ABSTRACT. Mammogram images play a crucial diagnostic role in the early detection of breast cancer. However, the limited availability of medical data, particularly the scarcity of tumor-related images, remains a significant barrier to training artificial intelligence models. To address this challenge, a generative artificial intelligence approach is proposed—specifically, an algorithm for generating synthetic mammogram images based on the CycleGAN (Cycle-Consistent Generative Adversarial Network) model. The proposed approach enables the generation of new synthetic samples through bidirectional image-to-image translation between two classes of images—healthy and tumor-affected mammograms. CycleGAN can operate on any unpaired image dataset and is capable of producing synthetic images that closely resemble real ones, exhibiting statistically consistent characteristics. The increase in the number of synthetic images enhances the generalization capability of deep learning models and improves the accuracy of medical AI systems. In the experimental phase, the CBIS-DDSM and INbreast datasets were utilized, and the visual quality as well as the statistical realism of the generated synthetic images were evaluated. The results demonstrate that the CycleGAN-based algorithm can serve as an effective auxiliary tool for training diagnostic models in scenarios with limited availability of medical images.

ABSTRACT. Breast cancer is one of the most common malignancies among women. Early diagnosis plays a critical role in increasing the effectiveness of treatment. Nowadays, artificial intelligence (AI) and deep learning methods are widely used for detecting cancer based on mammogram images. However, due to the privacy and distributed nature of medical data, traditional centralized training approaches face practical limitations. This paper proposes a federated learning approach for detecting breast cancer based on mammogram images. In the proposed method, medical data is stored locally across multiple centers (e.g., hospitals), and each center trains a model locally. Subsequently, the updates from these local models are aggregated on a central server to produce a global model. This approach enables large-scale model training on real clinical data without compromising data privacy. Experiments were conducted on publicly available datasets such as CBIS-DDSM and INbreast. The proposed federated model achieved competitive accuracy and sensitivity compared to conventional centralized models. The results demonstrate that federated learning is a reliable and privacy-preserving alternative for breast cancer detection.

ABSTRACT. Abstract—The study's objective is the process of orthopaedic rehabilitation of patients with musculoskeletal disorders using mobile sensor systems for gait analysis. The addressed problem comes from the need to create an affordable, portable, and informative tool for long-term monitoring of patient recovery dynamics that enables unbiased biomechanical indicators to be recorded without requiring a clinical environment. During the research, a mobile system for collecting and analysing pododynamometric data was developed, which combines a sensor insole with FSR sensors, an ESP microcontroller, along with data visualisation in the form of stabilograms and pressure or weight graphs. Experimental studies of different gait types were conducted, which allowed us to identify pressure or weight patterns characteristics for further use in diagnostics and monitoring the effectiveness of rehabilitation. The features of the developed system are portability, modularity, autonomy, energy efficiency, low cost of components, and the ability to personalise for different sizes of the patient's foot. The field of practical application of the results covers clinical and home rehabilitation after injuries, amputations, orthopaedic surgeries, as well as medical support for military and civilian personnel in the process of returning to an active lifestyle.

ABSTRACT. The paper presents an intelligent platform for scientific research in the field of computer vision, aimed at comprehensive comparative analysis of the accuracy, performance, and robustness of modern models under variable lighting and weather conditions. The primary objective of the development is to create a scalable system that supports experiments with a wide range of models (YOLOv8–YOLOv11, DETR, Mask R-CNN, SSD, RetinaNet, etc.) using interactive analytics, complete logging, and automated report generation. Unlike conventional testing environments, the proposed platform accounts for the characteristics of images from various application domains. For each scenario, log files are automatically stored with key metrics (mAP@0.5, mAP@0.5:.95, Precision, Recall, F1-score, IoU) and FPS, ensuring a comprehensive evaluation of both accuracy and processing speed. A distinctive practical novelty of the system is the implemented module for studying the impact of video stream preprocessing methods (CLAHE, gamma correction, median filtering, etc.) on object detection accuracy. Experimental results demonstrated, for example, that in low-visibility conditions the combination of Zero-DCE + gamma correction + CLAHE yielded an increase in mAP@0.5 in the range of 4–6%, depending on the model architecture. The platform also provides tools for assessing the effect of CUDA acceleration on the performance of each model. The developed system has undergone validation in a series of experiments, including tasks of adapting computer vision models within hardware-software complexes for assisting visually impaired individuals under limited computational resources, confirming its efficiency and practical applicability in various applied domains

