measuring eye-movement patterns on a smartphone for medical...
TRANSCRIPT
Hsin-Yu Lai, Gladynel Saavedra-Peña
Measuring Eye-Movement Patterns on a Smartphone for Medical Applications
Prof. Charlie Sodini, Prof. Thomas Heldt, Prof. Vivienne Sze
• Motivation & background
• Measurement system
• Eye movement data
• Conclusions
Outline1
• Motivation & background
• Measurement system
• Eye movement data
• Conclusions
Outline2
Neurodegenerative Disorders Affect Millions3
• Neurodegenerative disorders affect millions of people worldwide
• 5 million Americans live with Alzheimer’s Disease [NIH]
• 500,000 Americans live with Parkinson’s Disease [NIH]
• No cure for neurodegenerative diseases
• No accurate way to measure disease stage
• Need for accessible tools that can accurately measure disease progression
• Objective measurements of disease progression lead to effective therapies[Newsweek, 2016]
Current Tests to Track Disease Progression
• Repeat medical assessments are sparse and suffer from high retest variability
• Tests are mostly qualitative, time consuming and require a trained specialist.
Mini-Mental State Examination (MMSE)
Q1. What is the year? Season? Date?
Q2. Where are you now? State? Floor?
Q3. Could you count backward from 100 by sevens? (93, 86, …)
Clock-drawing test
[Agrell et al. Age and Ageing, 1998]
Current cognitive tests are not ideal tools to measure disease progression.
4
Based on a physician’s assessment of patient performance.
We want to track neurodegenerative disease
progression by measuring quantitative
physiological data with a consumer-grade device.
Goal of our Research5
Why consumer-grade devices?• Non-obtrusive, low-cost, higher availability• Does not require trained specialist, less sparsity in data
Alzheimer’s Disease or Parkinson’s Disease
Smartphone or a tablet
Visual Reaction Time in Patients 6
Visual reaction time is larger and more variable in patient populations.
[Garbutt et al. Brain, 2008]
Patients with frontotemporal
degeneration & dementia
Mean v
isual re
action t
ime (
ms)
[Heuer et al. Neurology, 2013]
Mean v
isual re
action t
ime (
ms)
NE: Normal Elderly
MCI: Mild Cognitive Impairment
AD: Alzheimer’s Disease
Reaction time is the time delay between the presentation of a stimulus and when the eye starts to move towards said stimulus.
Visual Tasks to Measure Reaction Time7
Central Fixation PointStimulus
x
TimeEye P
osit
ion
Reaction time
No
rma
lize
d c
ou
nt
Reaction time (ms)
Measurements in Clinical Environments8
Substantial head support
SR EYELINK 1000 PLUS
IR illumination
[Reulen et al., Med. & Biol. Eng. & Comp, 1988.]
Clinical measurements are done in constrained environments that rely on specialized, costly equipment.
High-speed camera
PHANTOM V25-11
Measurements in Clinical Environments
Substantial head supportHigh-speed camera IR illumination
9
… to this!
GOAL: Go from this…
• Motivation & background
• Measurement system
• Eye movement data
• Conclusions
Outline10
The Pipeline Consists of an Eye Tracking Algorithm and Saccade Onset Detection
11
Goal: Find a robust eye-tracking algorithm that is able to show clear saccade onsets.
Approach: Improve current state-of-the-art gaze estimation algorithms for our purpose and analyze their robustness.
[Lai et al. ICIP, 2018]
Reaction Time (ms)
iTracker is a CNN-based Gaze Estimation Algorithm Trained on 30-fps Recordings
12
[Krafka et al. CVPR, 2016]
To Apply iTracker for Our Purpose, We Manually Annotated the Facial Landmarks
13
Identify the face
location in the grid
We Propose iTracker-face for Clear Onset Presentation
14
• To measure saccade latency, we took recordings at 240 fps.
We propose iTracker-face: iTracker with only face-related layers.
30 fps 240 fps
Images become blurry.
• Our goal is to estimate when the gaze changes.i.e. Accuracy ⇒ Clear saccade onset
Eye Gaze (x,y)
• Test 1: Chinrest + manual face annotation on the first frame of each video
Compare iTracker and iTracker-face based on Whether a Clear Onset can be Seen
15
iTracker: Grey + Blue
iTracker-face: Blue
Video Recording Setup16
• Subject: Subjects were recorded while viewing a visual task that was displayed on a laptop.
