The advantage of the interval priority weight estimation from a crisp pairwise comparison matrix has been shown over the crisp priority weight estimation. However, the performances of estimation methods for interval priority weights diminish a little when the widths of the envisaged interval priority weights decrease with decreasing the centers. Therefore, the exploration of better estimation methods is still required. In this presentation, we propose modified estimation methods based on the minimum possible ranges by introducing center estimation to the auxiliary models. The better performances of the proposed methods over the previous estimation methods for interval priority weights are demonstrated in numerical experiments of ranking alternatives in all prepared experimental settings.

Three group opinion aggregation methods developed in the interval Analytic Hierarchy Process (AHP) are improved in this paper. Individual interval priority weights are estimated from a pairwise comparison matrix given by each member of the group in the interval AHP. Three aggregation methods were proposed for obtaining group interval priority weights. One includes all individual interval priority weights of members, another one is included in all individual interval priority weights of members, and the remaining one intersects all individual interval priority weights of members. They show the range excluding commonly unacceptable values, the maximal agreeable range, and the minimal range including acceptable values for all members. Those aggregation methods are improved by introducing the non-(breakpoint)uniqueness of the solutions to the individual estimation problems of interval priority weights. It is shown that the aggregation methods are performed simply by solving linear programming problems and that the obtained group interval priority weight vectors are more efficient than the previous ones.

A pattern is set by the relative size of the input values, and the pattern is represented by sets of set function representations. A set function representation is obtained from a set of input values, a pattern is determined by whether each set of values is positive or not, and the function value of the set is the degree of the pattern. As an example, the conditions for the pattern of an egogram (TEG II) are represented by a set function representation. The rules of each pattern are expressed in terms of fuzzy measures. The Choquet integral with respect to a set functionis defined as the sum of the product of the set function and the set function obtained from the input values. This computation allows us to determine the degree of fit to the rule.

This paper proposes an educational system for understanding information theory. This system allows learners to experience the process of sending messages from the transmitter to the receiver through a noisy communication channel. The transmitter compresses the message string using Huffman coding and adds redundancy using Hamming coding. The result is saved as a text file simulating a noisy communication channel by randomly flipping 0s and 1s. The receiver opens this text file, performs error correction using Hamming decoding, and decompresses the message using Huffman decoding to restore the original string. By rewriting the Python program that implements these processes, learners can practically learn the basic concepts of information theory.

Noise Clustering based on Local Outlier Factor (NCLOF) is a robust clustering method against outliers that adjusts dissimilarity according to the value of Local Outlier Factor (LOF), which indicates the degree of data outliers. In this study, we propose an extension of NCLOF to a sequential extraction method. The sequential extraction method is a method that extracts clusters one by one, eliminating the need to specify the number of clusters in advance. Numerical experiments were conducted on several datasets, and the results were compared with those of existing methods. The results of the numerical experiments show that the proposed method is as good as or better than the existing methods.

This study proposes an automatic estimation method for the number of clusters in network clustering. Based on the x-means algorithm, the proposed method estimates the number of clusters by using a partition criterion for network data. Numerical experiments were conducted to compare the proposed method with the Louvain method. The numerical experiments suggest that the proposed method achieves the same results as the Louvain method when Modularity is used.

Fuzzy Compactness and Separation (FCS) is an extension of Fuzzy c-Means (FCM) for finding compact but separate clusters with the combined objective function of the FCM aggregation criterion and the cluster separation measure. This paper further extends FCS to linear fuzzy clustering by replaceing FCM prototypes with lines such that the clustering criterion of Fuzzy c-Lines (FCL) is combined with the degree of cluster distortion measuring the distances among each data object and the global centroid. Experimental results demonstrate that the initialization sensitivity can be improved by introducing the separation criterion in linear fuzzy clustering.