ABSTRACT. In the field of personalized footwear and or thopedic sole manufacturing, acquiring accurate lower limb measurements is essential. While numerous image-based tech niques—such as OpenCV with a reference object—have been explored for capturing foot dimensions, these methods often suffer from limitations including occlusion, perspective dis tortion, and missing or inaccurate parameters. To address these challenges, this study introduces an optimized predictive pipeline using the Gradient Boost algorithm to estimate foot length and horizontal breadth with accuracy. Unlike traditional predictive models that rely on generalized features such as BMI, gender, height, or weight, our model leverages refined anthropometric parameters from the ANSUR II dataset to achieve higher accuracy. The predicted values are used to generate a 2D outline of the sole, which is then converted into G-code compatible with CNC machines for contactless, precision-cut insole manufacturing. This approach significantly streamlines the production of custom orthopedic soles, offering scalable, non-invasive, and clinically relevant solutions for individualized foot care. Clinical relevance—Accurate determination of foot dimen sions is critical to diagnosing and treating flat feet, diabetic foot ulcers, plantar fasciitis, and limb-length disparities. Traditional casting and manual mensuration methods are slow, error-prone, and relying on clinical judgment and skill, and commercial image-based solutions are commonly marred by imprecision due to occlusions and perspective distortions. The new Gradient Boost predictive pipeline offers a non-invasive, rapid, and accurate alternative by predicting horizontal breadth and foot length from anthropometric inputs. With automatic generation of CNC-compatible G-code from predicted mensuration, the system enables rapid production of customized orthopedic soles and insoles with minimal manual intervention. This elevates patient comfort, reduces risk of pressure-related injuries and damage, and enhances mobility outcomes especially among diabetic, geriatric populations and post-surgical individuals. The solution is interoperable with modern telemedicine and digital orthotics workflows and enables scalable deployment within rehab units, home-based environments, and hospital settings.

ABSTRACT. —Parkinson’s Disease (PD) presents unique rehabil itation challenges due to fluctuating motor symptoms, freezing of gait (FOG), and variability in medication responses. Current rehabilitation tools, such as canes and external cueing systems, provide limited real-time monitoring and adaptive support. This paper presents a novel wearable device designed to enhance gait rehabilitation in PD patients by integrating real-time gait analysis, visual cueing, and vibratory actuation. The device incorporates an Inertial Measurement Unit (IMU) to track lower limb orientation and gait patterns, an Electromyography (EMG) sensor to monitor muscle fatigue and activation phases, and multi-point pressure sensors to map foot placement and force distribution. Unlike traditional assistive devices, this wearable solution provides vertical and horizontal visual cues for step alignment and deploys micro vibrations to mitigate FOG episodes. Additionally, a Range of Motion (ROM)-adjustable knee brace offers controlled mobility support, reducing muscle rigidity and preventing falls. The device’s continuous monitoring capabilities facilitate personal ized rehabilitation by identifying gait irregularities, shuffling tendencies, and fatigue-induced variations. By addressing the limitations of conventional rehabilitation tools, this device offers a comprehensive, patient-centric approach to PD mobility management. Preliminary evaluations indicate its potential to improve step initiation, gait stability, and rehabilitation adherence. Future work will explore machine learning-based adaptation for personalized therapeutic interventions. This wearable system represents a significant advancement in Parkinson’s rehabilitation by providing real-time, data-driven insights for clinicians and enhancing mobility for patients

ABSTRACT. With each new CMOS technology node, interconnect reliability rather than transistor performance increasingly defines the limits of System-on-Chip (SoC) architectures. Among the most critical challenges, crosstalk has evolved into a primary threat to timing integrity and functional correctness due to shrinking wire spacing, faster transition edges, and reduced supply voltages. Conventional mitigation techniques—including wire spacing, shielding, differential signaling, and coding—have consistently proven insufficient in modern dense designs where routing congestion, strict timing margins, and low-power constraints dominate. As a result, industry practice relies heavily on output-driver upsizing, a universal countermeasure that suppresses crosstalk by lowering the effective impedance of victim interconnects and attenuating injected disturbances. However, this brute-force approach introduces prohibitive penalties in both static and dynamic power consumption, aggravating IR-drop, thermal density, and simultaneous switching noise. The present work addresses these trade-offs by proposing a conditional reinforcement mechanism that selectively strengthens victim drivers only during aggressor transitions. This targeted activation achieves robust crosstalk suppression while minimizing power overhead, thus providing a scalable and energy-efficient solution that is well aligned with the requirements of advanced nanoscale SoC implementations.

