Research

Original Investigation

Automated Ptosis Measurements From Facial Photographs

Zachary M. Bodnar, MD; Michael Neimkin, MD; John B. Holds, MD

Supplemental content at
IMPORTANCE Measurements of the margin reflex distances 1 and 2 are crucial for the surgical jamaophthalmology.com
planning of ptosis repair and blepharoplasty. Facial photographs annotated with automated
measurements of eyelid position could provide objective, accurate, and reproducible
documentation of these features.

OBJECTIVES To describe a software algorithm for determining the margin reflex distances 1
and 2 from facial photographs and to evaluate its agreement with manual measurements of
the margin reflex distances 1 and 2.

DESIGN, SETTING, AND PARTICIPANTS Observational study at a single-surgeon oculoplastic

private practice among 55 eyes of 28 adult volunteers. The study dates were July 30, 2014, to
September 12, 2014. The dates of our analysis were October 12, 2014, to June 18, 2015.

MAIN OUTCOMES AND MEASURES Agreement between manual and automated

measurements of the margin reflex distances 1 and 2.

RESULTS Among 55 eyes of 28 participants, automated margin reflex distance 1

measurements were strongly correlated with manual measurements (r = 0.97; 95% CI,
r = 0.95 to r = 0.98; P < .001). The bias of automated margin reflex distance 1 measurements
was 0.03 mm (95% CI, −0.06 to 0.12 mm), with 95% confidence limits of −0.66 and 0.71 mm.
Automated margin reflex distance 2 measurements were strongly correlated with manual
measurements (r = 0.96; 95% CI, r = 0.93 to r = 0.98; P < .001). The bias of automated
margin reflex distance 2 measurements was 0.13 mm (95% CI, 0.03-0.22 mm), with 95%
Author Affiliations: Department of
confidence limits of −0.54 and 0.80 mm.
Ophthalmology, Saint Louis
University, St Louis, Missouri (Bodnar,
CONCLUSIONS AND RELEVANCE Automated ptosis measurements produced by our Holds); Department of
software algorithm compare favorably with manually performed clinical measurements. Ophthalmology and Visual Sciences,
An automated, photography-based system could provide an archival and highly Washington University in St Louis,
St Louis, Missouri (Neimkin).
reproducible means for obtaining the margin reflex distances 1 and 2 and other facial
morphometric data. Corresponding Author: Zachary M.
Bodnar, MD, Department of
Ophthalmology, Saint Louis
JAMA Ophthalmol. 2016;134(2):146-150. doi:10.1001/jamaophthalmol.2015.4614 University, 1755 S Grand Blvd,
Published online November 25, 2015. St Louis, MO 63104
(bodnarz@slu.edu).

T
he standard clinical evaluation of ptosis includes manual
measurements of eyelid positions to quantify the degree
of ptosis and its effect on a patient’s vision. This value is
most often expressed as the margin reflex distance 1 (MRD1), de-
fined by the vertical distance between the upper eyelid margin
and the corneal light reflex, which is the specular reflection at
the corneal apex from a light source that is aligned with the vi-
sual axis (eg, a penlight). Similarly, the margin reflex distance 2
(MRD2) is the vertical distance between the corneal light reflex
and the lower eyelid margin. Another way that ptosis is com-
monly quantified is to plot the area of visual field deprivation that
results from eyelid malposition by performing Goldmann perim-
etry, once with the eyelid in its natural position and again with
the eyelid taped up to simulate the results of surgical correction.
Problems that limit the usefulness of these methods include
operator dependence, subjectivity, patient movement, and cog-
nitive inability to participate in testing (eg, in children or cogni-
tively impaired adults). The capability to objectively extract mea-
surements of the MRD1 and MRD2 from a single photograph of
a patient, obtained in the clinic with an inexpensive digital cam-
era, could mitigate many of these difficulties.
A software algorithm developed by one of us (Z.M.B.) can
identify the corneal light reflex and eyelid margins in frontal
photographs of human faces and automatically calculate the
MRD1 and MRD2. The objectives of this study were to intro-
duce our method for computing automated ptosis measure-
ments from digital photographs and to compare it with stan-
dard manual clinical measurements.
Software Algorithm
Our software algorithm accepts as input a single digital
photograph in any one of several standard file formats (JPEG,
TIFF, etc) and an optional set of calibration matrices as follows:
fx 0 cx
C= 0 fy cy
0 0 1

146 JAMA Ophthalmology February 2016 Volume 134, Number 2 (Reprinted) jamaophthalmology.com

Copyright 2016 American Medical Association. All rights reserved.

