This paper discusses the challenges of maintaining consistent image sequences during patient examinations. One of the key issues that arises during MRI is the patient’s inability to maintain a constant position, thus resulting in image shifts that impede image fusion and subsequent diagnostic quality. To address this, coregistration is employed, which involves aligning two or more images based on a reference image to minimize shifts between them. This is a crucial step in preparing data for analysis and can lead to previously overlooked images becoming important components of medical solutions. In this study, a novel measurement of coregistration effectiveness is presented that includes normalization, difference, Top-hat transformation filtering, and global thresholding to achieve image similarity. Mean squared error and the structural similarity index measure are utilized as references against which the described method is compared. The outcomes of the measurements are presented and discussed. The results demonstrate the efficacy of the proposed metrics of coregistration effectiveness in facilitating a more accurate and comprehensive analysis of MRI data. Due to the achieved results, the main purpose is feasible.
The anticipated expansion of TinyML-related technologies will require methods for efficient ML model development. In this paper, we present a method for obtaining a set of ML models for TinyML systems that satisfies the assumption that a more efficient model is also more complex and therefore consumes more energy. The results show that our method is capable of providing numerous diversified sets of ML models.
A novel approach to syntactic pattern recognition based on multi-string parsing is introduced in the paper. A methodological motivation of the research, which has resulted from application-oriented requirements, is presented. Formal foundations, an architectural model and a generic recognition algorithm of the approach are introduced. The approach allows one to recognize complex objects which should be represented by a series of images instead of one image. The series of images is represented with a sequence of aspectual patterns which are treated as a holistic aspect-integrated description of a recognized object.
Computed tomography (CT) images are widely used in medical examination applications. They are constructed from the raw data using different kernels. However, due to the technology used, the software finds problems with image segmentation in the presence of noise or blurred edges. In this paper, we use quaternion mathematics for image de-noising by applying the bilateral filtering algorithm to reconstruct a pure soft image from the images reconstructed with sharp kernels, without using the raw data. Our output is compared to reconstructed images with different denoising techniques, such as non-local means and wavelet denoising. The reconstructed images using other techniques are assessed using structural index similarity (SSIM) and peak signal-to-noise ratio (PSNR). Results show high similarity by comparing the outputs of sharp kernels to that of the B30f. Furthermore, the results show that the proposed work can efficiently increase the similarity between the images reconstructed with hard and soft kernels.
Relative representations allow the alignment of latent spaces which embed data in extrinsically different manners but with similar relative distances between data points. This ability to compare different latent spaces for the same input lends itself to knowledge distillation techniques. We explore the applicability of relative representations to knowledge distillation by training a student model such that the relative representations of its outputs match the relative representations of the outputs of a teacher model. We test our Relative Representation Knowledge Distillation (RRKD) scheme on supervised and self-supervised image representation learning with MNIST and show that an encoder can be compressed to 47.71 https://github.com/Ramos-Ramos/rrkd .
Classifier ensembles have shown the ability to classify drifted data streams. The following paper proposes an ensemble consisting of a single hidden layer feedforward neural network and an Extreme Learning Machine. For this purpose, a new incremental version of the Extreme Learning Machine is also proposed. Motivations behind such an approach have been precisely described and supported by conducted research. The achieved results show when the architecture might be most useful and what are the possible directions for future development of this method.
Image texture analysis is ubiquitous as it finds applications in many scientific fields of interest, including biomedical and material science. The detection of meaningful texture properties such as directionality remains a challenging task, due to the complexity of texture. We build upon our past work on the design of convolutional neural networks (CNNs) for texture directionality detection. The CNNs are trained on a library of synthetic textures with known directionality and varying perturbation levels. The present effort focuses on enhancing the training data through a new perturbation procedure and a more diverse set of synthetic textures. We study the performance of new CNN architectures, such as grouped CNNs, on the enhanced synthetic texture library. The results yield novel insight into CNN-based texture directionality detection. Shallow and grouped CNNs show better performance than deep CNNs, unlike previously. We discuss this performance shift and its implications, and suggest possible future work directions.