• Recording device: iPhone 6 with a framerate of 240 fps and an image resolution of 1280 x 720 pixels
• Synchronized monitor: Duplicates the laptop’s display and allows for the calculation of reaction time.
iTracker-Face Produces Traces with Clearer Saccade Onsets Compared to iTracker
17
How about robustness?
We tested • under four lighting conditions (highest to lowest) • on five subjects • with and without glasses.⇒ For this experiment, we acquired 9,600 saccades in total.
Sample Frames from Four Lighting Conditions 18
278 Lux 220 Lux 54 Lux 26 Lux
0
20
40
60
80
100
High Light Room Light Medium Light Low Light
Without Glasses
iTracker-Face is more Robust than iTracker19
0
20
40
60
80
100
High Light Room Light Medium Light Low Light
With Glasses
• Two expert annotators visualized 9,600 traces and labeled them as good/bad saccades (clear/unclear onset).
• It took each annotator approximately 10-15 hours to do so.
iTracker-face is much more robust to lighting conditions and glasses.
Automatic Rejection of Bad Saccades is Crucial.
Ag
reed
Go
od
(%
)
iTracker
iTracker - face
[Lai et al. JBHI, 2019]
We Choose iTracker-face to be the Eye Tracking Algorithm
20
iTracker-face
Goal: Find a saccade onset detection method and a way to automatically detect bad saccades.
Approach: Model for saccade traces.
Reaction Time (ms)
Tanh Fitting can Measure Saccade Onset and Detect “Bad” Saccades
21
• The normalized root-mean-square error (NRMSE) between the model and the eye movement trace can be used to reject bad saccades.
Using an adequate NRMSE threshold, we can classify traces as “good” or “bad” with an average true-positive rate of 0.87 and an average false-positive rate of 0.18.
Recall that we have 9,600 annotated traces from the robustness test.
• Horizontal saccades are modeled using tanh.Saccade
latency
iTracker-face
Tanh model
3% of the
model
amplitude
Reaction
Time
We Choose Tanh Fitting to Measure Saccade Onset and Determine Good Saccades
22
iTracker-faceTanh fitting
with automated rejection
How do smartphone recordings compare to recordings made with a research-grade camera?
Reaction Time (ms)
High-Speed Video Recording Set-up23
• Subject: Subjects were recorded while viewing a visual task that was displayed on a laptop.
• High-speed video recordings: In a subset of recordings, subjects were simultaneously recorded with a high-speed video camera at 500 fps and an image resolution of 1280 x 720 pixels.
• Recording device: iPhone 6 with a framerate of 240 fps and an image resolution of 1280 x 720 pixels
• Synchronized monitor: Duplicates the laptop’s display and allows for the calculation of reaction time.
We Achieved Similar Statistics Compared with a Research Grade Camera
24
High-speed research-grade cameraPhantom v2511,cost $100k, 720p resolution, 500 fps28 µm pixel size, global shutter
Consumer grade cameraiPhone 6, cost <$1k,
720p resolution, 240 fps1.5 µm pixel size, rolling shutter
[Lai et al. ICIP, 2018]
Reaction Time (ms)
Our Pipeline is Almost Automated25
iTracker-faceTanh fitting
with automated rejection
Can we remove manual face annotation?
Reaction Time (ms)
• Test 1: Chinrest + manual face annotation
• Test 2: Chinrest + Viola-Jones face detector
Can Our Pipeline be Fully Automated?26
Approach: We compare the reaction time acquired from manual face annotation with the reaction time acquired
from the Viola-Jones face detector.
Manual Face Annotation and Face Detector Attained Almost Identical Reaction Time
27
The distribution shows the absolute difference in mean reaction time calculated with manual face annotation and the Viola-Jones face detector.
Automating face-detection does not affect the estimated reaction time.
