Tensor networks provide efficient representations for high dimensional data, although selecting suitable neural architectures for their integration remains challenging. Neural Architecture Search (NAS) offers a promising solution by automating architecture design through structured search spaces and algorithmic optimization, often matching or surpassing human-crafted models. In this work, we evaluate the selected NAS methods on CIFAR-10 using common metrics including clean and adversarial accuracy, F1-score, parameter count, and search cost. The experimental results show that DARTS delivers the highest clean accuracy with minimal computational load, whereas HA-ENAS achieves the strongest adversarial robustness due to its multi-objective evolutionary strategy. REP consistently enhances robustness within differentiable NAS frameworks, especially on structured benchmarks, while NAS-RL demonstrates significant scalability limitations and limited architectural diversity. Overall, findings confirm that no single NAS method dominates across all evaluation criteria. Instead, effectiveness depends on the trade-off between robustness, efficiency, and architectural complexity. The study concludes that REP provides the most balanced performance, DARTS excels in efficiency, and HA-ENAS is best suited for robustness-oriented applications, offering a practical foundation for future NAS research in tensor-network–based model design.

Breast cancer is one of the most common diseases affecting women around the world, with millions of new cases each year and a high rate of mortality. Although advances in detection and treatment have led to better outcomes for many, some regions still struggle with limited access to early diagnosis. Early detection of breast cancer significantly improves treatment success, emphasizing the need for accessible, accurate diagnostic tools. This study focuses on improving the segmentation and classification of breast tumors in mammographic images by examining the impact of various preprocessing techniques on the performance of advanced object detection models. We use the Digital Database for Screening Mammography (DDSM), a benchmark dataset for mammography analysis, to evaluate three state-of-the-art object detection models: Faster R-CNN, YOLOv8, YOLOv9, and YOLOv11. These models are trained on images processed with different methods to observe how each technique affects the accuracy of tumor segmentation and classification into benign or malignant categories. The primary contribution of this study is the comparative analysis of preprocessing effects on AI model performance in tumor detection, aiming to identify an optimal preprocessing approach that maximizes accuracy and reduces misclassification rates. We systematically assess preprocessing effects on model accuracy, focusing on improvements in tumor visibility and classification performance.

This work detects predators in extensive breeding farms, especially those dedicated to free-range chicken production. In this environment, the main threats are birds of prey and poachers. The proposal focuses on the first case to study how using deep learning models can help detect and label images. Specifically, we evaluate the performance of several object detection models on a dataset focused on birds of prey in outdoor farm settings. The study compares architectures from two families: Faster R-CNN and YOLO. Results show notable differences in performance between model families. Faster R-CNN shows a better performance than YOLO. This considerable difference marks the importance of selecting the model to be used, especially in cases such as this one, where the objects to be detected and classified are small in size.

The automatic detection of anomalies in medical images is a significant challenge in the assisted diagnosis of neurodegenerative diseases such as Alzheimer's. This paper presents an anomaly detection model based on Transformers for the analysis of brain magnetic resonance images. The proposed architecture combines a Vision Transformer as an encoder with a memory bank module that allows modeling the distribution of healthy brains and detecting deviations through reconstruction error. The model is trained using a one-class learning approach, using only images considered normal, with the aim of learning the representation of normality and automatically flagging atypical structural patterns. To adapt volumetric studies to the architecture, a preprocessing procedure is designed that transforms three-dimensional information into a two-dimensional representation compatible with the model. The results obtained demonstrate a solid ability to characterize normality and generate reliable predictions, confirming the viability of Transformer-based architectures for unsupervised anomaly detection in neuroimaging. This approach lays the foundation for future extensions in clinical settings and other medical imaging applications.

This study focuses on an industrial facility’s time-slot allotment (TSA) problem, where loading and unloading docks are assigned to the incoming lorries, reducing the number of them waiting for service. Several constraints apply to the different dock stations, including disparate timetables and task duration, various capacities of simultaneous services, etc. Evolutionary algorithms cope with the enormous variability of the combinatorial problem, maximizing the number of accepted lorries while decreasing the queue at the entrance. Interestingly, the problem’s structure led to unconventional operator probabilities, which also analyzes the evolutionary techniques and operators included in this study. Comparative analysis with state-of-the-art methods highlights the algorithm’s effectiveness, though computational demands rise with population size.

