Developing an inexpensive optical measurement device to estimate ocular media density.
Selection of HDR camera, deep learning software stack for Purkinje image detection and segmentation, NVIDIA Jetson as the embedded computer triggering either LED or LASER lights to be projected on the cornea
Alternative download link:
https://www.dropbox.com/s/fh7r8szuc2pctfr/purkinje_imaging_inPractice.pdf?dl=0
1. Petteri Teikari, PhD
Singapore Eye Research Institute (SERI)
Visual Neurosciences group
http://petteri-teikari.com/
Version “Thu 6 September 2018“
Multispectral
Purkinje Imaging
developing an inexpensive
measurementdeviceto
estimateocular media density
1938CarlMydans
4. Purkinje
Images
Image
Processing
Rextractautomatically from
the HDR frame(s) the
diference insignal intensity
(region proposal / localization
Fourth image isformed after adouble passthrough the crystalline lens
1) REGIONPROPOSAL
https://medium.com/@smallfshbigsea/faster
-r-cnn-explained-864d4fb7e3f8
2) Pixel-by-pixel segmentation,
a.k.a. “SemanticSegmentation”
https://github.com/GeorgeSeif/Semantic-Segmentation-Suite
The raw A/D output was then converted to a radiance
profle using calibration data for the camera/frame-grabber
system. Custom software allowed automated analysis of the
averaged radiance profles for each interference flter. A least
squares linear regression line was ftted to the flanns of each
radiance profle (representing bacn-scattered light from
various layers in the crystalline lens, etc.), and the radiance for
the point on the regression line lying directly underneath the
peak of the IVth image was calculated. The radiance for
this position on this downward sloping scatter shoulder was
subtracted from the gross radiance at the pean of the radiance
profle. Thus, the net reflected radiance of the IVth Purninje
image alone, or output, was computed separately for
eachwavelength.
The subsequent calculation of individual lens
density consisted of taning the common log of the ratio of
the input radiance to the net refected output
radiance of the IVth image. The result was divided by two
to compensate for the double-pass through the anterior
segment of the eye. Several such density measurements were
made, and the average density was calculated for each eye of
every subject. No adjustment was made for the path length
diference (about 3%) caused bythe beam beingan average of
about 15° of-axis or forpupillary area.
5. Applications
Ocular Media Density
van deKraatsandvan Norren2007
Teinari and Najjaret al. 2012
Cataractassessment
deCastro etal.2018
DohyeunKim et al. 2018
Intraocular lens(IOL)
issues
CRSToday
deCastroetal.2007
Schaefel2008
Nishi etal.2010
Sun et al.2014
Maedeletal.2017
Kumaretal.2018
+Laboratorio deOptica,UniversidaddeMurcia
+ InstitutodeÓptica Daza deValdés,Consejo Superior
deInvestigacionesCientífcas,Madrid
+Instituteof Ophthalmology,University College
London,andMoorfeldsEyeHospital NHS Foundation
Trust
+Vienna Institutefor ResearchinOcular Surgery
+Institutefor OphthalmicResearch,Tubingen
(a) Three-dimensional (3-D) optical coherence tomography (OCT)
image (example is for S#3, OD); (b) en face OCT image showing the
pupil merging, the locations of PI, PIII,and PIV in red and pupilcenter in
yellow; (c) PI (left), PIII (center), and PIV (right) in OCT B-Scan images.
-Sun et al.2014
Kumaretal. 2018
The clinical feasibility of
the Spanish Purkinje
meter system was better
than that of the German
Purninjemeter.
In pseudophanic eyes, the
German Purninje meter
measured signifcantly higher
IOL tilt and decentration values
than the Spanish Purninje
meter.
7. Camera
Multiframe HDRconstruction
or/and HDRcamera
The faster the framerate,
lessthere will be motion
artifactsbetweensuccessive
frames (eachwith diferent
wavelength)
e-CAM20_CU0230_MOD (NVIDIAJetson TX2/TX1)
See3CAM_CU20
HDRUSB3.1 Gen1 CameraBoard
(Color)
CMOSsensor 12-Bit RGB Bayer
Dynamic Range:Max upto105dB
SensorStrobeoutputto connectexternal flash
ElectronicRolling Shutter
Fixed Focuswith M12lensholder
VOLUME PRICE
STARTFROM
US$59-89
Petteri: Monochrome
camera withouttheBayer
arraywould bebetterfor
thisproject
Ximea MQ003MG-CM
CMOSISCMV300
https://www.ximea.com/en/products/usb3-vi
sion-cameras-xiq-line/mq003mg-cm
https://www.ximea.com/downloads/usb3/manuals/xiq_technical_manual.pdf
Digitizationwith12bits
HDR:KneepointbasedHDR
Mane your life easier with a proper
camera, and when you have the idea
worning tryto optimize the cost→
C-mount [or with C/CSmount moduleB]
20. Ximea
MQ003MG-CM
500fpsatVGA
- 500 spectral bandsinasecond?
More line lessbandsreally fast
toreduce ocular motion during
acquisition withHDR mode.
Ifyou automate thiswell,theacquisition+
processing couldbeiterated duringone
secondmultiple timestomanesurethatthe
qualityisgood,andyou do nothaveto
realizeduring ofineanalysis(avoidthis)that
you wouldneedtore-measurethatperson.
Ximeasupportsthefollowing
exposuretimes:
Exposuretime(EXP)
54µsto1sec
instepsof7.56µs
Soifyoucannothave
brightenough illumination
for54µs,youcanexpose
longer (orthatbrightness
isabove thesafetylimits)
Thesedependon
yourilluminants
(intensity and
wavelength)
For example 400 nm would need to be 2x
brigher than 500 nm with the same
exposure time
25. Ifyoumoney, and
fast switching
times with
narrow bands
gowithNKTFianium
WhiteLase Class3B
supercontinuumlaser
and AOTF flters
(Goochand Housego)
https://www.nntphotonics.com/lasers-fbers/product/fanium-whitelase-micro/
https://goochandhousego.com/product-categories/tunable-flters-aotf/
26. Back toReal
World
Easiertouse quasi-
monochromatic
LEDs
Use more bandson
shortwavelenghts
asthe ocularmedia
variesthere more
between subjects
OcularAbsorbers Function ofage
80yr
20yr
Opto-electronics > LEDLighting > LEDEmitters
https://www.mouser.sg/Optoelectronics/LED-Lighting/LED-Emitters/_/N-8usfd
Select as many wavelengths that you can “comfortably” use and ft the template provided by
van deKraatsandvan Norren (2007) for your measured data to estimate the full spectral characteristics of theocular
media.
van deKraatsandvan Norren (2007)