論文 - 河野　英昭

An artificial neural network−pharmacokinetic model and its interpretation using Shapley additive explanations 査読有り

Ogami C., Tsuji Y., Seki H., Kawano H., To H., Matsumoto Y., Hosono H.

CPT: Pharmacometrics and Systems Pharmacology   10 ( 7 )   760 - 768   2021年07月

　詳細を見る

記述言語：英語   掲載種別：研究論文（学術雑誌）

We developed a method to apply artificial neural networks (ANNs) for predicting time-series pharmacokinetics (PKs), and an interpretable the ANN-PK model, which can explain the evidence of prediction by applying Shapley additive explanations (SHAP). A previous population PK (PopPK) model of cyclosporin A was used as the comparison model. The patients’ data were used for the ANN-PK model input, and the output by ANN was the clearance (CL). The estimated CL value from the ANN were substituted into the one-compartment with one-order absorption model, the concentrations were calculated, and the parameters of ANN were updated by the back-propagation method. Kernel SHAP was applied to the trained model and the SHAP value of each input was calculated. The root mean squared error for the PopPK model and the ANN-PK model were 41.1 and 31.0 ng/ml, respectively. The goodness of fit plots for the ANN-PK model represented more convergence to y = x compared with that for the PopPK model, with good model performance for the ANN-PK model. The most influential factors on CL output were age and body weight from the evaluation using Kernel SHAP, and these factors were incorporated into the PopPK model as the significant covariates of CL. The ANN-PK model could handle time-series data and showed higher prediction accuracy then the conventional PopPK model, and the scientific validity for the model could be evaluated by applying SHAP. Study Highlights WHAT IS THE CURRENT KNOWLEDGE ON THE TOPIC? A black-box property of an artificial neural network (ANN) decreases the scientific confidence of the model, and making it difficult to utilize the ANN in the medical field. Moreover, difficulty in handling the time-series data is a significant problem for applying the ANN for pharmacometrics study. WHAT QUESTION DID THIS STUDY ADDRESS? How can we apply the ANN for predicting the time-series pharmacokinetics (PKs), and confirm the scientific validity of the ANN model? WHAT DOES THIS STUDY ADD TO OUR KNOWLEDGE? Using the ANN in combination with a conventional compartment (ANN-PK) model enabled to handle the time-series PK data, and the predicting performance of the model was higher than that of the population PK model. Furthermore, we could evaluate the scientific validity of the ANN model by applying the Shapley additive explanations. HOW MIGHT THIS CHANGE DRUG DISCOVERY, DEVELOPMENT, AND/OR THERAPEUTICS? We expect that our study will contribute to develop the interpretable ANN model, which can predict the time-series PKs, drug efficacies, and side effects with high prediction performance.

DOI： 10.1002/psp4.12643

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85106594489&origin=inward

Detecting System Alzheimer's Dementia by 1d CNN-LSTM in Japanese Speech 査読有り

Nishikawa K., Hirakawa R., Kawano H., Nakashi K., Nakatoh Y.

Digest of Technical Papers - IEEE International Conference on Consumer Electronics   2021-January   2021年01月

　詳細を見る

記述言語：英語   掲載種別：研究論文（国際会議プロシーディングス）

Dementia patients have increased year by year in many developed countries. We need to solve this problem. The questionnaire method is generally used for present dementia screening. However, this method requires time for 10 to 15 minutes with a doctor and a clinical psychologist, which puts a burden on hospitals and test subjects. For this reason, we propose a detecting method using deep learning by means of 1d CNN-LSTM to construct a dementia discriminant system using speech voice. Our method obtained the Accuracy of 0.908 and the F-measure of 0.897.

DOI： 10.1109/ICCE50685.2021.9427692

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85106001096&origin=inward

Speech Enhancement System using SVM for Train Announcement 査読有り

Kinoshita Y., Hirakawa R., Kawano H., Nakashi K., Nakatoh Y.

