Mohamed A. Abd El Ghany
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Mohamed A. Abd El Ghany has a total of 99 co-authors. Here are the top co-authors by number of shared publications:
Klaus Hofmann
Integrated Electronic Systems Lab, TU Darmstadt, Germany
11 shared publicationsMohammed Ismail
The Ohio State University, Columbus, USA, currently with khalifa University of Science, Technology and Research (KUSTAR), The UAE
9 shared publicationsDiana Goehringer
Chair of Adaptive Dynamic Systems, TU Dresden
7 shared publicationsSalma Hesham
Ruhr University Bochum, Germany, German University in Cairo, Egypt
7 shared publicationsMagdy A. El-Moursy
Electronics Research Institute, Cairo, Egypt, Mentor Graphics Corporation, Cairo, Egypt
5 shared publications- Mohammed Abdel Megeed Salem
German University in Cairo
5 shared publications
Latest Publications
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Embedded Systems Letters
- Vol. 18 (5) , 328-331
SFLGuard: Model Poisoning Mitigation in Split Federated Learning via Activation-Level Attestation
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Embedded Systems Letters
- Vol. 18 (5) , 328-331
SFLGuard: Model Poisoning Mitigation in Split Federated Learning via Activation-Level Attestation
- Open Access
- 27 March 2026
- Published by Springer Nature in Journal of King Saud University Computer and Information Sciences
- Vol. 38 (3)
Review on federated optimization in multi-tier architectures
Federated Learning (FL) enables collaborative model training without exposing sensitive data, making it a cornerstone for privacy-aware AI. However, bringing FL from theory to practice in multi-tier architectures that contain hierarchical edge-fog-cloud systems remains difficult, challenged by security vulnerabilities, resource constraints, and system heterogeneity. This paper reviews secure and adaptive optimization techniques that address these challenges in hierarchical environments. We focus on three main areas: (1) security-enhanced FL, covering threat models, adversarial attacks, and defenses tailored to layered architectures; (2) adaptive optimization, including dynamic client selection, resource-aware aggregation, and context-sensitive privacy management; and (3) systems integration, examining architectural designs, communication protocols, and orchestration methods for scalable deployment. By synthesizing recent work, the survey highlights trade-offs between privacy, performance, and scalability, and proposes a taxonomy of current challenges and solutions. It also evaluates approaches to trustworthiness in FL, including fairness, accountability, and robustness, while considering practical issues such as regulation, benchmarking, and deployment constraints. The paper concludes with open research directions, emphasizing the need for secure, adaptive, and production-ready FL systems.
- 27 March 2026
- Published by Springer Nature in Journal of King Saud University Computer and Information Sciences
- Vol. 38 (3)
Open AccessReview on federated optimization in multi-tier architectures
Federated Learning (FL) enables collaborative model training without exposing sensitive data, making it a cornerstone for privacy-aware AI. However, bringing FL from theory to practice in multi-tier architectures that contain hierarchical edge-fog-cloud systems remains difficult, challenged by security vulnerabilities, resource constraints, and system heterogeneity. This paper reviews secure and adaptive optimization techniques that address these challenges in hierarchical environments. We focus on three main areas: (1) security-enhanced FL, covering threat models, adversarial attacks, and defenses tailored to layered architectures; (2) adaptive optimization, including dynamic client selection, resource-aware aggregation, and context-sensitive privacy management; and (3) systems integration, examining architectural designs, communication protocols, and orchestration methods for scalable deployment. By synthesizing recent work, the survey highlights trade-offs between privacy, performance, and scalability, and proposes a taxonomy of current challenges and solutions. It also evaluates approaches to trustworthiness in FL, including fairness, accountability, and robustness, while considering practical issues such as regulation, benchmarking, and deployment constraints. The paper concludes with open research directions, emphasizing the need for secure, adaptive, and production-ready FL systems.
