The increasing incidence of cancer has spurred research across various fields. The digitization of tissue samples stained with hematoxylin-eosin (H&E) enables the acquisition of histological images, which are used to develop computational systems to support pathologists in cancer diagnosis. However, the limited availability of public images has been a well-documented obstacle to advancing these systems. This work presents a systematic review of published studies on the preprocessing of H&E images, focusing on identifying public databases. Through this review, we mapped the countries of origin of these images, the most common lesions, preprocessing techniques, and evaluation metrics. This study provides valuable insights for future research in Biomedicine and Computer Science.
Android malware has emerged as a major threat to mobile device security. While machine learning techniques have proven effective in detecting malicious applications, they remain vulnerable to attacks based on adversarial examples. This study investigates the robustness of five widely-used classifiers (Random Forest, Logistic Regression, Support Vector Machines, Naïve Bayes, and K-Nearest Neighbors) against adversarial machine learning in malware classification. Through systematic evaluation using a gradient-based perturbation method, the analysis reveals critical vulnerabilities where adversarial samples cause misclassifications, particularly converting malware into benign classifications. Random Forest and K-Nearest Neighbors exhibited greater robustness compared to other classifiers, but their accuracy still degraded under high-severity attacks. Linear models exhibit more severe degradation, with Logistic Regression showing consistent vulnerability to adversarial perturbations across all malware families. The results demonstrate that current detection systems can be dangerously deceived by adversarial examples, potentially allowing malicious applications to evade detection.
Mental health problems are becoming more common. Unfortunately, there are still not enough health specialists able to assist them. As a way to provide accessible and immediate support to those in need, digital interventions are receiving increased attention. Conversational agents based on large language models (LLMs) are increasingly used in mental health support. To help psychiatric patients, we have developed a therapeutic dialogue system called Terabot. Earlier experiments revealed several situations that paused the natural flow of the conversation with the patients. We created a new LLM-based dialogue system with an eye tracker as an additional input signal. We propose a feedback loop where the eye tracker gives the patient’s real-time gaze data into the dialogue system. Thanks to this, the patient can receive a response from Terabot, which is more suitable for the current dialogue situation. The dialogue agent’s responses can be improved, resulting in a more human-like flow of conversation.
Robot-human interaction is rapidly increasing in work environments and patient/elderly care centers. Human-robot collaboration is becoming increasingly prevalent in modern production facilities. In addition, robotic systems have been actively used to enable patients and people needing assistance to continue their daily lives. This study aims to determine the estimated position of the human's hand so that the mobile manipulator carrying or grasping an object can deliver the object from hand to hand to the designated human. For this, the humans in the environment and their body poses are determined using deep learning on the image taken live from the depth sensor camera on the mobile manipulator. Then, the hand positions of the humans are determined using the determined poses. The distances between the mobile manipulator and each detected person are calculated using the signals received from the depth sensor camera on the mobile manipulator. Thus, the mobile manipulator can create a position vector for each detected human. Using these position vectors, it will be possible for the mobile robot to deliver the object it carries to the desired human more quickly by using different shortest-path algorithms. Moreover, it will be easier for robots that use the proposed approach to assist humans in work environments or patient/elderly care centers.
With the rapid growth of video streaming services, particularly under the advent of 5G, ensuring optimal enduser service quality has become a pressing challenge for mobile network operators (MNOs). The increasing demand for high-bandwidth, low-latency video applications, such as live streaming and ultra-high-definition content, requires new approaches to detect and mitigate service quality degradation. This paper investigates the use of machine learning models for estimating end-user service quality in 5G networks for adaptive video streaming performance. We analyze real-world data collected from a 5G standalone network, combining in-network metrics, client-side data, and radio environment measures. A key aspect of this study is comparing the effectiveness of the currently available mobile core network data from the User Plane Function with high-resolution (1-2 second) client-side data. The client-side measurements were obtained by the GoDASH framework and was used for proposing improvements on core network user plane probing function. Although client side measured service quality measurement are the ideal, our results demonstrate that core network transport probing data can be used for estimating the end user service quality or detecting video stall but the probing metrics should be synchronized with the video chunks. Including radio environment data in the input dataset of the model only slightly improves the performance. Higher resolution core network data improves the model accuracy but they would be needed for covering mobility scenarios and changing traffic conditions. This research provides actionable insights for MNOs to enhance network analytics and address potential service quality issues in 5G environments.
