This thesis addresses the critical challenge of real-time spike extraction from high- throughput voltage imaging data, often overwhelming existing analysis pipelines de- spite advancements in genetically encoded voltage indicators and imaging hardware. A novel hybrid processing pipeline is presented, integrating the strengths of state-of- the-art systems, designed for unified offline and online analysis with enhanced mod- ularity and reproducibility. Through architectural optimizations, including streaming- compatible design, algorithmic enhancements like GPU-accelerated filtering, and a robust calibration framework, processing speed and signal fidelity was significantly improved. Our evaluation demonstrates real-time throughput of 500 frames per sec- ond at 680×680 with 8 ms latency and up to 1000 frames per second at 480×480 with 16 ms latency using a Nvidia Tesla V100, notably reducing startup times and improv- ing deployability via a containerized environment. While revealing motion-correction as a persistent bottleneck and the inherent latency-throughput trade-off, this work pro- vides a scalable, accurate, and user-friendly solution that bridges the gap between fast data acquisition and real-time analytical capability, paving the way for next-generation closed-loop neuroscience experiments.

Neuromodulation is a promising treatment methodology for disorders of the brain. However, current clinical applications are primarily focused on a rudimentary form of neuromodulation, exclusively targeting individual regions in the brain. This has led to limitations in effective treatment options for patients suffering from neurological and neuropsychiatric diseases, with brain stimulation yielding a success rate of only 50% in 50% of patients after five years. In recent years, it has been strongly theorized that most brain disorders are emergent properties of pathological network connectivity rather than dysfunction of isolated brain regions. Targeting brain connectivity via neuromodulation is, therefore, a promising approach for improving treatment efficacy. This study aims at bridging this therapeutic gap by advancing treatment strategies towards network-based neuromodulation. To achieve this goal, a computational framework was designed and implemented that analyzes brain activity to identify biomarkers of abnormal connectivity networks. Furthermore, software pipelines for the construction of patient-specific,
virtual brain models were designed and implemented which enable analysis of simulated neural activity and connectivity networks, and enable in silico neuromodulation. The framework is tested and validated using an EEG dataset consisting of neural-activity recordings belonging to patients diagnosed with neuropathic pain and a separate healthy control group. Our computational framework shows high ability in identifying disease-specific abnormalities in network connectivity and in constructing pathological network hubs, with accuracies up to 98.6% in explaining neuropathic pain. Patient-specific,
brain-model construction is also able to accurately represent empirical patterns of connectivity in simulation. Furthermore, the simulated neuromodulation with personalized virtual brains demonstrates the ability to accurately model dynamic modulation of brain connectivity networks. These findings illustrate the effectiveness of our overall framework and highlight the potential for integration into clinical decision-making and closed-loop implantable neuromodulation systems, enabling more effective treatment of brain disorders.

The exponential growth in the scale and complexity of neural recording systems demands increasingly efficient computational frameworks to manage the vast data generated by modern multi-electrode ar- rays. This thesis explores the feasibility of integrating Graphics Processing Unit (GPU) acceleration with the Open Ephys ONIX system to enhance its support for next-generation neural interfaces and compu- tational algorithms. ONIX, a modular and open-source electrophysiology platform, currently relies on CPU-based computations that may struggle to meet the rising demands of closed-loop experiments. This work investigates PCIe peer-to-peer (P2P) communication between ONIX and GPUs to by- pass CPU involvement, aiming to reduce latency and increase throughput. A secondary data channel was implemented on the ONIX FPGA, accompanied by a custom kernel driver integrating GPUDirect with the existing RIFFA driver. Performance evaluations compared unpinned and pinned GPUDirect transfers to traditional CPU-mediated methods. Pinned transfers reduced transfer times by up to 30% for small data and 14% for larger data sizes, though significant variance in transfer time was observed. Application-level benchmarks revealed that GPU acceleration provides advantages for larger data sizes and higher computational loads but the effectiveness of GPUDirect is limited by ONIX’s PCIe through- put. The findings highlight that while GPUDirect offers theoretical benefits, its practical impact is con- strained by hardware limitations, including the PCIe bandwidth of the ONIX FPGA. This thesis out- lines necessary hardware improvements, such as higher-speed SerDes modules and advanced FPGA cards, to fully realize the potential of GPU acceleration in electrophysiology systems. A decision tree framework is proposed to guide researchers in determining scenarios where GPU and GPUDirect in- tegration is beneficial. While this study demonstrates the feasibility of GPU integration with ONIX, GPUDirect integration is currently not a beneficial technology for the ONIX ecosystem.