- 3 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
- Vol. 34 (4) , 1179-1191
A First-Settle, First-Convert Scheduling Scheme for Low-Latency AI Inference in Crossbar Arrays
Resistive random access memory (RRAM) crossbar arrays enable analog in-memory computing, promising high parallelism and reduced data transfer for deep neural networks (DNNs) inference. However, when many columns share a single time-multiplexed analog-to-digital converter (ADC), readout latency, dominated by serialized conversion, can become a performance bottleneck. This article presents the First-Settle, First-Convert (FSFC) readout architecture, a readiness-driven ADC scheduling scheme for RRAM-based fully connected neural networks (FCNNs). Conventional crossbar readout relies on global settling, where all bitlines must wait for the slowest column to stabilize before sequential conversion begins. FSFC replaces this schedule with per-column readiness detection and a priority-queue encoder that continuously routes the earliest-settled column to the shared ADC, overlapping bitline settling and conversion without adding extra ADCs or changing the crossbar. Software experiments and SPICE-level circuit simulations in SkyWater 130 nm show that FSFC reduces inference latency by 35%–67% across 1- to 8-bit quantized MNIST FCNNs, with no loss of classification accuracy. A transistor-level implementation of the differentiator, hysteresis comparator, and digital priority-queue encoder incurs about 80 pJ of control energy per matrix-vector multiplication (MVM) for a $30\times 10$ array, scaling to approximately 1 nJ per MVM and about 5%–8% tile area overhead for a 128-column tile. Because FSFC shortens ADC active time while introducing only modest peripheral circuitry, it provides a low-cost, drop-in latency improvement for RRAM compute-in-memory (CIM) tiles. The technique is orthogonal to existing ADC pooling, precision-reduction, and analog-cascading schemes, and can be integrated as a readout front-end in future crossbar-based accelerators.
- 3 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
- Vol. 34 (4) , 1179-1191
A First-Settle, First-Convert Scheduling Scheme for Low-Latency AI Inference in Crossbar Arrays
Resistive random access memory (RRAM) crossbar arrays enable analog in-memory computing, promising high parallelism and reduced data transfer for deep neural networks (DNNs) inference. However, when many columns share a single time-multiplexed analog-to-digital converter (ADC), readout latency, dominated by serialized conversion, can become a performance bottleneck. This article presents the First-Settle, First-Convert (FSFC) readout architecture, a readiness-driven ADC scheduling scheme for RRAM-based fully connected neural networks (FCNNs). Conventional crossbar readout relies on global settling, where all bitlines must wait for the slowest column to stabilize before sequential conversion begins. FSFC replaces this schedule with per-column readiness detection and a priority-queue encoder that continuously routes the earliest-settled column to the shared ADC, overlapping bitline settling and conversion without adding extra ADCs or changing the crossbar. Software experiments and SPICE-level circuit simulations in SkyWater 130 nm show that FSFC reduces inference latency by 35%–67% across 1- to 8-bit quantized MNIST FCNNs, with no loss of classification accuracy. A transistor-level implementation of the differentiator, hysteresis comparator, and digital priority-queue encoder incurs about 80 pJ of control energy per matrix-vector multiplication (MVM) for a $30\times 10$ array, scaling to approximately 1 nJ per MVM and about 5%–8% tile area overhead for a 128-column tile. Because FSFC shortens ADC active time while introducing only modest peripheral circuitry, it provides a low-cost, drop-in latency improvement for RRAM compute-in-memory (CIM) tiles. The technique is orthogonal to existing ADC pooling, precision-reduction, and analog-cascading schemes, and can be integrated as a readout front-end in future crossbar-based accelerators.