In clustering-based collaborative filtering (CF), clusters of users with similar preference patterns are extracted, and highly preferred items within these clusters are recommended. Since the data used in CF tasks contain uncertainties due to human sensitivities, rough clustering based on rough set theory, which handles these uncertainties, is considered effective. Thus, RMCM-CF, a CF method based on rough membership C-means (RMCM) clustering, a type of rough clustering, has been proposed. In this study, we examine NRMCM-CF, a CF method based on noise RMCM, which incorporates a noise rejection mechanism into RMCM. In NRMCM-CF, uncertainty is considered using rough membership values, which are the proportions of clusters in the neighborhoods of objects, and objects far from any cluster center are removed as noise to achieve robust recommendations. Additionally, we verify the recommendation performance of the proposed method through numerical experiments using real-world datasets.

Various fuzzification techniques have been applied to clustering algorithms for vector data, such as Yang-type and eQ-type fuzzification. However, only a few such techniques have been applied to fuzzy clustering algorithms for relational data. In this study, we propose two fuzzy clustering algorithms for relational data based on two fuzzification techniques: Y-type and eQ-type. Numerical experiments are conducted to evaluate the classification features and performance of the proposed algorithms.

In many variants of the fuzzy c-means (FCM) clustering algorithms, two variants are focused in this work: Yang’s FCM (YFCM) and the extended q-divergence-based FCM (EQFCM). These pro-(breakpoint)cedures of fuzzification have not been utilized in conjunction with certain fuzzy clustering algorithms that incorporate any dimension reduction methods. proposed, based on two procedures of fuzzification: Yang-type and extended q-divergence regularization, In this study, ten fuzzy clustering algorithms are along with five types of dimension reduction methods: principal component analysis (PCA), probabilistic PCA (PPCA), t-distribution-based PPCA, factor analysis (FA), and t-distribution-based FA. Numerical experiments conducted on one artificial dataset and two real datasets demonstrate that the combination of extended q-divergence regularization and t-FA outperforms the others, including conventional methods, in terms of clustering accuracy.

This paper reports the effect of constraint selection on constrained clustering using subordinate cluster generation based on Bayesian Information Criterion (BIC). Constrained clustering has the difficulty in handling pair-wise constraints located far away from each other. To address this difficulty, constrained clustering using subordinate cluster generation is proposed. This method aims to handle the difficulty using subordinate clusters that are temporally generated and finally merged into their master cluster during clustering process. However, there is a case that small clusters are generated by subordinate cluster generation, which can have a negative influence to a final clustering result. To investigate the influence by constraints, this paper reports the effect of constraint selection on a final clustering result using synthesis datasets.

This study examines methods for individual identification of Holsteins raised in freestall format for the purpose of efficient individual management of dairy cows, in response to the aging of dairy farmers in Japan and the increase in the number of dairy cows being raised. In particular, we aim to develop a machine learning model that generates individual names and bounding boxes in real time from camera images. In order to identify individual Holsteins using their patterns, we will build a model to estimate their orientation using images taken from eight different directions of the Holstein and real-(breakpoint)time video. In the experiment, object detection using YOLOv8, keypoint detection using AnimalPose, and instance segmentation will be used to develop a machine learning model to estimate orientation based on these data. The results show that even overlapping cows can be detected and identified with high accuracy.

Anticlustering is the partitioning objects into groups such that intergroup similarity and intra-(breakpoint)group heterogeneity are high. In this report, we propose several optimization problems for anticlustering, and show the results obtained from a heuristic of the proposed problems.

For fuzzy clustering for numerical data, not only Bezdek-type fuzzification but also entropy-regularization and q-divergence-regularization are adopted. Furthermore, cluster size controller was introduced. For fuzzy clustering for mixed numerical and categorical data, only Bezdek-type fuzzification is adopted based on fuzzy k-modes for categorical data, and cluster size controller was not introduced. In this report, six algorithms are proposed as variants of fuzzy clustering for mixed numerical and categorical data, where all the proposed algorithms introduce cluster size controller, three of these are based on fuzzy k-modes for categorical data, and the other three are based on fuzzy k-partitions for categorical data. Through numerical experiments using an artificial dataset, properties of proposed algorithms are observed.

