Detecting microsleep in real time is crucial to facilitating the transition from semi-autonomous systems to completely autonomous driving technologies. Integrating sophisticated detection algorithms with vehicle control systems enables the provision of prompt corrective measures, such as driver alerts or temporary vehicle control. Minimizing the probability of accidents caused by driver fatigue not only improves the safety of users, but also promotes the overall security of the road network. By integrating Kalman filtering, Voronoi tessellation, and adaptive coverage control algorithms, the study seeks to find a feasible and sophisticated methodology to enhance the spatial and temporal resolution of EEG data analysis, leading to more robust and reliable detection of microsleep episodes. The paper presents a framework as a significant advancement in sleep technology, offering a new method to diagnose and understand microsleeps, characteristics, and patterns of brain activity and sleep disorders.

The fuzzy-possibilistic product partition c-means (FPPPCM) algorithm was proposed as a robust solution to the c-means clustering problem, in which outlier data behave similarly to distant objects in gravity systems. Although FPP-PCM reliably provides fine partitions when its parameters are well chosen, things can be difficult when it is not initialized properly. To avoid such cases, this paper proposes a self-tuning version of the FPPPCM algorithm, which incorporates some cluster size controlling variables into the objective function that allow for the adjustment of the so-called possibilistic penalty terms during the alternative optimization process. The proposed method was evaluated using four standard test datasets in three different scenarios: (1) no added noise; (2) a single outlier added; (3) multiple noisy items added. The partitions provided by the proposed algorithm were evaluated based on cluster purity, normalized mutual information and adjusted Rand index, and was compared with the outcome of previous clustering models. The proposed method performed better or at least at the same quality level at previous ones, while reducing the number of parameters the user is responsible for.

sRPE and ACWR are valuable tools for controlling fatigue levels and minimizing injuries in performance sports. Their ability to provide individualized assessment, integrate subjective and objective measures, and inform data-driven decision-making makes them essential components of a comprehensive sports safety and performance monitoring system. The goal of this study was to provide first a computer-assisted solution to predict the fatigue level, and then to expand this with an artificial intelligence-supported solution for a more advanced pipeline in performance sports, to minimise the injury level.

Automated brain tumor classification is an intensively investigated problem, which recently attracted significant attention. Convolutional neural networks (CNN) and deep learning represent the standard for the foundation of any recent solution. This paper proposes two simplified VGG architectures and investigates their capabilities and limitations, in comparison with state-of-the-art CNN networks deployed via transfer learning. Various parameter settings are involved in the evaluation process, including different kernel sizes, dropout rules, loss functions, etc. Networks are trained and tested on a public brain tumor classification data set consisting of 3064 images and three tumor classes (meningioma, glioma and pituitary tumor). The thorough evaluation process revealed that the proposed CNN models can achieve competitive performances with regard to state-of-the-art methods in several scenarios. The best achieved accuracy benchmarks are 98.2% overall Dice similarity score and correct decision rate, and AUC values over 99.6% for each of the three tumor classes.

Reinforcement learning (RL) has shown promise in controlling blood glucose levels in a personalized way in type 1 diabetic patients. In this study, we investigate the impact of different activation functions and layer numbers on RL performance in blood glucose control. We train RL agents with various combinations of activation functions and layer numbers on a virtual patient model. The RL agents are evaluated based on their ability to maintain blood glucose levels within a target range while minimizing the frequency and magnitude of hypoglycemia and hyperglycemia events. Our results show that the choice of activation function and layer number significantly affects the RL performance. Specifically, the agents with ReLU activation functions and two or three hidden layers outperform the other agents, achieving a higher percentage of time in the target range and fewer hypoglycemia and hyperglycemia events. These findings provide valuable insights for the development of RL-based blood glucose control systems in type 1 diabetic patients.

Brain tumor segmentation has been a widely researched topic for decades, and it intensified ten years ago as a consequence of the Brain Tumor Segmentation Challenges (BraTS), which provided and yearly updated a standard multi-spectral brain tumor MRI data set and a unified evaluation framework to the research community. This paper proposes a procedure for brain tumor segmentation, which uses a spatial histogram enhancement method to preprocess the data, and two identical cascaded U-net networks that work with 3D convolution. The first U-net accomplishes an intermediary segmentation of the brain volume, while the second one reevaluates the labels given to pixels based on the labels of neighbor pixels. The output of both U-nets are evaluated using statistical accuracy benchmarks. The proposed procedure achieved an average Dice score of 88.8% on the high-grade glioma records of the BraTS 2019 training data set. Post-processing increased the average Dice score by 1.1%, but in case of typical small high-grade tumor lesion it can achieve an improvement of up to 5%.

One of the most challenging area of diabetes research is to provide such automated insulin delivery systems – so called artificial pancreas systems – that have robust and adaptive capabilities in a highly sophysticated way. I.e. they are able to provide robust insulin delivery actions at the beginning of the therapy to satisfy the requirements of the patients without knowing the users daily lifestyle and preferences however adaptive on the short-term to learn these patient specifics to increase the quality of the therapy. One possible solution is the closed-loop systems that have self-learning features. In the present study, we have examined a glucose regulatory problem using direct reinforcement learning based controller. The approach represents the fully automatic insulin administration as the timepoint and the carbohydrate content of the meals were unknown and randomized. We constructed a virtual environment of the patient with type 1 diabetes by applying a mathematical model. Proximal policy optimization learning model with continuous action space was used. Furthermore, we evaluated the effect of different training lengths on the test scenario.