The human brain is arguably the most complex system we know of. For centuries, understanding the brain and the way it works has been of great concern to scientists. In 1952, the first mathematical model of a neuron was developed by Hodgkin and Huxley. This work pioneered the field of computational neuroscience, which studies the organization of the brain and the way it processes information. Modeling and simulation of neurons and small brain regions at the cellular level offer great insights into the details of brain mechanics, however, they are far away from clinical applications.

The Virtual Brain (TVB) is a simulation framework that takes a different approach to brain modeling. TVB reduces complexity at the micro level to achieve the macro level of brain modeling and simulation. This approach in brain modeling is referred to as Large-scale Brain Network or Neural-Mass Modeling. Additionally, TVB incorporates individual brain-imaging data with the models to achieve personalized, patient-specific digital brain twins. The digital brain twins can be used in clinical settings such as neurosurgeries and real-time brain-interface systems. However, building and utilizing such large-scale brain models requires high performance and low latency in terms of computation.

Today, heterogeneous architectures, due to their enhanced performance, energy efficiency, and better flexibility are being utilized increasingly in computing systems ranging from low-power edge devices to high-performance cloud infrastructures. Versal Adaptive Compute Acceleration Platform (ACAP) is a high-performance, heterogeneous computing platform developed by Xilinx/AMD. Versal ACAP offers an array of vector processors, specialized in artificial intelligence (AI) and signal processing workloads, next to the traditional programmable logic with extensive memory resources. The large-scale neural-mass simulation problem at hand has heterogeneous computational needs, which makes it interesting to see how it maps to the Versal ACAP. In this work, we explore the benefits and challenges of using heterogeneous systems (specifically, the Versal ACAP) in the high-performance workload of large-scale neural-mass simulation.

We designed and implemented a dataflow-style system for large-scale brain network simulation on the Versal ACAP. Two different versions of this system, namely AIE-Only and Heterogeneous, were developed in order to explore the capabilities of the Versal ACAP when accelerating our application. Compared to the current GPU version of TVB, the Heterogeneous implementation performed on average around 4× slower in terms of throughput. However, the Versal ACAP Heterogeneous implementation delivered around 13× lower latency, two orders of magnitude better energy efficiency, and around 2× better power efficiency compared to the GPU version of TVB. Although the Heterogeneous implementation falls short in terms of throughput compared to the GPU version, its lower latency and energy efficiency make it suitable for real-time applications that might use large-scale neural-mass modeling such as virtual surgery or brain interface devices.

This thesis aims to evaluate the viability of using a Field Programmable Gate Array (FPGA) in a neural Implantable Medical Device (IMD). The primary motivation for incorporating FPGAs is their potential to support future functionalities, such as running neural networks for medical condition analysis but also advanced cybersecurity algorithms. These algorithms are compute-intensive, and accelerators like FPGAs offer advantages in terms of speed and efficiency. To assess the effectiveness of such a device, state-of-the-art Microcontroller Units (MCUs) commonly used in similar applications are employed as a reference. Comparisons are made between MCU-only platforms and hybrid platforms integrating both an MCU and an FPGA. Feasibility analysis considers operational modes and use cases based on various realistic scenarios. The results show mixed outcomes across scenarios. Under a 100% duty cycle, the FPGA demonstrates higher efficiency, consuming less active power than the MCU. However, at lower duty cycles, MCUs are generally more effective on average. The use of an FPGA becomes practical when power-gating techniques are applied to minimize power consumption during inactive periods.

