Charles-Antoine Brunet

Sensitivity in PET: Neural networks as an alternative to compton photons LOR analysis

In high-resolution small-animal positron emission tomography (PET), sensitivity remains an active issue. Sensitivity can be increased by lowering the energy threshold to include more Compton-scattered events, but then computation of the correct annihilation line-of-response (LOR) proves problematic. The complexity of Compton-kinematics analysis, compounded with finite energy resolution and detection position quantization of finite-size detectors, yields unaffordable methods with rather poor success rates. As an alternative, this paper proposes an artificial neural network (ANN) approach, which forfeits all explicit handling of equations at the expense of a priori statistical training, and which has the potential to better handle the previous measurement impairments. The method first consists in a preprocessing step involving geometrical transformations, which simplifies the actual use of the neural network, in the second step. This paper presents the methods proof-of-concept. It focuses on a simple yet prevalent inter-crystal scatter scenario, where a 511-keV annihilation photon is detected coincidently with two inter-crystal-scattered photons whose energy sum accounts for the whole 511 keV annihilation energy. It shows, in preliminary simulations, a promising correct LOR computation rate in the range from 90 to 94%. Finally, it discusses the steps and requirements for the eventual implementation of the method, including further validation, hardware requirements, system- level issues and possible other applications.

A multi-agent architecture for a driver model for autonomous road vehicles

A look at the current state of driver models and driver related activities (parking, avoiding collisions, path selection, etc.) shows that driving is broken up in a series of small tasks. These must be executed in a coordinated manner by the driver, Their different nature suggest that each could be implemented by an artificial intelligence paradigm (for example: fuzzy logic, neural nets and knowledge based systems). We think that a multi-agent system integrating autonomous agents and enabling them to cooperate to solve problems that are beyond their individual capabilities is the architecture for a more complete driver model. In the long run, the driver model must integrate all characteristics of driver and vehicle, but for the moment current driver models are dedicated to some specific tasks executed by the driver, lateral and longitudinal control, parking and collision avoidance are examples. In this paper we propose an architecture for a generic multi-agent system used to develop the driver model.

Sensitivity Increase Through a Neural Network Method for LOR Recovery of ICS Triple Coincidences in High-Resolution Pixelated- Detectors PET Scanners

Scanner sensitivity is often critical in high-resolution Positron Emission Tomography (PET) dedicated to molecular imaging. In neighboring pixelated detectors with individual readout, sensitivity decreases because of multiple coincidences produced by Compton scattering. Correct analysis of those coincidences would enable a substantial sensitivity increase. However, including scattering byproducts in the image often lead to image quality degradation because of inaccurate Line-of-Response (LOR) assessment. In such scanners, to support high count rates, multiple coincidences are usually discarded when image degradation is not acceptable, or blindly accepted for a low computational burden. This paper presents a new, real-time capable method that includes Inter-Crystal Scatter (ICS) triple coincidences in the image without significant quality degradation. The method computes the LOR using a neural network fed by preprocessed raw data. As a proof of principle, this paper analyzes the simplest ICS scenario, triple coincidences where one photoelectric 511-keV event coincides with two more whose energy sum is also 511 keV. The paper visits the algorithm structure, presents Monte Carlo assessment with the LabPET model, and displays images reconstructed from real data. With an energy window of 360-660 keV and a singles energy threshold of 125 keV, the inclusion of triple coincidences yielded a sensitivity increase of 54%, a resolution degradation similar to that of other sensitivity-increasing methods, and only a slight contrast degradation for real LabPET data, with potential for numerous further improvements.

Monte Carlo results from neural networks as an alternative to Compton photons LOR analysis

Sensitivity is critical in small-animal PET, and including more of the discarded detections would increase it. However lowering the energy threshold compromises the spatial resolution. This paper is an update on a method to include triple coincidences in the image without significant image degradation. With the energy threshold set as low as 50 keV, the triple coincidences analyzed are the simplest inter-crystal Compton scatter scenario where one photoelectric 511-keV detection coincides with two detections whose energy sum is also 511-keV. The method uses neural networks instead of traditional Compton interaction mathematical models to compute the proper Line-of-Response (LOR) for that coincidence. The paper revisits the algorithm structure, and in particular the preprocessing steps required to simplify the data fed to the network, preprocessing which improves the LOR computation significantly. The paper then presents Monte-Carlo analysis of the method with various point and cylinder sources. The simulated scanner geometry is purposely made to encompass the very worst-case conditions seen in PET scanners today: small diameter, poor photoelectric fraction, poor 35% FWHM energy resolution. LOR identification error is around 20 to 25% and the sensitivity increase ranges from approximately 70 to 100%. Finally, images were recently obtained, with overall very good quality.

Automatic Channel Fault Detection on a Small Animal APD-Based Digital PET Scanner

Fault detection and diagnosis is critical to many applications in order to ensure proper operation and performance over time. Positron emission tomography (PET) systems that require regular calibrations by qualified scanner operators are good candidates for such continuous improvements. Furthermore, for scanners employing one-to-one coupling of crystals to photodetectors to achieve enhanced spatial resolution and contrast, the calibration task is even more daunting because of the large number of independent channels involved. To cope with the additional complexity of the calibration and quality control procedures of these scanners, an intelligent system (IS) was designed to perform fault detection and diagnosis (FDD) of malfunctioning channels. The IS can be broken down into four hierarchical modules: parameter extraction, diagnosis, channel fault detection and fault prioritization. Of these modules, parameter extraction and fault detection have previously been reported and this paper focuses on diagnosis, improved fault detection and fault prioritization. The status diagnosis module will diagnose all channels and propose an explanation of the reasons that lead to the diagnosis. The purpose of the fault prioritization module is to help the operator to zero in on the faults that need immediate attention. The FDD system was implemented on a LabPET avalanche photodiode (APD)-based digital PET scanner. Experiments demonstrated a FDD Sensitivity of 99.9% (with a 95% confidence interval (CI) of [99.6, 100.0]) for major faults. Globally, the balanced accuracy of the diagnosis for varying fault severities is 91%. This suggests the IS can greatly benefit the operators in their maintenance task.

