Damjan Zazula

Multichannel Blind Source Separation Using Convolution Kernel Compensation

This paper studies a novel decomposition technique, suitable for blind separation of linear mixtures of signals comprising finite-length symbols. The observed symbols are first modeled as channel responses in a multiple-input-multiple-output (MIMO) model, while the channel inputs are conceptually considered sparse positive pulse trains carrying the information about the symbol arising times. Our decomposition approach compensates channel responses and aims at reconstructing the input pulse trains directly. The algorithm is derived first for the overdetermined noiseless MIMO case. A generalized scheme is then provided for the underdetermined mixtures in noisy environments. Although blind, the proposed technique approaches Bayesian optimal linear minimum mean square error estimator and is, hence, significantly noise resistant. The results of simulation tests prove it can be applied to considerably underdetermined convolutive mixtures and even to the mixtures of moderately correlated input pulse trains, with their cross-correlation up to 10% of its maximum possible value.

Correlation-based decomposition of surface electromyograms at low contraction forces

The paper studies a surface electromyogram (SEMG) decomposition technique suitable for identification of complete motor unit (MU) firing patterns and their motor unit action potentials (MUAPs) during low-level isometric voluntary muscle contractions. The algorithm was based on a correlation matrix of measurements, assumed unsynchronised (uncorrelated) MU firings, exhibited a very low computational complexity and resolved the superimposition of MUAPs. A separation index was defined that identified the time instants of an MUs activation and was eventually used for reconstruction of a complete MU innervation pulse train. In contrast with other decomposition techniques, the proposed approach worked well also when the number of active MUs was slightly underestimated, if the MU firing patterns partly overlapped and if the measurements were noisy. The results on synthetic SEMG show 100% accuracy in the detection of innervation pulses down to a signal-to-noise ratio (SNR) of 10 dB, and 93±4.6% (mean± standard deviation) accuracy with 0 dB additive noise. In the case of real SEMG, recorded with an array of 61 electrodes from biceps brachii of five subjects at 10% maximum voluntary contraction, seven active MUs with a mean firing rate of 14.1 Hz were identified on average.

Combined edge detection using wavelet transform and signal registration

This paper presents a novel edge detection algorithm, using Haar wavelet transform and signal registration. The proposed algorithm has two stages: (a) adaptive edge detection with the maximum entropy thresholding technique on time-scale plane and (b) edge linkage into a contour line with signal registration in order to close edge discontinuities and calculate a confidence index for contour linkages. This index measures the level of confidence in the linkage of two adjacent points in the contour structure. Experimenting with synthetic images, we found out the lower level of confidence can be set to approximately e^-^2. The method was tested on 200 synthetic images at different signal-to-noise ratios (SNRs) and 11 clinical images. We assessed its reliability, accuracy and robustness using the mean absolute distance (MAD) metric and our confidence index. The results for MAD on synthetic images yield the mean of 0.7 points and standard deviation (std) of 0.14, while the mean confidence level is 0.48 with std of 0.19 (the values are averaged over SNRs from 3 to 50dB each in 20 Monte-Carlo runs). Our assessment on clinical images, where the references were experts annotations, give MAD equal 1.36+/-0.36 (mean+/-std) and confidence level equal 0.67+/-0.25 (mean+/-std).

Automated analysis of a sequence of ovarian ultrasound images. Part I: segmentation of single 2D images

Abstract An improved algorithm is presented for ovarian follicle detection in ultrasound images. This fully automated recognition algorithm is composed of three successive steps. First, initial homogeneous regions are determined. Then, these initial regions are grown. The growing is controlled by average grey-level and by a newly introduced weighted image gradient. In the last stage, those regions are extracted that probably correspond to the follicles. The algorithm has been tested on 50 ovarian ultrasound images. The recognition rate of follicles using this procedure was around 78%. A possible extension of the algorithm deals with the entire information in the ultrasound image sequence, which is covered in Part II of this paper.

Multiscale entropy-based approach to automated surface EMG classification of neuromuscular disorders

We introduce a novel method for an automatic classification of subjects to those with or without neuromuscular disorders. This method is based on multiscale entropy of recorded surface electromyograms (sEMGs) and support vector classification. The method was evaluated on a single-channel experimental sEMGs recorded from biceps brachii muscle of nine healthy subjects, nine subjects with muscular and nine subjects with neuronal disorders, at 10%, 30%, 50%, 70% and 100% of maximal voluntary contraction force. Leave-one-out cross-validation was performed, deploying binary (healthy/patient) and three-class classification (healthy/myopathic/neuropathic). In the case of binary classification, subjects were distinguished with 81.5% accuracy (77.8% sensitivity at 83.3% specificity). At three-class classification, the accuracy decreased to 70.4% (myopathies were recognized with a sensitivity of 55.6% at specificity 88.9%, neuropathies with a sensitivity of 66.7% at specificity 83.3%). The proposed method is suitable for fast and non-invasive discrimination of healthy and neuromuscular patient groups, but it fails to recognize the type of pathology.