Fuzzy c-means (FCM) is the most fundamental fuzzy clustering algorithm for vectorial data, and several variants of this algorithm have been proposed, such as q-divergence based FCM, Yang-type FCM, and extended q-divergence-based FCM. However, only Yang-type fuzzification technique has been applied to fuzzy c-directions (FCD) clustering for spherical data. In this regard, this study presents two fuzzy clustering algorithms for fuzzy c-directions. The first algorithm is referred to as the q-divergence based FCD, which is constructed by replacing the object-cluster dissimilarity used in q-divergence-based FCM with that in FCD for spherical data. The second algorithm is referred to as the extended q-divergence-(breakpoint)based FCD, which is also constructed by replacing the object-cluster dissimilarity used in extended q-(breakpoint)divergence-based FCM with that in FCD for spherical data. Based on numerical experiments with an artificial data set, the characteristics of each method are presented.

Traffic accidents between pedestrians and vehicles occur frequently during twilight. I have been researching this issue from the pedestrian’s perspective, focusing on their visual behavior. As a result, it has become clear that pedestrians are slower to notice vehicles during twilight than during daytime and nighttime. This is an unconscious perception for the pedestrian. With this in mind, this study analyzes pedestrians’ unconscious impressions and compares them during daytime, twilight, and nighttime.

The mainstream of technology has shifted from mechanical to information technology and is now transitioning to AI. The same applies to waste incineration plants. Waste incineration plants can be viewed as an integrated plant that combines various mechanical technologies. For example, crane technology, the incineration furnace technology, power generation technology using steam, and technology for neutralizing pollutants. This paper discusses the challenges posed, and possible application of AI to overcome the challenges.

In Waste-to-Energy plants, combustion is dependent on many complex factors and a sign of unstable states appears in other locations. We have proposed a fuzzy relational map of sensors (FuRMS) for expressing relationships between many sensors as relations between fuzzy sets of sensor values. We calculate importance degrees between fuzzy sets of sensors of the same combustion states. For data of unknown states, we calculate an evaluation value as weighted average of importance degrees of all pairs by the products of matching degrees of fuzzy sets of sensors to get a degree how close the data is to the combustion state of the map. The number of pairs of sensors with strong relations get small when the number of weak relations get large. We have proposed a method to subtract a fixed value from fuzzy counts of data in a calculation of importance degrees. It decreases the importance degrees of pairs with strong relations, but the change of evaluation values becomes like those without subtractions. In this paper, we propose a method to subtract the average entropy of relations of fuzzy sets from the importance degrees. However, a change of evaluation values with this method are not significantly difference from the method without subtractions. We also propose a method to ignore weak relations whose importance degree below a threshold value. It can slightly identify combustion states compared to the previous method.

The authors have been developing a design, training and building application with a user-(breakpoint)friendly operation interface for CNN (Convolutional Neural Network), CAE (Convolutional Autoencoder), SVM (Support Vector Machine), YOLO, FCN and so on, which can be used for the defect detection of various kinds of industrial products even without deep skills and knowledges concerning information technology. In those models, images are basically used for training data. In this presentation, intelligent anomaly diagnosis system for numerical control (NC) machine tools is considered, i.e., what structures of neural networks should be applied. Mechanical sound and vibration generated from a machine tool itself or machining sound and vibration generated from a router bit, i.e., end mill cutter is recorded and used for training data. For experimental evaluation, nine kinds of mechanical sounds (.wav) are collected from several machine tools, and then training datasets consisting of sound blocks are prepared. Each sound block is time series data extracted from wave files (.wav). For example, if a wave file is recorded with a sampling rate 44100 [Hz] and an extracted time for forming a sound block is set to 0.005 [s], then the data length of the sound block becomes 44100×0.005≒220. The extracted sound blocks from a wave file are employed for training three type of NN models. As for the NN models for comparison, conventional shallow NN, RNN and 1D CNN are designed and trained. Classification results of test sound blocks by the three models are shown. Furthermore, an autoencoder is designed and considered for anomaly detection by training it using only normal sound blocks of a machine tool.

In the dairy farming sector, health status of dairy cows significantly affects milk productivity. Early detection and appropriate response to diseases are essential for ensuring stable productivity. In this study, we investigate a method to detect diseases by extracting the activity levels of cows, which have been confirmed to be related to cow diseases, through image processing and analyzing their temporal changes using a Hidden Markov Model.

