Daegu Gyeongbuk Institute of Science and Technology
DGIST Library Institutional RepositoryNot a member yet
12664 research outputs found
Sort by
A Study on the Effects of Power Loading Profile in Discrete Multitone Wireline Serial-Data Transceiver with Fixed-Point DSP-SerDes Simulator
No full textThis paper presents the method to obtain an optimized set of power loading coefficients (PLC) considering the transmitter's power- and area-efficiency for discrete multitone (DMT) modulation wireline transceivers (TRX). A simple way to estimate the PLCs is to invert the channel loss profile; however, to achieve an optimal signal-to-noise ratio (SNR), the PLCs should be multiplied by an additional value calculated by the peak-to-average power ratio (PAPR) of the TX output signal. In this work, we show how to calculate this value using the PAPR. Also, the TX DSP datapath where the PLCs are applied has an optimal bit-width in terms of the area given the bit-error-rate (BER) target, but the register-transfer-level (RTL)-based BER simulation iterating with various bit-widths of the DSP's internal datapath is significantly time-consuming. To address this, we designed a fixed-point DMT TRX simulator in MATLAB that allows us a fast iteration-based performance-area optimization of the DSP. With the design parameters including the PLCs obtained by the developed fixed-point simulator, the BER of the DMT TRX achieves 1.8E-5 with only 4-bit PLCs to communicate over a channel exhibiting 10 dB insertion loss at Nyquist without considering the analog front-end. © 2024 IEEE
Magnetically Actuated Capsule with Multiple Sampling Capability for Gut Microbiome Research
No full textUnderstanding the relationship between gut microbiota and gastrointestinal (GI) diseases through longitudinal analysis is vital for advancing diagnosis and treatment methods. To achieve this, ingestible devices capable of multipoint gut microbiota sampling are necessary. Conventional devices move passively through the GI tract, relying on physiological factors and are unable to perform multipoint sampling. This study introduces a novel magnetically actuated capsule robot designed to collect gut microbiota from various GI tract locations while minimizing cross-contamination. The capsule includes a body, a driving unit, six sampling tools, a central rod, and two heads. It uses electromagnetic field control for precise orientation and position of the capsule, ensuring the sampling channel faces downward for accurate collection. The capsule can collect six distinct microbiota samples, protecting the tools from contamination throughout the process. Its active locomotion and sampling capabilities were tested through ex-vivo tests. The feasibility of clinical application was demonstrated in preclinical trials using a porcine model, confirming the capsule's potential for integration and use in medical settings. © 2024 ICROS
30.1 A 40nm VLIW Edge Accelerator with 5MB of 0.256pJ/b RRAM and a Localization Solver for Bristle Robot Surveillance
No full textTiny surveillance robots need to efficiently compute a perception front-end workload, consisting of a neural network inference stack, and a localization back-end workload implementing a set of state-space equations. Miniaturization and low-power actuation make bristle robots [1] attractive locomotion platforms, but size limits lead to stringent energy constraints. The edge accelerator needs low leakage for long retentive stretches and efficient matrix compute for active bursts. We present a 0.84TOPS/W, 110μW retentive-sleep-capable resistive random-access memory (RRAM)-based accelerator in 40nm with 10 very long instruction word (VLIW)-controlled nonvolatile memory (NVM) matrix units (NMUs) with, in total, 5MB of RRAM, combined with a 10T SRAM-based state-update accelerator enabled by in-place memory updates. At VMIN, the design improves NVM access energy to 0.256pJ/b and peak NVM bandwidth to 12.8GB/s. © 2024 IEEE
