Light Neural Network Research Articles

Particle Size Analysis of Nanopowders Using Neural Networks and Electron Microscopy

The purpose of the research is to develop an application capable of automatically determining the particle size distribution of nanopowder using neural network technology in order to simplify the process of preparing documentation during its manufacture.Methods. To determine the physical properties of nanopowders during their fabrication, it is necessary to analyze the particle size distribution. A methodology for determining the size distribution of nanopowder particles based on light neural networks is proposed. Images obtained by electron microscopy are used for processing, which allows to speed up the preparation of manufactured powders for sale. The dataset collected for training contains real images of samples of different powders, augmented data and generated images. The Python language, LabVIEW graphical programming environment, YOLO convolutional neural network and various Python language libraries were used in the development.Results. The study resulted in a model trained on the collected dataset that is capable of recognizing particles in images. A software interface was created to work with the model to analyze nanopowder samples.Conclusion. The developed application allows to automatically determine the size of each powder particle on the basis of the obtained images, as well as to build graphs of their size distribution. This greatly simplifies the work of nanopowder producers and facilitates the preparation of the necessary documentation for the produced product.

Decoupled learning for brain image registration.

Image registration is one of the important parts in medical image processing and intelligent analysis. The accuracy of image registration will greatly affect the subsequent image processing and analysis. This paper focuses on the problem of brain image registration based on deep learning, and proposes the unsupervised deep learning methods based on model decoupling and regularization learning. Specifically, we first decompose the highly ill-conditioned inverse problem of brain image registration into two simpler sub-problems, to reduce the model complexity. Further, two light neural networks are constructed to approximate the solution of the two sub-problems and the training strategy of alternating iteration is used to solve the problem. The performance of algorithms utilizing model decoupling is evaluated through experiments conducted on brain MRI images from the LPBA40 dataset. The obtained experimental results demonstrate the superiority of the proposed algorithm over conventional learning methods in the context of brain image registration tasks.

A neural network-based algorithm for simultaneous event positioning and timestamping in monolithic scintillators

Objective. Monolithic scintillator crystals coupled to silicon photomultiplier (SiPM) arrays are promising detectors for PET applications, offering spatial resolution around 1 mm and depth-of-interaction information. However, their timing resolution has always been inferior to that of pixellated crystals, while the best results on spatial resolution have been obtained with algorithms that cannot operate in real-time in a PET detector. In this study, we explore the capabilities of monolithic crystals with respect to spatial and timing resolution, presenting new algorithms that overcome the mentioned problems. Approach. Our algorithms were tested first using a simulation framework, then on experimentally acquired data. We tested an event timestamping algorithm based on neural networks which was then integrated into a second neural network for simultaneous estimation of the event position and timestamp. Both algorithms are implemented in a low-cost field-programmable gate array that can be integrated in the detector and can process more than 1 million events per second in real-time. Results. Testing the neural network for the simultaneous estimation of the event position and timestamp on experimental data we obtain 0.78 2D FWHM on the (x, y) plane, 1.2 depth-of-interaction FWHM and 156 coincidence time resolution on a LYSO monolith read-out by 64 Hamamatsu SiPMs. Significance. Our results show that monolithic crystals combined with artificial intelligence can rival pixellated crystals performance for time-of-flight PET applications, while having better spatial resolution and DOI resolution. Thanks to the use of very light neural networks, event characterization can be done on-line directly in the detector, solving the issues of scalability and computational complexity that up to now were preventing the use of monolithic crystals in clinical PET scanners.