The verification of identity documents is crucial for institutions such as banks, legal entities, and academia. However, manual processing is costly, which makes automation highly beneficial for both efficiency and accuracy. In this context, we propose a pipeline for classifying and extracting information from different Brazilian identity documents, combining computer vision and machine learning techniques. The Brazilian Identity Document Dataset was used to train CNN models for document classification, and U-Net models were employed to perform fields segmentation. We used Tesseract OCR in conjunction with pre- and post-processing techniques to extract textual information. In our experiments, VGG-16 and ResNet-101 achieved the best overall performance, while U-Net also delivered excellent results in field segmentation, reaching 97% Dice score and 94% IoU. Moreover, OCR errors were significantly reduced through post-processing — down to 1% for CPF numbers in the CPF1 Front class. These findings emphasize the crucial role of post-processing in improving document recognition accuracy.
Emotions are a vital component of human interaction and can impact the performance of speech technologies such as automatic speech recognition significantly. Therefore, this paper analyzes how emotional speech affects the correctness of the Slovenian end-to-end automatic speech recognizer. The open-data Slovenian Artur speech database was employed as the speech resource for training the speech recognizer, and the Slovenian part of the Interface emotional speech database was used to analyze the impact of emotions. The emotion test set contains 4,960 recordings, totalling 6.25 hours, and comprises six emotions and neutral speech. First, we performed a baseline evaluation of the end-to-end speech recognizer on the Artur speech database test set. The word error rate was 6.69%. The end-to-end speech recognizer was evaluated on the Interface speech database neutral emotion, achieving 33.62% WER. Then, the emotional speech was recognized. The word error rate for emotional speech increased between 35.36% and 40.02%.
This work investigates the use of machine learning models for identifying Colorectal Carcinogenesis (CRC) based on changes in liver tissue, as represented in microphotographs. It was provided by the Enteric Neural Plasticity Laboratory of the State University of Maringá, which is involved in this investigation, analyzing the impact of CRC on other tissues such as the liver. There is a high demand for image analysis, and research for identifying CRC using rat liver microphotographs is limited to visual inspection, which is subjective when made by humans. Therefore, automation could expedite this process by utilizing machine learning models. For this purpose, the Deep Learning models ResNet50, InceptionV3, DenseNetB0, and EfficientNet121 were evaluated as feature extractors, along with the classifiers K-Nearest Neighbors (K-NN), Support Vector Machine (SVM), Random Forest, Logistic Regression, and two Multilayer Perceptrons (MLP) for the classification process. The study has demonstrated a significant accuracy of 92.5% in the goal task.
Optimal energy management in a microgrid is studied by an energy management system supported by artificial intelligence (AI). Two stage optimization process is organized. First, a bicriterion optimization model is proposed to determine the optimal charging/discharging schedule of the energy storage system. Historical data is used for the optimization. Second, a nonlinear Reinforcement learning (RL) model is proposed to optimize the MPPT (Maximum Power Point Tracking) mode of the photovoltaic system and the wind turbines in a real-time. Artificial intelligence uses a Soft ActorCritic (SAC) agent to perform the energy flows optimization. The agent is trained offline. Finally, it is used in a real-time to continuously learn (self-tuning) and optimize tracking. The simulation and optimization are implemented in the MATLAB environment.
