mina emad azmy research assistant – signal and image processing lab
TRANSCRIPT
Toward Automated Assessment of Heart Regional Function
Mina Emad AzmyResearch Assistant – Signal and Image Processing Lab
Significance
Coronary artery disease is the leading
cause of death
Early diagnosis can help in preventing
heart attacks by providing better
diagnosis and treatment
Assessment of Regional Function using Cardiac MRI
Comprehensive method for
assessment of cardiac
regional function.
Provides great flexibility in
imaging the structure and
anatomy of the heart.
Considered the gold
standard for assessing Left
Ventricle regional function.
Tagged MR Parallel Planes of saturated
magnetization Applied orthogonal to the
imaging plane Move with the motion of the
heart• Permit quantification of its mechanical strain in three-coordinate directions (circumferential, longitudinal, and radial)
Harmonic Phase (HARP) MethodAutomatic tracking of cardiac
motion from a tagged MR image
sequence
Based on the idea that the phase of
each point is invariant with time
Easy to compute the strain after
acquiring these points with time
HARP Output
Peak Strain
Systolic Rate
Diastolic Rate
Objective: To extract diagnostic information from the strain curvesHARP Images Analysis
(Raw Strain Measurements)
Images Acquisition(Cardiac MRI)
Flow Diagram for Diagnostic Information
HARP Images Analysis(Raw Strain
Measurements)
Information ExtractionDecision / Info about Patient’s
condition
Images Acquisition(Cardiac MRI)
Measurements Inaccuracies
Denoising the Strain Measurements
Images Acquisition(Cardiac MRI)
HARP Images Analysis(Raw Strain
Measurements)
Information ExtractionDecision / Info about Patient’s
condition
Signal Denoising(Artifacts
Removing)
Understanding the Causes of the Errors
Method I
Method II
Images Acquisition(Cardiac MRI)
HARP Images Analysis(Raw Strain
Measurements)
Signal Denoising(Artifacts
Removing)
Causes of
Errors:
1- Noise
2- Tag
Overlapping
3- Twisting
4- Tag Fading
Method - 1
Apply non-linear filter on the strain curve
Ɛin = input strain curve
Ɛfilter = output of the filter
• Each peak in Ɛfilter corresponds to a noisy point
• It computes the difference between the input strain curve and its linear approximation
Method 1 – Cont’dActual Strain Curve
Filter Result
Curve fitting - Second order Fourier series
Time Derivative
Actual strain curveSpiky Locations
Denoised strain curveCurve fittingSecond order Fourier SeriesTime Derivative
Method 2 Based on observing real
strain curves, we noticed the spikes are most likely +ve spikes
They are considered as peaks in the strain curve having
Ɛin (t ) ˃ Ɛin (t + 1)AND
Ɛin (t) > Ɛin (t – 1)
• Strain at these locations is averaged
• This process is repeated till there are no spikes remaining
Evaluation of the techniques
Method I
Method II
Images Acquisition(Cardiac MRI)
HARP Images Analysis(Raw Strain
Measurements)
Signal Denoising(Artifacts
Removing)
Causes of
Errors:
1- Noise
2- Tag
Overlapping
3- Twisting
4- Tag Fading
Simulated Images
Experiments
Three experiments to evaluate the performance of the developed techniques
Simulated Tagged MR images were generated (DICOM format)
Simulation parameters: Max Contraction = 25% Tag Separation = 7mm to 8mm
Performance Evaluation: Root-Mean-Square (RMS) and Maximum error
for the strain curves and the extracted features
Experiment 1
Evaluate the techniques for different SNR levels
Noise Variance = 0.001 to 0.1 of the max intensity
SNR = 20 to 60 dB
For Systolic RateFor Diastolic RateFor Strain CurvesFor Peak Strain
Experiment 1 - Results
Experiment 2
Evaluate the technique for tags overlapping
This was done by removing the tags from
several time-frames around the max contraction
To simulate the aliasing of the tags, a tag line is
removed for 1 to 7 time-frames around end-
systole.
For Strain CurvesFor Peak StrainFor Systolic RateFor Diastolic Rate
Experiment 2 - Results
Combining noise effect and tag overlapping (Experiment 1 and 2, combined)
For Strain CurvesFor Peak StrainFor Systolic RateFor Diastolic Rate
Experiment 3
Summary and Conclusions Two methods were developed to denoise the
strain curves and accurately extract features Three experiments were performed to
evaluate the performance of the techniques Method I proved to robust to different
SNR levels Method II proved to efficiently recover
the strain curves when the overlapping tags occur
Artifacts appearing in the strain curves are more likely due to overlapping tags more than the SNR
Future Work
Images Acquisition(Cardiac MRI)
Images Analysis(Raw Strain
Measurements)
Information ExtractionDecision / Info about Patient’s
condition
Signal Denoising(Removing Artifacts)
Machine Learning
Thank You