Yohei Fukumizu

Signal and noise separation in medical diagnostic system based on independent component analysis

The possibility to study graphic recording (PCG -Phonocardiography) of auscultator findings is a helpful diagnostic tool for the clinician and forms the basis of early detection of the heart problems. Due to its dispersed nature and overlapping with breathing sounds Heart Sound Signals (HSS) is difficult to detect and comprehend in conventional PCG. We present a Hardware system utilizing Frequency-Domain Independent Component Analysis (ICA) deploying Direction of Arrival (DOA) and Beamforming (BF) techniques for the suppression of noise which will enhance the quality of HSS and aid the physicians. Such techniques of HSS extraction have been rarely studied in the past.

Detecting mass and its region in mammograms using mean shift segmentation and Iris Filter

In recent years, many Computer Aided Diagnosis (CAD) systems are suggested. Those systems can diagnose instead of a doctor, thus they are expected to reduce heavy burdens on the doctor during screening. The purpose of this study is to improve detection sensitivity for masses reducing the number of false positives as well as to extract mass regions accurately. In the proposed method, we focused on brightness and density of masses, thus we applied mean shift segmentation. After the segmentation, we obtained concentration of gradient vectors using Iris Filter and detected mass regions. According to the field test with a doctor, the proposed system was tested with 398 mammograms containing 193 masses. In the result of a performance test, a sensitivity of 81% was obtained at 5.0 false positives per image and 75% masses are detected at Area Overlap Measure (AOM) of more than 60%.

Local Gabor directional pattern histogram sequence (LGDPHS) for age and gender classification

This paper proposes means to classify age and gender in facial images using novel features. First, Gabor magnitude pictures are obtained by convolving the image with Gabor filters in several scale and orientation, followed by encoding with Local Directional Pattern (LDP) operator which enhances information. Then, the maps are divided into several blocks, and histograms are extracted from each block. The histograms are concatenated to a vector. Then Principal Component Analysis (PCA) is used to reduce the dimensions. Finally, the feature vector is classified by Support Vector Machine (SVM). The experimental results demonstrate that the algorithm proposed in this paper is effective method, compared to other similar methods.

Biosignals separation method for medical diagnostic system

In this paper, we will propose method to separate biosignals such as breath, blood, heart signal from mixed signals in body. As a result, we could get only target signal which is breath, blood and heart signals from actual mixture signals of carotid artery sound that extracted from a healthy human subject in a real environment using our algorithms and microphone sensors. Although this method is first trial to separate biosignals from sound in body, we get a significant result.

Robustly separating sound components in human body based on 2-ch ICA and EM algorithm with dirichlet distribution

An algorithm to separate breath sounds (BS), blood stream sounds (BSS), and heart sounds (HS) from sound components in the human body (biosignals) is introduced as a pre-process for detecting circulatory disease such as auricular fibrillation (AF), arteriosclerosis and apnea syndrome. Existing methods in the time-frequency model have been proposed to analyze biosignals with microphone sensors to obtain BS, BSS and HS. However, these methods have negative points. Thus, we propose band pass filter, 2-ch independent component analysis (ICA) and expectation-maximization (EM) algorithm with Dirichlet distribution to solve these problems. Experimental results show that our method performs better than existing methods.

Adaptive BP-RRC mixture model for background subtraction

In the real-time segmentation of the moving objects in a image sequence, methods based on background subtraction are used widely. Bi-polar Radial Reach Correlation (BP-RRC) has realized robust background subtraction by evaluating a local texture per pixel, suppressing the influence of brightness change. However, this method has a fault that is affected by the local illumination change of a background. This paper proposes a robust and stabilized background subtraction algorithm which can cope with texture change of various illumination changes by improving this texture background model. We verified the proposed method using the image sequence including a loose lighting change and a rapid lighting change, and adaptability higher than the conventional method on the sudden illumination changes was attained as a result of verification.

Visual categorization method with a Bag of PCA packed Keypoints

Visual categorization is one of a key function in the next generation of a driving assist system, which is expected to reduce a traffic accident. This paper proposes a high performance visual categorization method, which is based on Feature Accelerated Segment Test (FAST) feature point detectors, Histograms of Oriented Gradients (HOG) feature descriptors and Bag-of-Keypoints (BoK). Each feature descriptors were orthogonalized by applying the Principal Component Analysis (PCA) to reduce the size of dimension. As a result, our proposed method has achieved the recognition rate of 69.5% and the performance of 43.1 ms on a PC in order to categorize one object in an image into traffic related categories, i.e. pedestrians, cars, bikes, bicycles, and so on. The comparison with conventional methods will be also discussed.

Higher frequency components estimating method using codebooks based on formants

A talk with communication device, such as a cell phone, loses information of language because of a limited frequency range. Currently, we research algorithm estimating higher frequency components from the limited signal. We propose method to estimate codebooks using formants based on Hilbert transform. Firstly we prepare temporary limited signals from the original signals from raw voice sound sources. And then we derive the estimated signals from limited signal using proposed method. The experimental results successfully demonstrate that the method restores a part in higher frequency. As shown in the characteristic frequency, the proposed method is better than only using Hilbert transform, comparing to formant of the based signal. The estimated signal can be heard clearly when the formant in the original signal becomes closer.