Functional ultrasound is by now a well-established technique in the neuroscience community to measure brain activity. However, the transcranial application of functional ultrasound on humans, with the exception of the acoustic windows in the skull, remains a huge challenge because of the properties of the cranial bone the acoustic waves will reflect, refract, attenuate or will cause aberration. Therefore, the signal-to-noise ratio (SNR) of the incoming signal is too low for transcranial functional ultrasound (TCfUS). This thesis tries to overcome the SNR problem of TCfUS with the design of a 64-channel ultrasound acquisition system that is focused on achieving the highest SNR possible. This is achieved by placing the analog front-end (AFE) chips as near to the transducer elements as possible and by oversampling the incoming signal with a factor of 25, a theoretical increase of 15 dB of the signal-to-quantization-noise ratio over the current state-of-the-art is estimated. The proposed design is a receive-only system where the transmission of the ultrasound pulses is carried out by a separate ultrasound system. The design splits the acquisition system into a front-end and a back-end subsystem, where the front-end system is implemented using four AFE58JD48 analog front-end chips from Texas Instruments. From here the data samples are transported over fiber optics to the back-end subsystem, which consists of a VCK190 FPGA board from Xilinx, where the samples are processed and/or transported to a workstation for storage. Because the system organization differentiates from more conventional research ultrasound systems, a trade-off is introduced between processing and throughput, resulting in three processing configurations for the FPGA with each a different focus: raw RF data sampling, real-time processing, and hardware processing. Due to time and resource constraints, no measurements and results are available on the SNR and decimation. However, a theoretical exploration has been done on the expandability of the number of channels which was found to be 192 channels for the VCK190.

Medical Body Area Networks (MBANs) are a cluster of possibly heterogeneous devices, communicating with each other in, on or around the human body. Through these devices, medical data is collected, processed in some way and transferred outside of the network. The IEEE 802.15.6 standard aims to govern communications between such devices. It includes a set of constraints for physical features and communication on the PHY and MAC level, as well as association/ disassociation protocols and security services that applications need to comply with. Given the high sensitivity of the medical data transmitted via MBANs, network security is crucial. This thesis consists of three main contributions: (a) a structured procedure to analyse the security features of the IEEE 802.15.6 standard by using realistic hypothetical scenarios is introduced (b) a thorough security analysis of the standard is conducted (c) recommendations on how to improve the security posture of the standard are given.

Epilepsy is a medical condition which is caused by excessive or synchronous neuronal activity of the brain cells. These activities can lead to attacks where the patient can lose conciseness or experiences random muscle cramps at seemingly any point in time. Using implantable on body sensors these seizure attacks could be detected and even prevented. These sensors would form a Medical Body Area Network (MBAN) which interconnects all of the sensors. This project looks at a proof of concept implementation of such an MBAN and focuses on a secure connection between an implant and a gateway device, which is a mobile phone. The implant and mobile phone will communicate with each other using Bluetooth Low Energy (BLE). This form of communication does not provide a secure pairing method for devices that lack in- and output capabilities, such as an implant. To set up a secure connection the data will be encrypted with an encryption key, which has to be shared between the implant and mobile phone. In order to do this in a secure way, an Out Of Band (OOB) channel will be used to pair the two devices. This thesis looks at three different OOB channels, Near Field Communication (NFC), ultrasound and galvanic coupling and compares them in therms of security, health safety, data rate, power consumption and feasibility.

Whisker tracking in rodents is an ongoing research in neuroscience. Neuroscientists have recorded experiments with high-speed cameras in which untrimmed, head-restrained mice are provided with air stimuli. These videos required the development of algorithms to reliably track whisker movement. Recently, a Whisker Tracking System, WhiskEras, emerged, which is able to detect and track whiskers over the course of such videos in a more accurate way than pre-existing methods. WhiskEras is slow, processing less than 1 frame per second. Additionally, it involves a great number of parameters which need tuning for different experimental setups and recording settings which were used for this experiment. This thesis addresses these two problems. First, the algorithm was examined and its shortcomings were exposed. A more accurate whisker point detection algorithm was suggested and implemented, among a range of alternative solutions which were studied. Furthermore, its Stitching stage was modified to replace a range of hard-to-tune parameters with more robust ones. Finally, the improved WhiskEras was accelerated by porting the MATLAB code to C++ and using advanced parallelization techniques with CUDA and OpenMP to achieve a speedup of 74.96. Overall, the improvements yielded better tracking results in our benchmarks, while the parameters were much easier to tune and remained constant under different video setups of whisking experiments.