Towards Organic Photodetectors Functioning Under Strong Sunlight. Machine-learning-assisted Design of Diarylethene n-type Dopants to Mix with p-type Organic Semiconductor P3HT
No full textLinear dynamic range of organic photodetectors, which is typically narrow due to low mobility of organic semiconductor, has been extended by diarylethene (DAE) photochromic switches doped to a poly-3-hexylthiophene photoactive layer. A speculated mechanism is that DAE acts as n-type electron traps only in its aromatic closed form, which is predominant only under strong sunlight, addressing the early saturation problem on sunny days. We herein identified two optimal DAE derivatives out of ~100 candidates, using the TDDFT calculations on the HOMO-LUMO energies of their open-closed isomers (~400 data). Since this is only a small subset of ~105 candidates, we predicted the HOMO-LUMO energies of the remaining candidates by machine learning with various artificial neural network models and molecule representation methods. We were able to identify additional optimal candidates, which were screened by machine learning prediction, and were confirmed by TDDFT calculations
A Discrete Multitone Wireline Transceiver Using Optimal Loading Over Reflective Channel For ADC-Based High-Speed Serial Links
No full textIn this paper, we validate the effectiveness of the bit and power loading (BPL) of discrete muti-tone (DMT) modulation over a reflective wireline channel, exhibiting notches in the signal-to-noise ratio (SNR) profile. The loading profile is optimized via the margin-adaptive Levin-Campello (LC) water-filling algorithm based on the measured per-subchannel SNR distribution. The aggregate bit-error rate (BER) is minimized by allocating only a small number of bits for low SNR subbands. The experiment is conducted with the M8196A arbitrary waveform generator (AWG) for the transmitter (TX) and the DSAX96204Q digital signal analyzer (DSA) for the receiver (RX) with the captured data and signal processing on MATLAB software. The measurement results have demonstrated that the DMT with the optimal BPL achieves 28.2 Gb/s data rate and 4.73E-6 BER when communicating over a reflective channel, showing 16.3 dB loss at 5 GHz Nyquist frequency. © 2024 IEEE
Bioelectronic Sutures with Electrochemical pH-Sensing for Long-Term Monitoring of the Wound Healing Progress
No full textThe physiological pH level at wound sites is one of the fundamental factors for monitoring wound conditions in clinical practice. To continuously assess the wound conditions, a variety of bioelectronic pH sensors are extensively developed. However, despite significant advances in bioelectronics for wound monitoring, the application of existing bioelectronic devices, primarily designed as bandages or patches, remains challenging for monitoring pH levels in deep wounds. Here, a flexible pH-sensing suture is introduced that can be simultaneously used as both a precise pH sensor for wound monitoring and a conventional medical suture. The electrochemical pH-sensing suture comprises Au nanoparticle-based flexible electrodes functionalized with polyaniline for the working electrode and Ag/AgCl for the reference electrode, seamlessly integrated onto a standard medical suturing thread. This dual-function sensing suture offers a reliable and high sensitivity of 58.9 mV pH−1, negligible hysteresis, high stability, and excellent selectivity in pH sensing. The biocompatibility of the sensing suture is systematically verified for its in vivo use. To demonstrate the capabilities of the pH-sensing suture, it is successfully applied to an incision and chronic wound model of mouse to perform continuous and accurate monitoring of the inflammation and healing progress of the wound throughout the healing period. © 2024 Wiley-VCH GmbH.FALSEsciescopu