Downloaded From: on 10/10/2018

 Automated Ptosis Measurements From Facial Photographs
and K = [k1, k2, k3, k4, k5], where fx and fy are focal lengths, cx, and
cy is the image center in pixels, and [ki] are radial distortion co-
efficients specific to the camera used to acquire the image. These
matrices may be used to remove any significant distortion intro-
duced by the camera lens using a method previously described
by Zhang.1 Our software algorithm can be used to estimate the
matrices C and K for a particular camera. First, a set of at least 10
checkerboard-pattern images in different orientations is acquired
by the camera undergoing calibration. These images are analyzed
by the software algorithm to estimate C and K using another
method developed by Zhang1 (eFigure 1 in the Supplement).
The camera calibration method is required only once per
lens and camera combination of a particular design and manu-
facture and may not be necessary for high-quality lenses that
introduce little radial distortion. eFigure 2 in the Supplement
shows the results of camera calibration and undistortion on a
lens with a large amount of radial distortion.
Image Analysis
The undistorted image is segmented using standard face
detection algorithms to first identify a region corresponding
to the face.2,3 A search is performed within this region to iden-
tify the eyes. Facial feature edges are then detected within the
regions corresponding to the eyes using the standard edge de-
tection algorithm developed by Canny.4 The corneal light re-
flex of each eye is determined by the specular reflection of the
camera flash and is identified by the software algorithm as the
area enclosed by the edge nearest to the highest-intensity pixel
of the corresponding eye region. In the vertical dimension, the
distances from the centroid of the corneal light reflex to the
detected feature edges are reported.
To enable the software algorithm to determine the scale of
the image plane in millimeters per pixel, a circular marker of
known radius (in millimeters) is placed on the individual’s fore-
head before image capture. The marker is automatically detected,
and its radius (in image pixels) is determined using another well-
known computer vision algorithm.5 Figure 1 shows a sample in-
put photograph and the software algorithm’s output, consisting
of a re-rendering of the original photograph with the detected
edges and calculated measurements overlaid. Processing one
such photograph takes only a few seconds.
Methods
This observational study was performed at a single-surgeon
(J.B.H.) oculoplastic private practice. The study dates were July
30, 2014, to September 12, 2014. The dates of our analysis were
October 12, 2014, to June 18, 2015. Clinical photographs were ob-
tained at 4288 × 2848 (12.3 megapixel) resolution using a com-
mercially available camera system (Figure 2). The software al-
gorithm was implemented using the Java programming language
and an open-source computer vision library (OpenCV, version
2.4.6; Itseez).6 Institutional review board approval of the study
was obtained from Saint Louis University. Written informed con-
sent was obtained from study participants in accord with the Dec-
laration of Helsinki,7 and the study was performed with Health
Insurance Portability and Accountability Act of 1996 compliance.
jamaophthalmology.com
Copyright 2016 American Medical Association. All rights reserved.
Downloaded From: on 10/10/2018
Original Investigation Research
At a Glance
• We designed and tested a software algorithm for determining the
margin reflex distances 1 and 2 (MRD1 and MRD2, respectively)
from facial photographs.
• The software algorithm allows rapid, automatic, archival, and
objective quantization of eyelid position.
• Automated measurements compared favorably with standard
manual measurements, with minimal bias (0.03 mm for MRD1 and
0.13 mm for MRD2) and excellent agreement in this single-surgeon
study, although manual measurement of the MRD1 and MRD2
should remain standard practice given the limitations of this study.
Manual measurements of the MRD1 and MRD2 in each in-
dividual were obtained by us using a penlight to produce the cor-
neal light reflex and a ruler to judge the distance. A single cam-
era system (camera, lens, and flash model) was used to obtain
frontal facial photographs of all study participants at the time of
manual measurements at a fixed focal length of 1 m, with the
camera positioned directly in front of the patient at eye level at
a distance of 1 m as determined by the focal plane of the camera.
At this distance, small translational and rotational variation in
camera position should have minimal effect on the corneal light
reflex and eyelid measurements. These variables were chosen
to simulate image acquisition in a realistic clinical setting.
The camera was calibrated using 16 images. A marker with
a diameter of 19.05 mm was placed on the forehead of each
individual to provide a reference of scale. The images were
processed using our software algorithm, and the MRD1 and MRD2
measurements for each photograph were recorded. Because the
large distance from the optical center of the lens to the camera
flash (108.00 mm) causes some displacement in the corneal light
reflex, measurements from photographs with the flash positioned
above, below, to the left of, and to the right of the camera lens were
averaged for comparison with manual measurements and with
each other to assess the effect of the flash position. Statistical
analyses were performed using a software package (R, version
3.1.2; The R Foundation). Bland-Altman analysis was used to com-
pare agreement between the manual measurements and the au-
tomated photometric measurements. P < .05 was considered sta-
tistically significant. The r values are by Pearson product moment
correlation.
Results
In total, 55 eyes of 28 volunteers 24 to 80 years old (mean age,
57 years) were included in the study (one eye with an absent cor-
neal light reflex because of a negative MRD1 was excluded). Five
participants (18%) were male, and 23 participants (82%) were fe-
male. Twenty-six participants (93%) were of white race/ethnicity,
and 2 participants (7%) were African American. The manual
MRD1 measurements ranged from 0 to 6.00 mm (mean [SD], 2.87
[1.56] mm). The automated MRD1 measurements ranged from
0.23 to 5.64 mm (mean [SD], 2.91 [1.48] mm). The manual MRD2
measurements ranged from 4.00 to 9.00 mm (mean [SD], 5.68
[1.19] mm). The automated MRD2 measurements ranged from
4.34 to 9.46 mm (mean [SD], 5.81 [1.19] mm).
(Reprinted) JAMA Ophthalmology February 2016 Volume 134, Number 2 147
 Research Original Investigation Automated Ptosis Measurements From Facial Photographs

