Presentation on theme: "SEGMENTATION ON PHONOCARDIOGRAM"— Presentation transcript:
1SEGMENTATION ON PHONOCARDIOGRAM Professor KuhEE645 FINAL PROJECTSBy Rebecca Longstreth
2DEFINITIONS Phonocardiogram Cardio cycle S1 S2 S3 and S4 Graph representing sounds made by the human heartCardio cycleRepresented by S1, S2, S3 and S4S1Orignates at the closure of the atrioventricular valvesrecordable between 91hz and 179 hzS2Originates at the aortic an pulmonic valvesCan reach 200 hzS3 and S4Represent cardiac wall vibrationsNot all audible, low density
3DEFINITIONSS1 and S2Sound goes left to rightS3 and S4Right to left
4Reason for Segmentation Tool to assist doctors in diagnosing heart malfunctions accuratelyPrevents mistakes in diagnosesImproved quality care for patientsDoctors experience benefits everyone
5Choices of Segmentation Fourier transformLaplace transformWavelet decomposition
6Fourier transform Reasons not used: Phonocardiograms are not smooth wave signalsNot linear time invarianceNon-stationaryNon-harmonic
7Laplace transformReasons not used:Initial condition does not exist
8Wavelet decomposition Preferred method of analyzing phonocardiogramsTime limited and frequency limitedUtilize digital filters and down sampling
9Wavelet decomposition conditions Normalize the amplitude of all artifacts before transformation to avoid amplification errorsShort segmentation window for high frequencyLong segmentation window for low frequencyShort lengths of data are considered due to file size considerations
10Phonocardiogram anomalies Fetal acoustic signals inconsistent from one sound to the nextWave shape can changeFrequency could shiftAmplitude, duration and position in the cardiac cycle can changeBackground noise of the mother and the shielding affect of the womb
11GOAL Isolate one cycle for analysis Identify S1 and S2 Boundaries of S1 and S2Systolic and diastolic periodsSystolic: time interval from beginning of S1 to the beginning of S2Diastolic: time interval from beginning of S2 to the beginning of S1Associate frequency spectrum
12Research According to Cardiologists interviewed There is no way to recognize S1 and S2 using only a phonocardiogramLocation of the probe can radically affect the signal strengths of S1 and S2 on the same individual
13Finding S1 and S2: Method 1Compare EKG and the phonocardiogram simultaneously to ensure accurate designation of S1 and S2S1 occurs shortly after the EKG peakThe first PCG signal following the EKG peak is S1S2 occurs shortly after the 2nd EKG peak
14Finding S1 and S2: Method 1 Disadvantage Need EKG requires expensive bulky and fragile equipment to performNeedRobust, portable, easy to use low cost equipment
16Wavelet Decomposition Heart sounds (sampled at an 8kHz sample rate, 16 bits/sample) are first hand segmented into 4096 sample segments, each consisting of a single heartbeat cycle.
17Feature Reduction & Denoising Figure 1. A Simple Heart Sound Classification System.Each segment is transformed using a 7 level wavelet decomposition, based on a Coifman 4th order wavelet kernel (relative symmetry and fast execution).The resulting transform vectors, 4096 values in length, are reduced to 256 element feature vectors by discarding the 4 levels with shortest scale.Neural network in the classifier reduces noise. The magnitudes of the remaining coefficients in each vector are calculated, then normalized by the vector’s energy.
18ClassificationEach feature vector is classified using a three layer neural network (256 input nodes, 50 hidden nodes, and 5 output nodes).
19Results And Discussion The system was evaluated using heart sounds corresponding tonormalmitral valve prolapse(MVP)coarctation of the aorta (CA),ventricular septal defect (VSD),and pul-monary stenosis (PS).The classifier was trained using 10 shifted versions (over a range of 100 samples) of a single heartbeat cycle from each type.