Multimodal data processing has recently become popular due to technological advances and easier access to real video, audio, images, or text data. This data type is often processed using deep neural networks associated with high time and computational complexity. The present work addresses the problem of classifying a multimodal MM-IMDb dataset, representing the problem of recognizing a movie genre based on a poster and a brief description of the plot. For experiments, 20 binary subsets were separated, from which features were then extracted. Features from the text were obtained using the tf-idf method, while the posters were reduced to a single color. Computer experiments on the resulting tabular data were conducted separately on both modalities, as in the concatenated feature space. The results confirmed that classical approaches to feature extraction could allow satisfactory quality classification of multimodal data to be obtained even when using relatively simple pattern recognition algorithms.
Medical simulations based directly on CT data are appreciated for their realism in reflecting anatomy. On the other hand, such an approach introduces a number of complications when it comes to the possibilities of constructing interventional simulators. In this paper an approach to CT-based volumetric data deformation is presented using the Obi simulation system and Unity framework. This method allows for easier and faster development, while yielding realistic outcomes built upon established and reliable solution. The method was implemented within the MrTEEmothy simulation system, where deformations of the heart tissues are visible in the simulated ultrasound images, during the transeptal puncture procedure.
Growing share of Renewable Energy Sources (RES) in Polish electricity system has motivated the Polish government to make changes in the law. Those changes decrease the economic profit of using photovoltaic power plants, but incentivize the use of hybrid systems equipped with batteries. In this paper, the state-of-the-art battery management system, which maximizes self-consumption has been compared to our own time-based proposal, which is focused on economic profit. We used real-life measurements from a 6kWp residential power plant to compare both battery management systems. The results show that controlling batteries with our model allows us to save more money.
The empirical risk minimization approach of contemporary machine learning leads to potential failures under distribution shifts. While out-of-distribution data can be used to probe for robustness issues, collecting this at scale in the wild can be difficult given its nature. We propose a novel method to generate this data using pretrained foundation models. We train a language model to generate class-conditioned image captions that minimize their cosine similarity with that of corresponding class images from the original distribution. We then use these captions to synthesize new images with off-the-shelf text-to-image generative models. We show our method’s ability to generate samples from shifted distributions, and the quality of the data for both robustness testing and as additional training data to improve generalization.
Actinic Keratosis (AK) is a type of skin lesion that typically appears on skin areas that are exposed to the sun. It is considered a precursor to squamous cell carcinoma (SCC), which is a common form of skin cancer worldwide. Early detection and treatment of AK are essential for effective management of SCC. Recent developments in deep learning (DL) have shown significant promise in improving the detection and diagnosis of AK and SCC. This study aimed to evaluate the use of YOLOv7, a state-of-the-art object detection model, in identifying AK lesions in skin images. We compared the accuracy and efficiency of YOLOv7 and YOLOv7-tiny models to determine the most suitable model for AK lesion detection. The results of the experiment were promising and showed that YOLOv7 can effectively identify AK lesions in skin images with high accuracy and speed. Furthermore, the study utilized the Grad-CAM technique to gain a deeper understanding of how the DL models detect AK lesions in skin images. We hope that this technology can assist dermatologists in making clinical decisions, leading to early treatment and prevention of AK, and ultimately preventing the development of skin cancer. Overall, the findings of this study highlight the potential of DL models in dermatology and their usefulness in improving clinical practice. With the increasing prevalence of skin cancer worldwide, the use of DL models may play a critical role in the early detection and prevention of skin lesions.
With the rise of health consciousness in recent years, air pollution has become an issue of public concern. The Air Quality Index (AQI), which takes into account various air pollution factors has become an indispensable part of daily life. Of these factors, PM2.5 has a particularly serious impact on the human body. The final Wind-and-Rainfall sensitive LSTM (WRLSTM) can effectively predict PM2.5 concentrations in the next 6–18 h. The results are better than those of common deep learning algorithms, such as MLP regression.
