U. Kulsoom, M. Bendechache, and F.G. Glavin,School of Computer Science, University of Galway, Ireland
Pathology reports are essential documents physicians use to establish a diagnosis and for- mulate a treatment plan for a specific health condition or disease. The significance of these reports is particularly pronounced in the context of cancer. The accurate classification of these reports is essential for optimising clinical decision-making, ensuring timely interventions, and maintaining high-quality patient care. In this work, we propose a custom Bio+Clinical BERT-based multiclass classification approach for pathology reports, showing a strong performance in accurately distinguishing between 32 different cancer tissues. Furthermore, we examined the interpretability of the local model’s decisions using LIME and iden- tified the domain-specific features that influence the classification results. Our results demonstrate that high-performance transformer models can maintain transparency in clinical settings. Our interpretable framework enables pathologists to assess model outputs against established diagnostic criteria, facilitating the responsible integration of clinical language processing systems into clinical workflows.
BERT, NLP, Pathology Reports, Interpretability, Text Classification, LIME. ex
Vincent Kouakou, Nangui Abrogoua University, Ivory Coast
Let K = Q( ) be a quartic number Öeld, where Irr( ) = X4 + pX2 + qX + r, stands for the irreducible polynomial of . Let stand R( ; X) = X3 2pX2 + (p 2 4r)X + (4pr q 2 ) for its ps-cubic resolvent, and R(( ; X) = X3 pX2 4rX = q 2 for its sp-cubic resolvent. In this contribution, by using bothly some results of Kappe and Warren, we give new su¢ cient and necessary conditions using a 4-parametersíconditions (involving the both case, q = 0 and q 6= 0); for the well known but important quadratic subÖeld Q( p p 2 4r) to be or not, a subÖeld of K: We give then, through explicite expressions, the complete family of non galoisian quartic number Öelds containing the subÖeld Q( p p).
ywords. Resolvent Cubic, Irreducible Polynomial, Galois group, Non Galoisian ex
Hung-Manh Nguyen, Khang-Hoang Nguyen Vo, Thuan-Van Tran, Tuan-Minh Le, Cam-My Nguyen Thi, and Thu-Vo Le Minh, FPT University, Ho Chi Minh City, Vietnam
This paper presents a novel deep learning framework for low-light image enhancement, inte- grating Retinex theory with an extended U-Net architecture. The proposed model decomposes images into reflectance and illumination components, leveraging residual learning, attention mechanisms, and multi- scale fusion for superior brightness restoration and detail preservation. A four-channel input, including a maximum intensity channel, enriches feature representation, while spectral normalization ensures stable training. Additionally, adversarial training with GANs enhances illumination realism. To further refine enhancement quality, we introduce a multi-objective loss function optimizing structural consistency and color fidelity. Extensive evaluations on the LOLv1 and LOLv2 datasets demonstrate that our method out- performs state-of-the-art approaches in PSNR and SSIM, achieving robust low-light enhancement. Future research will focus on real-time efficiency and broader real-world applications.
Nhat Minh Nguyen, Phuc Minh Hoang Lai, Linh Van Nguyen, Tuan Minh Le, and Thu Minh Vo Le, FPT University, Ho Chi Minh Campus, Vietnam
Accurate identification of medicinal plants is crucial for preserving traditional knowledge, supporting pharmaceutical research, and promoting sustainable agriculture. In this study, we introduce Advanced-Leaf-CNN (ALC), a deep learning model designed for high-precision classification of medici- nal plant species. We integrate convolutional neural networks (CNNs) with an attention mechanism to enhance feature extraction. This approach refines spatial representations and improves classification per- formance. The model architecture consists of multiple convolutional layers with batch normalization and ReLU activation, followed by attention layers that selectively highlight key features critical for distinguish- ing between plant species. A global average pooling layer reduces dimensionality, while fully connected layers with dropout regularization prevent overfitting. To improve generalization, we apply extensive data augmentation, including random rotation, flipping, color jittering, and random erasing. To address class imbalance, we employ Random OverSampling, ensuring a well-balanced dataset. The model is trained using the AdamW optimizer combined with a Cosine Annealing learning rate scheduler for optimal con- vergence. Experimental results—evaluated through accuracy, precision, recall, F1-score, and AUC-ROC - demonstrate that our approach achieves high classification accuracy, making it a promising solution for automated medicinal plant identification. The proposed model can be seamlessly integrated into mobile or IoT-based applications for real-world deployment in agriculture and healthcare.
