient = 0.96
Results – Maximum Height of RPE Elevation: NOA Compared to Human Grader High Correlation Coefficient
Correlation coefficient = 0.92
Height measured in microns
Gra
der m
ax R
PE e
leva
tion
heig
ht [µ
]
NOA max RPE elevation height [µ]
Max RPE elevation height, CC=0.92, 146 clips
Maximum Height of RPE Elevation
Correlation coefficient = 0.92
Results – Time Series Analysis of Fluid Volume, Each Eye, Single Patient Benefit of Early Detection
0 500 1000 1500 2000 2500 3000 35000
2
4
6
8
10
12tot fluid 2451722C_OS
days from first scan
Aver
age
fluid
leve
l [µ]
0 500 1000 1500 2000 2500 30000
2
4
6
8
10
12#1, tot fluid 2451722C_OD
Days from first scan
Aver
age
fluid
leve
l [µ]
● Left eye was diagnosed on initial visit to the clinic with CNV and pre-existing fluid (“CNV at start” or late detection). A higher maximal fluid volume was reached even with treatment ● Right eye with dry AMD initially, which converted to CNV with fluid, detected during regular monitoring visits (“dry at start” or early detection). A lower maximal fluid volume was reached with treatment.
CNV at Start Dry at Start
MaxFluidVolume
0 20 40 60 80 100 1200
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
CD
F
max fluid found in 1 scan µ
CNV at start, n = 41, median = 17.02dry at start, n = 20, median = 3.25
Pvalue KS = 0.0001Pvalue MW = 0.0006
CNV at start Dry at start0
20
40
60
80
100
120
Bars
at
min
>1,
25,
50,
75,
max
<99
per
cent
iles max fluid at 1 scan
Pvalue KS = 0.0001 Pvalue MW = 0.0006
CDF = Cumulative Distribution Function
Max
imum
Flu
id H
eigh
t [µ]
Results – Time Series Analysis of Fluid Volume, Population Benefit of Early Detection
C
umul
ativ
e D
istri
butio
n Fu
nctio
n
KS - Two-sample Kolmogorov-Smirnov testMW - Mann-Whitney test
Subjects / Methods ● 41 eyes diagnosed on initial visit to the clinic with CNV and pre-existing fluid (“CNV at start” or late detection) ● 20 eyes with dry AMD initially which then converted to CNV with fluid, detected during regular monitoring visits (“dry at start” or early detection). The fluid in these eyes was validated by a human grader.
Key Findings● The maximum volume of fluid was significantly higher in CNV eyes with late detection compared to those with early detection.
0 20 40 60 80 100 1200
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
CD
F
corrected fluid level [µ]
begin, n = 245, median = 1.19end, n = 245, median = 0.19
Pvalue KS = 2.3e-010 Pvalue MW = 1.6e-008
180 200 220 240 260 280 300 320 340 360 3800
5
10
15
20
25
30
35
40
45#
scan
s
Retina average height [µ]
begin, n = 245, median = 248.16end, n = 245, median = 235.75
Pvalue KS = 4.1e-006Pvalue MW = 4.1e-008
KS - Two-sample Kolmogorov-Smirnov testMW - Mann-Whitney test
Key Findings● The average retina thickness and fluid volume were significantly higher at the beginning of treatment than at the end of treatment ● The fluid volume was also significantly higher at the beginning of treatment than at the end of treatment for the sub-groups of IRF and SRF (not shown)
IRF, MW p= 1e-004 SRF, MW p= 6e-005
Average retinal thickness [µ]
Results – Time Series Analysis of Retinal Thickness & Fluid Volume, Population Beginning vs. End of Treatment
Subjects / Methods ● 47 wet AMD patients diagnosed with retinal fluid ● >10 OCT cubes obtained over >1000 days ● First cube = at diagnosis with first treatment ● Last cube = last (or most recent) treatment ● “Begin”/”end” series = first/last 5 cubes if 10 or 11 cubes in total, first/last 6 cubes if > 12 cubes in total ● 245 cubes in each arm (“begin” series vs. “end” series)
Average fluid height [µ]
C
umul
ativ
e D
istri
butio
n Fu
nctio
n
Conclusions
• Notal OCT Algorithm (NOA) is able to detect retinal fluid with high sensitivity and specificity when compared to human graders
• NOA volume quantifications of fluid and RPE elevations are highly correlated with those by human graders
• NOA’s ability to provide volume quantifications of fluid and RPE elevations over time may be a clinically useful tool in evaluating retinal disease status and progression over time
References and Contact Information
References: [1] Chakravarthy, Usha, et al. "Automated identification of lesion activity in neovascular age-related macular degeneration." Ophthalmology 123.8 (2016): 1731-1736. [2] Guymer, Robyn H., et al. "Tolerating subretinal fluid in neovascular age-related macular degeneration treated with ranibizumab using a treat-and-extend regimen: FLUID study 24-month results." Ophthalmology 126.5 (2019): 723-734.
Author Contact Information: Michaella Goldstein, MD, Tel-Aviv Medical Center, [email protected] Moshe Havilio, PhD, Notal Vision, [email protected] Omer Rafaeli, MD, Notal Vision, [email protected] Anat Loewenstein, MD, Tel-Aviv Medical Center, [email protected]