ABSTRACT. This work proposes an algorithm for converting Finite State Machine diagrams from specification documents into Verilog HDL code. The design utilizes multiple YOLOv3 models for object detection, the Hough Transform for straight arrow detection, and k-nearest neighbours for number detection. The proposed methodology uses object recognition techniques to identify states/circles, state names, edges (arrows - straight and curved, arrowheads), source and destination states of an edge, input/output numbers on edges, etc. The validation test sets are machine printed images. Validation results show 98% accuracy and runtime under 30 seconds. The final accuracy represents the quality of the generated RTL.

ABSTRACT. In recent decades, multimedia compression has become a crucial area of research within the industrial sector, aiming to enhance compression methods. This field often faces a compromise between file size and quality, as the increase in image, audio, and video data significantly surpasses the advancements in compression rates achieved to date. Our research investigates the feasibility of employing neural networks for video compression, leveraging multilayer neural networks to provide a more effective solution. This paper introduces a lossy compression framework that capitalizes on the benefits of convolutional autoencoders (CAEs) to supersede traditional transformation techniques. Experimental findings indicate that our approach surpasses conventional coding methods by delivering improved compression ratios compared to existing studies.

ABSTRACT. Field-programmable gate arrays (FPGAs) are increasingly used in both academic and industrial research to evaluate new designs and systems due to their ability to deliver realistic and reliable power and performance metrics. They are especially valuable for validating simulators and calibrating models. Among these metrics, power consumption is critical—particularly for low-power designs—but measuring it accurately outside of simulation environments presents challenges in terms of cost and complexity.

To address this, FPGA boards often rely on built-in sensors and monitoring systems. However, such tools are not universally available across all platforms. In this paper, We revisit power measurement using a simple shunt resistor method and demonstrate its effectiveness across FPGA platforms. Using this setup, we analyze how different workloads and memory hierarchies influence both memory and system power. Our study highlights trade-offs often missed in DRAM power models, such as the cache power overhead that can outweigh memory savings, and the shifting efficiency of on-chip memory under contention. Together, these results provide practical guidance for power-aware FPGA evaluation and model calibration.

ABSTRACT. Physics-Informed Machine Learning (PIML) represents a transformative approach that bridges the gap between data-driven methods and traditional physics-based modeling. Unlike conventional machine learning algorithms, which demand extensive datasets and risk generating non-physical predictions, PIML embeds governing equations, conservation principles, and domain-specific constraints directly into the training process. By doing so, it ensures that model outputs remain consistent with established physical laws while reducing dependence on exhaustive simulation or experimental data. In high-performance electronic design, and particularly in the optimization of die-to-die (D2D) interconnects, PIML provides a powerful framework for addressing the dual challenges of power efficiency and signal integrity. By incorporating reduced-order models derived from Maxwell’s equations, RLGC transmission line parameters, and circuit-level telegrapher equations, PIML learns compact surrogate models capable of rapidly exploring vast design spaces. These surrogates not only preserve physical accuracy but also enable real-time predictions of key performance metrics such as driver and receiver power consumption, eye opening, jitter, and edge rates. Comparative studies indicate that PIML consistently outperforms conventional techniques such as Bayesian Optimization, delivering up to 5–6% reductions in power consumption alongside measurable improvements in timing stability and noise margins. The adaptability of PIML extends beyond interconnect design, offering scalability to diverse domains where physics-based constraints are critical, such as materials science, fluid dynamics, and structural optimization. Its hybrid architecture—integrating machine learning’s predictive strength with the rigor of physics—positions PIML as a cornerstone methodology for next-generation design automation. Ultimately, PIML accelerates the convergence of design cycles, enhances energy efficiency, and provides a physically consistent and computationally efficient pathway toward more reliable and sustainable integrated systems.