Previous reports have shown the validity of using thermal images taken by UAVs (Unmanned Aerial Vehicles), or aerial drones, for solar panel fault detection. In order to use AI (Artificial Intelligence) algorithms for automatic fault detection, it is necessary to convert the thermal video to still images. When converting the video to still images, if the images can be stitched together to create one large image, it will be easier to determine the exact location of the faulty panel. But since the glass surface of the solar panel is smooth and flat, light is reflected off the surface of the glass and the temperature reading of the thermal camera changes depending on the angle of incidence. Since standard stitching algorithms do not account for images which change color depending on angle, the created thermal image does not show temperature readings correctly for this reason. Furthermore, similar looking solar panels provide few feature points to distinguish between different images, which cause standard stitching algorithms to have a high stitching failure rate. In this research, a stitching algorithm to solve the above problem is proposed. By using the proposed algorithm, an aerial thermal video of photovoltaic power plants can be converted to a digital panoramic photograph, which can then be used for AI fault detection.

This paper investigates displacement detection methods using image processing in order to reduce the time and effort required to install displacement gauges when conducting durability and creep tests in the building field. Using a creep test of a door installed in a box-shaped frame as an example, the detection of structural displacement by image processing technology and the correction method of images using vectors for changes in the angle of view were investigated. The results suggest that displacement detection by image processing is useful in the field of architecture by applying image correction even when the angle of view changes.

In this study, we propose an index for determining the contribution of factors and the strength of interactions between factors, based on fuzzy measures and the Shapley Interaction Index. These measures are derived from the weights obtained after training an explainable neural network model, the Inclusion-Exclusion Integral Neural Network (IEINN), which utilizes fuzzy measures as non-additive measures. Additionally, we propose an automated visualization representation of this index.

In this study, we propose a similarity calculation method using λ fuzzy measure instead of cosine similarity as a method to calculate similarity between documents. Generally, cosine similarity is used to examine the similarity between documents, but it only considers the orientation of document vectors and does not take into account the length of the documents or the magnitude of the vectors. Fuzzy measures are non-additive and allow for the flexible assignment of weights to each word, including considering synergy effects in word co-occurrence. In this study, we use the λ-fuzzy measure, a specific type of fuzzy measure. The λ-fuzzy measure calculates similarity by determining the area of the common part of the two document vectors being compared. The calculation of this area employs the Choquet integral, one of the fuzzy integrals.

In this study, we propose the use of the inclusion-exclusion integral model as the explanatory model in LIME, a widely used local interpretation method for machine learning models. The inclusion-(breakpoint)exclusion integral model allows for a more comprehensive local interpretation by taking into account the interactions between explanatory variables, which traditional methods may overlook. By incorporating this model, we aim to enhance the interpretability and reliability of local explanations, providing deeper insights into the behavior and decision-making processes of complex machine learning models.

In the implementation of various questionnaires, it is conceivable that the survey items may not reflect the respondents' thoughts. This situation is regarded as "questionnaire dysfunction," and currently, efforts are being made to design a process for reviewing questionnaire items. In this presentation, as part of the questionnaire redesign, we attempt to detect non-functional questionnaire items using a quantitative approach. Specifically, by employing the binary logistic model adopted in Item Response Theory (IRT), we classify the characteristics of the survey items based on the estimated parameters and enumerate the candidates for the relevant items.

In recent years, generative AI, which is a mechanism for automatically generating content that meets human needs, has been attracting attention. There are a wide variety of types of generated content, and ChatGPT is a typical example of a generated AI that responds to requests from humans using natural language. Generative AI acquires generation methods based on automatic learning of past data. On the other hand, people have been gathering their knowledge, experience, and creativity to solve various problems on the Web. Q&A sites are a typical example. Furthermore, in the field of evolutionary computation, there is human-based evolutionary computation, in which a group of humans solves problems by performing evolutionary computation. Now that generative AI has appeared, it is necessary to identify problems for which problem-solving methods based solely on the power of human groups are effective. Therefore, in this study, we compare the problem-solving performance of ChatGPT, a generative AI, and human-based evolutionary computation based only on people’s power, for two types of problems. The two types of problems are problems for which the Web is full of information that can help solve problems, and problems for which there is little such problem on the Web. The experimental results suggest that ChatGPT is more suitable for the former type of problem and the human-based evolutionary computation is more suitable for the latter type of problem.