Digest of Technical Papers - IEEE International Conference on Consumer Electronics   2021-January   2021年01月

　詳細を見る

記述言語：英語   掲載種別：研究論文（国際会議プロシーディングス）

On the train, announcement speech is one of the important information to know the operation status. However, clarity of announcement speech is not enough because announcement speech is masked by noise. Therefore, new system is proposed to improve clarity of announcement speech based on sound classification. By classifying several sounds, users can select the announcement speech only they want to hear. In this paper, classifier is developed that distinguishes announcement speech using SVM. In the experiments, this classifier obtain that the accuracy of developed classifier is 96%.

DOI： 10.1109/ICCE50685.2021.9427600

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85106052569&origin=inward

Eye Fatigue Prediction System Using Blink Detection Based on Eye Image 査読有り

Kuwahara A., Hirakawa R., Kawano H., Nakashi K., Nakatoh Y.

Digest of Technical Papers - IEEE International Conference on Consumer Electronics   2021-January   2021年01月

　詳細を見る

記述言語：英語   掲載種別：研究論文（国際会議プロシーディングス）

Today, information devices such as smartphones have become indispensable in our daily life. As a result, the increased use of information devices has led to the accumulation of eye strain and the development of various eye diseases. As the accumulation of eye-fatigue is manifested in the change in the number of blinks, blink detection technology is necessary. In the previous paper, we proposed a highly accurate blink detection method called EARM. However, there is a problem with this technique, which is that the noise generated by the movement of the face can reduce the accuracy of blink detection. In this study, we used face image normalization to improve the blink detection system by reducing the noise associated with face movement. We found that the proposed system outperformed the results of blink detection accuracy using EARM in all subjects.

DOI： 10.1109/ICCE50685.2021.9427681

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85106066793&origin=inward

Sitting posture assessment method for back pain prevention system 査読有り

Joumori D., Hirakawa R., Kawano H., Nakashi K.

Advances in Intelligent Systems and Computing   1253 AISC   351 - 355   2021年01月

　詳細を見る

記述言語：英語   掲載種別：研究論文（国際会議プロシーディングス）

Back pain is widespread in Japan in recent years. The economic loss from back pain in Japan was 82.1 billion yen in 2011. Therefore, it is necessary to construct a low back pain prevention system that can be easily used. First, system estimate the body part from the image using OPENPOSE. It extract the five coordinates (neck, shoulders, and buttocks) from the output. It calculate the z coordinate of each point. Finally, It estimate the front-back tilt and the left-right tilt. We compare three distances (2 m, 3.5 m, 5 m) and three angles (20°, 0°, −20°) in experiments. Review the results. In the distance, the discrimination rates are 2 m: 83%, 3.5 m: 99%, 5 m: 66%. Next, regarding the angle, the discrimination system was good in the order of 20°, 0°, 20°.

DOI： 10.1007/978-3-030-55307-4_53

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85089628285&origin=inward

Recognition of emotional speech with cnn and visual explanations using grad-cam 査読有り

Otsuki T., Sarakon P., Kawano H., Orii H.

ICIC Express Letters, Part B: Applications   11 ( 8 )   767 - 772   2020年08月

　詳細を見る

担当区分：責任著者   記述言語：英語   掲載種別：研究論文（学術雑誌）

Nowadays there are many opportunities to communicate in situations where there is only speaking voice, such as over the phone. So we are necessary to understand the other’s emotions from only audio information in such situations. A service recognizing the other’s emotions from audio information has developed. However, its accuracy is not so high and it cannot classify many types of emotions. Therefore, we introduce the Convolutional Neural Network (CNN) architectures that can classify 9 types of emotions. In this paper, we propose and compare 3 different CNN architectures. Especially, the C-NN architecture with convolutions of time and frequency axis separately achieved the best performance. Then we applied Grad-CAM to this architecture, and we found that our CNNs predicted emotions from speech intonations.

