The rapid growth of scientific repositories has made large-scale semantic analysis increasingly dependent on automated methods for discovering latent knowledge structures. This work proposes a scalable hybrid framework that integrates probabilistic topic modeling with contextual embeddings and efficient neural adaptation. The approach combines Latent Dirichlet Allocation (LDA) with BERTopic and SciBERT embeddings. Experiments on arXiv subsets (2,000–10,000 documents) show that BERTopic + SciBERT achieves the best performance ( c_v = 0.65 , Macro-F1 = 0.85), outperforming Top2Vec and CTM. Bootstrap validation confirms statistical significance ( p = 0.001 ). Additionally, LoRA-based summarization improves ROUGE and BERTScore, supporting semantic consistency. Scalability analysis shows near-linear runtime growth and stable performance.
Despite significant technological advances in thermography, its use for early breast cancer detection still faces significant challenges. Indeed, there are static binary classification models that do not reflect a clinical process for cases in which abnormal findings in tissues progressively refine into a benign or malignant diagnosis. In a search for useful applications in this area, in this paper, we introduce the first hierarchical incremental learning framework for breast thermography, which progresses from several learning tasks: T_1 identifies normal/abnormal classes using the database called DMR-IR; then task T_2 learns normal/benign classes using the database called Mendeley, and finally task T_3 which is able to identify classes normal/benign/malignant. Achieved accuracies are: for T_1 96.7% ± 0.75% , for T_2 96.8% ± 0.35% and for T_3 92.48% ± 0.5% . We implemented a foundational RadImageNet deep convolutional neural network (CNN), tuned to classify thermal images, with a dynamic classifier expansion and iCaRL (Incremental Classifier and Representation Learning) rehearsal, to mitigate catastrophic forgetting; a pre-processing of the images was applied to enhance angiogenesis, which allowed that even with a small amount of malignant images (only 4 test images), T_3 achieved a weighted F1=0.92 , prioritizing robust normal/benign cases ( F1>0.93 ) and imitating clinical practice conditions. Ablation experiments showed that each component is necessary for ensuring model stability under limited data (9–60 images per task), providing insight into feasible deployment in clinical conditions without the need for full data replay. This paper proposes a solution for evolving diagnostic tasks that preserves vascular pattern recognition across datasets.
The integration of multi-stage vehicle perception pipelines such as object detection and license plate recognition (LPR) into low-cost embedded systems poses significant challenges in software optimization and computational efficiency. This paper presents a distributed edge-assisted architecture for an Advanced Driver Assistance System (ADAS), targeting deployment on a Raspberry Pi 5 equipped with a Hailo-8 AI accelerator. The system interconnects three specialized nodes via MQTT: an edge perception node, a central processing node for LPR and OCR, and a cognitive node integrating a Large Language Model for natural language scene description. The perception pipeline employs a coarse-to-fine strategy, where a general-purpose detector first identifies vehicle regions of interest, narrowing the search space for a specialized high-precision license plate detector. Model optimization via a PyTorch-ONNX-TFLite-HEF conversion pipeline yielded an 88
Alzheimer’s disease (AD) affects speech, producing disfluencies that can serve as early, non-invasive biomarkers of cognitive decline. This work proposes a graph-based approach that models impaired speech as token-level co-occurrence graphs enriched with special nodes capturing filled pauses, repetitions, and self-corrections. Transcripts are obtained automatically, encoded with fine-tuned BERT embeddings, and transformed into graphs that explicitly highlight disfluent patterns. Experiments on the ADReSSo 2021 dataset demonstrate that GNN models outperform a fine-tuned BERT as well as other unimodal baselines, with a GCN achieving the best performance. Explainability analysis using Integrated Gradients confirms that the model successfully leverages disfluent nodes and specific lexical patterns for classification, allowing us to conclude that explicitly encoding local disfluency structures within graph representations provides a more effective and interpretable approach for AD detection than using sequential models.
Subcutaneous mycoses are chronic infections that may lead to disability and delayed treatment when diagnosis is not timely. This work explores deep transfer learning as a reproducible baseline for multiclass classification of clinical images into four categories (control, chromoblastomycosis, sporotrichosis, and mycetoma). Starting from 99 original images, we generated mild augmented variants to increase variability and trained three ImageNet-pretrained CNN backbones (VGG16, ResNet50, and EfficientNetB0) under a shared protocol with a fixed stratified split and imbalance-aware training (balanced sampling or class-weighted loss). On the test set (N = 62), the models achieved accuracies of 89 F_1 scores up to 87
