Breast cancer remains a significant global health concern, where early and accurate diagnosis is paramount for improving patient survival rates. This paper presents a comparative analysis of two deep learning architectures, the Convolutional Neural Network (CNN) based ResNet-101 and the Vision Transformer (ViT), for the classification of breast ultrasound images into benign, malignant, and normal categories. Addressing the common challenge of limited data, we employed a data augmentation strategy to expand a benchmark dataset of 780 images to over 10,000 images, creating a robust training set. Both models were trained on this augmented dataset, achieving test accuracies of 98.64% for the Transformer model and 97.57% for Resnet-101 model. The result indicates that the ViT model achieved higher accuracy than the ResNet-101. Furthermore, the existing Deep learning models are black box models. To enhance model transparency and build clinical trust, Gradient-weighted Class Activation Mapping (Grad-CAM), an Explainable AI (XAI) technique, is utilized to generate visual heatmaps, highlighting the specific regions in the ultrasound images that were most influential in the models’ diagnostic decisions. The proposed model harnesses GPU-based parallel infrastructure.
Currently, data from Hitachi machines' Digital Imaging and Communications in Medicine (DICOM) reports is used manually, which is time-consuming and error-prone. This limits the effectiveness of the analyses needed to manage radiation doses and improve medical practices. This paper therefore proposes an automated solution for the rapid and reliable extraction of key information, including CTDI (Computed Tomography Dose Index) and PDL (Product of Dose Length) indices, to facilitate the determination of Diagnostic Reference Levels (DRL). The methodology combines several steps: collecting real data from Picture Archiving and Communication System (PACS) reports; studying the structure of DICOM reports; developing an extraction algorithm in Python; and visualising the results via a web interface. The solution's architecture is based on the Django framework for data processing and Angular for the interactive presentation of results. The results obtained demonstrate the effectiveness of the platform in automating the extraction and analysis of DICOM data, significantly reducing processing time while providing healthcare professionals with clear, intuitive visualisation. This solution represents a significant advance in radiation dose management, transitioning from manual processes to an automated system that is faster and aligns with international radiation protection recommendations.
<strong><em>Abstract—Skin diseases and disorders impact a significant portion of the global population, representing a nonfatal but substantial disease burden. Accurate and timely diagnosis is often challenging, particularly in low-resource settings with little access to specialists. To address this, an image-based skin disease detection system utilizing an ensemble of deep learning models YOLOv8, YOLO-NAS, and EfficientDet was developed. The system classified five common skin conditions Acne Vulgaris, Eczema, Melasma, Rosacea, and Shingles using a publicly available, annotated dataset enhanced by preprocessing and augmentation. Outputs from individual models were reviewed by a dermatologist for clinical reliability. The ensemble-based approach reached high levels of precision, recall, and mean average precision @0.5(mAP@0.5), mean average precision @0.5:0.95(mAP@0.5:0.95) demonstrating robust performance in screening applications. The solution was successfully deployed as a proof-of-concept mobile application for early detection and support, especially in underserved areas. Ethical considerations regarding data privacy and dataset bias were addressed throughout the study.</em></strong>
TB is still among the major global health problems especially in low- and medium-income economies whereby access to speedy and precise diagnosticities is still poor. We present a Multi-Modes Stacking Ensemble on Tuberculosis (MMSE-TB), a model that combines three modalities which are diverse and complementary that are used to detect tuberculosis; these include chest X-ray, cough audio, and clinical text. The modalities are modeled with separate architectures of deep learning: a Feature-Map-Normalized CNN which extracts radiological features, a Capsule Network which predicts patterns with space-temporal correlations of a cough spectrogram and a BioBERT-generated encoder which predicts features of clinical text with semantic meaning behind them. Models are combined using dynamically-optimized weighting program using Mayfly Optimization Algorithm to contribute dynamically and confidently and reliably with all modalities. Experimental analysis has demonstrated that this tri-modal ensemble has a drastic positive effect on the accuracy of diagnostic performance as well as a decrease in false negative rate and a high quality of robustness even in heterogeneous data sets. This architecture has a scaled, clinically flexible way of screening TB through artificial intelligence.<strong> </strong>
UML is a general modeling language that offers standardized modeling notations. Some specific domains require to be modeled by specific notations other that proposed in standard UML. Variability managing is one of these specific domains that necessitate specific modeling notations. By this contribution we aim to expand our previous works in extending UML notations in variability management domain. Our main goal is to create a complete UML profile with extended notations of all UML diagrams for modeling variability. In this paper we present new UML “variability” notations (stereotypes, constraints and tagged values) obtained by extension of standard UML ones. These new notations will be used in constructing adapted and appropriate “variability” diagrams. The list of studied UML diagrams contains: four types of structural diagrams (Object diagram, Component diagram, Package diagram, Profile diagram) and three types of behavioral diagrams (Statechart diagram, Communication diagram, Timing diagram).