Figure 1. Image Acquired With the Flash Positioned to the Left Figure 3. Margin Reflex Distance 1 (MRD1)
of the Camera Lens
A Scatterplot of manual vs photometric MRD1 measurements
A Sample input image
6.0

4.5

Photometric, mm
3.0

1.5

0
B Software algorithm output image
0 1.5 3.0 4.5 6.0
Manual, mm

B Bland-Altman plot of the MRD1 measurements

1.0

1.96σ = 0.71

Photometric Minus Manual, mm

0.5

μ = 0.03
The flash positioning results in lateral displacement of the corneal light reflex 0
(in this case to the patient’s right). The effect is asymmetric, with the corneal
light reflex of the left eye being more off center than that of the right eye.
The numerals in B are in millimeters.
–0.5

Figure 2. Camera Systems Used

–1.96σ = –0.66

–1.0
0 1.0 2.0 3.0 4.0 5.0 6.0
(Photometric Plus Manual)/2, mm

Photometric measurements were obtained with the software algorithm.

σ Indicates SD; μ, arithmetic mean.

correlated with manual measurements (r = 0.97; 95% CI, r = 0.95

A commercially available camera system (D90 camera body with a 60-mm
f/2.8G AF-S Micro Nikkor autofocus ED lens at F16; Nikon) was used. Another
system (R1C1 wireless close-up Speedlight system for i-TTL single-lens reflex
camera; Nikon) was used for illumination.
Automated MRD1 measurements were normally distributed
according to the Shapiro-Wilks test (P = .05). They were strongly
to r = 0.98; P < .001). The bias of automated MRD1 measurements
was 0.03 mm (95% CI, −0.06 to 0.12 mm), with 95% confidence
limits of −0.66 and 0.71 mm as determined by Bland-Altman
analysis. Figure 3 shows a Bland-Altman analysis of the MRD1
results.
Automated MRD2 measurements were normally distributed
(P = .01). They were strongly correlated with manual measure-
ments (r = 0.96; 95% CI, r = 0.93 to r = 0.98; P < .001). The bias
of automated MRD2 measurements was 0.13 mm (95% CI, 0.03-
0.22 mm), with 95% confidence limits of −0.54 and 0.80 mm.
Figure 4 shows a Bland-Altman analysis of the MRD2 results.
Displacement of the corneal light reflex occurred as a result
of the distance between the center of the camera lens and the
flash. The effect is asymmetric when the flash is oriented hori-
zontally, with the corneal light reflex being more off center of the
eye contralateral to the flash, as in Figure 3. This translation of
the corneal light reflex was averaged in this study by obtaining
clinical photographs with the flash positioned in 4 quadrants. We

