AI, VR/MR
In today's world where customers' needs are constantly changing, we are constantly challenged to enter new areas. We at INTACT are committed to creating innovative solutions and providing optimal services to our customers through such challenges.
The use of cutting-edge artificial intelligence technologies, such as machine learning and generative AI, is also part of our approach. By utilizing these technologies, we are able to address complex issues and come up with effective and efficient solutions. We are sensitive to the evolution of technology and actively introduce new methods and tools to stay ahead of the curve.
Contents
Initiatives with Yamaguchi University AISMEC
Yamaguchi University established the AI Systems Medicine Research and Training Center (AISMEC / AI Systems Medicine Research and Training Center) in 2018, which promotes both AI and systems biology at the Graduate School of Medicine and the Hospital attached to the School of Medicine. In collaboration with Yamaguchi University, INTACT is building systems and services using various advanced technologies such as AI and VR/MR.
Toward the Introduction of Medical AI into Clinical Practice
In particular, in the area of AI, we are developing a medical information/AI integration system for the introduction of medical AI into clinical practice. This will serve as a framework for linking medical AI to medical information systems, including electronic medical record systems and clinical decision support systems (CDSS / Clinical Decision Support System).
For example, in the verification on the subject of “estimation of drugs causing adverse reactions,” a system was established to predict the drugs causing adverse reactions based on prescriptions and injections in the past six months for patients with “suspected adverse reactions,” and to notify them on the medical information system.
In the future, we will continue to integrate other types of medical AI developed at Yamaguchi University Graduate School of Medicine and Yamaguchi University Hospital into the collaborative system, and based on the knowledge gained in the process, we aim to further generalize the procedures for integrating new medical AI into the collaborative system and the common framework to be used in this process. We are also developing other types of medical AI to be integrated into the collaborative system.
Utilization of hand recognition in video
We are developing software that uses Media Pipe Hands, which detects hand and finger joints in videos of hand movements based on a machine learning model, to record or play back the estimated coordinates of each hand and finger joint in 3D space as a time series, with the aim of applying it to estimating the type of disease caused by some kind of nerve damage running through the arm based on abnormal hand movements. We are developing software that records or plays back the estimated coordinates of each hand joint as a time series.
Recording of hand joint coordinates
You can set the number of skipped video frames and process multiple videos at once.
Loading and displaying coordinate data
Time-series data of hand coordinates can be read, and a hand stick model based on the data can be superimposed on the hand animation and displayed. This is useful for intuitive confirmation of data after processing, such as removing noise from coordinate data, and for creating presentation materials.
Use of VR (virtual reality)/MR (mixed reality)
We are developing a system that allows multiple physicians to share 3D models of the human body and organs using VR/MR technology, with the aim of allowing physicians to review surgical procedures prior to surgery, etc. 3D models can be displayed/operated in virtual space using Meta Quest 2 or HoloLens 2, and at the same time, people without the devices can observe them on a PC monitor or iPad. People can also observe them on a PC monitor or iPad.
Session Creation
Multiple sessions can be created in parallel. Participants log in by designating a session and share 3D models according to the session.
Manipulation of 3D models
Participants to a session can select a 3D model to be shared in the session and perform equilibrium movement, rotation, scaling, and other operations on the selected 3D model. Participants can also share more detailed locations by pointing to specific locations on the 3D model or placing markers at specific locations.
Device-specific display and control
When participants in a session use Meta Quest 2, they can manipulate the 3D model according to the Meta Quest 2 controller, and when participants use HoloLens 2, they can manipulate the 3D model through hand recognition. Even if participants do not have VR devices, they can share the situation of the session by displaying the session on a PC monitor or tablet.
Switching perspectives
Participants in a session can switch between “individual viewpoint mode,” in which each participant has his or her own viewpoint, and “viewpoint sharing mode,” in which all participants share a particular participant’s viewpoint.
New AI OCR
We developed a new AI OCR based on the paper ” TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models ” published in 2021, which was developed from a customer’s inquiry about AI OCR for handwriting recognition. We have built a new AI OCR that employs We leverage advanced implementations based on this paper and incorporate our clients’ own training data to further enhance the performance of the model.
New AI OCR Features
1. adoption of TrOCR technology
Our AI OCR deploys TrOCR technology based on the latest paper published in 2021; the Transformer-based model has demonstrated excellent performance in natural language processing and provides high accuracy in handwriting recognition.
2. use of proprietary training data
By incorporating training data from customer-provided handwriting, the model can be optimized for specific industries and applications to achieve high generalization performance. This enables personalized AI OCR tailored to your needs.
3. flexible customizability
Our AI OCR is designed to be flexible and scalable, allowing you to customize the model and functionality according to your requirements. It is also available on-premise, whereas typical AI OCR is a cloud service. This enables us to provide tailor-made OCR solutions for various industries and business environments.
The current base model was created using a total of 2.5 million training data using multiple fonts for the following data
- Japanese First name + Last name 54,000
- Westerners First name + Last name 30,000
- 230,000 Chinese texts
- Approx. 6,000 Chinese regular Kanji characters
- About 2,600 Japanese regular-use Kanji characters announced by the Agency for Cultural Affairs
- 1.35 million pages of the 01/01/2023 version of Japanese Wikipedia