Search

Aim To estimate the number of cephalograms needed to re-learn for different quality images, when artificial intelligence (AI) systems are introduced in a clinic. Settings and sample population A total of 2385 digital lateral cephalograms (University data [1785]; Clinic F [300]; Clinic N [300]) were used. Using data from the university and clinics F and N, and combined data from clinics F and N, 50 cephalograms were randomly selected to test the system's performance (Test-data O, F, N, FN). Materials and methods To examine the recognition ability of landmark positions of the AI system developed in Part I (Original System) for other clinical data, test data F, N and FN were applied to the original system, and success rates were calculated. Then, to determine the approximate number of cephalograms needed to re-learn for different quality images, 85 and 170 cephalograms were randomly selected from each group and used for the re-learning (F85, F170, N85, N170, FN85 and FN170) of the original system. To estimate the number of cephalograms needed for re-learning, we examined the changes in the success rate of the re-trained systems and compared them with the original system. Re-trained systems F85 and F170 were evaluated with test data F, N85 and N170 from test data N, and FN85 and FN170 from test data FN. Results For systems using F, N and FN, it was determined that 85, 170 and 85 cephalograms, respectively, were required for re-learning. Conclusions The number of cephalograms needed to re-learn for images of different quality was estimated.

OBJECTIVES To determine whether AI systems that recognize cephalometric landmarks can apply to various patient groups and to examine the patient-related factors associated with identification errors. SETTING AND SAMPLE POPULATION The present retrospective cohort study analysed digital lateral cephalograms obtained from 1785 Japanese orthodontic patients. Patients were categorized into eight subgroups according to dental age, cleft lip and/or palate, orthodontic appliance use and overjet. MATERIALS AND METHODS An AI system that automatically recognizes anatomic landmarks on lateral cephalograms was used. Thirty cephalograms in each subgroup were randomly selected and used to test the system's performance. The remaining cephalograms were used for system learning. The success rates in landmark recognition were evaluated using confidence ellipses with α = 0.99 for each landmark. The selection of test samples, learning of the system and evaluation of the system were repeated five times for each subgroup. The mean success rate and identification error were calculated. Factors associated with identification errors were examined using a multiple linear regression model. RESULTS The success rate and error varied among subgroups, ranging from 85% to 91% and 1.32 mm to 1.50 mm, respectively. Cleft lip and/or palate was found to be a factor associated with greater identification errors, whereas dental age, orthodontic appliances and overjet were not significant factors (all, P

Although cervical lymph node metastasis is an important prognostic factor for oral cancer, occult metastases remain undetected even by diagnostic imaging. We developed a learning model to predict lymph node metastasis in resected specimens of tongue cancer by classifying the level of immunohistochemical (IHC) staining for angiogenesis- and lymphangiogenesis-related proteins using a multilayer perceptron neural network (MNN).We obtained a dataset of 76 patients with squamous cell carcinoma of the tongue who had undergone primary tumor resection. All 76 specimens were IHC stained for the six types shown above (VEGF-C, VEGF-D, NRP1, NRP2, CCR7, and SEMA3E) and 456 slides were prepared. We scored the staining levels visually on all slides. We created virtual slides (4560 images) and the accuracy of the MNN model was verified by comparing it with a hue-saturation (HS) histogram, which quantifies the manually determined visual information.The accuracy of the training model with the MNN was 98.6%, and when the training image was converted to grayscale, the accuracy decreased to 52.9%. This indicates that our MNN adequately evaluates the level of staining rather than the morphological features of the IHC images. Multivariate analysis revealed that CCR7 staining level and T classification were independent factors associated with the presence of cervical lymph node metastasis in both HS histograms and MNN.These results suggest that IHC assessment using MNN may be useful for identifying lymph node metastasis in patients with tongue cancer.

In this work, an optimization method for 3D container loading problem with multiple constraints is proposed. The method consists of a genetic algorithm to generate an arrangement of cargoes and a fitness evaluation using physics simulation. The fitness function considers not only the maximization of container density or value but also a few different constraints such as stability and fragility of the cargoes during transportation. We employed a container shaking simulation to include the effect of the constraints to the fitness evaluation. We verified that the proposed method successfully provides the optimal cargo arrangement to the small-scale problem with 10 cargoes.

To determine the dominant sources for cosmic reionization, the evolution history of the global ionizing fraction, and the topology of the ionized regions, we have conducted a deep imaging survey using four narrow-band (NB) and one intermediate-band (IB) filters on the Subaru/Hyper Suprime-Cam (HSC), called Cosmic HydrOgen Reionization Unveiled with Subaru (CHORUS). The central wavelengths and full-widths-at-half-maximum of the CHORUS filters are, respectively, 386.2 nm and 5.5 nm for NB387, 526.0 nm and 7.9 nm for NB527, 717.1 nm and 11.1 nm for NB718, 946.2 nm and 33.0 nm for IB945, and 971.2 nm and 11.2 nm for NB973. This combination, including NB921 (921.5 nm and 13.5 nm) from the Subaru Strategic Program with HSC (HSC SSP), are carefully designed, as if they were playing a chorus, to observe multiple spectral features simultaneously, such as Lyman continuum, Ly$\alpha$, C~{\sc iv}, and He~{\sc ii} for $z=2$--$7$. The observing field is the same as that of the deepest footprint of the HSC SSP in the COSMOS field and its effective area is about 1.6 deg$^2$. Here, we present an overview of the CHORUS project, which includes descriptions of the filter design philosophy, observations and data reduction, multiband photometric catalogs, assessments of the imaging quality, measurements of the number counts, and example use cases of the data. All the imaging data, photometric catalogs, masked pixel images, data of limiting magnitudes and point spread functions, results of completeness simulations, and source number counts are publicly available through the HSC SSP database., Comment: 17 pages, 6 figures, PASJ open access (https://academic.oup.com/pasj/advance-article-abstract/doi/10.1093/pasj/psaa100/5959796), all the data products are available at https://hsc-release.mtk.nao.ac.jp/doc/index.php/chorus