- Open Access
- 13 December 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 22 (4) , 1-29
An End-to-End Framework for Compiling Dense and Sparse Matrix-Vector Multiplications for FPGA-HBM Acceleration
The bandwidth improvement provided by high-bandwidth memory (HBM), and the capability of FPGAs to customize the processing and memory hierarchy, results in a considerable performance increase for memory-intensive workloads such as graph processing, sorting, machine learning, and database analytics. Modern systems integrating 3D-stacked DRAM memory can be leveraged to realize the Near-Memory Computing (NMC) paradigm by offloading some computations to accelerators placed near the HBM. Matrix-vector multiplication (MVM) kernels, which are memory-bound, can significantly benefit from being executed on FPGA-HBM platforms. MVM kernels can be broadly categorized into two types: dense (General Matrix-Vector Multiplication, GEMV) and sparse (Sparse Matrix-Vector Multiplication, SpMV). Recent literature has predominantly focused on optimizing SpMV for FPGA-HBM, leaving a unified solution relatively unexplored. In this work, we introduce an end-to-end framework for compiling MVM kernels for FPGA-HBM Acceleration. It consists of a software and a hardware components. The software component introduces the MATIO compiler, a novel toolflow for detecting MVM and matrix multiplication (MM) kernels in C or C++ code, and replacing MVM kernels with a call to our FPGA accelerator. MATIO is capable of detecting 90% of MVM and MM kernels in real-world benchmarks collected from Github. Additionally, it is faster than state-of-the-art detection methods by at least 45x. On the hardware side, we introduce VecMADS, a novel FPGA architecture designed to efficiently handle both GEMV and SpMV operations. Our architecture leverages the high memory bandwidth of HBM to overcome memory bottlenecks, providing a comprehensive solution for accelerating matrix-vector multiplication on FPGAs. Evaluation results show that VecMADS delivers 1.5x higher throughput and 4.8x higher energy efficiency compared to cuSPARSE library on GPU. Considering dense benchmarks, VecMADS achieves 1.26x higher throughput than the hipBLAS library running on GPU.
- 13 December 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 22 (4) , 1-29
Open AccessAn End-to-End Framework for Compiling Dense and Sparse Matrix-Vector Multiplications for FPGA-HBM Acceleration
The bandwidth improvement provided by high-bandwidth memory (HBM), and the capability of FPGAs to customize the processing and memory hierarchy, results in a considerable performance increase for memory-intensive workloads such as graph processing, sorting, machine learning, and database analytics. Modern systems integrating 3D-stacked DRAM memory can be leveraged to realize the Near-Memory Computing (NMC) paradigm by offloading some computations to accelerators placed near the HBM. Matrix-vector multiplication (MVM) kernels, which are memory-bound, can significantly benefit from being executed on FPGA-HBM platforms. MVM kernels can be broadly categorized into two types: dense (General Matrix-Vector Multiplication, GEMV) and sparse (Sparse Matrix-Vector Multiplication, SpMV). Recent literature has predominantly focused on optimizing SpMV for FPGA-HBM, leaving a unified solution relatively unexplored. In this work, we introduce an end-to-end framework for compiling MVM kernels for FPGA-HBM Acceleration. It consists of a software and a hardware components. The software component introduces the MATIO compiler, a novel toolflow for detecting MVM and matrix multiplication (MM) kernels in C or C++ code, and replacing MVM kernels with a call to our FPGA accelerator. MATIO is capable of detecting 90% of MVM and MM kernels in real-world benchmarks collected from Github. Additionally, it is faster than state-of-the-art detection methods by at least 45x. On the hardware side, we introduce VecMADS, a novel FPGA architecture designed to efficiently handle both GEMV and SpMV operations. Our architecture leverages the high memory bandwidth of HBM to overcome memory bottlenecks, providing a comprehensive solution for accelerating matrix-vector multiplication on FPGAs. Evaluation results show that VecMADS delivers 1.5x higher throughput and 4.8x higher energy efficiency compared to cuSPARSE library on GPU. Considering dense benchmarks, VecMADS achieves 1.26x higher throughput than the hipBLAS library running on GPU.