Artifacts are noisy signals that commonly contaminate electroencephalographic (EEG) recordings, mixing with underlying brain activity and degrading the quality of neurophysiological data. Previous research on epileptic Anomaly Detection has shown that this approach is also sensitive to unlabelled artifacts, often leading to an increased False Positive rate. While most methods focus on detecting or removing a single type of artifact, this work proposes a unified multi-class framework to classify several artifact types alongside normal and pathological brain activity within a recording using a single, simple Machine Learning classifier. A set of spectral, temporal and statistical features commonly associated with different artifact types is extracted. KNN and XGBoost classifiers are trained and evaluated under a cross-validation scheme. The results demonstrate strong performance for both models, achieving approximately 90% sensitivity across all classes while maintaining a 100% specificity. These findings highlight the effectiveness of a unified and interpretable approach for multi-class EEG artifact classification.

This paper investigates the application of machine learning methods for anomaly detection of both physical and cyber threats in Industrial Internet of Things (IIoT) environments, with a novel method of separating different threat classes, performing delegation of computationally inexpensive threshold-based metrics to a simple rules-based alerting system, while performing anomaly detection of the more complex behavioural-based metrics in a machine learning model. This hybrid approach of separating threshold-based and behaviour-based detection methods is validated on the Edge-IIoTset2023 and CICIoT2023 public research datasets. As a new contribution, this hybrid methodology is validated against both tree-based classifiers and artificial neural network (ANN) classifiers. Experimental results indicate that while ANNs can be very effective, marginally higher accuracy (similar to 3%) and significantly faster predictions can be achieved with less computationally expensive tree-based algorithms such as Decision Trees and Random Forests, thereby optimizing the price-performance trade-off for the operators of IIoT environments.

Photosensitivity is a neurological disorder where the brain produces abnormal epileptic reactions to visual stimuli known as Photoparoxysmal Responses (PPR), which can sometimes result in epilepsy seizures. The Intermittent Photic Stimulation (IPS) protocol is used to diagnose this condition by exposing the patient to flashing lights -firstly, at increasing frequencies and, secondly, at decreasing frequencies- while recording the brain activity using an Electroencephalogram (EEG). Neurophysiologists observe the EEG signals to identify PPR, taking care to prevent triggering an epileptic seizure and halting the process if necessary. Because of the nature of the stimulation and the low prevalence of photosensitivity, automatically detecting these events is challenging because PPR activity represents minority events of unusual brain activity amidst a large volume of regular recordings. In previous research, a Variational AutoEncoder (VAE) was used to label EEG recordings from IPS sessions; with the encoder and the decoder incorporating recurrent neural networks and dense layers to deal with the EEG channels’ time series. The VAE outperformed the current state of the art and a battery of unsupervised anomaly detection methods. However, training the VAE involved Leave-One-Subject-Out cross-validation with a short number of records (gathered from up to 9 diagnosed photosensitive patients); thus, this approach needs testing on patients who do not have photosensitivity. This study tackles this issue by gathering data from 5 patients who, although suffering from epilepsy, are not photosensitive; in accomplishing this, the VAE training uses data from 9 patients from previous research. The resulting model then evaluates each non-photosensitive patient, identifying anomalous EEG channel sequences of values. Surprisingly, the VAE showed an unexpected rate of False Positive due to artefacts detected in the signal, especially for subject 3. An extra experiment evaluated the capabilities of the VAE in filtering any EEG anomaly. In this second experiment, the VAE showed a high performance, suggesting this configuration could be valid not only for labelling PPR but also for any possible artefact in the EEG signals, reducing the human effort in filtering these events. Nevertheless, a more in-depth study is needed to validate the proposal and to adapt the procedure to this new labelling task.