Adjustments Differ Among Low-Threshold Motor Units During Intermittent, Isometric Contractions

We investigated the changes in muscle fiber conduction velocity, recruitment and derecruitment thresholds, and discharge rate of low-threshold motor units during a series of ramp contractions. The aim was to compare the adjustments in motor unit activity relative to the duration that each motor unit was active during the task. Multichannel surface electromyographic (EMG) signals were recorded from the abductor pollicis brevis muscle of eight healthy men during 12-s contractions (n = 25) in which the force increased and decreased linearly from 0 to 10% of the maximum. The maximal force exhibited a modest decline (8.5 +/- 9.3%; P < 0.05) at the end of the task. The discharge times of 73 motor units that were active for 16-98% of the time during the first five contractions were identified throughout the task by decomposition of the EMG signals. Action potential conduction velocity decreased during the task by a greater amount for motor units that were initially active for >70% of the time compared with that of less active motor units. Moreover, recruitment and derecruitment thresholds increased for these most active motor units, whereas the thresholds decreased for the less active motor units. Another 18 motor units were recruited at an average of 171 +/- 32 s after the beginning of the task. The recruitment and derecruitment thresholds of these units decreased during the task, but muscle fiber conduction velocity did not change. These results indicate that low-threshold motor units exhibit individual adjustments in muscle fiber conduction velocity and motor neuron activation that depended on the relative duration of activity during intermittent contractions.

Cellular automata and follicle recognition problem and possibilities of using cellular automata for image recognition purposes

Cellular automata are discrete dynamical systems whose behaviour is completely specified in terms of a local relation. Guided by a suitable recipe, they can simulate a whole hierarchy of structures and phenomena. While investigating the problem of follicle recognition in ultrasonic images of womens ovaries, we became increasingly interested in using cellular automata for this purpose. We were very successful, which encouraged us to further investigate the use of cellular automata for image recognition purposes in general. This paper presents the results of our research in this area, along with the details of how we solved the follicle recognition problem.

Heartbeat and Respiration Detection From Optical Interferometric Signals by Using a Multimethod Approach

In this paper, a multimethod approach for heartbeat and respiration detection from an optical interferometric signal is proposed. Optical interferometer is a sensitive device that detects physical changes of optical-fiber length due to external perturbations. When in direct or indirect contact with human body (e.g., hidden in a bed mattress), mechanical and acoustic activity of cardiac muscle and respiration reflect in the interferometric signal, enabling entirely unobtrusive monitoring of heartbeat and respiration. A novel, two-phased multimethod approach was developed for this purpose. The first phase selects best performing combinations of detection methods on a training set of signals. The second phase applies the selected methods to test set of signals and fuses all the detections of vital signs. The test set consisted of 14 subjects cycling an ergometer until reaching their submaximal heart rate. The following resting periods were analyzed showing high efficiency (98.18 ± 1.40% sensitivity and 97.04 ± 4.95% precision) and accuracy (mean absolute error of beat-to-beat intervals 22±9 ms) for heartbeat detection, and acceptable efficiency (90.06 ± 7.49% sensitivity and 94.21 ± 3.70% precision) and accuracy (mean absolute error of intervals between respiration events 0.33 ± 0.14 s) for respiration detection.

SEGMENTATION OF OVARIAN ULTRASOUND IMAGES USING CELLULAR NEURAL NETWORKS

Segmentation of ovarian ultrasound images using cellular neural networks (CNNs) is studied in this paper. The segmentation method consists of five successive steps where the first four uses CNNs. In the first step, only rough position of follicles is determined. In the second step, the results are improved by expansion of detected follicles. In the third step, previously undetected inexpressive follicles are determined, while the fourth step detects the position of ovary. All results are joined in the fifth step. The templates for CNNs were obtained by applying genetic algorithm. The segmentation method has been tested on 50 ovarian ultrasound images. The recognition rate of follicles was around 60% and misidentification rate was around 30%.

An approach to surface EMG decomposition based on higher-order cumulants

We are addressing a possible approach to the decomposition of surface electromyograms (SEMGs). It is based on higher-order cumulants implemented in a two-step procedure. Firstly, a multivariate version of the w-slice method is applied in order to extract coarse approximations of motor-unit action potentials (MUAPs) out of the measured SEMGs. Secondly, these coarse estimates are refined by modified Newton-Gauss iteration to achieve an optimum fit of the model-based and the observation-based cumulant estimates. All the necessary conditions are derived theoretically and, afterwards, implemented in simulation runs in order to prove the decomposition power of the proposed approach on synthetic SEMGs. The first-norm difference between the original and the decomposed MUAPs, obtained at the signal length of 102400 samples and expressed in percentage of the MUAP amplitude span, yields 5.4% in the noise-free case, 6.0% with a signal-to-noise ratio (SNR) of 10dB, and 6.5% with a SNR of 0 dB.