Select Field
Select Ministry
14 results
A million-pixel image might only have one annotation from a professional pathologist, posing a significant challenge for all AI models. Our team has designed the world＇s first weakly supervised multiple instance AI learning framework. This framework can analyze million-pixel pathology slides to predict gene expression and prognosis in colorectal cancer patients. Additionally, it has been successfully validated across patient datasets from multiple generational cohorts. A more significant challenge arises as each complete million-pixel digital pathology slide image often has only one set of annotated answers from a physician, which cannot be applied to all the smaller segmented images. This issue is exactly what top AI experts worldwide, NVIDIA, and our team aim to solve. Our research team has developed innovative AI algorithms to address this problem. We designed a novel attention-based multiple instance learning model embedded in a convolutional neural network, overcoming the limitations of traditional machine learning methods while effectively highlighting lesion features. This pioneering weakly supervised AI model can automatically analyze and accurately integrate the best prediction after processing a large volume of image data. Furthermore, our unique AI model enables computers to recognize gene features that pathologists cannot detect with the naked eyes. If successfully implemented in clinical practice, this achievement could reduce the workload of physicians＇ manual annotations by at least 75%. Remarkably, we developed the new deep learning model with an accuracy of up to 99.38%. This novel AI algorithm framework not only identifies specific gene expressions from million-pixel digital pathology images of colorectal cancer but also performs prognostic prediction tasks for patients. Additionally, the model＇s performance has been validated across different cohort study.
Future Tech | Information ＆ CommunicationsThis developed technology integrated big data analytics, meta model, and XAI algorithm to develop more accurate demand and cycle time prediction for supply chain and production planner to reduce production costs in semiconductor industry and enhance network resilience via analyzing big data of supply chain, capacity portfolio and production data enabling automated feature engineering. The developed solution (Advanced Artificial-intelligence based forecasting technology) integrates big data analytics, metamodels by incorporating interpretable AI algorithms, advanced cycle time prediction models for future new process platform introductions and new capacity portfolios. The fundamental goals are as follows: (1) Achieve flexible decision-making, enabling models to self-adjust and learn in different planning scenarios. (2) Achieve system resilience in production environments facing discrete events through action selection, monitoring of fluctuations, and adjustment. (3) Achieve real-time planning and control of dynamic systems. The technology includes (1) automated feature engineering techniques, (2) capacity configuration classifier, (3) integration of metamodels with XAI to interpret advanced prediction models, and (4) hybrid strategies to adapt linear models applicable to planning scenarios. By collecting supply chain information, capacity portfolio and production data, the technology automatically identifies key factors in semiconductor supply chain and manufacturing processes and provide more accurate forecasted demand and manufacturing cycle time (Enhance 30%, 23% accuracy), and thus enhances semiconductor supply chain resilience, while reducing manufacturing costs (30% inventory costs and 20% production costs).
Future Tech | Information ＆ CommunicationsWe have developed innovative AI models, particularly for Extreme Gradient Boosting models, to integrate genetic information derived from whole-genome SNPs, medical imaging information from four types of medical images, and demographic variables (including age, sex, and family history of disease) obtained from the Taiwan Biobank. This integration allows for the identification of high-risk subgroups and early detection of T2D. Additionally, a T2D risk assessment system has been established.
Future Tech | Information ＆ CommunicationsWe developed two methods to screen for white-coat hypertension and masked hypertension. First method uses machine learning classifiers such as Random Forest. Features were ranked with SHAP values, and supervised feature selection was performed. Second method utilizes dimensionality reduction and deep learning with PCA and Tab-PFN.
Future Tech | Information ＆ CommunicationsThis technology utilizes Retrieval-augmented Generation (RAG), Agentic Reasoning, and Knowledge Graphs to implement a generative AI computation service (AI as a Service, AIaaS) that facilitates the code compliance and generation of construction inspection forms based on construction specifications and drawings. It offers an innovative solution to longstanding challenges faced by engineering consulting firms and significantly enhances productivity.
Future Tech | Information ＆ CommunicationsWe develop a low-temperature plasma-assisted selenization technique to grow different 2D materials for high-sensitivity gas sensing applications. It overcomes the current requirement for growth temperatures above 700°C, meeting semiconductor standard processes. The advantages include large-scale production, and tunability of thickness and phase. 2D materials exhibit enhanced corrosion resistance and material stability against corrosive and toxic gases like H2S and NH3.
