Carotid artery thermography and thermal image processing with data - - PowerPoint PPT Presentation

Carotid artery thermography and thermal image processing with data management (MAE03)

NEO JUN WEI ANGLO-CHINESE SCHOOL (INDEPENDENT) RESEARCH MENTOR: PROF. NG YIN KWEE, EDDIE SCHOOL OF MECHANICAL AND AEROSPACE ENGINEERING NANYANG TECHNOLOGICAL UNIVERSITY

Project Objectives

Determining pulse rates using method

• f non-invasive

thermography

Using pulse rates and

• ther quantitative and

qualitative parameters to detect stenosis within carotid arteries

Unable to complete due to time constraints and methodological limitations

Introduction: Significance of Project

 Cardiovascular diseases (CVDs) are the largest cause of death

worldwide

 17.7 million deaths caused by CVDs in 2015

 Current methods used for CVD often invasive and/or time-consuming.

 ECGs take from 10-30 minutes and MRIs can take from 25 minutes to 1 hour

 Aid in devising a method to preliminarily screen a subject before further

examinations

 Could give rise to potential methods of contactless pulse rate detection

Introduction: Background Information

 Thermography

 The use of infrared thermal imaging

to study temperature fluctuations and blood flow within the body

 Passive thermography is utilised

in this study – does not require an energy (heat) source

 Subject is placed under regular

conditions, not subject to heating and cooling cycles

Introduction: Theoretical basis of project

 Thermal images of left and right

superficial temporal arteries and external carotid arteries are taken

 Temperature fluctuations within

these arteries should correspond to cardiac cycle since they are caused by start-stop flow of blood with each pump of the heart

Methodology

 1) Thermal Imaging Collection

 39 male human subjects were recruited  VarioCAM Infrared (IR) camera was placed 0.5m from each

subject

 Thermal images of 240 × 320 pixels were captured at a rate of 25

frames per second for 20 seconds

 The body parts observed were left and right sides of the neck and

foreheads.

Methodology

 2) Image Pre-processing  Thermal images were

cropped such to include only the features of interest (blood vessels)

 Filters were applied to

enhance contrast and to minimise noise from ambient heat.

Superficial temporal artery

Methodology

 3) Image Post-processing  Step 1: Rectangular region R containing the major blood

vessel of the body part of interest was selected.

ROI

 Step 2: The multiple thermograms are treated as a

three-dimensional matrix (x, y, t).

t=20

Methodology

 3) Image Post-processing

 Step 3: This is then reduced into a two dimensional matrix

by averaging one of the spatial dimensions, which is chosen as x in this case. This matrix is mathematically expressed as

Step 4: Fast Fourier Transform (FFT) is then applied to plot the signal in the frequency domain. The values obtained from this transformation are known as the power spectra P, and they are averaged along the y dimension in order to obtain a composite power spectrum, ത 𝑄. This is expressed by the equation on the left.

time x y

1 ( , ) ( , , )

x

R x x

y t x y t R







1

y

R y y y

P P R







Methodology

 3) Image Post-processing

 Step 5: The power spectrum at any frequency f in time ത

𝑄0 is is convolved with the average power spectrum (over the existing period of time M) in order to obtain the historic power spectrum at any value of f, as expressed mathematically on the right.

 This amplifies the signal as a result of heat from blood-flow and

minimised any “noise” due to heat from surrounding tissue

 Step 6: H(f) is then averaged to obtain the historic power

spectrum ഥ 𝐼 before it is then convolved with ത 𝑄0 to obtain the peak frequency which is designated to be the pulse frequency fpulse. This is multiplied by 60 to obtain the pulse rate in beats per minute, bpm.

1 1 1

( ) ( ) ( )

M i i M F i i j

P f H f P j

  

 



Methodology

 3) Image Post-processing

 Step 7: The calculated pulse rates were compared to heart rates obtained from NHCS.