By exploring the feasibility of medical imaging applicable to periodontal disease, we have designed a MapReduce-like deep learning model for the severity assessment by estimating the pocket depth from oral images. However, deep learning typically relies on supervised training with a large annotated dataset, and medical data often faces an insufficiency in quantity and variety. Furthermore, obtaining patient data and annotating such data by experts still remain a challenge. To overcome the insufficiency in the data, we propose random cropping and GAN-based augmentation methods on tooth pocket region images extracted from oral images. We verify that the proposed methods successfully increase the number of training data and its variety, and these synthetic data contribute to improving the estimation accuracy from 78.3% to 84.5%, and sensitivity from 50.4% to 74.0%, with specificity of around 90%, compared to the MapReduce-like model without the augmentation.

In recent years, an evolutionary change occurred in the digital world when various social networking sites became popular. This led to a big increase of users who share their activities in digital form. A huge amount of digital information has become available, providing researchers with a unique insight into the behavior and activities of entire population. The use of Geo-tagged social media data became an emerging trend to represent user actions and behaviour at specific geographical locations. However, the inaccuracies of Geo-tagged data from social media often limits the utility of this data source in micro scale analysis (at the street or place of interest level (POI)). Our study supports how social media data could be matched accurately with specific physical locations. More specifically, our investigation includes Geo-tweet data and image data from flickr to map accurately physical location using machine leaning and deep learning techniques. In this paper, a preliminary discussion of this work is provided, using Geo-tagged data from Twitter and Open StreetMap in cities such as San Francisco and London as well as for image mapping using Flickr data in cities such as San Francisco and Kyoto.

The ability to guess the posture of a body is an essential process for activity recognition. OpenPose, a deep learning-based model, detects body feature points that form the posture of a subject from moving images.OpenPose is often used for action recognition. However, there are cases where the feature points cannot be detected due to image blurring or a part of the human body being shielded, for example. Conventional interpolation methods estimate missing values, but the estimation accuracy of these methods is not high when a particular feature point is missing in a large number of consecutive frames. We propose a deep learning-based point estimation model that performs data completion of missing body feature points, which cannot be captured by OpenPose. The model is constructed based on Long Short-term Memory (LSTM) that takes consecutive frames in a video. In order to improve estimation accuracy, we revised the model by considering the following two extensions, namely, the division of body parts and the use of relative coordinates. Experimental results show that the proposed model successfully completes body feature points within an error of 8.15 pixels, which is three times better than the conventional interpolation methods and five times better than a LSTM without the proposed extensions.

In today’s big data era, there is a tremendously huge amount of data available. Layman users lack not only the knowledge and experience in data analytics to make sense of these data but also the computational resources for executing the analytics. In this paper, we propose and develop a layman analytics system (LAS), which provides the layman users with a scalable and ready-to-use analytics tool to automatically generate analytics workflows for classification tasks. The LAS is designed to benefit from existing open-source data analytics tools using generic ontological modeling of analytics operators from these tools as well as adaptive constraint refinement for metadata learning. Moreover, the LAS can be deployed on both public and private clouds to cater to the need of scalable computing and easy maintenance. To demonstrate the performance of the LAS, we conducted experiments with 114 data sets obtained from the University of California Irvine Machine Learning Repository. The workflows generated by the LAS were benchmarked against the OpenML whereby each data set has a range of classification accuracy obtained using classifiers designed and fine-tuned by data experts. The comparisons showed that 87 out of 114 data sets have exceeded the 50th percentile of the benchmark data. Among these 87 data sets, the LAS outperforms the 90th percentile of the benchmarks on 49 data sets.

In this paper, we propose a method to predict the future shot direction in a tennis match using pose information and player position. As far as we know, there is no work that deals with such a predictive task, so there is no shot direction dataset as yet. Therefore, using a YouTube tennis match video, we construct an time of impact and shot direction dataset. To reduce annotation costs, we propose a method to automatically label the shot direction. Moreover, we propose a method to predict the future shot direction using the constructed dataset. The shot direction is predicted using LSTM(long short-time memory), from sequential pose information up to the time of impact and the player position. We employ OpenPose to extract the position of skeleton joints. In the experiment, we evaluate the accuracy of shot direction prediction and verify the effectiveness of the proposed method. Since there are no studies that predict future shot direction, we set four baseline methods to evaluate the effectiveness of our proposed method.

Cloud use for High Performance Computing (HPC) and High Performance Data Analytics (HPDA) is increasing. The data are transferred to the cloud and usually left there even after the data being processed. There is security concern for such data being left online. We propose secure staging system to prepare not only data but also computing platform for processing the data dynamically just while the data is processed. The data plane of the secure staging system has dynamic reconfigurability with several lower-than-IP-layer partitioning mechanisms. The control plane consists of a scheduler and a resource provisioner working together to reconfigure the partitioning in the data plane dynamically. A field trial system is deployed for treating secure data in dental school to be processed in the computer center with the location distance of 1km. The system shows high score in the Common Vulnerability Scoring System (CVSS) evaluation.