DOI： 10.24507/icicelb.11.08.767

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85090736361&origin=inward

Shrinking model and random-dot sensor on embedded systems 査読有り

Sarakon P., Kawano H., Shimonomura K., Serikawa S.

ICIC Express Letters   14 ( 7 )   703 - 710   2020年07月

　詳細を見る

担当区分：責任著者   記述言語：英語   掲載種別：研究論文（学術雑誌）

© 2020. We address trade of between low latency and high accuracy for tactile image recognition using a random-dot sensor on embedded systems. Convolutional neural networks are developed deeper and more complicated to achieve higher accuracy, but realworld applications such as mobiles and embedded systems have a fixed resource budget. The resource-constrained platforms need to be driven in a timely fashion and keep high correctness to build an efficient convolutional neural network. We propose an efficient model based on a shrinking technique in order to manage the multi-objectives, that are low latency in embedded systems and an accurate model. The shrinking technique takes advantage of high accurate pretrained networks and reduces the number of parameters and number of operations. We not only provide faster and smaller model but also increase accuracy. We measured performance on tactile image for accuracy and Jetson TX2 for latency. The result showed that latency decreased by 28.3 percent, but high accuracy. ICIC International

DOI： 10.24507/icicel.14.07.703

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85085710494&origin=inward

A proposal of walking support system for visually impaired people using stereo camera 査読有り 国際誌

Takefuji H., Shima R., Sarakon P., Kawano H.

ICIC Express Letters, Part B: Applications   11 ( 7 )   691 - 696   2020年07月

　詳細を見る

担当区分：責任著者   記述言語：英語   掲載種別：研究論文（学術雑誌）

© 2020 ISSN. The number of visually impaired tends to increase in the future. They have many accidents when walking outdoors and the demand for walking support system development is high. They walk on a Braille block by relying on a white stick or guide dog. However, there is a problem that they collide without noticing the obstacle on Braille block. In this paper, we propose the construction of a system that detects and notifies obstacles by combining existing devices and models to prevent collision accident. The proposed device is built on Jetson TX2 and uses a stereo camera ZED Mini that provides depth image. The object detection part is real-time object detection system that is the learned model. Voice from a wearable speaker through a smartphone notifies the location of obstacle. Demonstration test confirms the usefulness of this device.

DOI： 10.24507/icicelb.11.07.691

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85085703159&origin=inward

Factor analysis using artificial intelligence in patients with chronic kidney disease 査読有り

Fukai K., Kawano H., Enomoto J.

ICIC Express Letters   14 ( 6 )   585 - 591   2020年06月

　詳細を見る

担当区分：責任著者   記述言語：英語   掲載種別：研究論文（学術雑誌）

© 2020. In recent years, the number of patients with chronic kidney disease in Japan has reached 13.3 million (1 in 8 of adults over 20 years old). The disease is said to be a new country illness. The important factors of the chronic kidney disease have been clarified by long-term clinical studies. In this paper, we propose a new method to discover important factors inherent to the chronic kidney disease patients composed by Topological Data Analysis (TDA) and Layer-wise Relevance Propagation (LRP). To verify the effectiveness of the proposed method, we compared the result obtained by our proposed method with the known important factors by the clinical studies. ICIC International

DOI： 10.24507/icicel.14.06.585

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85085711685&origin=inward

Examination of judgment items of liver disease patients by fuzzy input using LRP 査読有り

Enomoto J., Kawano H., Fukai K.

ICIC Express Letters, Part B: Applications   11 ( 6 )   601 - 606   2020年06月

　詳細を見る

担当区分：責任著者   記述言語：英語   掲載種別：研究論文（学術雑誌）

© 2020, ICIC International. All rights reserved. Nowadays, the technological development of neural networks is used in the medical field. With the progress of medicine, the average life span of people is on the rise. In order to investigate the cause of the disease, the patient’s data is obtained by blood collection and then the treatment is performed. From the data obtained by blood collection, it is judged whether it is sick or not. By comparing physical characteristics between patients with specific symptoms and healthy people, it is possible to find out the cause of symptoms and to early detect symptoms. In this paper, we use data on liver disease patients. We construct an over-learning model of binary classification, and create two membership functions for each input size. And it determines the important factors that become the basis of classification by LRP method.

