In order to develop an AI model that can accuratelycompletely segment coronary arteries, our team, the TW-CVAI, has established a training dataset composed of strictly verified annotations of coronary lumen boundaries in coronary CT angiography (CCTA). We designed a deep learning model, two-channel 3D-UNet, with a priori prerequisite (vesselness prior) to facilitate identification of vascular structures. The final model, the TaiCAD-Net, greatly shortens the CCTA interpretation time from 6 hours to 10 minutes, with the overall segmentation accuracy of 86 by Dice similarity coefficient.

The HeaortaNet is developed by the TW-CVAI team. The HeaortaNet is a deep learning model trained by 70,000 axial images from 200 patients with verified annotations of the pericardiumaorta. It shortens the time for data processing from 60 minutes, by manual segmentation of both pericardiumaorta, to 0.4 seconds. The segmentation accuracy, as assessed by dice similarity coefficient, is 94.8 for the pericardium,91.6 for the aorta. The imaging-based Cardiovascular Risk Prediction module was constructed by analyzing data from the National Health Insurance Databank.

As we continue to face the rapid increase in confirmed Coronavirus cases around the world, we created an AI-based pneumonia detection platform for COVID-19. The system is able to automatically detect high-risk patients with pneumonia that will then send information to doctors. With that information, the doctors are then able to make follow-up decisions and provide a treatment plan after the diagnosis. In specific, doctors from the Department of Medical Imaging provided us thousands of positive and negative chest x-rays for pneumonia as a training set. Our system has already been tested with and adopted by doctors at the NCKU Hospital. The system achieved 95% accuracy to detect the pneumonia symptom, based on 1400 test images.

With the rise in computing power, deep-learning based computer-aided diagnosis systems have gained interest in the research community. Our system process the images to assist doctors to determine whether the patients have nodules in lungs. Meanwhile, we utilized the Feature Pyramid Network to extend the receptive field on the convolutional kernel, which improved the performance on the nodule detection with various locations in CXR. The semi-supervised learning mechanism also achieves the way of soft-annotation to reduce human effort in medical image annotation.

The AR-assisted endoscopic navigation system for brain surgery combines CT/MRI images, 3D cerebrovascular/nerve models,endoscopic images for surgical planningnavigation. AR glasses can display the 3D navigation inside the patient＇s skull, providing surgeons with intuitive 3D surgical navigation. In addition, the cerebrovascular/nerve model is also superimposed on the endoscopic image, allowing the surgeon to predict the upcoming surgical situation. The AR display endoscopic images can be simultaneously transmitted to the remote site for assistance through 5G communication.

The proposed TOA-STFM is a spatiotemporal image fusion technology that can preserve top-of-atmosphere (TOA) reflectance. By fusing high spatial (Landsat-8SPOT-6)high temporal resolution (Himawari-8) images, the fused 10-minute-6~30-meter-resolution images can solve the problem of existing air quality monitoring techniqueseffectively capture the dynamic changes of air quality in near-real-time.

1. CNN model architecture optimization techniques: Propose an Agilev3-L architecture, with a performance index of mAP@50:95FPS:FP16 up to 171.79 2. GOP-mode acceleration scheme for real time inference: employ GOP-modeTRACKING algorithms to enhance the processing rate. 3. Propose a rapid data labeling system HiTag.

This study built an automatic insect pest image identification system based on tiny Yolov3 deep learning model. By optimizing the tiny Yolov3 detection model, images of insect pests on scanned sticky paper can be automatically identified. The system achieves a testing accuracy of 0.93, 0.90 for whitefliesthrips respectively.

Our team construct an AIoT smart aquaculture management system. The management system mainly consists of: (1) Image Behavior MonitoringAnalysis Subsystem (2) Smart Feeding Subsystem (3) IOT Subsystem including underwater sensors, ROV,Drone (4) Cloud Subsystem (5) Big Data Analysis Subsystem.