These were taken to be the ground truth readings

Results



Table 1: Pulse rates of left and right necks and heads, mean and actual pulse rates and percentage accuracy

 Average percentage accuracy of

the pulse rates however were 83.37%.

 The Pearson product moment

correlation test was applied yielding an R2 value of 0.01, indicating a weak correlation.

Results

 Poor correlation for heart rates derived from both necks and heads  Heart rates for necks show better correlation  Mean accuracy for Necks = 79.78%, whereas for Heads = 84.23%

y = -0.0467x + 72.177 R² = 0.0016 20 40 60 80 100 120 20 40 60 80 100

Mean HR (Neck) Actual HR

Mean HR(Neck) against Actual HR

y = 0.2178x + 52.712 R² = 0.0415 20 40 60 80 100 120 10 20 30 40 50 60 70 80 90 100

Mean HR(Head) Actual HR

Mean HR(Head) against Actual HR

Results

RMSE = 17.06 bpm RMSE = 13.86 bpm



Proportional error for both detected, where bias increases proportionally to mean



Good agreement despite poor correlation as fell within 95% confidence intervals



Bias for necks deviates from 0 more than for heads



Could be due to greater contour contrast offered by head, as less obstruction by skin and fat on head compared to neck

Discussion

• Repetition of the pulse-rate values 58.776 and 88.164.
• Reasons for these two values constituting the majority of values displayed are still unknown
• Could be due to a bug within the Python code or the FFT package installed
• These two values appear at random for various body parts
• The mean yielded from the average pulse rate of each patient appears to be accurate to

a certain degree

• However, the extremely low R2 between the derived and real pulse rates indicates that this is likely to

be by coincidence, and that the reliability of this method is questionable

• When compared to Bin and Li utilising discrete wavelet transform,
• They obtained accuracies varying from 85.2% to 98.5%, with a mean accuracy rate of 94.5%
• While the functions used were different, the methods both approached pulse extraction through

power spectral density and obtained significantly more accurate results

Conclusion

• Results obtained were accurate to a certain extent
• However, reliability of this method remains unknown given the

weak trend between extracted and true data

• The study does showcase a possible method for contactless pulse

rate extraction that could serve in a future medical extensions if the method is further improved upon

• This would be applicable to many scenarios where short-term non-

invasive pulse-rate determination would be desirable, for sports training studies, sleep studies and psychological evaluations

Further work

• Addition of motion-tracking algorithm to fix ROI
• Use of thermal imaging to determine temperature of blood vessels

and other parameters that could aid in cardiovascular disease diagnosis

• Use DFT instead of FFT due to high-frequency and infinite-period

nature of waves FFT is suited to

References



Cardiovascular diseases (CVDs). (2018, September 26). Retrieved from https://www.who.int/cardiovascular_diseases/en/



Heart disease. (2018, March 22). Retrieved from https://www.mayoclinic.org/diseases- conditions/heart-disease/diagnosis-treatment/drc-20353124



Blood Vessels In The Head Blood Vessel Back Of Head This Diagram Shows The Blood Vessels In (n.d.). Retrieved September 3, 2018, from https://anatomyclass01.us/blood- vessels-in-the-head/blood-vessels-in-the-head-blood-vessel-back-of-head-this- diagram-shows-the-blood-vessels-in/. Accessed 3 Sept. 2018.



Khan, S. U., MD, DePersis, M., DO, & Kaluski, E., MD, FACC, FESC, FSCAI. (n.d.). Figure 1. Course of ulnar artery in forearm and palmar blood flow with puncture sites. [Digital image]. Retrieved September 3, 2018, from https://citoday.com/2017/10/ulnar- access-for-catheterization-and-intervention/. Accessed 3 Sept. 2018.



Sun, N., Garbey, M., Merla, A., & Pavlidis, I. (2005, June). Imaging the cardiovascular

• pulse. In Computer Vision and Pattern Recognition, 2005. CVPR 2005. IEEE Computer

Society Conference on (Vol. 2, pp. 416-421). IEEE. Accessed 3 Sept. 2018.