This study aims to apply ChatGPT for evaluation in Interactive Evolutionary Computation (IEC). This paper discusses whether ChatGPT can initially evaluate a design by comparing the free version of ChatGPT 3.5 with the paid version of ChatGPT 4.0 as a preliminary study.

In pattern classification problems, it is becoming increasingly important to have a highly transparent classifier that enables us to understand the process and basis for classification. A fuzzy classifier has high transparency and can make decisions considering the uncertainties of the real world. Evolutionary computation has been actively used in fuzzy classifier design under the name of evolutionary fuzzy systems. MAP-Elites, an algorithm inspired by evolutionary computation, can search for optimal solutions while maintaining diversity in a predefined feature space. In this paper, we study fuzzy classifier design using MAP-Elites, which searches for accurate classifiers in the feature space based on the transparency-related complexity measures. We try to obtain a set of fuzzy classifiers with high accuracy and high diversity of transparency and further investigate the relationship between accuracy and transparency.

In cases where misclassification can have serious consequences, such as healthcare and finance, classification systems must ensure high interpretability and reliability. Fuzzy classifiers can provide lin-(breakpoint)guistic explanations for their classification results, thus offering high interpretability. To improve their reliability, we have proposed fuzzy classifiers with a threshold-based reject option, which rejects classifica-(breakpoint)tions with low confidence. A traditional method optimizes the thresholds using constrained single-objective optimization. However, there is a possibility of getting trapped in local optima. In this study, we formu-(breakpoint)late the threshold optimization as a bi-objective optimization problem to minimize misclassification and rejection rates. We apply an evolutionary multi-objective optimization algorithm to this problem in order to find the optimal thresholds to improve the reliability and classification performance of fuzzy classifiers.

In this paper, we aim to quantitatively classify the vast number of swarm intelligence optimiza-(breakpoint)tion methods, and propose to express the similarity between two swarm intelligence optimization methods as the quantitative difference between the environments (fitness functions) that make the two solution search processes similar. We also propose a framework for measuring the similarity of methods based on this idea. Furthermore, we use this framework to quantitatively compare the two swarm intelligence opti-(breakpoint)mization methods and quantitatively show the differences between the methods. However, the comparison shows that the similarity is not high between similar methods with slightly different parameter values. The reason for this result is thought to be due to the method for measuring the similarity between the two solution search processes, and we show that improving the method for measuring similarity is a future challenge.

Learning Fuzzy-Classifier Systems (LFCSs), which leverage evolutionary algorithms to opti-(breakpoint)mize multiple local classification fuzzy rules, have been extensively studied as a form of eXplainable AI due to their ability to provide linguistically interpretable decision-making processes. However, because LFCSs discretely partition the input space, a single fuzzy rule may lack the necessary granularity to ac-(breakpoint)curately approximate class boundaries within a given region. To address this limitation, we propose a Neural-Assisted LFCS (NFCS), which divides the entire input space into linguistically explainable and unexplainable regions. Each region is then assigned either a fuzzy rule or a neural rule, thereby achieving both model interpretability and enhanced performance. This approach mitigates the issue inherent in LFCSs, where users are compelled to rely on fuzzy rules with low reliability (i.e., low classification accu-(breakpoint)racy) for regions that cannot be adequately explained linguistically. Experimental results on real-world classification problems demonstrate that NFCS significantly outperforms Fuzzy-UCS, a well-known LFCS, in terms of test accuracy without overfitting, which is often a concern when incorporating neural rules.