148 JAMA Ophthalmology February 2016 Volume 134, Number 2 (Reprinted) jamaophthalmology.com

Copyright 2016 American Medical Association. All rights reserved.

Downloaded From: on 10/10/2018

 Automated Ptosis Measurements From Facial Photographs Original Investigation Research

Figure 4. Margin Reflex Distance 2 (MRD2) Table. Effect of the Flash Position on the Measurement of MRD1 and MRD2

MRD1, mm P Value MRD2, mm P Value

A Scatterplot of manual vs photometric MRD2 measurements
Δ(Left, right) 0.23 .69 0.25 .32
10.0 Δ(Left, mean) OD -0.31 <.001 0.3 .09
OS -0.1 .23 0.2 .23
Δ(Right, mean) OD -0.3 .002 0.35 .03
7.50
OS -0.2 .23 -0.01 .02
Photometric, mm

Δ(Up, mean) -0.53 .001 0.42 .004

Δ(Down, mean) 0.86 <.001 -0.96 <.001
5.00
Abbreviations: MRD1, margin reflex distance 1; MRD2, margin reflex distance 2.

2.50
surements was 0.30 mm OD and 0.20 mm OS, with neither dif-
ference being statistically significant (P = .09 and P = .23,
0 respectively). The mean absolute difference between MRD2 mea-
0 2.25 4.50 6.75 9.00 surements with the flash in the right position and the averaged
Manual, mm measurements was +0.35 mm OD and −0.01 OS, neither of which
was statistically significant (P = .03 and P = .23, respectively).
B Bland-Altman plot of the MRD2 measurements
Vertical flash placement causes significant variation in mea-
1.0 surements. Positioning the flash superiorly underestimated the
1.96σ = 0.80
MRD1 by a mean of 0.53 mm (P = .001) and overestimated the
MRD2 by a mean of 0.43 (P = .004). The Table summarizes these
results.
Photometric Minus Manual, mm

0.5

μ = 0.13

0 Discussion
–0.5
–1.96σ = –0.54
4.0 5.0 6.0
(Photometric Plus Manual)/2, mm
Photometric measurements were obtained with the software algorithm.
σ Indicates SD; μ, arithmetic mean.
compared the differences between measurements made with the
flash in the 4 different positions with each other and the aver-
aged measurements and then evaluated these variations for sta-
tistical significance using a 2-tailed t test.
When in a horizontal plane, the effect of the flash position was
minimal. The mean absolute difference between MRD1 measure-
ments with the flash in the left and right positions was 0.23 mm,
which was not statistically significant (P = .69). The mean abso-
lute difference between MRD1 measurements with the flash in
the left position and the averaged measurements was −0.31 mm
OD and −0.10 mm OS, which was statistically significant for right
eyes only (P < .001 and P = .23, respectively). The mean absolute
difference between MRD1 measurements with the flash in the
right position and the averaged measurements was −0.30 mm OD
and −0.20 mm OS, which was statistically significant for right eyes
only (P = .002 and P = .23, respectively). The mean absolute dif-
ference between MRD2 measurements with the flash in the left
and right positions was 0.25 mm and was not statistically signifi-
cant (P = .32). The mean absolute difference between MRD2 mea-
surements with the flash in the left position and the averaged mea-
Manual measurements of the MRD1 and MRD2 are used in the
clinical evaluation of ptosis and the surgical planning of ptosis re-
pair. In addition, these measurements, along with photographic
documentation of ptosis, are typically required by insurers to
prove medical necessity. However, interobserver variability, re-
producibility, patient movement, and poor cooperation with
7.0 8.0 9.0 10.0
testing present a challenge to current methods of preoperative
evaluation. In conjunction with the MRD1 and MRD2, additional
considerations in aesthetic eyelid surgery include measurements
of the brow fat span and tarsal platform show, as described by
Goldberg and Lew.6 Because the relevant facial features that
define these metrics can be identified by our method of edge de-
tection analysis, indirect measurements of the brow fat span and
tarsal platform show can be obtained from our software algo-
rithm’s output.
Computer-assisted analysis of facial photographs for mea-
surement of the MRD1, MRD2, eyelid contour, and palpebral fis-
sure has been previously described.8-10 However, the methods
used rely on significant user and computer interaction after im-
age acquisition and depend on an observer to identify edges and
facial features. To our knowledge, ours is the first software al-
gorithm capable of obtaining eyelid measurements with com-
pletely automated image processing.
One key issue was displacement of the corneal light reflex by
the position of the flash relative to the lens. For data collection
in this study, we chose to average readings from all 4 quadrants
(superior, left, right, and inferior). In clinical practice, one would
most likely choose to place the flash as close to the center of the
lens as possible or position it to the left or right of the lens for all
photographs. In our data set, choosing only one position altered
the measurements by approximately 0.20 mm, which is well