H&E 조직병리학 이미지에서 향상된 종양 미세환경 분석을 위한 정확한 핵 영역화 및 분류
No full textComputer aided diagnosis, Deep learning, Histopathology, Nuclei segmentation, Tumor micro-environmentPerforming accurate analysis of the tumor micro-environment in H&E histopathology images requires precise detection, segmentation and classification of multiple types of cells. In oncology clinical practice, some prognostic indicators are directly derived from the presence, abundance or ratio between specific types of cells. Other diseases required more in depth analysis as some types of immune cells can help to reduce the proliferation and migration of tumor cells but, the presence of other types might suppress their anticancer activities. Therefore, the correct identification of individual cells and their type are fundamental steps to reliably analyze histopathology samples and to compute clinically relevant scores. However, manually labeling and analyzing large images with thousands of cells per region of interest is a tedious and time consuming task for clinicians, highlighting the need for automatic tools to ease the burden in clinical practice. Applying machine learning to leverage data to learn models for detection, segmentation and classification has demonstrated great potential in a wide variety of domains. However, in the case of histopathology images, tumor induced abnormalities usually change the morphology of nuclei as well as the tissue structures resulting in irregular agglomerations that makes the automatic detection and separation of individual instances difficult. In addition, the non-uniform distribution of cell types also introduces limitations during the learning process leading to limited detection of rare types of cells. In this thesis we propose methodologies to address the correct separation and segmentation of neighboring nuclei as well as cell type classification under a long-tailed distribution. In the first work, we propose a set of Siamese networks to correctly find the boundaries between adjacent nuclei given an initial, but inaccurate segmentation map. In the second work, we propose an end-to-end segmentation model that considers overlapping between neighboring cells to resolve ambiguous boundaries via amodal masks prediction. In the final work, we leverage the high level learned representation of the Segment Anything Model (SAM) to improve classification on rare types of nuclei effectively mitigating the effects of the natural long-tailed distribution of cell types in histopathology images. Overall, we present several techniques to aid the correct detection, segmentation and classification of individual nuclei in order to provide a strong foundation for tumor micro-environment analysis in H&E histopathology images. Keywords: Computer aided diagnosis, Deep learning, Histopathology, Nuclei segmentation, Tumor micro-environment.|H&E 조직병리 영상에서 종양 미세 환경을 정확하게 분석하려면 다양한 종류의 세포를 정밀하게 감지, 영역화 및 분류해야 합니다. 종양학 임상 시험에서는 특정 세포의 존재 여부, 밀도 또는 세포 간의 비율로부터 여러 예후 지표가 유도됩니다. 일부 면역 세포는 종양 세포의 증식과 이동을 억제할 수 있지만 다른 세포는 항암 활동을 저해할 수 있으므로 더 심층적인 분석이 필요한 질병도 있습니다. 따라서 개별 세포와 그 유형을 올바르게 식별하는 것은 조직병리학 샘플을 안정적으로 분석하고 정확한 측정항목을 계산하기 위한 기본 단계입니다. 그러나 대형 이미지에는 관심 영역 당 수천 개의 세포가 포함되어 있어 이를 수동으로 분석하는 것은 시간이 많이 걸리고 지루한 작업이므로, 이를 문제를 해결하기 위한 자동화 도구가 필요합니다. 기계 학습을 적용하여 탐지, 세분화 및 분류 모델을 학습하는 것은 매우 다양한 영역에서 큰 잠재력을 입증하였습니다. 그러나 종양으로 인한 변화는 종종 핵의 형태와 조직 구조를 변형하며, 이는 자동 감지와 분리를 어렵게 만듭니다. 또한, 세포 유형의 불균일한 분포로 인해 학습 과정에서 한계가 발생할 수 있습니다. 이 논문에서 우리는 긴 꼬리 분포 하에 세포 유형 분류 및 근접한 세포 핵의 분리 및 영역화를 다루는 방법을 제안합니다. 첫째, 부정확한 영역화 맵이 주어진 경우 인접한 핵 사이의 경계를 올바르게 찾기 위한 시아미즈 네트워크 세트를 제안합니다. 둘째, 근접 세포 간의 중첩을 고려하는 엔드 투 엔드 영역화 모델을 제안합니다. 최종 작업에서는 SAM(Segment Anything Model)이 학습한 잘 최적화된 분할 패턴을 사용하여 희귀 유형의 핵에 대한 분류를 개선하고 조직병리학 이미지에서 세포 유형의 자연적인 긴 꼬리 분포의 영향을 효과적으로 완화합니다. 우리는 H&E 조직병리 영상에서의 각 세포핵의 검출, 영역화 및 분류 기술을 제안하여 종양 미세 환경 분석을 위한 강력한 기반을 제공합니다. 핵심어: 컴퓨터 지원 진단, 딥러닝, 조직병리학, 핵 영역화, 종양 미세 환경.I. Introduction 1