Surrogate-assisted evolutionary algorithms (SAEAs) are useful optimizers for solving expensive multiobjective optimization problems (EMOPs). Recent works have shown that decomposition-based SAEAs perform well for medium-dimensional EMOPs, i.e., 30-100 dimensional problems. However, they evaluate a solution in each subproblem, making it hard to solve high-dimensional problems, i.e., 100 or more dimensional problems, because the search efficiency is reduced. To deal with this challenge, we employ a multifactorial evolutionary algorithm into a decomposition-based SAEA for solving similar subproblems simultaneously. Experimental results show that our proposed algorithm outperforms some state-of-the-art SAEAs with up to 300 dimensional EMOPs.

Surrogate-assisted evolutionary algorithms (SAEAs) are effective approaches for solving expen-(breakpoint)sive multi-objective optimization problems (EMOPs). Many existing SAEAs have focused on improving the quality of surrogate models, but less attention has been dedicated to effectively reducing search spaces. In this paper, we propose an extension of our space reduction technique and incorporate it into a typical SAEA, K-RVEA. Specifically, the proposed algorithm uses both good and bad rules derived from rough set theory to reduce the search area and facilitate more efficient search. Experimental results show that the performance of K-RVEA is improved by incorporating search area reduction by rough sets.

The deterioration of mobility caused by qualitative deterioration of the locomotor system with aging is called a locomotive syndrome, and the risk of falling increases with the severity of the syndrome. This study aims to develop a motion classification method based on the Gaussian Process Dynamical Model (GPDM) to evaluate gait function and fall risk by measuring daily walking motions. The proposed method probabilistically evaluates the likelihood that a new movement arises from the learned dynamics in GPDM, using a generative model that relates latent variables to body postures and dynamical models that represent the behavior of latent variables.

Sit-to-stand motion is one of the most crucial activities in daily life because its difficulty can significantly impact the quality of life. Since this action involves various physical characteristics such as muscle strength of the trunk and lower limbs, joint mobility, and sense of balance, its measurement data may contain a great deal of information to evaluate the qualitative deterioration of the locomotor system that may increase the risk of falls and other problems. In this study, we analyze sit-to-stand motion data using the Gaussian Process Dynamics Model (GPDM) to identify and quantify which aspects of the body movements show differences in physical characteristics. GPDM can summarize motion data into low-dimensional latent variables and allows us to examine the effects of physical attributes on movement conditions.

In this study, we propose a method for identifying walking balance states by using plantar pressure sensors to measure changes in plantar pressure in both balanced and unbalanced walking conditions. Specifically, we collected plantar pressure data during walking and exercise, divided the data into walking cycles, and quantified the intensity of the plantar pressure. Furthermore, we applied a K-(breakpoint)Nearest Neighbor (KNN) model to classify the quantified data into three walking conditions: left foot loaded, right foot loaded, and normal walking. We then evaluated the classification accuracy of this method.

In this study, we propose a system that leverages peer learning to enhance exercise motivation by facilitating mutual instruction and feedback between users and conversational agents on exercise perfor-(breakpoint)mance. The system aims to address the challenge of maintaining motivation for continuous exercise, which is recommended for preventing lifestyle-related diseases that have become an increasing health concern. The proposed system implements a peer learning environment where the user and agent take turns acting as an exercise buddy - performing exercises and offering advice to improve each other’s form. Over 80% of participants reported enjoying exercising more with this reciprocal system compared to one with unilateral instruction from the agent. These findings suggest peer learning between users and conversational agents could effectively boost motivation for exercise.

Continuous exercise is crucial for achieving health and physique goals, yet maintaining moti-(breakpoint)vation can be difficult when physical results take time to become visible. In this paper, we propose a novel exercise support system leveraging interactive evolutionary computation (IEC) to generate self-evolving avatars personalized to users’ ideal body shapes. The system employs motion capture technology to synchronize users’ exercise movements with avatars that undergo progressive evolution towards their de-(breakpoint)sired physical ideal based on IEC feedback. An experimental evaluation of the system revealed 81.8 % of participants reported heightened motivation to continue exercising after interacting with the self-evolving avatars. These results suggest that providing an embodied visualization of potential long-term outcomes could offer a compelling incentive to maintain exercise routines.