Future Tech | Information ＆ CommunicationsOur team developed a TCN model using publicly available ICU datasets and externally validated it by data from NTUH. The results show that TCN can predict over 90% of sudden cardiac arrest cases six hours before the cardiac arrest events, with an AUC of 0.96, outperforming NEWS, which has an AUC of 0.87 and a sensitivity of 81% six hours before cardiac arrest occurs. Furthermore, we utilized generative AI to predict vitals of the next hour based on patient past vital signs and medical records.
Future Tech | Information ＆ CommunicationsThis technology introduces an integrated smart TENS within a robotic platform for material identification via contact electrification. TENS uses a self-powered tactile sensing mechanism to accurately identify material composition and amino acids. It features micro-pyramidal structures made from Ecoflex material, with NaCl solution as the electrolyte conductor, enhancing stretchability and sensitivity. Integrated into robotic fingertips, TENS generates triboelectric output signals during contact and separation, which are used for material identification and wirelessly transmitted to a computer. Elevated temperatures were found to affect electronic structure changes, altering triboelectric output signals. In stability tests, TENS showed excellent response behavior under various stretching conditions, maintaining stable performance even when stretched to 150% of its original size. After 5000 bending cycles, TENS continued to produce stable output signals. Market analysis indicates a growing demand for multifunctional, efficient sensors driven by the rapid expansion of smart robots and IoT applications. It is estimated that over 30 billion sensors will be needed in the IoT sector alone. Applications in bionic prosthetics, virtual societies, and big data processing are increasing, driving the need for sensors that identify materials, temperature, pressure, shape, size, and humidity. Traditional sensors rely on batteries, posing issues like high energy consumption, limited lifespan, bulkiness, and environmental threats. TENS can harvest environmental energy, powering its multifunctional sensing capabilities. This technology shows great potential in self-powered sensing, applicable in smart manufacturing, environmental monitoring, healthcare, and human-machine interfaces. According to market research, the global smart sensor market is projected to reach $65 billion by 2024, with self-powered sensors being significant.
Future Tech | Information ＆ CommunicationsWe leverage extensive mobile telecom data, integrated with sparse vision data, to provide comprehensive traffic insights. First, we utilize telecom data from widespread road sections as a novel traffic indicator while ensuring user privacy. Next, for the first time, we enhance traffic prediction accuracy by fusing telecom data with camera-based vision data. To further optimize performance, we propose a multi-modal framework that balances the influence of different data modalities, enabling precise cross-modal predictions. This research progress has been recognized by top conferences such as AAAI, WWW, and CIKM.
Future Tech | Information ＆ CommunicationsStroke causes healthcare burden. Neuroimaging-based AI integration for in-depth stroke analysis helps stroke care. The Semantic Segmentation Guided Detector Network (SGD-Net) applies 2-stage models to accurately segment and quantify ischemic lesions on brain MRI images. Lesion-based semantic and radiomic analysis and feature fusion precisely predict outcomes, stroke etiology, and large artery stenosis. SGD-Net integrates neurology and neuroimaging, making it highly valuable and applicable.
Future Tech | Information ＆ CommunicationsThis project integrates cognitive-behavioral models with innovative technology to create a virtual reality platform for clinical training on essential workplace social skills in autistic adults. To accommodate the low tolerance of low-resolution images in autistic adults, we use high-resolution specifications and advanced displays to ensure feasibility. For various scenarios, we apply speech analysis, content analysis, coherence analysis, and biofeedback to assess the effectiveness of training.
Future Tech | Information ＆ CommunicationsThe ToFEye includes ToF sensors, ToF imaging dataset (ToF Dataset), AI edge computing platform, ToFPose AI software model (ToFPose) and ToF abnormal posture recognition algorithm. ToFEye uses ToF sensors to capture spatial human images in Bathroom and Toilet, and applies AI and DL technologies to identify the abnormal posture action of ToF images. It can be applied to any space in the nursing place and achieve de-identification real-time abnormal posture notification purpose.
Future Tech | Information ＆ CommunicationsThis research develops innovative hyperspectral satellite imagery sensing, fusion, and secure transmission technologies. By combining the advantages of multispectral and hyperspectral images, it enhances both spatial and spectral resolution. It also integrates deep learning and deepfake image recognition mechanisms with the technology published in top-tier journals like IEEE TGRS. This technology not only improves image transmission efficiency but also ensures image authenticity.
Future Tech | Information ＆ CommunicationsThis work is an anti-theft and warning system that combines artificial intelligence, object detection, computer vision and virtual fence algorithms. This technology allows users to select and draw virtual fence protection areas according to site requirements and create three warning lines to detect objects such as vehicles and underwater divers approaching the warning area. Based on the degree of intrusion, corresponding warning signals will be immediately sent to the management personnel.
Future Tech | Information ＆ CommunicationsComing soon!
