Automated Method for Doppler Echocardiography Analysis in Patients

Automated Method for Doppler Echocardiography Analysis in Patients with Atrial Fibrillation O. Shechner H. Greenspan M. Scheinowitz The Department of Biomedical Engineering and M. S. Feinberg The Heart institute, Sheba Medical Center, Tel Hashomer Tel Aviv University, Tel Aviv, Israel

Presentation structure Introduction Methods Results Conclusions

Introduction Ø Doppler echocardiography: l Non invasive modality for the assessment of cardiac function l Blood flow velocity tracing through the heart valves can be obtained by transthoracic Doppler echocardiography. l Extracted data: • Maximal Velocity Envelope (MVE) • Peak velocity • Peak and mean pressure • Velocity-time integral (VTI)

Transvalvular blood flow patterns Ø MV signals: “M” shape E A Ø TV signals: Gauss shape

Atrial Fibrillation (AF) is the most common sustained cardiac arrhythmia Ø AF characterized by irregular heart rate, electrogram and haemodynamic changes. Ø Ø MV signals: only E-wave present due to the loss of atrial contraction E E E Ø TV signals: inter-beat amplitude variability

Manual methods l l l Time consuming Inter and intra observer variability Difficulties when dealing with AF patients Early work Ø Doppler image analysis l l Ø MVE estimation by averaging points and fitting into a kinetic model (Hall et al, 1995 -1998) Edge detection-based algorithm for Brachial artery Doppler tracings (Tschirren et al, 2000) Validation using phantoms, simulations and normal patient groups

Our work Ø Automated analysis of MV and TV Doppler signals Ø Validation on a large dataset of both AF and non-AF patients

Proposed Framework Input Image separation into ECG and Signals Signal enhancement ECG analysis: segmentation into cardiac cycles Signal processing: Edge detection Rough MVE extraction Point linking Parameters Parameter curve fitting Parameter extraction

Methods Image separation Ø Dividing the image into region of interest (ROI) and ECG signal: l l The ECG signal is extracted by its color The location of the horizontal axis is found using horizontal projection – ROI extraction Original Image ROI ECG

Methods Image enhancement Segmentation of ROI pixels by their gray level into three clusters (K-means) Ø Contrast stretching improves image contrast and suppresses noise Ø High threshold background weak signal Low threshold strong signal

Image enhancement

Methods Signal processing: Edge detection Ø Combining the Sobel operator with the non- linear Laplace operator (NLLAP): d(x, , y) – Neighborhood of (x, y) l l l NLLAP introduces adaptive orientation of the Laplace operator Edge is detected at places of zero crossings Thresholding is applied on the edge strength

Methods Edge processing Sobel NLLAP Sobel + NLLAP + Post processing

Methods Rough MVE extraction Ø MVE vector is extracted from the edge image: l Using the biggest-gap algorithm a pixel is selected from each column

Methods Linking Ø The linking process is done beat-wise l maximal vertical value taken as anchor l Ascending and descending slopes are detected l Vertical “Noise level” is determined “noise level” l Starting slopes are determined; slopes are interpolated from starting slope to anchor point Anchor point

Methods Parameter fitting Ø The MVE is fitted into a parameter model using the Levenberg-Marquardt algorithm (MSE criteria) Ø Partial Fourier series model is used (TV: n=4; MV: n=5) Ø Parameter extraction

Methods Experimental Setup Ø Dataset: 467 beats from 121 images that were taken from 45 patients (25 AF, 20 non-AF) Ø Validation: l Beat-by-beat comparison between the automatically extracted parameters and the manually extracted parameters (two technicians) l Via Average-beat (manual vs calculated)

Results Ø MV results Non-AF AF Ø TV results Non-AF AF

Results: Technicians vs. Automatic Automated Vs Technician 1 non-AF AF MV : peak velocity 0. 9927 0. 9911 MV : VTI 0. 9892 TV : peak velocity 0. 9526 Automated Vs Technician 2 Automated Vs Technician avg non-AF AF MV : peak velocity 0. 9853 0. 9751 MV : peak velocity 0. 9925 0. 9891 0. 9812 MV : VTI 0. 9780 0. 9541 MV : VTI 0. 9896 0. 9754 0. 9445 TV : peak velocity 0. 9678 0. 9426 TV : peak velocity 0. 9628 0. 9434 Technician 1 Vs Technician 2 non-AF AF MV : peak velocity 0. 9895 0. 9759 MV : VTI 0. 9816 0. 9726 TV : peak velocity 0. 9703 0. 9537 non-AF AF

Results: Technicians vs. Automatic (cont. ) MV signals y = 1. 02 x + 5. 50 Peak velocity TV signals y = 0. 95 x + 0. 097 AF y = 1. 12 x + 7. 75 y = 1. 16 x + 0. 39 non-AF

Averaged Beat Experiments Ø Comparing the error between manual average and automated average to the error between manual average and representative beat Automated / Manual Mean error MV: peak velocity Non -AF MV : VTI TV : Peak Pressure MV: peak velocity AF MV : VTI TV : Peak Pressure Representative / Manual Mean error 2. 9% 6. 2% 6. 3% 13. 4% 4. 9% 6. 8% 9. 7% 8. 5% 4. 6% 9. 3% 13. 0% 6. 0%

Conclusions Ø The possibility of automated system for MV/TV Doppler image analysis was shown Ø The system is robust and manages to deal with both AF and non-AF signals with different morphology Ø Parameters are extracted from all the beats in the image, allowing the computation of an accurate average