[Lai et al. JBHI, 2019]
Absolute Reaction Time Difference (ms)
Can We Remove the Chinrest?28
• Test 1: Chinrest + manual face annotation
• Test 2: Chinrest + Viola-Jones face detector
• Test 3: Without chinrest + Viola-Jones face detector
Recordings with Chinrest and without Have Comparable Signal-to-Noise Level
29
*The traces were intentionally chosen to have similar reaction time for easier comparison.
[Lai et al. JBHI, 2019]
We Completely Automate the Pipeline for Reaction Time Measurement
30
iTracker-faceTanh-fitting
with automated rejectionViola-Jones face
detector
Reaction Time (ms)
• Motivation & background
• Measurement system
• Eye movement data
• Conclusions
Outline31
Eye movement Data Breakdown32
We recorded 19,200 eye movements across 160 recording sessions in 29 subjects.
Average fraction of good data per session: 77%Average fraction of bad data due to noise: 13%
Data can be used
Data cannot be used
[Lai et al. JBHI, 2019]
Reaction Time in Healthy Individuals33
• We recorded over 19,000 saccades in 29 healthy subjects.
• Among the 29 subjects, 10 subjects were recorded on multiple ocassions.
[Lai et al. JBHI, 2019]
Saccade latency (ms)
No
rma
lize
d c
ou
nt
Reaction time (ms)
No
rma
lize
d C
ou
nt
Reaction time distributions have distinctive shapes and parameters. Reaction time variability is substantial within a subject and across subjects.
• Motivation & background
• Measurement system
• Eye movement data
• Conclusions
Outline34
• We have enabled the measurement of visual reaction time outside of the clinical environment by proposing a robust, novel system that combines deep-learning and modeling techniques.
• We have eliminated the need for research-grade cameras, IR illumination, and chinrest by building algorithms that can measure reaction time using recordings from a smartphone camera.
• We collected over 19,000 measurements across 29 healthy individuals and observed that their reaction time distributions have distinctive shapes and parameters.
Conclusions35
Questions?
36
• Neurodegenerative Diseases
- “The new offensive on Alzheimer’s Disease: Stop it before it starts,”
https://www.newsweek.com/2017/02/24/stopping-alzheimers-disease-it-starts-557221.html,
Accessed on 04-26-2019.
- “Neurodegenerative diseases”,
https://www.niehs.nih.gov/research/supported/health/neurodegenerative/index.cfm, Accessed on
04-26-2019.
- B. Agrell and O. Dehlin, “The clock-drawing test,” Age and Ageing, vol. 27, pp. 399 – 403, 1998.
- H. Heuer, J. Mirsky, E. Kong, B. Dickerson, B. Miller, J. Kramer, and A. Boxer, “Antisaccade task
reflects cortical involvement in mild cognitive impairment,” Neurology, vol. 81, no. 14, pp. 1235 –
1243, 2013.
- S. Garbutt, A. Matlin, J. Hellmuth, A. Schenk, J. Johnson, H. Rosen, D. Dean, J. Kramer, J.
Neuhaus, B. Miller, S. Lisberger, and A. Boxer, “Oculomotor function in frontotemporal lobar
degeneration, related disorders and Alzheimer’s Disease,” Brain, vol. 131, pp. 1268 – 1281,
2008.
• Eye-Tracking Algorithm
- K. Krafka, A. Khosla, P. Kellnhofer, H. Kannan, S. Bhandarkar, W. Matusik, and A. Torralba,
“Eye tracking for everyone”, in Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), 2016, pp. 2176 – 2184.
References37
• Our Publications
- H.-Y. Lai, G. Saavedra-Peña, C. Sodini, T. Heldt, and V. Sze, “Enabling saccade latency
measurements with consumer-grade cameras”, in Proceedings of the IEEE International
Conference on Image Processing (ICIP), 2018, pp. 3169 – 3173.
- G. Saavedra-Peña, H.-Y. Lai, V. Sze, and T. Heldt, “Determination of saccade latency
distributions using video recordings from consumer-grade devices,” in Proceedings of the IEEE
Engineering in Medicine and Biology Conference (EMBC), 2018, pp. 953 – 956.
- H.-Y. Lai, G. Saavedra-Peña, C. Sodini, V. Sze, and T. Heldt, “Measuring saccade latency using
smartphone cameras,” Journal of Biomedical and Health Informatics (JBHI), 2019.
References38