- Open Access
- 26 September 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Embedded Computing Systems
- Vol. 24 (5s) , 1-26
Timekeepers: ML-Driven SDF Analysis for Power-Wasters Detection in FPGAs
As the integration of FPGAs into cloud computing platforms accelerates, the risk of fault injection attacks - especially through power-wasting designs - becomes increasingly critical. Malicious tenants can upload FPGA designs that, under specific input stimuli, generate excessive power consumption, jeopardizing the integrity of the shared power delivery network (PDN) and enabling denial-of-service or side-channel attacks. Traditional detection techniques relying on netlist and bitstream analysis struggle with generalization and can be evaded through circuit obfuscation and seemingly benign designs. In contrast to these netlist-based approaches, we introduce Timekeepers, a novel detection method that utilizes Standard Delay Format (SDF) timing data combined with machine learning to detect anomalous power behavior in synthesized FPGA designs. Our method trains a decision tree classifier on SDF files generated from both benign and malicious designs, focusing on timing characteristics such as propagation delays and setup/hold violations to identify power wasters at the primitive level. By abstracting away from circuit connectivity and emphasizing timing patterns, our framework is both scalable and robust across different FPGA architectures. The classifier independently evaluates each FPGA component and aggregates the results using a threshold-based voting system to improve detection granularity and reduce false positives. Timekeepers achieves 99.6% accuracy and demonstrates superior performance compared to state-of-the-art solutions. Furthermore, our approach is platform-agnostic and does not require access to netlists or bitstreams, preserving intellectual property confidentiality while enhancing pre-deployment security checks.
- 26 September 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Embedded Computing Systems
- Vol. 24 (5s) , 1-26
Open AccessTimekeepers: ML-Driven SDF Analysis for Power-Wasters Detection in FPGAs
As the integration of FPGAs into cloud computing platforms accelerates, the risk of fault injection attacks - especially through power-wasting designs - becomes increasingly critical. Malicious tenants can upload FPGA designs that, under specific input stimuli, generate excessive power consumption, jeopardizing the integrity of the shared power delivery network (PDN) and enabling denial-of-service or side-channel attacks. Traditional detection techniques relying on netlist and bitstream analysis struggle with generalization and can be evaded through circuit obfuscation and seemingly benign designs. In contrast to these netlist-based approaches, we introduce Timekeepers, a novel detection method that utilizes Standard Delay Format (SDF) timing data combined with machine learning to detect anomalous power behavior in synthesized FPGA designs. Our method trains a decision tree classifier on SDF files generated from both benign and malicious designs, focusing on timing characteristics such as propagation delays and setup/hold violations to identify power wasters at the primitive level. By abstracting away from circuit connectivity and emphasizing timing patterns, our framework is both scalable and robust across different FPGA architectures. The classifier independently evaluates each FPGA component and aggregates the results using a threshold-based voting system to improve detection granularity and reduce false positives. Timekeepers achieves 99.6% accuracy and demonstrates superior performance compared to state-of-the-art solutions. Furthermore, our approach is platform-agnostic and does not require access to netlists or bitstreams, preserving intellectual property confidentiality while enhancing pre-deployment security checks.
- 18 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. 72 (10) , 5825-5838
Domain-Specific Hyperdimensional RISC-V Processor for Edge-AI Training
Edge AI has become the cornerstone of many applications. Yet, progress is limited by the large complexity of training dnn. hdc is positioned as an alternative approach for Edge AI that is compact enough to enable training. The main challenge for hdc model is to maintain its key features while balancing high inference accuracy with efficiency. A simple binary () model lacks accuracy, while the computational complexity of a floating-point model is too high. This work presents FixedHD, a novel 16-bit fixed-point model enabling training at the Edge. FixedHD achieves an accuracy similar to floating-point model while lowering computational complexity. The model is supported by a customized RISC-V processor tailored to speedup both training and inference. The processor is extended with advanced -specific instructions, a vector unit to utilize ’s parallel nature, and, for the first time, approximate computing to exploit its robustness. Further, memory requirements are reduced by quantizing mathematical functions and reducing the large encoding matrix by up to 390. Compared to the baseline processor, inference and training are accelerated on average by 6.9and 3, respectively. The energy consumption is reduced by 4.6and 1.9at the cost of an increase in area by 45%. The inference accuracy remains at the high level of floating-point models despite the heavy quantization and approximation.