ABSTRACT. The increasing demand for high power and high frequency microwave integrated circuits (MICs) has confirmed Gallium Nitride (GaN) high electron mobility transistor (HEMT) devices as a key technology. However, accurate modeling of these transistors remains a main challenge due to their strong nonlinearities, trapping effects, and the presence of significant parasitic elements. Existing approaches often provide either static large-signal descriptions or small-signal equivalent models, but each alone is insufficient for reliable circuit design over a wide operating range. In this work, we propose an improved equivalent circuit model for enhancement-mode GaN HEMTs that combines both static and dynamic parameters. The model explicitly separates intrinsic and extrinsic elements, incorporates nonlinear capacitances, and accounts for parasitic resistances and inductances relevant at microwave frequencies. Compared to conventional models, the results show that the proposed GaN HEMT model keeps the I_DS prediction within about 5% of the reference values.

ABSTRACT. Traditional public key infrastructures (PKIs) struggle in IoT environments due to centralized certificate authorities and excessive computational demands. While recent fog-assisted approaches using DNS-Based Authentication of Named Entities (DANE) and DANCE have shown promise, single-cluster architectures face scalability limitations, handling only up to 10,000 devices with no mechanism for load distribution across multiple clusters. This paper presents a federated multi-cluster fog architecture extending DANE/DANCE-based key management for massive-scale IoT deployments. Our four-layer architecture introduces intelligent multiplexer gateways for request routing, coordinated fog node clusters for parallel processing, master nodes for cluster management, and an inter-cluster federation protocol enabling dynamic load balancing across geographically distributed clusters. Experimental evaluation using Raspberry Pi clusters and ESP32 devices demonstrates that our federated architecture supports over 50,000 concurrent devices (5x improvement), achieves 85% cache hit rates through intelligent distributed caching, and adds only 35-45ms federation overhead while preventing request dropping during overload conditions.

ABSTRACT. Duty cycle correction (DCC) plays a critical role in ensuring the reliable and stable operation of high-performance electronic and communication systems. By addressing signal imbalances and maintaining accurate timing relationships between clock edges, DCC techniques support consistent circuit behavior, improved data integrity, and reduced timing-related errors. These capabilities are especially essential as systems continue to scale in speed and complexity to meet the demands of modern digital technologies. A high-speed duty cycle correction approach has been proposed that utilizes capacitors and transmission gates to achieve accurate signal balancing. This method is capable of operating effectively at frequencies up to 8 GHz, demonstrating both high performance and robustness under demanding conditions. Post-layout simulation results confirm that the circuit can accurately correct input clock duty cycles ranging from 25% to 75%, restoring the output to a near-ideal 50% duty cycle. The correction is achieved with a very small error margin - less than ±1.2% - which is well within acceptable limits for high-speed applications. The versatility and precision of the proposed technique make it highly suitable for integration into input/output (I/O) circuits across a range of widely adopted high-speed interface standards. These include Double Data Rate (DDR) memory interfaces, Universal Chiplet Interconnect Express (UCIe), Universal Serial Bus (USB), and Peripheral Component Interconnect (PCI) systems. With its wide correction range, low error rate, and compatibility with high-frequency operation, the proposed DCC circuit presents a reliable solution for advanced electronic design challenges.

ABSTRACT. The reliability of integrated circuits (ICs) is a critical factor affecting their performance and longevity. Traditional physical design methodologies have struggled to keep pace with the increasing complexity and miniaturization of modern ICs. This paper presents a novel approach to enhancing IC reliability through the integration of machine learning (ML) techniques in physical design. By leveraging ML algorithms, the proposed method optimizes design parameters and identifies potential reliability issues early in the design process. It is demonstrated how ML models can predict failure points and suggest design modifications that significantly improve reliability while maintaining performance and power efficiency. This paper offers a comprehensive framework for implementing ML-driven reliability improvements in IC design and discusses potential future developments in this evolving field.

ABSTRACT. The rapid increase in integrated circuit (IC) component density has led to a significant rise in design complexity. To address this challenge within constrained timelines, various methodologies have been developed, one of them is the digital design flow. During the physical implementation phase, pin placement is one of the earliest but most important stages. Incorrect pin placement can make timing issues unfixable or design not routable, because of this the time to market time will be increased. To solve these issues the Pin Builder flow has been developed which is fully automated flow. With the use of these flows the runtime can be decreased by ~14%, Total negative slack can be decreased by ~5% but total Design rule violation count can be increased by ~17%.