Automatic channel fault detection and diagnosis system for a small animal APD-based digital PET scanner

Fault detection and diagnosis is critical to many applications in order to ensure proper operation and performance over time. This applies to positron emission tomography (PET) scanners which are complex systems that require regular calibrations by qualified scanner operators to ensure optimal performance. Furthermore, for scanners employing one-to-one coupling of crystals to photodetectors to achieve enhanced spatial resolution and contrast, the calibration task is even more daunting because of the large number of independent channels involved. To cope with the additional complexity of the calibration and quality control procedures of these scanners, an intelligent system (IS) was designed to perform fault detection and diagnosis (FDD) of malfunctioning channels. The IS can be broken down into four hierarchical modules: parameter extraction, channel fault detection, fault prioritization and diagnosis. Of these modules, the first two have previously been reported and this paper focuses on fault prioritization and diagnosis. The purpose of the fault prioritization module is to help the operator to zero in on the faults that need immediate attention. The fault diagnosis module will then diagnose the causes of the malfunction and propose an explanation of the reasons that lead to the diagnosis. The FDD system was implemented on a 8 cm axial length LabPET™ scanner located at the Sherbrooke Molecular Imaging Center and experiments demonstrated a FDD efficiency of 99.3% (with a 95% confidence interval (CI) of [98.7, 99.9]) for major faults. Globally, the balanced accuracy of the diagnosis for varying fault severities is 92 %. This suggests the IS can greatly benefit the operators in their maintenance task.

Results from neural networks for recovery of PET triple coincidences

High-resolution PET scanners with pixelated detectors have great sensitivity increase potential through the inclusion of multiple coincidences. Indeed, poor energy resolution and in-crystal detection mispositioning often prevent “traditional” Compton kinematics analysis from yielding high Line-of-Response (LOR) discrimination rates, while Bayesian methods are computationally expensive. Hence multiple coincidences are usually discarded when image degradation is not acceptable. This paper presents results from a new method to include Inter-Crystal Scatter (ICS) triple coincidences in the image without significant image degradation. The triple coincidences analyzed are the simplest inter-crystal Compton scatter scenario. Instead of mathematical models, the method employs geometry simplification of the raw energy and position measurements, which are then fed to a neural network. The paper quickly visits the algorithm structure, presents some Monte-Carlo validation results of the method with the LabPET model and shows images reconstructed from real data. The method achieves a 42% increase in sensitivity at the expense of a 10% degradation in contrast-to-noise ratio (CNR), with numerous potential improvements.

Real Time Artificial Neural Network FPGA Implementation for Triple Coincidences Recovery in PET

In small-animal Positron Emission Tomography (PET), spatial resolution improvements rely on detector minimization in size and often come at the expense of lowering the detector photoelectric fraction. As a result, Inter-Crystal Scatter (ICS) occurrences are increased and affect the overall PET detection efficiency. To reclaim some lost efficiency, previous work used an artificial neural network (ANN) to identify the true line of response (LOR) for the simplest multiple event detection case, three coincident singles known as triplets. Despite promising results, this method is limited to an offline processing which is impractical when a limited data bandwidth is present between the scanner and the PC. This paper demonstrates the capability of processing triplets in real time using an ANN implemented in the field-programmable gate array (FPGA). The ANN pipelined architecture can process over 1 million triplets/second using less than 6000 FPGA slices. Real time processing on the LabPET I scanner yielded an overall 39.7% increase in detection efficiency relative to traditional high resolution settings with a 360-660 keV energy window along with a slight Contrast-to-Noise Ratio ( CNR) degradation. Although improvements are still possible, the proposed FPGA implementation proves the usability of an ANN in real time PET applications in conditions where spare computational resources are limited and the data rate to be processed is high.

A New Satellite Attitude State Estimation Algorithm using Quaternion Neural Networks

Satellite attitude control based on state feedback techniques requires measurement of all the state variables describing the attitude dynamics. The Extended Kalman Filter (EKF) has been used for this task, and works quite well in the general cases. However, the EKF is computationally intensive and requires a significant design effort due to mathematical modeling, linearization and its quaternion-motion version requires the use of two different attitude models a . A number of estimation techniques based on neural networks have shown to be more accurate than the EKF, but none of them seem to have been applied to satellite attitude or to quaternion motion. In this paper, a neural networks based satellite attitude estimation algorithm is presented. The proposed approach is original by using a quaternion neural network. It also presents a new way of integrating the neural network into the state estimator and develops a training procedure which is easy to implement. The suggested algorithm is shown to provide the same accuracy as the EKF with significantly lower computational complexity.

COP: a simple way to integrate imperative programming and declarative programming

The paper proposes how to integrate two languages, C++ and Prolog, into one. The resulting language is COP (C++ Or Prolog). The motivation behind this work was to offer in one language two programming styles in order to simplify program writing. For example, an application programmer can use the COP language when it is necessary to program in a procedural or object oriented way (C++) and also with rules (Prolog). Our feeling is that a programmer could benefit from our approach because he or she has the choice to use a programming style adapted to the application needs. We present the COP language. In COP, we try to respect the syntax, the semantics and the philosophies of C++ and Prolog. We define how the two languages can work together. Our approach is to add some features allowing C++ to call Prolog goals. We give all the details that a COP programmer must know in order to use the language; that was possible because we kept our design as simple as we could.