Sueharu Miyahara

Classification between normal and abnormal respiratory sounds based on maximum likelihood approach

In this paper, we have proposed a novel classification procedure for distinguishing between normal respiratory and abnormal respiratory sounds based on a maximum likelihood approach using hidden Markov models. We have assumed that each inspiratory/expiratory period consists of a time sequence of characteristic acoustic segments. The classification procedure detects the segment sequence with the highest likelihood and yields the classification result. We have proposed two elaborate acoustic modeling methods: one method is individual modeling for adventitious sound periods and for breath sound periods for the detection of abnormal respiratory sounds, and the other is a microphone-dependent modeling method for the detection of normal respiratory sounds. Classification experiments conducted using the former method revealed that this method demonstrated an increase of 19.1% in its recall rate of abnormal respiratory sounds as compared with the recall rate of a baseline method. It has also been revealed that the latter modeling method demonstrates an increase in its recall rate for the detection of not only normal respiratory sounds but also for abnormal respiratory sounds. These experimental results have confirmed the validity of our proposed classification procedure.

Discrimination between healthy subjects and patients with pulmonary emphysema by detection of abnormal respiration

In this paper, we propose a robust classification strategy for distinguishing between a healthy subject and a patient with pulmonary emphysema on the basis of lung sounds. A symptom of pulmonary emphysema is that almost all lung sounds include some abnormal (i.e., adventitious) sounds. However, the great variety of possible adventitious sounds and noises at auscultation makes high-accuracy detection difficult. To overcome this difficulty, our strategy is to adopt a two-step classification approach based on the detection of “confident abnormal respiration.” In the first step, hidden Markov models and bigram models are used for acoustic features and the occurrence of acoustic segments in each abnormal respiratory period, respectively, to calculate two kinds of stochastic likelihoods: the highest likelihood for a segment sequence to be abnormal respiration and the likelihood for normal respiration. In the second step, the patients are identified on the basis of the detection of confident abnormal respiration, which is when difference between these two likelihoods is larger than a predefined threshold. Our strategy achieved the highest classification rate of 88.7% between healthy subjects and patients among three basic classification strategies, which shows the validity of our approach.

On-line cursive Kanji character recognition as stroke correspondence problem

This paper describes a stroke-number and stroke-order free on-line Kanji character recognition method by a joint use of two complementary algorithms of optimal stroke correspondence determination: one dissolves excessive mapping and the other dissolves deficient mapping. Also, three kinds of inter-stroke distances are devised to deal with stroke concatenation or splitting and heavy shape distortion. Only a single reference pattern for each of 2,980 Kanji character categories is generated by using training data composed of 120 patterns written with the correct stroke-number and stroke-order. Recognition tests are made using the training data and two kinds of resting data in the square style and in the cursive style written by 36 different people; recognition rates of 99.5%, 97.6%, and 94.1% are obtained.

On-line cursive Kanji character recognition using stroke-based affine transformation

This paper describes a distortion-tolerant online Kanji character recognition method using stroke-based affine transformation (SAT). The first part of the method determines one-to-one stroke correspondence between an input pattern and each reference pattern. The second part applies optimal SAT to each stroke of the input pattern to absorb handwriting distortion. The last part calculates the inter-pattern distance between the reference pattern and the SAT-superimposed input pattern. Only a single reference pattern for each of 2,980 Kanji character categories is generated by using training data written carefully with the correct stroke-number and stroke-order. Recognition tests are made using two kinds of test data in the square style and in the cursive style written by 36 different people; recognition rates of 98.4% and 96.0% are obtained.

