Álvaro Anzueto-Ríos

En este trabajo se analiza cinemáticamente el pereiópodo del cangrejo terrestre Gecarcinus lateralis para obtener los parámetros que describen su marcha. A partir de estos datos se construye un modelo matemático del ciclo de marcha promedio mediante una red neuronal multicapa, el cual se emplea para el diseño y construcción de un mecanismo de barras tipo Klann capaz de imitar su desplazamiento lineal. Finalmente, se compara el comportamiento del sistema mecánico con el del organismo biológico mediante métricas espaciales y angulares. Los resultados muestran una semejanza global del 73 % y del 65 % en condiciones de terreno irregular.

Computer models and protocols commonly used for superficial thermotherapy assume anatomically symmetric tissue geometries. However, anatomical variability, particularly the variability of the adipose tissue thickness, may have a significant impact on heat transfer and cause non-uniform thermal doses in the patient population. This paper studies the influence of anatomical asymmetry due to adipose tissue thickness on heat propagation during superficial thermotherapy based on combined computational and experimental analyses. A Finite Element Model of the lower limb was designed with anatomically representative layers and tissue thermal properties. Parametric simulations were conducted for medium (M) and extra-large (XL) models at hot-pack temperatures (38–44 °C) for 15 min. The simulation results were compared with experimental measurements on ex vivo porcine tissue. The results showed that the thickness of adipose tissue significantly affected deep-tissue heating. When the same heating condition was applied, the boundary temperature of the muscle was increased by ~1.38 °C for the 44 °C hot-pack in the M-size model and was restricted to 0.21 °C for the XL-size model. The experimental muscle temperature increased from 22.26 °C to approximately 23.3 °C after 15 min. Under the same initial temperature and nominal 38 °C heating condition, the perfused and zero-perfusion M-size simulations reached approximately 26.8 °C and 30.5 °C, respectively; these endpoint differences support a qualitative comparison rather than quantitative validation. These results show that anatomical asymmetry has a strong effect on heat transfer during superficial thermotherapy and challenge the precept that standard heating protocols provide similar thermal doses in different body morphologies.

Computer models and protocols commonly used for superficial thermotherapy assume anatomically symmetric tissue geometries. However, anatomical variability, particularly the variability of the adipose tissue thickness may have a significant impact on heat transfer and cause non-uniform thermal doses in the patient population. This paper studies the influence of anatomical asymmetry due to adipose tissue thickness on heat propagation during superficial thermotherapy based on combined computational and experimental analyses. A Finite Element Model of the lower limb was designed with anatomically representative layers and tissue thermal properties. Parametric simulations were conducted for medium (M) and extra-large (XL) models at hot-pack temperatures (38–44 °C) for 15 min. The simulation results were verified experimentally on ex vivo porcine tissue. The results showed that the thickness of adipose tissue significantly affected deep-tissue heating. When the same heating condition was applied, the boundary temperature of the muscle was increased by ~1.03 °C for the 44 °C hot-pack in the M-size model and was restricted to 0.21 °C for the XL-size model. The experimental measurements indicated an increase of 1.16 °C in the muscle temperature, with an average difference of 0.81 °C between numerical predictions and experimental measurements. These results show that anatomical asymmetry has a strong effect on heat transfer during superficial thermotherapy and challenge the precept that standard heating protocols provide similar thermal doses in different body morphologies.

Training recurrent neural networks with multiple feedback connections is challenging due to the complexity of backpropagation through time, gradient instability, and the need to redesign optimization routines when the recurrence topology changes. This paper introduces a metaheuristic framework for training recurrent neural networks with arbitrary topology and heterogeneous delays. The framework provides a unified parameter encoding and decoding, automatic identification of recurrent parameters, and topology-aware constraints for recurrent and feedforward weights. By separating the network model from the optimizer, it enables different metaheuristics to be applied with minimal modification. The approach is evaluated on the Adding problem, Thermostat control, Temporal function approximation and time-series prediction, using Simulated Annealing and Equilibrium Optimizer. Results show that the framework effectively supports different recurrence configurations and offers a competitive alternative when gradient-based training is unstable or difficult to implement.