Medicinal plant classification, deep learning, convolutional neural network, image augmen-tation, healthcare, argiculture, IoT integration.
Bao-Nguyen Quoc , Huy-Ho Huu , Khang-Nguyen Hoang Duy , Thu-Le Vo Minh, FPT University, Ho Chi Minh Campus, Vietnam
Medical Visual Question Answering (VQA) systems have the potential to revolutionize clinical decision-making by automat- ically interpreting medical images and providing accurate, contextually relevant answers. However, the complexity of medical language, variability in imaging modalities, and limited annotated data pose significant challenges. In this paper, we propose a unified multi-dataset framework for medical VQA that leverages state-of-the-art pretrained transformers to bridge the gap between vision and language. Our approach integrates a Vision Transformer (ViT)-based BLIP model for robust image feature extraction, a BERT-based encoder for high-quality clinical question understanding, and a BioGPT-based autoregressive decoder for domain- specific answer generation. To effectively fuse multimodal information, we employ dedicated projection layers that align visual and textual features into a common embedding space, guided by a question type classifier. The training process combines contrastive learning, image-text matching (ITM) loss, and an autoregressive language modeling (LM) loss to enforce cross-modal alignment and promote coherent answer generation. We pretrain our model on the PathVQA and VQA-RAD datasets and fine-tune it on downstream medical VQA tasks. Experimental results demonstrate competitive performance, with overall accuracies of 69.3% on VQA-RAD and 56.1% on PathVQA, underscoring the generalizability and robustness of our approach. Our framework advances the state-of-the-art in medical VQA and lays the foundation for more reliable AI-assisted diagnostic systems.
Vision Transformer (ViT), Medical VQA, Transformer, PathVQA, VQA-RAD.
Mario DeSean Booker, USA
The rapid digitalization of financial systems has precipitated a critical examination of cryptographic technologies within central banking infrastructures, with the United States Federal Reserve at the forefront of this technological transformation. This paper provides a comprehensive analysis of the potential implementation of cryptography in the Federal Reserve System, exploring the intricate intersection of technological innovation, monetary governance, and systemic financial security. By conducting a detailed comparative analysis between traditional central banking mechanisms and emerging cryptocurrency systems, it illuminates the complex challenges and unprecedented opportunities presented by cryptographic technologies. The study demonstrates that integrating cryptography within the Federal Reserve is not merely a technical upgrade but a fundamental reimagining requiring a holistic, collaborative approach balancing innovation with institutional integrity.
Le Vo Minh Thu, Nguyen Le Gia Bao, Nguyen Thi Minh Thu, Nguyen Cao Cuong, Nguyen Hong Nhung, FPT University, Ho Chi Minh, Vietnam
Traditional shoe recognition models, constrained by manual feature extraction, struggle with low accuracy and poor scalability. This paper introduces XAI-KickVision, a hy- brid AI-driven footwear recognition system that utilizes multimodal learning. Our approach combines ResNet18 for local feature extraction, Vision Transformer (ViT) for global image analysis, and DistilBERT for contextual text representation [1]. To enhance interpretability, we integrate Explainable AI (XAI) using Grad-CAM++, providing visual explanations for model decision. Evaluated on the UT Zappos50K dataset, our model outperforms the stan- dalone CNN and ViT architectures, achieving a classification accuracy of 96%, an F1 score of 94. 5% and a recall of 94. 1%. The system demonstrates robust performance in both product identification and recommendation tasks, offering a scalable and interpretable solution for AI-driven e-commerce applications.[12]
Footwear Recognition, Product Description, Hybrid AI system, Multimodal Learning, Explainable AI.