- 18 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. 72 (10) , 5825-5838
Domain-Specific Hyperdimensional RISC-V Processor for Edge-AI Training
Edge AI has become the cornerstone of many applications. Yet, progress is limited by the large complexity of training dnn. hdc is positioned as an alternative approach for Edge AI that is compact enough to enable training. The main challenge for hdc model is to maintain its key features while balancing high inference accuracy with efficiency. A simple binary () model lacks accuracy, while the computational complexity of a floating-point model is too high. This work presents FixedHD, a novel 16-bit fixed-point model enabling training at the Edge. FixedHD achieves an accuracy similar to floating-point model while lowering computational complexity. The model is supported by a customized RISC-V processor tailored to speedup both training and inference. The processor is extended with advanced -specific instructions, a vector unit to utilize ’s parallel nature, and, for the first time, approximate computing to exploit its robustness. Further, memory requirements are reduced by quantizing mathematical functions and reducing the large encoding matrix by up to 390. Compared to the baseline processor, inference and training are accelerated on average by 6.9and 3, respectively. The energy consumption is reduced by 4.6and 1.9at the cost of an increase in area by 45%. The inference accuracy remains at the high level of floating-point models despite the heavy quantization and approximation.
- Open Access
- 20 August 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 118339-118361
Triple Layer RGB Image Encryption Algorithm Utilizing Three Hyperchaotic Systems and Its FPGA Implementation
Image encryption is essential for protecting sensitive visual data during transmission and storage. This article introduces a novel image encryption technique that utilizes three hyperchaotic systems: a memristor system, a hyperchaotic 7D system, and an Erbium-doped fiber laser-based system. The algorithm takes advantage of the complex dynamics and random-like behavior of these systems to implement a robust three-layer encryption process. In each layer, a pseudo-random number generator is used to construct an S-box and an encryption key, which are subsequently applied to the image data. The proposed encryption scheme is implemented in software using Wolfram Mathematica® version 13.1 and in hardware using an AMD Virtex-7 FPGA VC707, achieving encryption rates of 9.7 Mbps and 0.67 Gbps, respectively. Experimental results demonstrate the scheme’s effectiveness in terms of encryption quality, resistance to various cryptanalysis attacks, and a high computational efficiency. The proposed method shows significant potential for secure image transmission and storage applications.
- 20 August 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 118339-118361
Open AccessTriple Layer RGB Image Encryption Algorithm Utilizing Three Hyperchaotic Systems and Its FPGA Implementation
Image encryption is essential for protecting sensitive visual data during transmission and storage. This article introduces a novel image encryption technique that utilizes three hyperchaotic systems: a memristor system, a hyperchaotic 7D system, and an Erbium-doped fiber laser-based system. The algorithm takes advantage of the complex dynamics and random-like behavior of these systems to implement a robust three-layer encryption process. In each layer, a pseudo-random number generator is used to construct an S-box and an encryption key, which are subsequently applied to the image data. The proposed encryption scheme is implemented in software using Wolfram Mathematica® version 13.1 and in hardware using an AMD Virtex-7 FPGA VC707, achieving encryption rates of 9.7 Mbps and 0.67 Gbps, respectively. Experimental results demonstrate the scheme’s effectiveness in terms of encryption quality, resistance to various cryptanalysis attacks, and a high computational efficiency. The proposed method shows significant potential for secure image transmission and storage applications.