The automatic segmentation of medical images represents a research domain of high interest. This paper proposes an automatic procedure for the detection and segmentation of gliomas from multi-spectral MRI data. The procedure is based on a machine learning approach: it uses ensembles of binary decision trees trained to distinguish pixels belonging to gliomas to those that represent normal tissues. The classification employs 100 computed features beside the four observed ones, including morphological, gradients and Gabor wavelet features. The output of the decision ensemble is fed to morphological and structural post-processing, which regularize the shape of the detected tumors and improve the segmentation quality. The proposed procedure was evaluated using the BraTS 2015 train data, both the high-grade (HG) and the low-grade (LG) glioma records. The highest overall Dice scores achieved were 86.5% for HG and 84.6% for LG glioma volumes.

The steadily growing amount of medical image data requires automatic segmentation algorithms and decision support, because at a certain time, there will not be enough human experts to establish the diagnosis for every patient. It would be a good question to establish whether this day has already arrived or not. Computerized screening and diagnosis of brain tumor is an intensively investigated domain, especially since the first Brain Tumor Segmentation Challenge (BraTS) organized seven years ago. Several ensemble learning solutions have been proposed lately to the brain tumor segmentation problem. This paper presents an evaluation framework designed to test the accuracy and efficiency of ensemble learning algorithms deployed for brain tumor segmentation using the BraTS 2016 train data set. Within this category of machine learning algorithms, random forest was found the most appropriate, both in terms of precision and runtime.

Offset compensation (OC) systems are indispensable parts of multi-Gbit/s optical receiver (RX) frontends. Effects of offset are addressed in this paper. The analytical expression for the highest lower-cut-off frequency of the OC with minimum impact on the sensitivity is found. Existing OC solutions are discussed. Then, a novel mixed-signal (MS) architecture is introduced which uses digital filtering of the signal, and current-digital-to-analog converters (IDACs) to compensate the static offset in the limiting amplifier (LA) and transimpedance amplifier (TIA), as well as continuously track and compensate the TIA offset. By using two feedback loops and a continuous tracking the presented solution offers more functionality than other existing MS architectures. Three RX implementations, with RC, switched-capacitor (S-C) and with the MS-OC architectures, in the same 28 nm bulk-CMOS are compared quantitatively with measurements. The presented MS design reaches a lower-cut-off frequency of under 9 kHz, a dynamic range of over 1 mA, 3. 2μA residual input offset-current and it is compensating the RX via two feedback loops. These are achieved using an area of only 1345 μm2, nearly half of RC-filter based architecture. Although the SC implementation needs less area, its residual offset is 8 times higher. Both conventional implementations have a higher high-pass characteristic of about 20 kHz and can compensate only the offset of the TIA. It is concluded, that the presented system offers a higher flexibility and functionality in implementation, as well as a very good compromise between area, precision and performance over the commonly used RC-filter and S-C filter based solutions.

The relaxation of the probabilistic constraint of the fuzzy c-means clustering model was proposed to provide robust algorithms that are insensitive to strong noise and outlier data. These goals were achieved by the possibilistic c-means (PCM) algorithm, but these advantages came together with a sensitivity to cluster prototype initialization. According to the original recommendations, the probabilistic fuzzy c-means (FCM) algorithm should be applied to establish the cluster initialization and possibilistic penalty terms for PCM. However, when FCM fails to provide valid cluster prototypes due to the presence of noise, PCM has no chance to recover and produce a fine partition. This paper proposes a two-stage c-means clustering algorithm to tackle with most problems enumerated above. In the first stage called initialization, FCM with two modifications is performed: (1) extra cluster added for noisy data; (2) extra variable and constraint added to handle clusters of various diameters. In the second stage, a modified PCM algorithm is carried out, which also contains the cluster width tuning mechanism based on which it adaptively updates the possibilistic penalty terms. The proposed algorithm has less parameters than PCM when the number of clusters is [Formula: see text]. Numerical evaluation involving synthetic and standard test data sets proved the advantages of the proposed clustering model.

As the global IP traffic and its demand for computation increase in a rapid and sustained manner, processor, server, and network architectures are also undergoing a considerable evolution. Two of the manifestations of this evolution are the integration of a large number of computing nodes in a single server and the interconnection of many servers via high-speed communication links. At present, however, the node-to-node communication bandwidth is one of the severest resource bottlenecks in massively parallelized applications. There is a concerted effort by the academia and the industry to achieve higher data rate by assembling multiple parallel links. This effort, however, is inherently limited by many constrains, including space. Optical interconnects, on the other hand, promise superior data rates, lower transmission losses, and less inter-channel crosstalk when compared to electrical interconnects. Development in this area promise data rates in the range of Tera bits per second per link and beyond. So far, however, little attention is given to the power adaptiveness of optical interconnects. In this paper, we present an optical interconnect concept which adjusts its power consumption in response to the change in the statistics of the incoming workload. The several components of the link have been designed and developed in hardware. Based on initial power and performance measurements of the components, a link model of our optical interconnect was created. The performance-power consumption characteristics of this model was simulated applying different workload statistics and the potential of the energy savings by the adaptivity have been evaluated. It is revealed that the power consumption of our optical interconnect reduces by up to 40% when its workload was exponentially distributed (signifying underutilisation) compared to a Weibull distribution workload (signifying full capacity workload). This study confirms the high potential for power saving in performance adaptive optical interconnects.