Searchlight is composed to high brightness redgreen LEDs. The message of Morse code is transmitted through LED colorlengths of ”on/off” time in human-machine interface. CCD image sensor integrated with LEDs receives light signal of LED from another ship,then identify the real message through the image recognition software.

We invented a non-destructive terahertz (THz) deep-learning computed tomography system based on time-domain spectroscopy. In the method, THz time-domain signals are profiled. Multiple features are retrieved from those profiles by a trained modeltransformed to the spatial domain to reconstruct a cross-sectional tomographic image. We have also invented a 3D THz tomographic system based on multi-scale spatio-spectral feature fusion in a multi-scale manner. We believe our work will stimulate further applicable research of THz tomographic imaging with advanced computer vision techniques.

Laser material recognition technologyintelligent color compensation lighting system, with AUV to realize the function of underwater rust detectionimage color restoration, real-time acquisition of rust information through spectral analysisdisplay of true rust color on the image, so that the algorithm can be accurate Determine the health of the cylinder. The vehicle will establish the health information of the wind turbine to help the inspector make the best maintenance strategymaintain the power generation efficiency of the wind turbine in the best condition.

"The IHC staining intensity grading algorithm combines a series of complementary processes of deep learningimage processing in order to overcome the difficulties in cell boundary segmentationgrade definition,provides a cell-based IHC staining intensity gradingquantification capability. The proposed method serves as a useful assistive tool for physicians in performing accurate staining quantification in tissue microscopic images. In addition, it analyzes panoramic pathological images in order to determine the factors most strongly associated with likely disease recurrence."

This theme is designed to implement the robot＇s appearancepractical functions. Apply PSPNet to detect the walkable planeYolo to detect obstacles, so that the robot has the autonomous obstacle avoidance function, informing more information about the environmental obstacles around the visually impaired,apply CNN to locate indoor position with self-built indoor database.

This project is used for advanced driver assistance systems (ADAS)automatic driving. At present, our project will use a more sensitive CMOS single photon detector array to reduce the number of photons required for detection to less than ten, in order to break through the cost barriers of applications of LiDARestablish low-cost, high-performance LiDAR modules.

We propose a semantic-guided framework (FashionOn+) that generates image-based virtual try-on results with arbitrary poses. FashionOn+ contains three stages: (I) conducts the semantic segmentation to have the prior knowledge of body parts for rendering the corresponding texture in stage (II). (III) refines two salient regions, i.e., faceclothes, to generate high-quality results. With the novel architecture, we win first place in the Multi-pose Virtual Try-on Challenge in CVPR, 2020. Further, we tackle the low-resolution limitation (256x192)achieve high-resolution results (640x480).

High-accuracy 3D sensingAI image processing system for constructing high-quality immersion 3D image for AR/VR. Chaotic Lidar with APD arrayTOF sensors supports millimeter-accuracyinterference-avoiding capability. High-performance CNN processor supports high-performancelow DRAM bandwidth computations for various image AI applications.

We developed a wearable ultrasound monitoring device to monitor the collapse of tongue roots of sleep respiratory patients throughout the night,completed clinical trials in the sleep center. The experimental results show that this device can effectively monitor OSA. The tongue structure changes at the timeintegrates ultrasound information into the commercial PSG system. It is expected

Two indoor positioning techniques are presented. The first one is a monocular vision based landmark matching scheme for identifying absolute indoor locations. The scheme requires just one photo shotmatches it with a landmark database to obtain the location. The landmark data base can be easily adapted to different fields. The proposed scheme is highly computing efficient. The correct landmark identification rate is up to 90the positioning accuracy is 1.5m. The second one is a relative positioning scheme based on ultra-wide band (UWB) technology. It can be employed on an automatic guided vehicle (AGV) to implement the trailing function. The accuracy of positioning is 80cm.

The chromium-doped zinc gallate, ZnGa2O4:Cr3+, material is viewed as a long-lasting luminescent phosphor which can avoid tissue autofluorescence interference for in vivo imaging detection. The cubic ZGC revealed a specific accumulation in liver0.5 Gy used at the end of X-ray excitation was sufficient for imaging of deep seated hepatic tumors.