Microgrids provide a resilient and efficient alternative to traditional power grids, yet they remain vulnerable to operational anomalies, electrical faults, and cybersecurity threats. This study proposes a machine learning-based anomaly detection system utilizing an autoencoder neural network to monitor and analyze microgrid sensor data, specifically current, voltage, and communication traffic. The system learns normal operating patterns and identifies deviations indicative of potential failures or cyber threats. The system undergoes preprocessing techniques such as outlier removal and normalization to ensure data quality, followed by autoencoder-based feature extraction and threshold-based anomaly classification. The proposed framework is implemented as a web-based platform, featuring a Flask-based backend for real-time anomaly detection, a React.js dashboard for intuitive visualization, and a REST API for seamless data exchange. The aim of the research is to enhance the microgrid resilience, reduce downtime, and improve predictive maintenance through the integration of machine learning, real-time data processing, and user-friendly visualization. Experimental results are proposed and serve to assess the accuracy in anomaly detection. The system identifies the deviations and classifies them into low, medium, and high-severity alerts.
This paper presents a method for drone detection that combines the state-of-the-art YOLO-v8 model with Doppler radar to improve robustness in challenging conditions such as low light or bad weather. The YOLOv8 model is implemented on the NVIDIA Jetson Orin Nano, a power-efficient embedded platform suitable for real-time applications. We provide a brief overview of the YOLO algorithm, its working principles, and the steps taken to train the drone detection model. In addition, we describe the integration of Doppler radar using the ADALM-PLUTO phased array platform to enhance detection performance through motion-based information. Experimental results demonstrate the potential of combining visual and radar data for reliable drone detection in diverse environments.
This paper presents an algorithm for depth map interpolation used to generate novel views from multiple camera positions. We investigate both a traditional stereo-based approach with two cameras and extend it to a four-camera setup in a rectangular configuration, approximating placement at monitor corners, enabling view synthesis at any point within the rectangular field. Our method employs state-of-the-art neural networks for disparity computation, followed by a pipeline that adapts to both configurations.
Frequency hopping spread spectrum (FHSS) signals span a widely varying frequency band, making it challenging to capture the entire spectrum in a non-cooperative context. Consequently, developing methods to blindly estimate FHSS parameters from only a partial observation becomes essential. This paper presents a robust algorithm for blind parameters estimation for FHSS signals under partial-band observation. The proposed algorithm first obtains a time-frequency image (TFI) of the partially observed FHSS signal and extracts its time-domain energy profile by marginalizing the TFI. By computing the derivative of this profile, the observed hop transitions are highlighted as positive and negative peaks, and correlation is used to reveal the repeating hop cycle. Subsequent peak detection isolates the hop duration, enabling the estimation of hop sequence length. Simulation results demonstrate that the proposed method achieves high estimation accuracy under partial-band observations, providing a valuable tool for spectrum monitoring and security applications in diverse non-cooperative contexts.
The detection and classification of banknotes can make life more accessible for visually impaired individuals. This paper aims to propose a method for segmenting Brazilian Real banknotes in images with varying backgrounds and lighting conditions, using an adaptive pre processing approach. The methodology includes adaptive histogram analysis and modification in the V channel, followed by the application of the Sobel filter to identify contours and segment the region of interest. A set of 72 new images made available was used to evaluate the effectiveness of the approach, that achieved an accuracy of 98.2% in detecting the banknote region under different backgrounds and lighting conditions. The research presents an effective approach to minimizing the effects of lighting while preserving the edge characteristics of the banknotes.
Alzheimer’s disease (AD) is a major global health challenge, and early diagnosis is crucial for effective treatment. However, traditional diagnostic methods struggle to extract meaningful patterns from medical data. This study proposes a deep learning approach that integrates brain magnetic resonance imaging and structured clinical data to improve classification accuracy. We fine-tune MobileViT, a hybrid model combining convolutional neural networks (CNNs) and transformers, to analyze sagittal MRI slices, which capture longitudinal brain structures but are underutilized due to lower classification performance and fewer labeled datasets. Additionally, we integrate TabNet to process structured clinical data, including patient age, Socioeconomic Status (SES), Mini-Mental State Examination (MMSE), Clinical Dementia Rating (CDR), and volumetric brain measures. By combining MRI and clinical data through multimodal learning, our approach achieves a peak test accuracy of 99% on the OASIS-1 (416 participants) and OASIS-2 (150 participants) datasets, using a 70/15/15 train-validationtest split. This outperforms unimodal methods, where MRI-only and clinical-data-only models achieve 94% and 97% accuracy, respectively.