The main issues with traditional 3D printing technology are human involvement, physical contact, and limited remote control capabilities. A drawback of this 3D printing is that it is not fully automated or accessible to the public, which slows down the process. This paper aims to design and develop a remote 3D printing system to improve the standard process. The system also includes wireless connectivity, allowing users to upload design files wirelessly to the printer, eliminating the need for USB flash drives or SD cards. Additionally, incorporating a camera module allows the printing process to be streamed in real-time via a dedicated website or mobile app. Our system consists of the original QIDI Tech 1 3D printer, a PC, a Raspberry Pi 4 Model B 4GB, and the Raspberry Pi camera module v2. The QIDI Tech 1 3D printer features a closed-source motherboard with an ATmega 2560 microcontroller. Software development is divided into frontend and backend. The frontend, visible to users, manages the user interface and interactions. It is built with HTML (HyperText Markup Language) for content structure, JavaScript for interactivity, and CSS (Cascading Style Sheets) for styling, all within the React JavaScript framework. The backend manages server-side operations and hardware interactions, including the 3D printer. Its core is the REST (REpresentational State Transfer) API (Application Programming Interface), which connects various services.
<div style="text-align:justify">Drones continuously generate flight log data containing valuable information about flight states, sensor readings, and system events. These logs are critical forensic artifacts for investigating incidents such as crashes or operational anomalies. However, previous forensic studies seldom explore the semantic context embedded within the human-readable flight log messages. This paper presents a Transformer-based named entity recognition (NER) framework to automatically extract meaningful entities such as events and issues from drone flight logs. To enhance generalization and simplify downstream analysis, the six original entity types defined in the DroNER dataset were merged into two higher-level classes Event and NonEvent to represent operational and anomalous contexts, respectively. We fine-tuned three pre-trained language models: BERT, DistilBERT, and SqueezeBERT, on this adapted dataset curated for drone forensic analysis. Experimental results show that SqueezeBERT, with only 51M parameters, achieves an F1 score of 96.79%, comparable to BERT’s 98.00%. This study is the first to benchmark. These findings suggest that lightweight Transformer architectures are highly promising for edge-level forensic NLP applications, enabling real-time log-based investigation automation on resource limited forensic platforms.</div>
Financial literacy is an essential life skill in today’s increasingly digital economy. This study investigates the relationship between financial literacy—measured through knowledge, behavior, and attitude—and students’ academic performance in finance-related courses at a major university in Bandung, Indonesia. Using a quantitative design and multiple linear regression analysis on 250 undergraduate respondents, the research explores how cognitive, behavioral, and affective components of financial literacy contribute to learning outcomes. The findings reveal that although financial literacy dimensions positively correlate with academic performance, their direct effects on GPA are limited. These results suggest that knowledge alone does not translate into improved academic outcomes without supporting behavioral and motivational factors. The study highlights the potential of IoT-based learning platforms, such as Android financial education applications, to strengthen engagement, financial awareness, and self-regulation. Integrating IoT-enabled tools into financial education could enhance experiential learning, making financial literacy more interactive and effective in improving students’ long-term financial competence.