Optical coherence tomography using an optical probe with a diameter of 0.9 mm, combined with the 14-18 gauge needle, is used for clinical needle puncture. Using the real-time image obtained from the tip of the needle in the tissue, the needle position can be identified. Combined with artificial intelligence can achieve objective, accurate and automatic identification of the tissue, which has been successfully verified in anesthesia and laparoscopic surgery.

"a.The Digital ArchivingConservation System of Cultural Relics combines various imaging techniques like catoptric imaging, 3D renderingmulti-spectral imaging, which provides multiple scientific images data for the inspection of cultural relics. b.Environment Detection System allows controlling the environmental device from a far distance."

This technology integrates 1000+ times of UAV imaging experiences with labeled rice lodging images for training. A rice lodging recognition model using deep learning reaches 90 accuracy. The recognition model can be deployed in a microcomputer mounted on UAVs to implement edge computing. While taking aerial images, the inference can be completedreveal lodging areadamage level in-time.

This technology integrates acoustic-optical sensing calculationrecognition information,uses hydrophones to estimate the state of dynamic targetsvisually recognizelocate the dynamic target for vehicle motion control. The fusion acoustic-optical sensing architecture can be used according to the target statecharacteristics. By calculatingadapting to the environmenttarget state characteristics, the exploration tasks will be performed.

This product is a novel risk assessment tool for carotid artery stenosis and stroke. This is a revolutionary healthcare technology using motion analysis and quantification to extract information from pulses for risk assessments. The entire process is completed by taking a short video clip aimed at the neck with only one simple click on any mobile devices or our apparatus, anywhere, anytime. In less than five minutes, the user receives an evaluation report indicating low to high stroke risk. Our product accuracy stands higher than 90% when compared to the clinical outcome. The future indications of this product can be extended to arrhythmia, venous fistula obstruction, etc. This product has the great potential to achieve our dream of “personalized mobile hospital” in the future world.

We developed a Multiview 3D capturing system based on camera array for generating stereoscopic 3D printing. Based on the framework, we are capable to create well aligned 3D contentcalibrated color images. The depth perception of content can be manually adjusted for preview,for generating quality 3D stereoscopic printing.

Our devices automate the microscopic examinations, including samplinginspecting of micro-objectsmotorized stacking micrography. Sampling rate is expected to increase more than 30 with an improved product quality by a standardize procedure. Motorized stacking micrography extends back to front sharpness of a photomicrographgreatly reduce the operation timeerrors more than 50.

Nanopharmaterial is a term by combining pharmaceuticalmaterial via nanotechnology. Here we demonstrate this concept by developing a novel nanopharmaterial, so called 177Lu-Gold nanostar (AuNS). 177Lu is a therapeutic radionuclide emitting moderate-energy beta particles as well as gamma rays for SPECT/CT-based imaging diagnosis. Nanostar can target tumors by enhanced permeability retention (EPR) effectsowns the photothermal therapeutic potent. The core technology is to integrate radionuclidenanomaterial to perform a new radio-nanopharmaterial for tumor target theranostic purpose.

The principle of the image-based AFib discriminating system is based on the consistency that irregular cardiac cycles can both be detected on the waveforms of electrocardiographyrPPG. The quantity of the blood varies from time to time is captured by a general camera. Signals from RGB channels are then synthesized through the core algorithm to eliminate noisegenerate stable rPPG signals. The processed signals are then classified as AFibnot by a sample-size model. According to the IRB evaluation in En Chu Kong Hospital, the AFib can be well detected with an accuracy of 97.1.

The developed Unmanned autonomous underwater vehicles (AUV) included 3-D modeling technology, wireless power transfer between sub-system in AUV, GPS underwater locate system, low complexity frequency domain on ocean floor analyze with AI network, underwater objects detection network, image dehaze network on groundunder water, underwater color correction network, underwater objects classification network, underwater optical characteristic algorithmLED color compensation system.