We have developed a computer-aided diagnosis scheme for the detection of clustered microcalcifications in mammograms. Using a filter bank, the mammogram image is first decomposed into eight sub-images for extracting nodular patterns and nodular & linear patterns at scales from 1 to 4. Many regions of interest (ROI) with 100×100 matrix sizes are then selected from the mammogram image, and eight features determined in each ROI are obtained from sub-images for nodular patterns and nodular & linear patterns. In the first step of this computerized method for identifying initial candidates, a classifier based on maximum likelihood with eight features is used to distinguish between clustered microcalcifications and normal tissues. In the second step, an artificial neural network with 32 features, including the root-mean-square variation and the first moment of the power spectrum, is employed to reduce false positives (FPs). We evaluated the detection performance of the new scheme using a database of 331 mammograms. The detection scheme had a sensitivity of 96.5% with the number of FPs being 0.69 per mammogram. This computerized method may be a useful tool to assist radiologists in the detection of clustered microcalcifications in mammograms.

View full abstract

Presently, doctors predict the condition of hepatitis C using blood examination data based on their professional experience, and patients are then diagnosed by performing a liver biopsy to obtain a definite diagnosis. However, liver biopsies are a high-risk procedure and can be troublesome. In this paper, we suggest a new method that is easier and more accurate. It uses the SVM (support vector machine), which is one of the most effective learning machines, and SFFS (sequential forward floating search), which is a feature selection. The combination of SVM and SFFS make it possible to eliminate the unnecessary examination of various items. It also helps to obtain high accuracy compared to using only SVM. Performance was drastically improved by applying our new method to the blood examination data for hepatitis C.

View full abstract

Humans have the ability to recognize objects regardless of the viewing angle. Event-related potential (ERP) was introduced in this study to measure the electrophysiological activity underlying view-invariant object recognition. Two tasks were modified based on a delayed matching-to-sample task. One was an object recognition task, in which subjects had to recognize the same 3D object appearing as a sample regardless of the viewing angle. The other, as a control, was an image identification task, in which the same images as used in the object recognition task were used and the subjects had to compare two images to judge if they were the same. In both of the tasks, a negative component at around 150 ms, N1, after the onset of the test image was observed at the electrodes on the surface of the occipito-parieto-temporal cortex. When the tasks were compared, no significant difference in the N1 peak was found between them at beginning, but they became significantly different after extensive training in view-invariant object recognition. Modulation was observed at almost all of the occipitoparieto-temporal electrodes. These results demonstrate that the posterior N1 is sensitive to view-variant object recognition.

View full abstract

We developed a new probe to examine the anisotropy of dermis tissue by ultrasonic velocity measurement and confirmed the reliability of the method. Skin has the characteristic of anisotropy caused by collagen fiber in the dermis tissue. Though the characteristics are thought to be useful for the diagnosis of skin disease, the examination of anisotropy is very difficult and no optimum method has been devised. The characteristics of anisotropy correspond to the arrangement of collagen fiber and the direction of maximum sound velocity is the same as that of fiber. But determining the maximum sound velocity direction is very difficult in a thin skin layer. We applied a couple of wedges conventionally used in the field of acoustic emission to the measurement of sound velocity in skin and designed a sound path in a skin layer with a small load. By examining a skin model made of gelatin, we confirmed the sound propagation path we had expected. We then measured the sound velocity of skin as a function of direction in a human forearm and estimated the fibrous direction.

View full abstract

We have been developing a classification scheme for breast masses on mammograms as a part of a computer-aided diagnosis (CAD) system. In this study, the problem of classifying masses into benign and malignant using shape features is addressed. The type of the mass contour, such as circumscribed and microlobulated, is one of the features, and the recognition of this type is very important for classification. We attempted to improve recognition accuracy by using a new method to detect the saw-tooth region of a mass contour. The mass contours analyzed using the proposed method were drawn manually and extracted automatically by a method based on an active contour model to label them as convex segments. The shape features were calculated from the change of the slope of the tangent to the contour in the polar coordinate. A total of 160 masses (127 circumscribed and 33 microlobulated) were extracted from digitized mammograms for shape recognition. Moreover, a total of 202 masses (124 benign masses and 78 malignant) were used for benign/malignant classification. The corresponding accuracy using manually drawn contours was 88% (141/160). As a result of benign/malignant classification using this method, the classification rate was 84% (169/202) and the value was high compared to our conventional method based on fractal dimension. Automatically extracted contours achieved an accuracy of 76% (121/160) for shape recognition and an accuracy of 73% (148/202) for benign/malignant classification. The results demonstrate the feasibility of using the saw-tooth region detection method in shape recognition for classifying of benign and malignant masses on mammograms in a computer-aided diagnosis scheme.

View full abstract

The development of an automatic electroencephalogram (EEG) interpretation method is desired to reduce the labor of electroencephalographers. We have already proposed an EEG model that consists of sinusoidal waves with Markov process amplitude to represent the characteristics of the background EEG quantitatively. An automatic interpretation method for the background EEG was previously developed by some of the authors, and the results of the method brought satisfactory coincidence with the results interpreted by a qualified electroencephalographer. However, when the EEG is recorded with trend artifacts, it is highly probable that using the method could result in misjudgment. In this study, an automatic EEG interpretation method using an EEG model is proposed to improve the accuracy of judgment. First, the “period of amplitude variability of dominant rhythm” and “organization of dominant rhythm”, which are important factors in EEG interpretation, are represented quantitatively using EEG model parameters. Using the relationships between the model parameters and the characteristics of dominant rhythm, an automatic EEG interpretation method using the EEG model is proposed. In the proposed method, an EEG model that takes account of trend artifacts is constructed to separate the trend artifacts from the background EEG, and specific EEG parameters for each judgment are calculated from the EEG model parameters. The proposed method was applied to actual EEG data, and the results compared with those obtained by the previous automatic EEG interpretation method. The proposed method provided improved judgment for EEG data contaminated with trend artifacts, although misjudgments were reported in the previous method. Furthermore, the proposed method can judge the “organization of dominant rhythm” much more accurately than the previous method.

View full abstract

The heart rate time series contains two main rhythms : RSA (respiratory sinus arrhythmia), which is synchronous with the respiration rate, and the Mayer wave, which is generally centered around 0.1 Hz. The purpose of this study is the estimation of respiratory variability, such as the variation in respiratory frequency or amplitude, from the RSA time series. To calculate the time series of RSA power and RSA frequency, a wavelet transform was applied to the R-R interval time series. There was a significant correlation between the average frequency of RSA and the average respiratory frequency (r = 0.94, p < 0.01). The correlation coefficient of the standard deviation of RSA amplitude and the standard deviation of the respiratory amplitude time series was 0.79 (p < 0.01). The dominant frequency of the RSA amplitude and that of the respiratory amplitude were equal within the error range of ±0.02 Hz. There is a possibility that the respiratory variability, such as the respiratory frequency or amplitude, can be estimated from the RSA time series. The heart rate variability, which can be measured by a non-restricted method, is used for long-term monitoring of respiratory time series in daily life.

View full abstract