1 Background and Motivations 1
2 Contributions and Outline 2
3 Publications 3
3.1 Excluded Research 3
II. Precise Separation of Adjacent Nuclei using a Siamese Neural Network 6
1 Introduction 6
2 Method 8
2.1 Extraction of Ambiguous Neighboring Instances 8
2.2 Proposed Network for Nuclei Separation 9
2.3 Implementation Details 11
3 Experimental Results 11
3.1 Separation Accuracy 12
3.2 Classification Accuracy 12
4 Conclusion 14
III. Attention guided multi-scale cluster refinement with extended field of view for amodal nuclei segmentation 15
1 Introduction 15
2 Related Works 17
2.1 Amodal instance segmentation 17
2.2 Nuclei segmentation 17
3 Method 18
3.1 Network architecture 19
3.2 Instance segmentation using Gaussian clusters 21
4 Metrics for Nuclei Segmentation 24
4.1 Unique matching 25
4.2 Extended metrics 26
5 Experiments 27
5.1 Datasets 28
5.2 Implementation details 28
6 Results 29
6.1 MoNuSeg dataset 29
6.2 CPM17 Dataset 31
6.3 TNBC Dataset 32
6.4 Running time 34
6.5 Ablations studies 34
6.6 Amodal segmentation of crowded nuclei 36
7 Discussion 38
8 Conclusion 38
IV. Enhanced Nuclei Segmentation and Classification via Category Descriptors in the SAM Model 40
1 Introduction 40
2 Related works 42
2.1 Nuclei segmentation 42
2.2 Segment Anything Model (SAM) 42
3 Method 44
3.1 Category descriptors 44
3.2 Domain alignment 45
3.3 Training objective 45
4 Experiments 46
4.1 Dataset 46
4.2 Experimental setup 46
4.3 Evaluation metrics 47
4.4 Comparison methods 47
4.5 Implementation details 48
5 Results 49
5.1 Ablation studies 52
5.2 SAM pretrained on medical images 53
5.3 Long tailed classification analysis 53
5.4 Manual prompts 54
6 Clinical case study 55
7 Limitations 56
8 Generalizability 56
9 Conclusion 56
V. Concluding Remarks 58
1 Conclusion 58
2 Future Work 59
VI. Acknowledgement 60
References 61
요약문 72DoctordCollectio
Performance Analysis of 4G/5G C-V2X Video Transmission
No full text차량 간 통신(Vehicle-to-Everything, V2X), 특히 셀룰러 V2X(C-V2X)는 원거리에서도 실시간 교통 상황 정보 공유가 가능하다는 점에서 자율주행차 및 스마트 교통 시스템의 발전에서 주목받고 있다. 현재 C-V2X를 활용한 협력 인식 메시지(Cooperative Awareness Messages, CAM), 안전 메시지(Basic Safety Messages, BSM)와 같은 주기적 저속 데이터 메시지 전송이 고려되고 이에 관련된 연구가 다양하게 진행되고 있지만, 원격주행 등의 서비스를 제공하기 위해서는 센서 데이터 등의 고속 데이터 메시지 공유가 필요하다. 이에 본 연구에서는 MATLAB 기반 시뮬레이션을 활용하여 차량-기지국 간의 실시간 카메라 센서 영상 공유를 위한 통신 상황을 가정하고, 4G LTE와 5G NR 환경에서의 영상 전송 성능을 분석하였다. 이 때, 영상 성능 전송 성능 평가를 위해 변조 코딩 구성(Modulation and Coding Scheme, MCS), 압축 여부, 해상도에 따른 지연 시간과 지원 가능한 초당 프레임(Frame Per Second, FPS)을 분석하였다.
Vehicle-to-Everything (V2X) communication, particularly Cellular V2X (C-V2X), is gaining attention in the development of autonomous vehicles and smart transportation systems due to its ability to share real-time traffic information over long distances. While current research and applications focus on the periodic low-speed data message transmission, such as Cooperative Awareness Messages (CAM) and Basic Safety Messages (BSM), utilizing C-V2X, high-speed data message sharing, such as sensor data, is essential for providing services like remote driving. Therefore, in this study, we assumed a communication scenario for sharing real-time camera sensor footage between vehicles and base stations using MATLAB-based simulations and analyzed the video transmission performance in 4G LTE and 5G NR environments. To evaluate the video transmission performance, we analyzed the latency and available frames per second (FPS) according to the modulation and coding scheme (MCS), compression, and image resolution.FALSEkc