Absence seizures have a real-life impact on epileptic subjects, as day-to-day tasks can by suddenly interrupted making for dangerous situations. Though a lot of work has been done on seizure detection, to limit the impact on epileptic patients, the true necessity lies in timely prediction of seizures before they manifest. Various attempts have been made using conventional algorithms to accurately predict seizures however, so far, results are not that encouraging. In this work, we applied various machine-learning algorithms, in an attempt to identify complex, multi-dimensional epileptic precursors in brain recordings. Three types of neural networks are used in this feasibility study, namely Multi-Layer Perceptron (MLP) networks, Long Short-Term Memory (LSTM) networks and Gated Recurrent Unit (GRU) networks. The used input data was annotated Electrocorticography (ECoG) data, recorded in living mutant rodents, containing epileptic events at an interval of about one minute. The data was pre-processed for better learning performance by data normalisation and by generating distinctive training features. The neural networks were configured as three-class classifiers, distinguishing among inter-ictal, pre-ictal and ictal periods. A grid-search approach was applied to determine the best set of parameters for the neural networks. Despite our best efforts, the relation between the input data and output data could not be learned in a reliable way. The maximum reached Average Prediction Rate (APR) was 0.57 with a prediction time of 3.1s when using the normalised data as input and 0.65 with a prediction time of 6.1s when using the distinctive features as input. These results essentially signify good detection but virtually no prediction of upcoming seizures. The evaluation of the experimental findings has revealed that the employed ECoG recordings were ill-selected for training our various neural-network models. Also, a non-conclusive exploratory experiment is performed by applying a Weibull Time-To-Event Recurrent Neural Network (WTTE-RNN) on a sub-set of the normalised input data. The experiment has yielded some positive results, a short-notice prediction of the upcoming seizure in some cases, encouraging for further exploration of this approach. Despite the limited success of this work, however, through its extended forensics analysis, it has paved the crucial, initial steps in the direction of seizure prediction.

In the field of computational neuroscience, complex mathematical models are used to replicate brain behavior with the goal of understanding the biological processes involved. The simulation of such models are computationally expensive and therefore, in recent years, high-performance computing systems have been identified as a possible solution to accelerate their execution. However, most of those implementations are model-specific and thus non-reusable for other modeling efforts, requiring a completely new development effort per model used. The challenge lies in offering high-performance and scalable libraries (so as to support the construction and simulation of large-scale brain models) while at the same time offering high degrees of modeling flexibility and parameterization. This thesis presents flexHH, a scalable hardware library implementing five accelerated and highly parameterizable instances of the Hodgkin-Huxley neuron model, one of the most widely used biophysically-meaningful neuron representations. As a result, the user is able to instantiate custom models using flexHH and immediately take advantage of the acceleration without the mediation of the engineer. The five flexHH implementations target the Maxeler Data-Flow Engine(DFE), an FPGA-based acceleration solution, and incrementally support a number of features such as custom ion channels, multiple cell compartments and inter-neuron gap-junction connectivity. Furthermore, for each of the five implementations it is possible to select either the forward-Euler, second, or third-order Runge-Kutta numerical method. A speedup between 14x-36x has been achieved compared to a sequential C implementation, when run on a 2.5-GHz Intel Core-i7 CPU, while no practical performance drop is observed when compared to a hard-coded version of a DFE, an Intel Xeon-Phi CPU, and a NVidia Titan X GPU. In this thesis, flexHH kernels are rigorously validated, an evaluation of the influence of the numerical methods is done, and a comprehensive resources usage, performance, and power-consumption evaluation of the various DFE implementations is presented.