jamaophthalmology.com (Reprinted) JAMA Ophthalmology February 2016 Volume 134, Number 2 149

Copyright 2016 American Medical Association. All rights reserved.

Downloaded From: on 10/10/2018

 Research Original Investigation Automated Ptosis Measurements From Facial Photographs

within the margin of error for clinical practice. Therefore, a single
photograph with the flash positioned in the horizontal plane
would prove adequate for routine clinical measurement. Accu-
racy could be further improved by minimizing the distance be-
tween the flash and the optical elements of the camera.
The 95% confidence limits of our method are within 0.71 mm
for MRD1 and 0.80 mm for MRD2, with a bias of 0.13 mm for
MRD2 measurements. To be considered equivalent to manual
measurements, it would be preferable that the 95% confidence
limits are within 0.50 mm. One reason for disagreement may in-
clude a difference in the reflex position secondary to displacement
of the flash from the optical axis. When there is a high-contrast
transition from the pupil to the iris, the pupillary border can be
delineated by the standard edge detection algorithm used by our
software algorithm, as shown in Figure 3. However, at this stage
of development, pupil detection is not reliable. Therefore, we rely
on the reflection of the flash as a basis for measurements. This ne-
cessity is a practical limitation of our present software algorithm
implementation because the use of the pupil centroid in lieu of
the corneal light reflex would render the analysis in the Table un-
necessary and simplify the measurements. Another limitation is
that negative MRD1 and MRD2 values cannot be estimated from
a single photograph (when the corneal light reflex or pupil center
would be obscured). Manifest strabismus will also cause signifi-
cant displacement of the corneal light reflex and is an additional
limitation, which may be overcome in some cases by using mul-
tiple photographs if the patient is capable of alternate fixation.
Some disagreement may also result from placement of the
scale marker in the plane of the forehead, which is slightly an-
terior to the plane of the eyelids. Estimating the scale by using
white-to-white measurements of the corneal limbus from the
photographs compared with the mean corneal diameter could
eliminate the need for the scale marker and might influence the
bias or disagreement of the automated measurements. Finally,
disagreement between the software algorithm and the manual
measurements may simply be due to the limited precision of the
manual measurements, which is 0.50 mm.
The participants in our study were recruited and photo-
graphed during a single routine clinic visit. Therefore, variabil-
ity of successive measurements (manual and automated) was not
quantified by this study but is an important factor to consider be-
fore our method could be used in clinical practice. All individu-
als were photographed with the head in frontal position. Addi-
tional measurements from different positions may help assess
the degree to which variability of head position and patient move-
ment degrades the reliability of measurements using our soft-
ware algorithm. Further work is necessary to render our software
algorithm reliable in the routine clinical setting because lighting,
facial positioning, and other variables can affect the photographic
appearance of ptosis and the results of digital image processing.11
Given these limitations, we do not expect our results to change
clinical practice in the near future, and manual measurement of
the MRD1 and MRD2 should remain the standard of care. How-
ever, we believe that we have demonstrated a proof of concept
for a potentially useful new approach to ptosis measurements.
Conclusions
An automated, photography-based system could provide an
archival and reproducible means for obtaining the MRD1,
MRD2, and other facial morphometric data while mitigating
potential sources of error, including movement and observer
bias. Automated ptosis measurements produced by our soft-
ware algorithm compare favorably with manual clinical mea-
surements. Software algorithm–based correction of lens dis-
tortion could make this technology available in inexpensive
handheld devices, including smartphones.