DOI： 10.24507/icicelb.11.06.601

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85084968268&origin=inward

Study on mistype correction support using attention in Japanese input 査読有り

Komatsu R., Hirakawa R., Kawano H., Nakashi K., Nakatoh Y.

Digest of Technical Papers - IEEE International Conference on Consumer Electronics   2020-January   2020年01月

　詳細を見る

記述言語：英語   掲載種別：研究論文（国際会議プロシーディングス）

© 2020 IEEE. How to input characters without mistakes is one of very important points for the interface, when we consider usability of electronic devices such as personal computers and smartphones. Therefore, in this paper, we examine whether we can classify mistypes for Japanese input and inform the users how they make mistakes using 'Attention' and 'LSTM' of machine learning methods. First, we classified the category of mistypes in Japanese input into seven types (Replacement, Removal, Exchange, Insertion, Involvement, Repetition). Second, we asked subjects to input the displayed words and collected their data. Next, we classified the data and labeled it. Finally, we made the Neural Network (LSTM + Attention) to classify the mistype data and inform how they make mistakes. As a result, the correct answer rate of the Neural Network (LSTM + Attention) was about 70% and we were able to know where point the Neural Network focus on by putting in 'Attention'. We are planning to use 'LSTM' to correct mistype directly and use 'Attention' to inform how they frequently make mistakes in the future.

DOI： 10.1109/ICCE46568.2020.9043138

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85082570013&origin=inward

Electrolarynx system using voice conversion based on WaveRNN 査読有り

Urabe E., Hirakawa R., Kawano H., Nakashi K., Nakatoh Y.

Digest of Technical Papers - IEEE International Conference on Consumer Electronics   2020-January   2020年01月

　詳細を見る

記述言語：英語   掲載種別：研究論文（国際会議プロシーディングス）

© 2020 IEEE. Voice is very important in our daily life. However, some people cannot speak because of throat surgery. Electrolarynx (EL) is a way for them to speak again but this voice is flat and noisy like a buzzer. Therefore, we propose a new Electrolarynx system based on WaveRNN. When we use it, the EL voice is converted into a voice close to the natural voice. We assume that this system will be provided as a smartphone application. In the experiment, we evaluated the converted voice by the system and found that naturalness improved.

DOI： 10.1109/ICCE46568.2020.9043135

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85082578662&origin=inward

視覚障がい者の歩行支援のための路上の段差検出方法 査読有り

柳原 康佑, 竹藤 輝, Sarakon Pornthep, 河野 英昭

日本知能情報ファジィ学会 ファジィ システム シンポジウム 講演論文集 ( 日本知能情報ファジィ学会 )   36 ( 0 )   395 - 398   2020年01月

　詳細を見る

担当区分：責任著者   記述言語：日本語   掲載種別：研究論文（研究会，シンポジウム資料等）

<p> </p>

DOI： 10.14864/fss.36.0_395

CiNii Article

その他リンク： https://ci.nii.ac.jp/naid/130007957963

ニューラルネットワークによるこどものストレス状態評価推定モデルの構築 査読有り

榎本 潤季, 河野 英昭, 山川 知美, 山川 烈

バイオメディカル・ファジィ・システム学会大会講演論文集 ( バイオメディカル・ファジィ・システム学会 )   33 ( 0 )   76 - 79   2020年01月