Evaluating our gait is extremely useful for estimating our health condition and physical abilities. Traditionally, gait has been evaluated mainly by visual inspection by experts or by camera images using a motion capture system. In this study, we focus on changes in sole pressure distribution during walking and propose a gait analysis method using data measured by a sock-type wearable sensor. This method is based on fuzzy inference and is evaluated by comparing the results with standard values measured by a fixed mat-type sensor. We evaluated the usefulness of this method by applying it to healthy male subjects.

In this paper, we discuss methods for efficiently generating a large amount of labeled data to improve model accuracy in the field of machine learning used for image analysis. In the domain of image analysis, precise and detailed annotations by experts are essential, posing a significant burden on the annotators. Therefore, we develop a novel annotation tool that leverages machine learning models, exemplified by random forest, aiming to achieve both accuracy and efficiency.

A large number of high-strength bolts are used to connect multiple steel members in steel bridge construction, and workers are currently visually confirming whether each bolt is fastened correctly or not. The workers have a heavy burden to confirm thousands of high-strength bolts, and there is a risk of missing inspections or personal injury accidents. Therefore, this paper proposes an automatic method to inspect whether the high-strength bolt is installed correctly or not. In the proposed method, images of multiple high-strength bolts are raster-scanned with a search window, and the histogram of oriented gradient features is extracted. After, a bolt presence probability map is generated from the high-strength bolt detection windows obtained by the support vector machine. For each detected bolt, a machine learning classifier inspects the bolt is installed correctly, installed incorrectly, or impossible to inspect due to the incorrect marking. Experiments on images acquired at the steel bridge construction sites showed that the proposed method is able to detect high-strength bolts with high accuracy, but there were some false positives in the background area. Due to the marker line detection errors and the difference in bolt acquisition angle, the fastening inspection accuracy was not high enough for practical use. In the future, we will improve the accuracy by taking into account the color, density, and thickness of markers, which can be easily extracted by general image processing techniques.

Fetal Growth Restriction (FGR) is an obstetric and gynecological disease. It causes fetal growth delay during gestation, resulting in a smaller fetal weight and increasing the risk of prenatal and postnatal health problems. Recently, some studies have reported that fetal growth is related to placental oxygenation. BOLD MRI visualizes oxygenation noninvasively and is expected to evaluate placental oxygenation and diagnose FGR. However, manual annotation of the placental region is a heavy burden on specialists to diagnose using BOLD MRI. This study proposes a method to reduce the annotation workload by applying image registration. The proposed method reduced the annotation workload to 1 time in 2D+t images, and the Dice coefficient of the aligned images and the ground truth mask was 0.779. This result beats the deep learning-based method without manual annotation with a Dice coefficient of 0.429. The proposed method is expected to be used for the evaluation of placental oxygenation and FGR diagnosis.

Acupuncture is a type of medical treatment in which needles are inserted into a patient's body. Although this treatment has many therapeutic benefits, the procedure is dangerous to some extents. There are no clear evidence have been identified. To tackle this problem, visual information such as CT are effective. But current systems using echo images for acupuncture face challenges in accuracy and execution time due to a lack of training data and not yet implemented the latest models. It is difficult to obtain the data, we implemented a DNN (Deep Neural Network) which conducts segmentation of blood vessels and nerves that can be easily identified with a small data set. Current system runs on a CNN (Convolutional Neural Network)-based U-net model, this study implemented different image segmentation method, ViT (Vision Transformer), to compare their performance. When a DNN was constructed with the same dataset and using Dice and IoU as evaluation metrics, it showed 87% accuracy for vessel segmentation, an 8% improvement in accuracy compared to existing systems. In conclusion, ViT showed a practical level of performance for vascular recognition, but existing methods were superior for neural recognition. A deep learning model combining CNN and ViT is expected to greatly improve accuracy. Transitional learning is effective for this purpose, and we plan to incorporate it in the future.

This study investigates effective action segmentation to identify the surgeon’s technique from surgical videos for use in the automatic evaluation of surgical skill in laparoscopic surgery. We adopt the MS-TCN and its enhanced version, MS-TCN++, as action classification models and explore the optimal number of prediction layers and refinement layers for both models. Evaluation experiments after optimization show that both models have nearly equivalent classification accuracies but exhibit different prediction tendencies. Based on this observation, we propose ensemble methods that integrates the classification results of both models. We demonstrate that by using a method based on the sum of the prediction probabilities of both models, the accuracy improves.