Bulk construction of pattern classifiers, whether for optimizing input data configurations or method hyperparameters, is a computationally highly complex task. The main problem is the prediction quality evaluation function based on estimation using the selected experimental protocol. In the case of iterative optimization algorithms, such an evaluation is computationally-intensive, runs in each iteration, and requires a separate data partition for quality estimation. So-called proxy models may be alternative solutions, which estimate classifier quality on data characteristics without the need to train the prediction model. There are some premises that the problem complexity measures can be used for this purpose. However, this paper negatively verifies this hypothesis – confirming the predictive potential of evaluating the effectiveness of models by complexity measures but also showing a relatively large measurement error in direct relation between quality metric and proxy measure.
Every day, the average Internet user perceives an abundance of content that is unintentionally consumed every day. We frequently hear the seemingly obvious remark that the modern world is full of data. We are bombarded with numerous links to amusing content circulated by our friends, various news and content providers, and social media. Unfortunately, an increasing amount of this information is only loosely related to the truth. Some of the low-quality content news could be automatically detected by modern large language models (LLM). Unfortunately, we need a large number of annotated articles to train such models. In this paper, we described our tool for news content annotation. In particular, we explain our research methodology, the tool architecture, and the analysis of the quality of the annotations. In our experiments, we engaged more than 100 volunteers, who annotated almost 4000 articles.
Words don’t come easy, which fosters the use of generative artificial intelligence models in ongoing popularity of widely available applications such as ChatGPT. The result is an even greater flood of online content that takes time to process. It is where Natural Language Processing tools for classification come in handy. Distinguishing fake news, event types, and other tasks can help process everyday information. In practice, such systems must work on data streams where fast prediction is needed. To achieve it, methods not based on neural networks can be used. Instead, they require feature extractions from the text to convert it to the model input. The primary methods used for this purpose are bag-of-words and n-grams, which allow converting the corpus of texts into a numerical format. This paper proposes a new strategy for creating n-grams – Hollow n-grams – which can be used to create classifiers ensembles with a higher generalization ability than models based on regular n-grams only.
Recognition of the surface myoelectric (sEMG) signal to decode the user’s intention is the most advanced approach to the control of a bionic upper limb prosthesis. This paper is devoted to the methods of distance-based recognition of the sEMG signals using the dynamic time warping (DTW) technique to determine the distance measures between the sEMG signals. Three methods are proposed and developed: a method in which distances are directly used in KNN classifier and two methods in which KNN classifier is fed with features extracted from distances using multidimensional scaling technique, and clustering-based approach, both methods in two versions. The proposed methods were experimentally compared for different K in the KNN classifier and for different quality measures. In addition, for methods with features extracted from distances, the quality of classification for different models of classifiers was experimentally tested. Experimental investigations were conducted using real sEMG signals coming from an amputee and from an able-body subject with an immobilised hand simulating an amputation.
Machine learning has become a key component of the effective detection of network intrusions. Yet, it comes with the lack of transparency - an issue which can be mitigated with the employment of explainable AI techniques. In this paper, the crucial role of explainability in intrusion detection is discussed, along with its benefits and drawbacks, followed by presenting and comparing the results of four main explainability techniques applied to an intrusion detection system.
The impact of air pollution on human health has been studied for decades. Unfortunately, each year brings new evidence of the harmfulness of living in a contaminated environment. The smallest particles of pollutants penetrate the entire body and affect a person already at the stage of the formation of reproductive cells. The spectrum of diseases that can develop from breathing polluted air is broad. Therefore, it is necessary to monitor the state of the air and warn of exceeding the standards, which will allow action in a specific place. In the article, we present a pollutant monitoring system that provides for their measurement and the propagation of information about their level among people in a given area. The system uses IoT technologies and the Internet.
This paper presents the real implementation of a fog computing environment for the execution of color tracking applications by using FogBus2 framework and an artificial intelligence based docker container scheduling. To be precise, an edge computing network has been developed by using a personal computer and several small computing devices such as Raspberry Pi and Nvidia Jetson Nano. Related to the scheduling policy, besides the existing policies in Fogbus2 framework, another one based on fuzzy rules-based system has been designed. Results demonstrate the proposed policy outperforms classical approaches, even when using, pavin the way to the use of knowledge acquisition techniques in order to improve the scheduling performance in terms of makespan and flowtime.
