Boris Mederos

Cerebral aneurysms are dilations of intracranial arteries whose early detection is essential for treatment planning and rupture prevention. Automated aneurysm segmentation remains challenging because of severe class imbalance and the small size of many lesions. This study presents a 3D U-Net incorporating deep supervision and an auxiliary classification branch, trained on the public ADAM dataset using skeleton-guided patch extraction. Fixed-size patches (64 × 64 × 64) were extracted from volumes resampled to 0.4 mm isotropic resolution, and performance was evaluated using stratified five-fold cross-validation. The proposed model achieved a Dice coefficient of 0.425 (95% CI: 0.362–0.488) and a false-positive (FP) patch rate of 0.026 (95% CI: 0.017–0.035) on the test set. For aneurysms larger than 2.5 mm, the Dice coefficient increased to 0.548 (95% CI: 0.485–0.612), while the FP patch rate was 0.030 (95% CI: 0.018–0.042), indicating improved performance on clinically relevant lesions. Statistical analysis confirmed that the proposed model was significantly superior to the compared models in both the Dice coefficient and FP patch rate. Although the framework produces patch-level rather than whole-volume segmentations, visual analysis demonstrates accurate localization of aneurysm regions. These results indicate that combining skeleton-guided patch extraction with deep supervision and auxiliary classification improves segmentation performance while providing reliable localization cues for automated aneurysm detection.

The increasing demand for materials with enhanced functionalities has driven the development of multiscale methodologies capable of predicting their effective properties from their heterogeneous microstructure and the interfacial interactions. In this work, we develop a coupled multiscale framework combining the two-scale asymptotic homogenization method with physics-informed neural networks to predict the effective behavior of one-dimensional periodic composites with imperfect interface. Physics-informed neural networks solve the local cell problems arising from the asymptotic homogenization method, and the resulting local fields are subsequently used to compute the homogenized coefficients and reconstruct the formal asymptotic approximation. To capture field discontinuities induced by the imperfect interface, a dual-network architecture with suitable optimization strategies is employed. The one-dimensional configuration serves as a reference framework for systematic numerical validation of the asymptotic homogenization method–physics-informed neural networks methodology through direct comparison with the analytical homogenization solution. Numerical experiments consider different interface parameters κ from 1 to 1 0 7 , stiffness contrasts E 1 ∕ E 2 from 1 to 100, with additional calculations up to 1000, and volume fractions V 1 from 0.1 to 0.9. The predicted effective Young’s modulus shows close agreement with the analytical solution, with relative errors ranging from 6 . 21 × 1 0 − 12 to 3 . 55 × 1 0 − 6 , while reconstructed displacement fields for ε = 0 . 01 − 0 . 5 reproduce the analytical solution with root-mean-square errors between 1 . 50 × 1 0 − 11 and 5 . 1 × 1 0 − 7 . Convergence histories show accurate solutions for all tested configurations, while higher stiffness contrasts increase the sensitivity of the physics-informed neural network accuracy to the combined material and interface parameters. These results demonstrate that the proposed methodology provides an accurate computational framework for predicting the effective behavior of one-dimensional composites with an imperfect interface.

La inteligencia artificial está transformando la interpretación de las mamografías y abre nuevas posibilidades para apoyar la detección temprana del cáncer de mama. Aunque la mamografía es fundamental para identificar la enfermedad antes de que aparezcan síntomas, factores como la densidad del tejido mamario, la fatiga visual y las diferencias de interpretación entre especialistas pueden dificultar el diagnóstico. En este contexto, la IA puede funcionar como una aliada del radiólogo al priorizar estudios, señalar regiones sospechosas y reducir la carga de trabajo. Este artículo presenta las principales formas en que esta tecnología se incorpora al análisis de mamografías, así como sus beneficios y desafíos. Si bien puede contribuir a detectar más casos y hacer más eficiente la atención, también plantea riesgos relacionados con falsas alarmas, sesgos, privacidad y responsabilidad ante posibles errores. Su implementación gradual, transparente y supervisada puede ayudar a aprovechar sus beneficios sin sustituir el juicio clínico y el componente humano de la atención médica.

This work presents an approach for recognizing isolated sign language gestures using skeleton-based pose data extracted from video sequences. A Graph-GRU temporal network is proposed to model both spatial and temporal dependencies between frames, enabling accurate classification. The model is trained and evaluated on the AUTSL (Ankara university Turkish sign language) dataset, achieving high accuracy. Experimental results demonstrate the effectiveness of integrating graph-based spatial representations with temporal modeling, providing a scalable framework for sign language recognition. The results of this approach highlight the potential of pose-driven methods for sign language understanding.

The adaptive filtering theory has been extensively developed, and most of the proposed algorithms work under the assumption of Euclidean space. However, in many applications, the data to be processed comes from a non-linear manifold. In this article, we propose an alternative adaptive filter that works on a manifold, thus generalizing the filtering task to non-Euclidean spaces. To this end, we generalized the least-mean-squared algorithm to work on a manifold using an exponential map. Our experiments showed that the proposed method outperforms other state-of-the-art algorithms in several filtering tasks.