Nicole Haddad, St Albans Secondary College, Victoria, Australia
This research paper uses an autoethnographic methodology to explore the ways in which Digital Natives, or learners, and Digital Immigrants, or educators, clash in classroom settings due to curricular demands and pre-conceived notions of how an education ought to be delivered. This, in turn, as this paper finds, causes a risk of student refusal, as well as increases the challenge in teacher retention. Reflecting on my use of visual literacies, introspection, as well as my recent experiences as a year-level coordinator, I revisit the ways in which trauma informed practice and the embrace of new technologies have enhanced my connection with my learners, and in turn, their pedagogical endeavours. This paper will posit that once an educator enters a classroom setting with an open mind, and a willingness to coalesce the microcosm that is the classroom, with the greater settings beyond it, the education system will be improved, and in turn will invite learners to improve themselves.
Digital Native, Digital Immigrant, Educators, Education, Multimedia, Japanese Philosophy, Ikigai, Flow, Multiple Intelligences, Differentiation, School Refusal, Procrastination, Visual Literacies, Teacher Education, Curriculum, Autoethnography, Learners.
Jahnavi Vellanki1,1Lab Validation Specialist, Indiana, USA
Middleware validation in life sciences is critical in ensuring data integrity, system interoperability, and compliance with regulatory standards. The survey paper on integrating AI technologies into middleware validation covers advancements, challenges, and emerging opportunities. It gives an overview of the functions of middleware and its applications in life sciences, including some unique challenges related to scalability, data security, and compliance issues. Major middleware reliability and performance improvements have already been demonstrated with machine learning, natural language processing, and automated testing techniques. However, the current paper evaluates extant AI-driven frameworks, shows their strengths and weaknesses, identifies gaps in current research and implementations, and lays out future directions for research related to quantum computing, AI advances, and the ethics of deployment. The paper concludes by urging collaboration toward making AI more adopted in middleware validation, hence scaling, compliance, and efficiency in life sciences and beyond.
Machine Learning, Risk Assessment, GxP Framework, Predictive Accuracy, Regulatory Compliance, Feature Importance.
Anh Thi Minh Tran, Dung Tri Nguyen, Anh Lam Nguyen, Cac Thoai Tran, and Thu Le Minh Vo, FPT University, Ho Chi Minh Campus, Vietnam
Early detection of plant diseases is critical to maintaining agricultural produc- tivity, yet implementing efficient solutions on mobile and web platforms poses significant challenges due to limited resources. This study presents an evaluation of the DINOv2-ViT model, specifically fine-tuned for plant leaf disease detection. For the comparative analysis, pretrained lightweight AI models, which have demonstrated robust performance on ImageNet and have been fine-tuned on the plant leaf disease dataset, were selected. The selected models include the DINOv2 base, DINOv2 small, CAS-ViT t, and the ResNet-50-based SimCLRv1. Model performance is evaluated using multiple metrics: Accuracy, AUC-ROC, F1 score, and inference time. The fine-tuning process is streamlined, focusing only on training with the plant disease dataset to maintain low system demands while achieving high classification accuracy. The results highlight the feasibility of deploying plant disease detection systems in resource-constrained environments, providing a practical tool for agricultural solutions.
Plant Disease Recognition, Deep Learning, Machine Learning, Computer Vi- sion, Vision Transformer, Self Supervised Learning Agricultural Technology.