ARTICLE INFORMATION REFERENCES 7. World Medical Association. World Medical

Submitted for Publication: July 6, 2015; final
revision received September 29, 2015; accepted
October 4, 2015.
Published Online: November 25, 2015.
doi:10.1001/jamaophthalmol.2015.4614.
Author Contributions: Dr Bodnar had full access to
all the data in the study and takes responsibility for
the integrity of the data and the accuracy of the data
analysis.
Study concept and design: Bodnar.
Acquisition, analysis, or interpretation of data: All
authors.
Drafting of the manuscript: Bodnar, Holds.
Critical revision of the manuscript for important
intellectual content: All authors.
Statistical analysis: Bodnar.
Administrative, technical, or material support:
Neimkin.
Study supervision: Neimkin.
Conflict of Interest Disclosures: All authors have
completed and submitted the ICMJE Form for
Disclosure of Potential Conflicts of Interest, and
none were reported.
1. Zhang Z. A flexible new technique for camera
calibration. IEEE Trans Pattern Anal Mach Intell.
2000;22(11):1330-1334.
2. Viola P, Jones M. Rapid object detection using a
boosted cascade of simple features. Published 2001.
https://www.cs.cmu.edu/~efros/courses/LBMV07
/Papers/viola-cvpr-01.pdf. Accessed October 9, 2015.
3. Lienhart R, Kuranov A, Pisarevsky V. Empirical
analysis of detection cascades of boosted classifiers
for rapid object detection. Revised December 2002.
http://www.multimedia-computing.de/mediawiki
//images/5/52/MRL-TR-May02-revised-Dec02.pdf.
Accessed October 8, 2015.
4. Canny J. A computational approach to edge
detection. IEEE Trans Pattern Anal Mach Intell.
1986;8(6):679-698.
5. Kimme C, Ballard D, Sklansky J. Finding circles by
an array of accumulators. Published 1975.
http://dl.acm.org/citation.cfm?id=360677. Accessed
October 9, 2015.
6. Goldberg RA, Lew H. Cosmetic outcome of
posterior approach ptosis surgery (an American
Ophthalmological Society thesis). Trans Am
Ophthalmol Soc. 2011;109:157-167.
Association Declaration of Helsinki: ethical
principles for medical research involving human
subjects. JAMA. 2013;310(20):2191-2194.
doi:10.1001/jama.2013.281053.
8. Burmann TG, Valiatti FB, Correa ZM, Bayer M,
Marcon Í. Margin reflex distance measured by
computerized image processing in rigid contact lens
wearers [in Portuguese]. Braz Arch Ophthalmol.
2008;71(1):34-37.
9. Cruz AA, Lucchezi MC. Quantification of
palpebral fissure shape in severe congenital
blepharoptosis. Ophthal Plast Reconstr Surg. 1999;
15(4):232-235.
10. Cruz AA, Coelho RP, Baccega A, Lucchezi MC,
Souza AD, Ruiz EE. Digital image processing
measurement of the upper eyelid contour in Graves
disease and congenital blepharoptosis.
Ophthalmology. 1998;105(5):913-918.
11. Flowers RS, Flowers SS. Diagnosing
photographic distortion: decoding true
postoperative contour after eyelid surgery. Clin
Plast Surg. 1993;20(2):387-392.

150 JAMA Ophthalmology February 2016 Volume 134, Number 2 (Reprinted) jamaophthalmology.com

Copyright 2016 American Medical Association. All rights reserved.

Downloaded From: on 10/10/2018