Malignant soft-tissue tumors are rare cancers, occurring in about 40 per million people per year. Rare cancers should be diagnosed by a pathologist who specializes in that organ, but there is a chronic shortage of pathologists in Japan. In order to reduce the burden on pathologists and improve the reliability of diagnosis, there is a need for pathological diagnosis of rare cancers using machine learning techniques. Although there are studies using machine learning to predict patient prognosis from pathology images, most of them use large data sets and networks such as CNNs (Convolutional Neural Network). However, soft-tissue malignant tumors are rare cancers, making it difficult to collect large amounts of training data, and existing methods are not expected to be used. Therefore, this paper proposes a method to predict the survival period using unsupervised machine learning. In the proposed method, the autoencoder extracts pathological features from small image patches extracted from pathological images, and the K-means clustering and K-nearest neighbor regression predict the survival period from extracted features. The experiment on pathology images of 27 patients showed a mean prediction error of 17.89 months.

Total Knee Arthroplasty (TKA) is a surgery to replace a deformed joint with an artificial joint due to osteoarthritis. Although the number of surgeries is increasing every year, patient satisfaction after TKA is reported to be 75% to 89%, which is lower than Total Hip Arthroplasty(THA). In this study, we developed and evaluated an algorithm to predict the achievement or non-achievement of the Minimal Clinically Important Difference (MCID) in patient satisfaction (KSS2011) one year after TKA. Three feature selection methods and three machine learning algorithms were evaluated on 62 knees (male: 15, female: 47) that underwent TKA at Nishinomiya Kaisei Hospital and Hyogo Medical University Hospital. As a result, logistic regression showed a maximum Area Under the Curve(AUC) of 0.90. Furthermore, we investigated the importance of features with SHAP. It indicated that preoperative satisfaction and expectancy and femoro-tibial angle (FTA) were important features to predict patient satisfaction.

Recently, the application of AI technology in the clinical field has advanced significantly. For instance, some diagnostic algorithms in ophthalmology can diagnose cases as accurately as specialists. However, constructing such algorithms requires plenty of high-quality data. In this study, the authors proposed a data augmentation method for retinal OCT images and discussed the effect of added generated images on classification performance. The experimental results indicated that classification accuracy was maintained when the generated images were appropriately proportioned. We confirmed that the proposed method is useful as a data augmentation method for the disease classification of OCT images. Our approach is expected to enable the construction of highly accurate disease classification algorithms with limited datasets.

Ureteral stone disease refers to mineral mass in the ureter, which extends from the kidney to the bladder. One of the least invasive and widely used treatments of ureteral stone is extracorporeal shock wave lithotripsy (ESWL). However, ESWL has a success rate of approximately 70%. Therefore, it is important to predict the treatment outcomes based on preoperative images and clinical findings. Although prior studies have identified several features effective in predicting ESWL outcomes (success or failure), these studies have not sufficiently addressed the impact of data bias and class imbalance on prediction accuracy. In this study, we perform stratified sampling on the features that were validated in previous studies to predict ESWL and assess its impact on prediction accuracy. We utilized CT/X-ray images and clinical findings from 162 patients who chose ESWL as their initial treatment. We predicted treatment outcomes by machine learning and compared the accuracy of predictions made with and without stratified sampling. Stratified sampling based on stone size achieved the highest AUC and accuracy of 0.850 and 0.897, respectively. These findings confirm that incorporating stratified sampling with effective features enhances the accuracy of ESWL outcome prediction models.

In infertility treatment, vacuolated structures called sERCs are sometimes found on oocytes obtained by egg retrieval. It is thought that oocytes with sERCs have a low fertilisation rate, but their specific characteristics are not yet known. In this study, DeepLab, a segmentation method using deep learning, is used to automatically extract sERC regions from videos during ICSI. The relationship between the size, shape and orientation of the extracted sERC regions and pregnancy is analysed from an information science approach.