The lungs are essential to human health and respiration, yet lung diseases cause significant morbidity and mortality worldwide. Accurate segmentation of pulmonary airways, including bronchi, is crucial for diagnosis―especially underscored by the COVID-19 pandemic. Automated segmentation from CT scans is challenging due to the complex, tree-like airway structure and fine details at terminal bronchioles. To address this, we propose W-Attention Net, a dual-encoder model that integrates CNNs and Swin Transformers via cross-attention, capturing both local and global features for fine-grained airway segmentation. Our method outperforms existing approaches on a private dataset from Shandong University and the public LIDC-IDRI dataset. In addition, we developed a mixed-reality system with Microsoft HoloLens 2 to enhance clinical and educational visualization of airways.

In this paper, we propose a GAN-based novel view synthesis (NVS) method that focuses on generating naturally varying viewpoint-dependent material appearances during viewpoint transitions. While NVS has gained increasing attention for its ability to generate images from arbitrary viewpoints, reproducing viewpoint-dependent material appearances, particularly for specular and transparent objects, remains challenging. The proposed method introduces a 3D convolutional discriminator that evaluates the naturalness of generated image sequences across consecutive viewpoints to reproduce material appearance caused by intensity changes during viewpoint transitions. Experimental results using CG datasets with specular and transparent objects demonstrate that the proposed method effectively generates view-consistent images while preserving fine surface details and maintaining consistency across viewpoint transitions.

In this study, we propose a method of enhancing the user experience through collaboration with a digital twin partner robot in order to connect between VR experience and real experience. In recent years, as the performance of head-mounted displays has been improved and their prices have become lower, virtual reality (VR) technology has been used not only in the entertainment field, but also in a variety of other fields such as education and medical care. Against this backdrop, many studies have been conducted to explore the potential applications of VR technology. However, conventional VR technologies only provide a one-time and personal experience, which makes it difficult to reinforce the experience by reflecting on and sharing with family or friends, which is often done in real life. By using our system, an avatar agent with the same appearance and voice as a real-life personal robot goes to a VR space and experiences VR contents together with a user. After returning to reality, the user recalls the VR experience together with a real robot that has the same personality as the avatar agent in the VR space. We aim to enhance the VR experience by working with the digital twin partner robot in both real world and virtual world to connect their experience.

In order to apply the arithmetic coding for multi-level Markov-model sources, most conventional methods were using binary arithmetic codes by decomposing multi-level Markov-model sources into plural binary sources, which we here call B-B coding. The advantages of B-B coding are the possibility of utilizing the study results of binary arithmetic codes but it may need longer time by the increase of coding/decoding the expanded binary sources. The usage of multi-alphabet arithmetic code to multi-level Markov-model sources, which will be called as M-M coding here, and the comparison with B-B coding will be very interesting but it has been impossible since there are almost no report concerning the practical multi-alphabet arithmetic coding. In this paper, we will try to propose the basic design and the practical guideline of multiplication-free multi-alphabet arithmetic code (MFMAC) which will make the comparison of the total performance of B-B coding and M-M coding, possible. The basic design of MFMAC is composed of model source assumption, symbol area determination rule, designing of model parameter set for static coding, model parameter estimation for dynamic coding, and symbol ranking detection and update. We will also evaluate its performance of MFMAC, based on a design example. As there are many types of multi-level Markov-model sources, the general comparison of B-B coding and M-M coding will be left for individual cases, in which specific coding specification will be determined based on this designing guideline.

One recent approach to unsupervised anomaly detection is training an Autoencoder and using the reconstruction error as an index. In this work, we adopt Sketch-RNN, a sequential Variational Autoencoder (VAE) model, to classify the online hand drawing pattern of healthy persons or Parkinson’s disease patients. We train the Sketch-RNN with no labeled data and evaluate the accuracy of the attribute recognition using the reconstruction error. The proposed method, online drawing data and Sketch-RNN model, outperforms the conventional method, recognition with still image data and CNN model, showing the pipeline for online data classification using VAE.

We introduce an interactive system for stylization of a photograph to match the level of abstraction of a given foreground illustration. The stylized photograph can then be used as the background. We employ an existing automatic stylization method but it often has parameters to control the degree of abstraction. Our system allows the user to interactively find the optimal parameters that produce the desired appearance. We conduct a simple user-study to investigate the usefulness of our system.