The increasing transistor density of Integrated Circuits (ICs) ever since their introduction, has scaled the computational performance of microprocessors. As a consequence of the gain in transistor density, the power dissipation density has also increased to a degree that has become a limiting factor in further performance scaling. The prevalent control flow computing paradigm is, especially in the context of High Performance Computing (HPC), faced with this power crisis. Data flow computing can achieve a better computations per unit power ratio and might therefore be the solution to achieving exa-scale computing, enabling progress in science and technology.
This work contributes in investigating methods for increasing the abstraction
level of data flow programming by using a Polyhedral Process Network (PPN) as abstraction to facilitate automated analysis and transformation of data flow applications. A Polyhedral Process Network (PPN) is a process network based on Polyhedral Compilation (PC) a mathematical framework for code analysis and optimisations.
A tool has been implemented to translate a PPN to a data flow implementation. The Maxeler Data Flow Engine (DFE) technology, a data flow hardware platform supported by a tool-flow, is used as target technology. For a number of specific challenges involved in the transformation from PPN to DFE-implementation a method has been implemented. This has shown that a PPN as well as the capabilities of PC are useful concepts that are applicable to further increase of the abstraction-level of data flow computing. However the transformation from PPN to DFE-implementation is complex. Therefore this work concludes that the development of a data flow tailored Polyhedral Compilation (PC) abstraction is the most promising future direction.

The movement of whiskers in head-fixed mice is of high interest for neurological research, as it allows scientists to learn more about learning processes during active touch. However, manual tracking of whiskers in thousands of frames is not feasible, and reliable tracking of individual whiskers is not possible with the best current software available. In this work, a new system for the automatic tracking of whiskers in top-view videos of untrimmed, head-fixed mice is presented. The design of this system is based on a variety of image processing and computer vision algorithms. For each step, several alternatives have been considered and assessed. The best options have been combined and implemented in MATLAB. This new system, the Full-Whisker Tracking System (FWTS), detects the centerline of whiskers on each frame along their whole length with sub-pixel accuracy. The whiskers are defined by their angle, position, shape and length. FWTS uses these features to track and recognize whiskers over multiple frames in video segments with a frame rate of 1000 Hz. The system was tested on representative, challenging video segments from two experiments using different mice, and showed itself superior to the pre-existing system BWTT, by detecting on average 21.2% more whiskers, and with twice as much precision. In contrast to earlier systems, FWTS has been demonstrated reliably track individual whiskers in untrimmed mice. Furthermore, FWTS's processing time is on average 41.7% lower than that of BWTT. This thesis work shows that it is possible to track whiskers in untrimmed mice, even in the face of partial occlusions, and makes it possible to study the functioning of the intact whisker system of mice in unprecedented detail.

Pavlovian eyeblink conditioning is a powerful experiment used in the field of neuroscience to measure multiple aspects of how we learn in our daily life. To track the movement of the eyelid during an experiment, researchers traditionally made use of potentiometers or electromyography (EMG). More recently, the use of computer vision and image processing alleviated the need for these techniques, but currently employed methods require human intervention and are not fast enough to enable real-time processing. We selected a combination of face and landmark-detection algorithms in order to fully automate eyelid tracking, and accelerated them to make the first step towards an on-line implementation. Various different algorithms for face detection and landmark detection (eyelid detection) are analyzed and evaluated. Based on this analysis, two algorithms are identified as most suitable for our use case: the Histogram of Oriented Gradients (HOG) algorithm for face detection and Ensemble of Regression Trees (ERT) algorithm for landmark detection. These two algorithms are accelerated on GPU and CPU, achieving speedups of 1753× and 11.49×, respectively. A combination of these algorithms is successfully implemented for a real neuroscientific use-case: eyeblink response detection, achieving an overall application runtime of 0.533 ms per frame, which is 1067× faster than the sequential implementation. Furthermore, this accelerated implementation was used to generate a database of 1440 eyeblink responses during the conditioning experiment. We made use of multiple machine-learning techniques to analyze this database and concluded that there is a correlation between the asynchronicity of the eyelids and the eyeblink-response performance during the conditioning experiment.