This paper provides an overview of the impact of certain types of galvanic faults that may occur in cables in drop wires on the transmission of IPTV services over VDSL2 lines. The study examines how severe serial and shunt galvanic faults affect the physical layer management parameters of VDSL2 transceivers and IPTV QoS parameters.
Modern tactical and emergency operations require secure and intelligent communication systems capable of adapting across complex, multi-domain environments. Legacy platforms such as SINCGARS, EPLRS, and CERT were not built to support real-time encrypted data exchange or large-scale coordination under modern threats. To address these challenges, this paper presents MobileSecureComm, a unified platform that integrates artificial intelligence, quantum-resilient encryption, and domain interoperability. Designed for defense, disaster relief, and peacekeeping, the platform offers dynamic routing, anomaly detection, and bandwidth optimization through edge AI. Unlike prior systems that address isolated features, MobileSecureComm delivers an integrated architecture validated through simulations. The results demonstrate its readiness for secure, scalable, and mission-critical communication in evolving operational landscapes.
Graffiti are omnipresent in many cities around the world. Depending on the type of graffiti, these can degrade the aesthetic appearance of urban environments (the vandalistic ones) or even help to promote tourism to a location (the artistic ones). The correct detection of different graffiti types and corresponding covered materials using urban images, will allow government entities to appropriately handling these environments. This paper uses a Deep Learning approach to accurately detect graffiti in complex images and classify them as vandalistic, artistic or wildstyle. Additionally, some common materials that are under the vandalistic graffiti are also classified. Two YOLO-based solutions are proposed and compared using a created dataset of urban graffiti images. This work also provides an experimental analysis on how some types of image distortions (in particular, impulse noise, image blur, random occlusions and image darkening, respectively) impact on the graffiti detection performance task.
This study evaluates the performance of traditional texture descriptors and dissimilarity-based methods for pulmonary disease classification. The dataset consists of 10,631 publicly available images categorized into four classes: pneumonia, tuberculosis, COVID-19, and healthy. Feature extraction was performed using Local Binary Patterns (LBP), and classification was conducted with K-Nearest Neighbors (KNN), Random Forest (RF), Support Vector Machines (SVM), Convolutions Neural Network (CNN) Resnet, Vgg16, EfficientNetBO, and dissimilarity-based approaches (spatial, vectorial, and embeddings). Additionally, a hierarchical process was applied to enhance traditional classification methods. The results indicate that spatial dissimilarity achieved the highest accuracy (0.977) and the most efective using only 500 images per class obtain a 0.972 accuracy, demonstrating its effectiveness in reducing data requirements while maintaining high performance.
A persistent challenge in machine learning is the identification of samples from unknown classes within test datasets—a task essential for improving class recognition and open set classification. This study investigates dissimilarity-based techniques as a means to address this issue, leveraging their conceptual alignment with open space scenarios, where samples may not belong to any known class. We propose a handcrafted approach to accurately detect such samples. Experiments were conducted on a synthetic feature dataset arranged in a grid, incorporating varying inter-class distances and noise levels. The method evaluates whether a sample exhibits sufficient similarity to known classes; lacking this, it is flagged as potentially unknown. For comparison, we also assess a probability thresholding strategy using conventional classifiers. The dissimilarity-based approach achieved a misclassification rate of approximately 16%, compared to nearly 37% using probability thresholds. Results indicate that dissimilarity methods are more robust in open set contexts, maintaining consistent performance despite variations in noise and class separation—an advantage not observed with probability-based method.