Comparing LDA with pLSI as a dimensionality reduction method in document clustering

In this paper, we compare latent Dirichlet allocation (LDA) with probabilistic latent semantic indexing (pLSI) as a dimensionality reduction method and investigate their effectiveness in document clustering by using real-world document sets. For clustering of documents, we use a method based on multinomial mixture, which is known as an efficient framework for text mining. Clustering results are evaluated by F-measure, i.e., harmonic mean of precision and recall. We use Japanese and Korean Web articles for evaluation and regard the category assigned to each Web article as the ground truth for the evaluation of clustering results. Our experiment shows that the dimensionality reduction via LDA and pLSI results in document clusters of almost the same quality as those obtained by using original feature vectors. Therefore, we can reduce the vector dimension without degrading cluster quality. Further, both LDA and pLSI are more effective than random projection, the baseline method in our experiment. However, our experiment provides no meaningful difference between LDA and pLSI. This result suggests that LDA does not replace pLSI at least for dimensionality reduction in document clustering.

Adaptive Subpixel Estimation of Land Cover in a Remotely Sensed Multispectral Image

Land surface corresponding to a pixel of remotely sensed image does not necessarily consist of only one category of objects. Several techniques of subpixel analysis have been developed which estimate the proportion of components of land cover in a pixel. However, when the available training data do not correctly represent the spectral characteristics of the categories in the pixel, large errors may appear in the results of estimation. The method of unsupervised estimation of component spectra has been presented to solve this problem. In this paper we present a method which apply the unsupervised analysis technique to subpixel estimation of land cover in an image in which spectral characteristics change with the location of the objective area. After partitioning the image into blocks, the number of categories and their component spectra are estimated in each block. Then the proportion of category are estimated for each pixel using the component spectra derived in the block. We confirmed the validity of this method by numerical simulation

Unmixing three types of lung sounds by convex optimization

We present a convex optimization technique for unmixing lung sounds to improve computer-aided pulmonary auscultation. An auscultatory sound of a patient with pulmonary disorder may be composed of continuous and discontinuous adventitious sounds as well as breath. Our technique exploits sparse and low-rank properties of these sounds in the Fourier, wavelet, and time-frequency domains, which can be quantified as convex functions. The optimization algorithm is derived from the alternating direction method of multipliers (ADMM). This approach enables the lung sound unmixing without training data for learning diverse structures of lung sounds in time-frequency domains. We show some experimental examples and discuss further improvements.

Subpixel estimation of land cover in a remotely sensed image using spectral information of surrounding pixels

Several techniques of subpixel analysis for remotely sensed image have been developed which estimate the proportion of components of land cover in a pixel. However, when the available training data do not correctly represent the spectral characteristics of the categories in the pixel, large errors may appear in the results of estimation. In this paper, we propose a semi-supervised method of subpixel estimation of land cover for remotely sensed multispectral image. First we provide small size of initial training data and determine pure pixels in the image. In the next step, component spectra are adaptively estimated for each mixed pixel using the surrounding pure pixels. Then the proportions of components in the mixed pixels are estimated based on the determined component spectra. We confirmed the validity of this method by numerical simulation and applied it to a remotely sensed multispectral image.

Sparse representation-based extraction of pulmonary sound components from low-quality auscultation signals

Toward assistance of respiratory system diagnosis, sparse representation of auscultation signals is utilized to extract pulmonary sound components. This signal extraction is a challenging task because the pulmonary sounds such as vesicular sounds and crackles are overlapping each other in the time and frequency domains, and they are so faint that the quality of recorded signals is quite low in many cases. It is experimentally shown that the pulmonary sound components are successfully extracted from low-quality auscultation signals via the sparse representation. This extraction method is confirmed to be highly robust against random noise and digital quantization.

Browsing and Similarity Searching Method for Videos Based on Cluster Analysis

We developed a browsing and similarity searching method for videos based on cluster analysis. First, videos are segmented into shots by fuzzy clustering of graph spectral methods. Second, the videos are represented as a sequence of symbols by grouping together shots. According to this representation, the directed graph of videos is formed based on the relationship between these symbols. Initial and terminal shots are extracted from the di-rected graph using fuzzy cluster extraction. The shots can be browsed from the initial shots to the terminal ones sequentially. Selected shots are used as a query video on similarity searches. The performance of the proposed method was evaluated using a video dataset from NASA.