This paper proposes a hybrid system for feature selection, allowing the degree of importance of different features to be determined. Different feature selection techniques, specifically the following four: Univariate feature ranking for classification using chi-square tests, Rank features for classification using minimum redundancy maximum relevance, Relief-based feature selection and Constrained greedy k-means with silhouette value ranking method are employed on the three well-known datasets from UCI Machine Learning Repository: Iris, Wine, and Ionosphere. Subsequently, with the variables ordered from highest to lowest importance as a result of each of the four feature selection methods, an iterative clustering process is performed, agglomerating features that are used to calculate accuracy. This accuracy is employed to determine the most relevant features. The best accuracy value for Iris dataset is 96

Photosensitivity refers to a neurophysiological condition in which the brain generates epileptic discharges known as Photoparoxysmal Responses (PPR) in response to light flashes. In severe cases, these PPR can lead to epileptic seizures. The standardized diagnostic procedure for this condition is called Intermittent Photic Stimulation. During this procedure, the patient is exposed to a flashing light, aiming to trigger these epileptic reactions while preventing their full development. Meanwhile, brain activity is monitored using Electroencephalography, which is visually analyzed by clinical staff to identify these responses. Hence, the automatic detection of PPR becomes a highly unbalanced problem that has been barely studied in the literature due to photosensitivity’s low prevalence. This research tackles this problem and proposes using Inception-based deep learning (DL) neural networks that, together with transfer learning, are trained in epilepsy seizure detection and tuned in the PPR automatic detection task. A data augmentation (DA) technique is also applied to balance the available data set, evaluating its effects on the DL models. The proposal outperformed state-of-the-art solutions in the literature, achieving higher ratios on standard performance metrics, and with DA significantly improving the Sensitivity without affecting Accuracy and Specificity. This project is currently being developed with patients from Burgos University Hospital, Spain.

Photosensitivity is a neurological condition in which the brain produces epileptiform reactions to visual stimuli known as Photoparoxysmal Responses (PPR). These events are typically diagnosed through Intermittent Photic Stimulation (IPS) while monitoring brain signals with electroencephalography (EEG). Manual analysis of PPR is time-consuming and subjective, which motivates the development of automated detection methods. In this work, we propose an unsupervised anomaly detection (AD) approach using a Variational Autoencoder (VAE) trained exclusively on normal EEG segments from non-photosensitive patients. The model is evaluated on EEG recordings from photosensitive patients to identify PPR activity as deviations from normal patterns. In previous research, this VAE model outperformed other unsupervised AD models in the literature for this task; however, it generated a large number of False Positives that were later confirmed as EEG anomalies that were not labelled. This research represents the first step in an EEG anomaly detection and multi-classification framework, aiming not only to detect PPR but to automatically label all Positive instances as the different types of anomalous EEG patterns in future work. Results reveal that the model performed well, reaching 83

Self-Organizing Map (SOM), an unsupervised learning method, is an artificial neural network (ANN) able to handle non-linear problems that can be used for exploratory data analysis, pattern recognition, and variable relationship assessment. Much more power ability is gained when the SOM-based model is merged with other clustering algorithms, creating hybrid frameworks and architectures. In this paper, hybrid frameworks and architectures are presented to improve the performance of a simple one-layer SOM in different domains such as image classification, assessing environmental pollutants, real-time scheduling, and other real-world problems.

This research demonstrates the effectiveness of combining an exact algorithm with an approximation algorithm to address a complex, multi-objective, multi-constrained problem. The findings show that, in isolation, each algorithm may struggle to find collision-free, optimal (or near optimal) paths for multiple robots under time constraints. However, a systematic and carefully integrated approach that leverages multiple algorithms can resolve the problem.

Function optimization represents a remarkable challenge in industry and society, aiming to find reasonable solutions-even if they are suboptimal-for everyday problems. Metaheuristics drive the optimization search towards the goals using a specific algorithm inspired by different concepts: from industrial processes to the behaviour of living beings in nature, from mathematical ideas to physics notions. This research proposes a new metaheuristic inspired by the Slime Mould and its foraging behaviours. On the one hand, an exploitation stage mimics the greedy amoeba's conduct when food is plenty. On the other hand, an exploration stage copies the fruity aggregation of the cells and the subsequent spore dissemination. This study compares the most cited metaheuristics and the Slime Mould Optimization in two different experimentation stages: on the one hand, the optimization of standard benchmarking functions; on the other hand, solving the robot path planning problem. Moreover, a hybridization of the SMO and the WOA is presented, which keeps the SMO's convergence speed and the WOA's good performance in finding the best solutions.