The rapid growth of coffee shops in Makassar has raised environmental concerns, shifting consumer attitudes toward sustainability. However, a gap remains between attitudes and behavior, where high environmental awareness does not always translate into loyalty. This study is structured to investigate the core factors driving customer loyalty by combining the roles of Environmental Awareness and Sustainability Practices and positioning Brand Image as a variable that mediates the relationship. Through a quantitative research approach, A total of 260 valid responses were obtained from specialty coffee shop customers in Makassar. The resulting structural model and hypothesis testing were performed using PLS-SEM (Partial Least Squares Structural Equation Modeling). The findings reveal different mechanisms in the formation of loyalty. Although both independent variables demonstrated significant direct effects on loyalty, the mediation analysis showed contrasting results. Sustainability practices were successfully enhancing customer loyalty via the mediating role of Brand Image, while environmental awareness had no effect on customer loyalty through brand image, indicating green skepticism. This suggests that for consumers in Eastern Indonesia, operational actions taken by companies are effective in building a credible image rather than mere awareness, which is susceptible to suspicions of greenwashing. These findings provide strategic insight for coffee shop businesses to prioritize tangible evidence of sustainability over superficial branding to gain customer loyalty.
Abstract— This study examines how digital entrepreneurial skills and growth mindset affect business performance of Generation Z entrepreneurs in Indonesia, with entrepreneurial burnout as a mediating factor. Using a quantitative survey of 100 Gen Z entrepreneurs, we applied Partial Least Squares Structural Equation Modeling (PLS-SEM) with 5,000 bootstrap resamples. Constructs (digital entrepreneurial skills, growth mindset, entrepreneurial burnout, and perceived business performance) were measured with validated Likert scales. Results show that both digital skills and growth mindset have significant positive direct effects on performance, and each is significantly negatively related to burnout. In turn, higher entrepreneurial burnout significantly predicts poorer performance. Mediation tests indicate that entrepreneurial burnout partially mediates the effects of digital skills and growth mindset on performance (all indirect confidence intervals exclude zero). In sum, Gen Z entrepreneurs with higher digital capabilities and a stronger learning-oriented mindset experience less burnout and thus achieve better venture outcomes. Theoretically, these findings extend entrepreneurship literature by linking digital competencies and psychological resources to performance outcomes. Practically, the results suggest that Gen Z entrepreneurs and support organizations should emphasize digital skills training and growth-mindset development (e.g. through targeted education and coaching) to enhance venture success and well-being. Appropriate policies might similarly integrate technical and psychosocial support to sustain youth entrepreneurship.
In this paper, novel triple-mode wideband filters based on a compact sixteenth-mode substrate integrated waveguide (HMSIW) resonator are presented. The proposed design achieves a remarkable size reduction of a conventional substrate integrated waveguide (SIW) circular cavity—while maintaining the same resonant frequency. The proposed filters demonstrate excellent potential for improving the performance of advanced telecommunication systems.In this paper, novel triple-mode wideband filters based on a compact sixteenth-mode substrate integrated waveguide (HMSIW) resonator are presented. The proposed design achieves a remarkable size reduction of a conventional substrate integrated waveguide (SIW) circular cavity—while maintaining the same resonant frequency. The proposed filters demonstrate excellent potential for improving the performance of advanced telecommunication systems.
<strong><em>With the growing application of Geographic Information Systems (GIS) in urban management, emergency response, and public services, traditional layer-based thematic map information services</em></strong><strong><em>—</em></strong><strong><em>relying on keyword retrieval and layer superposition</em></strong><strong><em>—</em></strong><strong><em>can no longer meet users</em></strong><strong><em>’</em></strong><strong><em>needs for semantic understanding and intelligent recommendation in cartography. To address this gap, this study proposes a knowledge graph (KG)-driven geographic information recommendation system framework. First, a geographic KG for thematic maps is constructed, focusing on geographic entities and their semantic relationships. High-quality entity relation extraction is achieved using the BERT+BiLSTM+CRF model, while graph embedding representation is implemented via random walk and word2vec to enable high-dimensional matching between user interest vectors and geographic information node vectors. On this basis, a hybrid recommendation model integrating KG semantic reasoning and graph embedding algorithms is designed, and a system prototype with recommendation, visualization, and feedback optimization capabilities is developed. Experimental results demonstrate that the system outperforms traditional methods in both recommendation accuracy(Precision@10: 0.78 vs. 0.58</em></strong><strong><em>–</em></strong><strong><em>0.67 for traditional methods) and processing speed(average response time: 250 ms vs. 370</em></strong><strong><em>–</em></strong><strong><em>410 ms for traditional methods), with strong practicality and scalability. This research achieves innovative breakthroughs in knowledge extraction, hybrid recommendation strategies, and system performance optimization, providing effective support for semantic scenario-oriented geographic information services.</em></strong><br />