This paper describes a new, robust system-architecture for common-cathode (CC) vertical-cavity surface-emitting laser (VCSEL) drivers for highly-scaled CMOS technologies with low supply voltages. The concept implies converting the input signal into a current which is transferred to an amplifier built in a floating well by the level-shifter. Setting the potential of the well as high as the parasitic diode break-down voltage, a high DC bias voltage is possible for the VCSEL, several times higher than the gate-oxide break-down of CMOS technologies. The architecture is demonstrated with the design of a VCSEL driver in 80 nm CMOS with 1.2 V breakdown. The VCSEL DC bias can go as high as 4.5 V. The fabricated chip was bonded to a CC VCSEL. Electrical, optical and robustness measurements were performed. The optical eye was open until 17 Gbps at a bit-error-rate (BER) of 10-12with only 60 mW power consumption including the VCSEL current. The driver met the electrical robustness evaluation offering a more reliable alternative to stacked CC architecture. The active area is of only 0.003 mm2, one of the smallest existing VCSEL diode drivers for this data-rate.

In this paper we propose an efficient color reduction framework that employs c-means clustering to extract optimal colors. The processing consists of three stages: preprocessing, c-means clustering, and creation of the output image. The main goal of the first stage is to transform the pixel matrix into a list of records, which indicates what colors are present in the image and how many times they appear. To achieve this, first we apply a static color quantization scheme that aligns the 16.7 million possible colors with 140 thousand grid points, and build the histogram of this quantized image. Then we mark least frequent quantized colors to be ignored during the clustering stage, the amount of such marks being controlled by the pixel inclusion parameter. Leaving out 2-5% of the image pixels can reduce the number of colors to 500-5000 in most images. This limited set of colors together with frequency information consists the input of the c-means clustering process performed in the second stage. Before creating the final output image, the marked quantized colors are mapped to the closest cluster. Thorough numerical tests were performed on 500 randomly chosen images using both fuzzy and hard c-means clustering. Evaluations revealed that hard c-means is more suitable than fuzzy c-means for the given problem, both in terms of accuracy and efficiency. The proposed method performs quicker 2-3 times than other recent reported solutions.

One of the main challenges in the field of clustering is creating algorithms that are both accurate and robust. This paper introduces a novel fuzzy-possibilistic shell clustering model aiming at accurate detection of circles, spheres, and multidimensional spheroids in the presence of outlier data. The proposed fuzzy-possibilistic product partition c-spherical shell algorithm (FP3CSS) combines the probabilistic and possibilistic partitions in a qualitatively different way from previous, similar algorithms. The novel mixture partition is able to suppress the influence of extreme outlier data, which gives it net superiority in terms of robustness and accuracy, compared to previous algorithms.

Although all three conventional c-means clustering algorithms, namely hard c-means (HCM), fuzzy c-means (FCM), and possibilistic c-means (PCM), had their merits in the development of clustering theory, none of them are generally good solutions for unsupervised classification. Several hybrid solutions have been proposed to produce mixture algorithms. Possibilistic-fuzzy hybrids generally attempt to get rid of the FCM's sensitivity to outliers and PCM's coincident cluster prototypes, while hard-fuzzy mixtures usually aim at quicker convergence while preserving FCM's accurate partitions. This paper presents a unifying approach to c-means clustering: the novel clustering model is considered as a linear combination of the FCM, PCM, and HCM objective functions. The optimal solution is obtained via evolutionary computation. Our main goal is to reveal the properties of such mixtures and to formulate some rules that yield accurate partitions.

In order to improve the accuracy, robustness, and computational load of c-means clustering models, a series of hybrid solutions have been proposed. Mixtures of fuzzy (FCM) and possibilistic c-means (PCM) clustering generally attempted to avoid the noise sensitivity of the former and the coincident clusters of the latter. On the other hand, mixtures of fuzzy and hard c-means (HCM) have been proposed to speed up fuzzy clustering without losing the quality of its partitions. In this paper, a novel hybrid c-means algorithmic scheme is proposed that unifies the objective functions of all three conventional clustering models. The strength of each component within the mixture is controlled by two tradeoff parameters. The optimization of the proposed objective function is achieved using the alternating optimization derived from zero gradient conditions and Lagrange multipliers. The novel hybrid's behavior is evaluated in terms of classification accuracy, cluster validity and execution time, using the IRIS data set. Suitably chosen tradeoff parameters enable the proposed algorithm to achieve better accuracy than previous models, while performing less computations.