Breast cancer remains one of the leading causes of mortality among women worldwide, making early detection crucial for improving survival rates. Deep learning-based approaches have shown remarkable potential in automating tumor detection from mammographic images; however, their effectiveness largely depends on the choice of data augmentation strategies and model architecture. In this study, we systematically investigate the impact of six data augmentation techniques and four object detection models on the task of breast tumor detection, using the publicly available DDSM dataset. Each model was trained on preprocessed mammographic images, with individual augmentation techniques applied separately to assess their isolated effects. The augmentations analyzed include Contrast Limited Adaptive Histogram Equalization, random rotation, random translation, horizontal flipping, Contrast Limited Adaptive Histogram Equalization combined with rotation, and Gaussian noise addition. Experimental results demonstrate that Faster Region-based Convolutional Neural Network consistently outperforms the YOLO variants, achieving the best detection result of 91.35

Breast cancer remains one of the most prevalent and life-threatening diseases worldwide, being the most frequently diagnosed cancer among women and the second leading cause of cancer-related mortality. Precise tumor segmentation is essential for breast cancer assessment, as it enables accurate estimation of tumor size, monitoring of disease progression, and evaluation of treatment effectiveness. Despite the importance of this task, the development of reliable automatic methods is hindered by the scarcity of fully annotated datasets, which makes manual labeling both time-consuming and subject to inter-observer variability. In this study, we propose an unsupervised 3D tumor segmentation method based on Fuzzy C-Means (FCM) clustering, specifically designed for volumetric Dynamic Contrast-Enhanced MRI (DCE-MRI) of the breast. Unlike supervised deep learning approaches, our method does not require manual annotations for training, making it especially valuable in scenarios with limited labeled data. The proposed pipeline combines preprocessing, region-of-interest extraction, and FCM-based clustering to generate accurate segmentation masks with minimal human intervention. We evaluated our approach using clinical data from the ACRIN-6698 dataset, comparing the automatic segmentations against expert manual annotations. The method achieved high performance across multiple metrics, including accuracy, precision, recall, specificity, Dice-Sorensen coefficient (DSC), and Jaccard index (IoU). These results demonstrate the feasibility of unsupervised clustering techniques for volumetric breast tumor segmentation, offering a promising alternative to supervised methods in clinical contexts where annotated data is limited.

The Vehicle Routing Problem (VRP) is concerned with the optimal design of routes to be used by a fleet of vehicles to serve a set of customers. Usually, VRP is based on a homogeneous fleet size problem which is complex when the vehicle fleet is heterogeneous. In this research, a new hybrid nearest neighbors - tabu search model for homogeneous fleet of vehicle optimization on VRP of logistics distribution is proposed. The proposed hybrid model in routing a homogeneous fleet VRP is tested on a realistic dataset in the Serbian company “Carnex”. The experimental results of this method, in two different scenarios, are compared to the empirical results, demonstrating that the experimental results obtained by the new hybrid nearest neighbors - tabu search model are better than empirical results in the company. Also, experimental results are better for scenario 2, where the starting point for optimization and delivery starts with a customer that has the smallest order.

The detection of recurrent prostate cancer following external beam radiotherapy relies on persistent increases in serum prostate-specific antigen (PSA) levels. However, this biochemical recurrence may take place after an extended period, delaying secondary treatment for patients with relapsing tumours. Recent research has identified situations where anticipating PSA value relapse, often linked with tumour recurrence, is possible using personalized mechanistic models for PSA forecasting. Yet, these models lack insight into false negatives and their susceptibility to relapse. To address this problem, we propose to determine an optimal threshold for detecting PSA outliers. This research analyses patient data, particularly PSA value sequences, to determine whether the mechanistic model’s parameters representing an outlier of the parameter distribution exhibit a correlation with the characteristics of the patient’s relapse.

This article reviews the metaheuristics published in the literature, emphasizing their usefulness in solving complex optimization problems. The review highlights inspiration's relevance in the metaheuristics design, being the main classification in multiple taxonomies existing in the literature. After reviewing the state of the art, six of the most relevant metaheuristics used to solve problems of various types (engineering, logistics, economics, data science, ...) were selected. This selection of metaheuristics will be subjected to an analysis of their performance using a set of problems selected from different authors. The problems selected for this analysis include problems with a single minimum or multiple minima, different sizes in terms of dimensions, and different types of mathematical functions such as polynomial, trigonometric, or exponential. The analysis offers a discussion of which scenarios are the best for each metaheuristic, analyzing aspects such as the ability of metaheuristics to explore and escape local minima. The article concludes by summarizing which metaheuristic is best for each type of problem. Keywords: Metaheuristics, benchmark, optimization problems, biological-based metaheuristics