- Open Access
MultiWave-Net: An Optimized Spatiotemporal Network for Abnormal Action Recognition Using Wavelet-Based Channel Augmentation
Human behavior is regarded as one of the most complex notions present nowadays, due to the large magnitude of possibilities. These behaviors and actions can be distinguished as normal and abnormal. However, abnormal behavior is a vast spectrum, so in this work, abnormal behavior is regarded as human aggression or in another context when car accidents occur on the road. As this behavior can negatively affect the surrounding traffic participants, such as vehicles and other pedestrians, it is crucial to monitor such behavior. Given the current prevalent spread of cameras everywhere with different types, they can be used to classify and monitor such behavior. Accordingly, this work proposes a new optimized model based on a novel integrated wavelet-based channel augmentation unit for classifying human behavior in various scenes, having a total number of trainable parameters of 5.3 m with an average inference time of 0.09 s. The model has been trained and evaluated on four public datasets: Real Live Violence Situations (RLVS), Highway Incident Detection (HWID), Movie Fights, and Hockey Fights. The proposed technique achieved accuracies in the range of to across the used benchmark datasets. Comprehensive analysis and comparisons between different versions of the model and the state-of-the-art have been performed to confirm the model’s performance in terms of accuracy and efficiency. The proposed model has higher accuracy with an average of , and higher efficiency by reducing the number of parameters by around 139.1 m compared to other models trained and tested on the same benchmark datasets.Open AccessMultiWave-Net: An Optimized Spatiotemporal Network for Abnormal Action Recognition Using Wavelet-Based Channel Augmentation
Human behavior is regarded as one of the most complex notions present nowadays, due to the large magnitude of possibilities. These behaviors and actions can be distinguished as normal and abnormal. However, abnormal behavior is a vast spectrum, so in this work, abnormal behavior is regarded as human aggression or in another context when car accidents occur on the road. As this behavior can negatively affect the surrounding traffic participants, such as vehicles and other pedestrians, it is crucial to monitor such behavior. Given the current prevalent spread of cameras everywhere with different types, they can be used to classify and monitor such behavior. Accordingly, this work proposes a new optimized model based on a novel integrated wavelet-based channel augmentation unit for classifying human behavior in various scenes, having a total number of trainable parameters of 5.3 m with an average inference time of 0.09 s. The model has been trained and evaluated on four public datasets: Real Live Violence Situations (RLVS), Highway Incident Detection (HWID), Movie Fights, and Hockey Fights. The proposed technique achieved accuracies in the range of to across the used benchmark datasets. Comprehensive analysis and comparisons between different versions of the model and the state-of-the-art have been performed to confirm the model’s performance in terms of accuracy and efficiency. The proposed model has higher accuracy with an average of , and higher efficiency by reducing the number of parameters by around 139.1 m compared to other models trained and tested on the same benchmark datasets. - 6 September 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 577-584
Auto-DOK: Compiler-Assisted Automatic Detection of Offload Kernels for FPGA-HBM Architectures
The bandwidth improvement provided by high-bandwidth memory (HBM), and the capability of FPGAs to customize the processing and memory hierarchy, results in a considerable performance increase for memory-intensive work-loads such as graph processing, sorting, machine learning, and database analytics. Modern systems integrating 3D-stacked DRAM memory can be leveraged to realize the Near-Memory Computing (NMC) paradigm by offloading some computations to accelerators placed near the HBM. Although numerous studies have investigated efficient accelerators for FPGA-HBM platforms, researchers have not proposed a systematic way for identifying which application kernels are suitable for execution near the HBM. In this article, we propose compiler support for recognizing offloading candidates without any burden on programmers. Auto-DOK analyzes an application code based on criteria derived from the hardware design goals of FPGA-HBM platforms, and automatically identifies kernels suitable for offloading. We evaluate Auto-DOK on benchmarks ranging from microbenchmarks to real-world kernels. Our results show that Auto-DOK can correctly identify kernels and input sizes suitable for execution near the HBM, and prevents slowdown caused by incorrect offloading decisions for other workloads. Moreover, Auto-DOK operates at compile time with negligible overhead and without the need for expensive profiling.