The segmentation of the vascular network in fundus images is a key step in the diagnosis and monitoring of ophthalmological and systemic pathologies, such as diabetic retinopathy (RD). However, the reliability of automated methods is challenged by inherent image issues, mainly non-uniform illumination, low contrast, and noise. This work proposes a methodology for the automated preprocessing and segmentation of fundus images, addressing these problems through an adaptation and regularization strategy. The method begins with the extraction of the green channel (G) to maximize contrast. Then, a Gaussian filter is applied, with its smoothing parameter (σ) optimized for each image by balancing noise reduction and detail preservation using the Variance of Laplacian (VoL) and High-Frequency Energy Residual (HFER). Next, a homomorphic filter is applied to correct illumination, with its parameters dynamically adjusted using a Particle Swarm Optimization (PSO) algorithm, employing Shannon entropy as the objective function. Finally, the Frangi filter is used to enhance tubular structures, followed by controlled binarization. The method was validated on the public DRIVE and STARE databases, where the homomorphic optimization stage showed a consistent increase in image entropy (e.g., from 5.7 to 6.2 bits/pixel), improving contrast. The proposed adaptive preprocessing strategy helps improve the reliability of vascular segmentation and, consequently, the diagnosis of related pathologies.

This paper presents a systematic design methodology for a soft pneumatic actuator inspired by the prehensile motion of the Octopus vulgaris tentacle. The approach integrates biomimetic analysis of the biological model, Finite Element Analysis (FEA), and experimental validation. A Yeoh hyperelastic model was used in ANSYS to simulate the mechanical behavior of EcoFlex 00-30 silicone. An iterative design process is detailed, progressing from a simple single-chamber actuator to a more complex, segmented multi-chamber model. A physical prototype of the initial design was fabricated and tested under a pressure of 12 psi ($\approx \mathbf{8 3 ~ k P a}$) to validate the simulation's predictive capabilities. Results demonstrate that the multi-chamber configuration provides enhanced control over bending curvature compared to the single-chamber design. The final proposed model, featuring a tapered geometry, achieves a controlled coiling trajectory under simulated sequential actuation, emulating the natural tentacle's wrapping motion. This work establishes a validated, simulationdriven framework for developing sophisticated biomimetic soft robots and provides a robust design for a functional octopusinspired manipulator.

This work presents an automatic system for the semantic segmentation of cutaneous macules using an architecture based on an Autoencoder. The dataset used was obtained from Kaggle and manually segmented by the author for the purpose of generating ground truth masks. The proposed architecture incorporates Inception V1 modules in the encoder and Batch Normalization layers with Leaky ReLU activation functions in the decoder, which improves training stability and reduces computational consumption. The model was evaluated using the Dice and IoU metrics, achieving values of 0.91 and 0.84 respectively, on a validation set. The results obtained suggest that automatic segmentation can support early melanoma diagnosis by highlighting regions of interest relevant for subsequent morphological and textural analysis.

This work presents a comprehensive approach for the automatic classification of heart sound signals as rhythmic or arrhythmic using signal processing and machine learning techniques. A balanced database was constructed from public recordings by applying a preselection process based on objective metrics such as signal-to-noise ratio, kurtosis, and spectral entropy to ensure acoustic quality. The signals were transformed into the spectral domain using the Fast Fourier Transform, and relevant features were extracted as input for a Multilayer Perceptron. The model demonstrated stable behavior during training, with consistent loss and accuracy curves across both training and validation sets. An accuracy of 95.65 % was reached, with 100 % of specificity and 96 % of precision. These results suggest the feasibility of developing efficient, low-cost computer-aided diagnosis systems based on phonocardiographic recordings, with potential applications in both clinical and research environments.

Este trabajo presenta una comparación entre dos métodos de entrenamiento de tipo forward propagation para la para la tarea de seguimiento de señales temporales, con el objetivo de ajustar sus parámetros sinápticos y la neurona intente aproximarse a una señal de referencia. Como alternativa al entrenamiento basado en el algoritmo de Backpropagation, que utiliza el descenso del gradiente para minimizar el error cuadrático medio (MSE), se implementa el algoritmo de Optimización por Enjambre de Partículas (PSO), una técnica de búsqueda global que opera sin necesidad de derivadas y realiza una exploración simultánea del espacio de solución. Para evaluar y comparar el desempeño de ambos métodos, se realizaron 30 ejecuciones para cada algoritmo. El rendimiento se midió con el MSE final, la tasa de éxito (definida como un MSE inferior a 0.005) y el tiempo de cómputo total. Adicionalmente, se analiza la superficie de solución generada por la red neuronal, representando la función de costo en términos de dos pesos seleccionados, lo cual permite ilustrar la presencia de múltiples óptimos locales y su impacto en la eficacia de cada algoritmo. Los resultados obtenidos demuestran que, aunque Backpropagation tiende a converger rápidamente, es susceptible a mínimos locales, mientras que PSO mostró mayor estabilidad frente a la topología no convexa de la superficie de solución, logrando convergencia hacia regiones de bajo error en la mayoría de los casos evaluados. Las observaciones realizadas permiten discutir las ventajas y limitaciones de cada enfoque en contextos donde las funciones objetivo son no diferenciables.