　詳細を見る

担当区分：責任著者   記述言語：日本語   掲載種別：研究論文（研究会，シンポジウム資料等）

<p><i>The blending of scents in current aromatherapy depends on the experience and knowledge of aromatherapy specialists, such as the selection and mixing ratio of fragrance components according to the purpose of use. However, in the near future, there is no doubt that the introduction of artificial intelligence technology will establish a "rational design method for essential oils" that takes into account the aromatic components of essential oils, their actions, contraindications and precautions. As part of this, in order to understand the stress status of children who are closed due to COVID-19 and refraining from going out, we used AI technology to respond to stress from the measured values of saliva amylase and the reaction to the scent of four types of essential oils. Build a system to estimate "total difficulty (TDS)" </i></p>

DOI： 10.24466/pacbfsa.33.0_76

CiNii Article

その他リンク： https://ci.nii.ac.jp/naid/130007980052

データ分布の偏りを持つ少数データからの 血中薬物濃度推定モデルの構築

中村 秀冴, 榎本 潤季, 河野 英昭, 折居 英章, 辻 康弘

バイオメディカル・ファジィ・システム学会大会講演論文集 ( バイオメディカル・ファジィ・システム学会 )   33 ( 0 )   9 - 13   2020年01月

　詳細を見る

担当区分：責任著者   記述言語：日本語   掲載種別：研究論文（研究会，シンポジウム資料等）

<p><i>Drug administration is performed by various treatments. Drug administration is not easy, and there is a high risk of side effects and sequelae if there are errors in the plans such as dosage and administration inter val. Therefore, it is important to estimate the blood drug concentration in advance. In previous studies, the use of neural networks has made it possible to exceed the prediction accuracy of drug motion analysis proposed in the field of pharmacy. However, the blood drug concentrations recorded in the data used are few and biased. </i></p><p><i>In this study, in order to estimate the blood drug concentration using a small number of data with a biased data distribution, a method to equalize the data by removing the data in the large part of the data distribution and A method is used to equalize the data by using GAN (Generative Adversarial Network) to generate pseudo data using the data in the area where the data distribution is small and adding the generated data as training data of the original data</i><i>. </i></p>

DOI： 10.24466/pacbfsa.33.0_9

CiNii Article

その他リンク： https://ci.nii.ac.jp/naid/130007980062

クライアントサーバーシステムに基づいた視覚障がい者向けの歩行支援装置の開発

工藤 啓史郎, 竹藤 輝, Sarakon Pornthep, 河野 英昭

日本知能情報ファジィ学会 ファジィ システム シンポジウム 講演論文集 ( 日本知能情報ファジィ学会 )   36 ( 0 )   389 - 394   2020年01月

　詳細を見る

担当区分：責任著者   記述言語：英語   掲載種別：研究論文（研究会，シンポジウム資料等）

<p> </p>

DOI： 10.14864/fss.36.0_389

CiNii Article

その他リンク： https://ci.nii.ac.jp/naid/130007957964

Efficient and Small Network Using Multi-Trim Network Structure for Tactile Object Recognition on Embedded Systems 査読有り

Sarakon P., Kawano H., Shimonomura K., Serikawa S.

IEEE Access   8   144277 - 144291   2020年01月

　詳細を見る

記述言語：英語   掲載種別：研究論文（学術雑誌）

Tactile object recognition (TOR) is critical in robot perception. However, as an embedded system, a robot brain has a fixed resource budget and is unsuitable for modern convolutional neural networks (CNNs). To bridge this gap, we present a simple network-compression approach that improves the accuracy-latency trade-off of the network. The multi-trim network structure (MTNS) is a robust combination of network compression (NC) techniques providing a lightweight network with no performance drop. Furthermore, as an optical tactile sensor, we present a random-dot sensor that obtains rich information with a single touch, thus avoiding modality fusion. The random-dot sensor captures the object shapes and inputs them to TOR. In an experimental evaluation, we compare the performances of the proposed MTNS approach with those of CNN filter pruning, the network quantization technique, an adaptive mixture of low-rank factorizations, and knowledge distillation. The MTNS better resolved the accuracy-latency trade-off in tactile object recognition than the modern NC methods. By combining the random-dot sensor and MTNS approach, TOR enhances the accuracy and processing time performances.