- 6 September 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 577-584
Auto-DOK: Compiler-Assisted Automatic Detection of Offload Kernels for FPGA-HBM Architectures
The bandwidth improvement provided by high-bandwidth memory (HBM), and the capability of FPGAs to customize the processing and memory hierarchy, results in a considerable performance increase for memory-intensive work-loads such as graph processing, sorting, machine learning, and database analytics. Modern systems integrating 3D-stacked DRAM memory can be leveraged to realize the Near-Memory Computing (NMC) paradigm by offloading some computations to accelerators placed near the HBM. Although numerous studies have investigated efficient accelerators for FPGA-HBM platforms, researchers have not proposed a systematic way for identifying which application kernels are suitable for execution near the HBM. In this article, we propose compiler support for recognizing offloading candidates without any burden on programmers. Auto-DOK analyzes an application code based on criteria derived from the hardware design goals of FPGA-HBM platforms, and automatically identifies kernels suitable for offloading. We evaluate Auto-DOK on benchmarks ranging from microbenchmarks to real-world kernels. Our results show that Auto-DOK can correctly identify kernels and input sizes suitable for execution near the HBM, and prevents slowdown caused by incorrect offloading decisions for other workloads. Moreover, Auto-DOK operates at compile time with negligible overhead and without the need for expensive profiling.
- 4 January 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control
- Vol. 70 (3) , 237-252
A Single-Shot Harmonic Imaging Approach Utilizing Deep Learning for Medical Ultrasound
Tissue Harmonic Imaging (THI) is an invaluable tool in clinical ultrasound owing to its enhanced contrast resolution and reduced reverberation clutter in comparison to fundamental mode imaging. However, harmonic content separation based on high pass filtering suffers from potential contrast degradation or lower axial resolution due to spectral leakage. Whereas nonlinear multi-pulse harmonic imaging schemes, such as amplitude modulation and pulse inversion, suffer from a reduced framerate and comparatively higher motion artifacts due to the necessity of at least two pulse echo acquisitions. To address this problem, we propose a deep-learning-based single-shot harmonic imaging technique capable of generating comparable image quality to pulse amplitude modulation methods, yet at a higher framerate and with fewer motion artifacts. Specifically, an asymmetric convolutional encoder-decoder structure is designed to estimate the combination of the echoes resulting from the half-amplitude transmissions using the echo produced from the full amplitude transmission as input. The echoes were acquired with the checkerboard amplitude modulation technique for training. The model was evaluated across various targets and samples to illustrate generalizability as well as the possibility and impact of transfer learning. Furthermore, for possible interpretability of the network, we investigate if the latent space of the encoder holds information on the nonlinearity parameter of the medium. We demonstrate the ability of the proposed approach to generate harmonic images with a single firing that are comparable to those from a multi-pulse acquisition.
- 4 January 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control
- Vol. 70 (3) , 237-252
A Single-Shot Harmonic Imaging Approach Utilizing Deep Learning for Medical Ultrasound
Tissue Harmonic Imaging (THI) is an invaluable tool in clinical ultrasound owing to its enhanced contrast resolution and reduced reverberation clutter in comparison to fundamental mode imaging. However, harmonic content separation based on high pass filtering suffers from potential contrast degradation or lower axial resolution due to spectral leakage. Whereas nonlinear multi-pulse harmonic imaging schemes, such as amplitude modulation and pulse inversion, suffer from a reduced framerate and comparatively higher motion artifacts due to the necessity of at least two pulse echo acquisitions. To address this problem, we propose a deep-learning-based single-shot harmonic imaging technique capable of generating comparable image quality to pulse amplitude modulation methods, yet at a higher framerate and with fewer motion artifacts. Specifically, an asymmetric convolutional encoder-decoder structure is designed to estimate the combination of the echoes resulting from the half-amplitude transmissions using the echo produced from the full amplitude transmission as input. The echoes were acquired with the checkerboard amplitude modulation technique for training. The model was evaluated across various targets and samples to illustrate generalizability as well as the possibility and impact of transfer learning. Furthermore, for possible interpretability of the network, we investigate if the latent space of the encoder holds information on the nonlinearity parameter of the medium. We demonstrate the ability of the proposed approach to generate harmonic images with a single firing that are comparable to those from a multi-pulse acquisition.