Emmanuel Agyei1, Zhang Xiaoling1, Ama Bonuah Quaye2, Odeh Victor Adeyi1, and Joseph Roger Arhin1, 1School of Information and Communication Engineering, University of Electronic Science and Technology, China, Chengdu, 610054, Sichuan, China, 2School of Public Administration, University of Electronic Science and Technology of China
Machine Translation (MT) for low-resource languages like Twi remains a significant challenge in Natural Language Processing (NLP) due to limited parallel datasets. Traditional methods often struggle, relying heavily on high-resource data, and fail to adequately serve low-resource languages. To address this gap, we propose a fine-tuned T5 model trained with a Cross-Lingual Optimization Framework (CLOF), which dynamically adjusts gradient weights to balance low-resource (Twi) and high-resource (English) datasets. This framework incorporates federated training to enhance translation performance and scalability for other low-resource languages. The study utilizes a carefully aligned and tokenized English-Twi parallel corpus to maximize model input. Translation quality is evaluated using SPBLEU, ROUGE (ROUGE-1, ROUGE-2, and ROUGE-L), and Word Error Rate (WER) metrics. The pretrained mT5 model serves as a baseline, demonstrating the efficacy of the optimized model. Experimental results show significant improvements: SPBLEU increases from 2.16% to 71.30%, ROUGE-1 rises from 15.23% to 65.24%, and WER decreases from 183.16% to 68.32%. These findings highlight CLOFs potential in improving low-resource MT and advancing NLP for underrepresented languages, paving the way for more inclusive, scalable translation systems.
Low-Resource Machine Translation; Twi Language; Federated Learning; Cross-Lingual Learning Fine-Tuning.
Xiaoqian Hu,Master of Information Technology, the University of New South Wales, Kensington,NSW
In Recent Years, Digital Technologies Have Made Significant Strides In Augmenting-Human-Health, Cognition, And Perception, Particularly Within The Field Of Computational-Pathology. This Paper Presents A Novel Approach To Enhancing The Analysis Of Histopathology Images By Leveraging A Mult- modal-Model That Combines Vision Transformers (Vit) With Gpt-2 For Image Captioning. The Model Is Fine-Tuned On The Specialized Arch-Dataset, Which Includes Dense Image Captions Derived From Clinical And Academic Resources, To Capture The Complexities Of Pathology Images Such As Tissue Morphologies, Staining Variations, And Pathological Conditions. By Generating Accurate, Contextually Captions, The Model Augments The Cognitive Capabilities Of Healthcare Professionals, Enabling More Efficient Disease Classification, Segmentation, And Detection. The Model Enhances The Perception Of Subtle Pathological Features In Images That Might Otherwise Go Unnoticed, Thereby Improving Diagnostic Accuracy. Our Approach Demonstrates The Potential For Digital Technologies To Augment Human Cognitive Abilities In Medical Image Analysis, Providing Steps Toward More Personalized And Accurate Healthcare Outcomes.
Image Captioning, Neural Network, Histopathology Images, Natural Language Processing.
Wallas Bruno S. Lira, Gilton José Ferreira da Silva, Silvio Mario Felix Dantas, Barbara Cristina Silva Rosa, Cassia Regina D’Antonio Rocha da Silva, Federal University of Sergipe (UFS), Brazil
The dynamic process of discovering and documenting software requirements demands effective approaches. This study explores, through a survey conducted via Google Forms using the snowball sampling technique, the application of Design Thinking (DT) stages in Requirements Engineering (RE). The findings indicate that integrating DT phases enhances Requirements Engineering activities, although maintaining architectural quality throughout the agile lifecycle remains challenging. It concludes that the synergy among DT, Requirements Engineering, and Software Architecture significantly improves the effectiveness of agile software projects.
Design Thinking, Requirements Engineering, Software Architecture, Software Development, Design Management.
Wallas Bruno S. Lira, Gilton José Ferreira da Silva, Silvio Mario Felix Dantas, Barbara Cristina Silva Rosa, Cassia Regina D’Antonio Rocha da Silva, Federal University of Sergipe (UFS), Brazil
This article investigates the feasibility of implementing a Decision Support System (DSS) in Public Higher Education Institutions, focusing on survey-based data collection. By integrating Business Process Model and Notation (BPMN) for process representation, Goal Question Metric (GQM^+) for defining objectives and indicators, and Business Intelligence tools for data analysis, the study provides a robust foundation for evidence based and strategic decision-making. The results show that DSS adoption optimizes both tangible and intangible asset management and generates valuable insights for the continuous improvement of educational operations. The survey methodology ensures reliable and comprehensive data, reinforcing the significant potential of DSS to enhance administrative processes and increase efficiency in higher education institutions.
Decision Support Systems, Asset Management, Higher Education Institutions, Survey Research, Operational Efficiency, Business Intelligence.