DOI： 10.1109/ACCESS.2020.3014879

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85090266854&origin=inward

Creation of random dot type contact sensor and estimation of pressure distribution by convolutional neural network 査読有り

Takahashi T., Sarakon P., Kawano H., Shimonomura K., Orii H.

ICIC Express Letters, Part B: Applications   11 ( 9 )   869 - 874   2020年01月

　詳細を見る

担当区分：責任著者   記述言語：英語   掲載種別：研究論文（学術雑誌）

There are dangerous tasks in the factory production process. Therefore, industrial robots are used. However, it is a reality that even either an injury or death is caused by an accident. Currently, contact-type sensors for directly pressing an object on a sensor unit have been developed. This sensor can recognize the direction of pressure or force applied to the object and the hardness of the object. By using the contact sensor, it is possible to grasp everything like a person. However, it cannot know at what position of the sensor how much pressure is applied. In this study, we developed a random dot type contact sensor that can estimate the pressure distribution. The heat map of pressure distribution was created using Convolutional Neural Network (CNN) for the change image of the taken sensor part.

DOI： 10.24507/icicelb.11.09.869

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85090663211&origin=inward

5G 環境に適応した視覚障がい者向けの外出歩行支援デバイスの開発

工藤 啓史郎, 竹藤 輝, 根本 昇, 柳原 康佑, Sarakon Pornthep, 河野 英昭

バイオメディカル・ファジィ・システム学会大会講演論文集 ( バイオメディカル・ファジィ・システム学会 )   33 ( 0 )   49 - 53   2020年01月

　詳細を見る

担当区分：責任著者   記述言語：日本語   掲載種別：研究論文（研究会，シンポジウム資料等）

<p><i>We are developing a walking support device for the visually impaired. Real-time detection based on camera images using 5G communication with the client server system. At this stage, it performs object recognition, deviation detection, step detection, and warns the user with voice output.</i> <i>We will investigate the recognition accuracy of the newly added deviation detection and step detection. Compare the 5G environment with 4G and examine the practicality of the development device.</i></p>

DOI： 10.24466/pacbfsa.33.0_49

CiNii Article

その他リンク： https://ci.nii.ac.jp/naid/130007980033

Surface-Defect Segmentation using U-shaped Inverted Residuals 査読有り

Sarakon P., Kawano H., Serikawa S.

BMEiCON 2019 - 12th Biomedical Engineering International Conference   2019年11月

　詳細を見る

担当区分：責任著者   記述言語：英語   掲載種別：研究論文（国際会議プロシーディングス）

© 2019 IEEE. Surface-Defect segmentation plays an important role. It is very necessary to detect on products during production process. Though there are several previous works in surface-defect segmentation, it needs high handcraft skill. We address automatic segmentation algorithm in surface using U-shape inverted residuals to achieve end-to-end learning network. A proposed method is data acquisition, surface-defect segmentation network creation and training. First, the experimental image is augmented by image processing technique, such as rotation, flip, translation, skew and zoom in, which is randomly augmented. Second, U-shape inverted residuals segmentation network is created by changing backbone of encoder and reconstructs decoder by inverted of encoder in order to improve performance of segmentation network. In the final step, the training step of the proposed network is set. To evaluate the performance of the proposed network, each plastic hose tip and dental caries 10, 000 image are used to compare between proposed network and Unet [15]. From the experiment, Dice score and IoU are 77.11% and 62.75% in plastic hose tip, respectively. In dental caries problem, Dice score and IoU are 84.16% and 72.65%, respectively. The results show that the proposed network is satisfactory and able to be improved for higher performance. Advantages of the method are that it avoids handcraft feature extraction and is automatically learning.

DOI： 10.1109/BMEiCON47515.2019.8990212

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85080069466&origin=inward