20Figure 2. Representative Heart Sounds (left to right) Without Added Noise, with Noise Variance 1000, and with Noise Variance 3000, x-axis is the number of sample segments
21Figure 2. Representative Heart Sounds (left to right) Without Added Noise, with Noise Variance 1000, and with Noise Variance 3000,x-axis is the number of sample segments
22Figure 3. Feature Vectors Corresponding to the Heart Sounds in Figure 2, x-axis is the number of feature vectors
23Figure 3. Feature Vectors Corresponding to the Heart Sounds in Figure 2, x-axis is the number of feature vectors
24Figure 3. Feature Vectors Corresponding to the Heart Sounds in Figure 2,x-axis is the number of feature vectors
25Figure 3. Feature Vectors Corresponding to the Heart Sounds in Figure 2,x-axis is the number of feature vectors
26Feature vectors with additive noise The feature vectors produced for these examples in Figure 3.key features remain relatively stable even with addiditive noise.
27Figure 4. Classification Accuracy (in Percent) as a Function of the Variance of the Added Noise
28Figure 5. Classification Accuracy as a Function of Signal-to-Noise Ratio (in dB)
29the sounds differ widely (e. g the sounds differ widely (e.g., by a factor of approximately 16:1 comparing a typical normal heartbeatwith one exhibiting VSD). Accounting for this variation, classification accuracy as a function of signal-to-noise ratio (SNR) is shown in Figure 5. For an SNR above 31dB (which is easily obtainable undermost practical circumstances) classification accuracy is 100%.
30REFERENCESBarschdorff, D., U. Femmer, and E. Trowitzsch (1995, Sept ). Automatic phonocardiogram signal analysis in infants based on wavelet transforms and artificial neural networks. In Computers in Cardiology 1995, pp. 753–756. IEEE, Vienna, Austria.Barschdorff, D., U. Femmer, and E. Trowitzsch (1995, Sept ). Automatic phonocardiogram signal analysis in infants based on wavelet transforms and artificial neural networks. In Computersin Cardiology 1995, pp. 753–756. IEEE, Vienna, Austria.Donnerstein, R. L. and V. S. Thomsen (1994, September). Hemodynamic and anatomic factors affecting the frequency content of Still’s innocent murmur. The American Journal of Cardiology 74, 508–510.Durand, L.-G. and P. Pibarot (1995). Digital signal processing of the phonocardiogram: review of the most recent advancements. Critical Reviews in Biomedical Engineering 23(3/4), 163–219.El-Asir, B., L. Khadra, A.H. Al-Abbasi, and M.M.J. Mohammed (1996, Oct ). Multireso-lution analysis of heart sounds. In Proc. of the Third IEEE Int’l Conf. on Elec., Circ., and Sys., Volume 2, pp. 1084–1087. Rodos, Greece.Rajan, S., R. Doraiswami, R. Stevenson, and R. Watrous (1998, Oct. 6-9). Wavelet based bankof correlators approach for phonocardiogram signal classification. In Proc. of the IEEE-SP Int’l Symp. on Time-Frequency and Time-Scale Analysis, pp. 77–80. Pittsburgh, PA.
31REFERENCESShino, H., H. Yoshida, K. Yana, K. Harada, J. Sudoh, and E. Harasawa (1996, Oct Nov. 3). Detection and classification of systolic murmur for phonocardiogram screening. In Proc. of the18th Int’l Conf. of the IEEE Eng. in Med. and Biol. Soc., Volume 1, pp. 123–124. Amsterdam, The Netherlands.THE ANALYSIS OF HEART SOUNDS FOR SYMPTOM DETECTION AND MACHINE-AIDED DIAGNOSIS, Todd R. Reed, Nancy E. Reed and Peter Fritzson, The NetherlandsH. Liang, S. Lukkarinen, and I Hartimo, Heart Sound Segmentation Algorithm Based on Heart Sound Envelogram, Helsinki, FinlandH. Liang, S. Lukkarinen, and I Hartimo, A Boundary Modification Mehtod for Sound Segmentation Algorithm, Helsinki, FinlandAbdelhani Djebbari, and Fethi Bereski Reguig, Short-time Fourier Transform Analysis of the Phonocardiogram Signal, Algiers