It is known that through the use of Convolutional Neural Networks (CNNs) classification problems can be solved, such as facial emotion classification. Fuzzy logic has allowed to reduce the uncertainty that may exist when making a decision. In our work, a convolutional neural network model is proposed that allows, through a fuzzy Mamdani inference system, to choose the optimal hyperparameters within Particle Swarm Optimization (PSO), adjusting variables such as inertial weight and cognitive and social acceleration coefficients, in order to determine the number of convolutional layers and filters to be used within the CNN model. We evaluate Type-1 and interval Type-2 fuzzy adaptation within an identical search space and training setup, focusing on how the treatment of uncertainty shapes the quality of the resulting architectures. The experimental results allow an evaluation of the proposed hybrid system for facial emotion classification using the FER2013 dataset. Spanish-language metadata / Metadatos en españolTítulo en español:Adaptación difusa de parámetros de tipo 1 y tipo 2 en PSO para la optimización de arquitecturas CNN aplicada al reconocimiento de emociones faciales Resumen:Es bien sabido que las redes neuronales convolucionales (CNN) permiten resolver problemas de clasificación, como la clasificación de emociones faciales. La lógica difusa ha contribuido a reducir la incertidumbre que puede existir durante la toma de decisiones. En este trabajo se propone un modelo de red neuronal convolucional que, mediante un sistema de inferencia difusa de Mamdani, permite seleccionar los hiperparámetros óptimos dentro de la optimización por enjambre de partículas (PSO). Para ello, se ajustan variables como el peso de inercia y los coeficientes de aceleración cognitiva y social, con el propósito de determinar el número de capas convolucionales y filtros que deben utilizarse en el modelo CNN. Se evalúan la adaptación difusa de tipo 1 y la adaptación difusa de tipo 2 por intervalos dentro de un espacio de búsqueda y una configuración de entrenamiento idénticos, prestando especial atención a la manera en que el tratamiento de la incertidumbre influye en la calidad de las arquitecturas resultantes. Los resultados experimentales permiten evaluar el sistema híbrido propuesto para la clasificación de emociones faciales mediante el conjunto de datos FER2013. Palabras Claves:Redes neuronales artificiales. Smart citations: https://scite.ai/reports/10.61467/2007.1558.2026.v17i4.1486Dimensions.Open Alex.

ABSTRACT The detection and classification of diseases have become a field of interest for artificial intelligence in recent years, where the development of methods and models that allow support for specialists in different health fields has allowed early detection of diseases and the provision of timely treatment to patients. This work proposes the classification of diabetic retinopathy: binary and multi‐class. The first classification consists of detecting whether a patient has diabetic retinopathy, whereas the second classification seeks to detect and determine the level of the disease. The classification is performed using convolutional neural networks, which are optimized using a grey wolf optimizer algorithm. The optimizer has the objective of finding the architecture of the convolutional neural network, as well as the number of convolutional layers, filters, hidden layers, and neurons. The combination of this optimization algorithm with fuzzy logic allows dynamic adjustment of parameters based on current information. This work presents a comparison with other optimization methods such as genetic algorithms and particle swarm optimization. Optimization allows the convolutional neural network model to achieve a maximum accuracy of 0.979536176 for the detection study case and 0.980900407 by adding fuzzy logic, while a maximum accuracy of 0.780354679 is obtained for the classification study case and 0.76807642 by adding fuzzy logic. The results obtained show that the application of optimization methods and fuzzy logic allow the convolutional neural networks to have an advantage over other optimized architectures.

One of the leading causes of blindness in young adults is diabetic retinopathy, a chronic eye complication that arises from diabetes mellitus. It occurs due to progressive damage caused by persistently high blood glucose levels. Convolutional neural networks are becoming a fundamental tool for analyzing and classifying medical images, making them a valuable aid in medical diagnosis. This work proposes a particle swarm optimization-based design of convolutional neural networks. During training, fuzzy logic estimates the learning rate based on the accuracy and loss of both the training and validation sets. The Asia Pacific Tele-Ophthalmology Society 2019 Blindness Detection dataset was used to test the proposed method. In the binary case, the best accuracy was 96.73%; in the multiclass case, it was 77.49%. The results demonstrate that the proposed method, when fuzzy logic is applied, improves diagnostic accuracy for diabetic retinopathy compared with a design that excludes fuzzy logic from the learning process and applies the adaptive moment estimation algorithm.

Detecting diseases in early stages allows patients to have a higher probability of success in their treatments, in addition to reducing treatment costs, which usually increase in advanced stages of various diseases. This is why intelligent techniques have recently become increasingly valuable for physicians, as they are capable of detecting patterns that a physician might not see or miss. This work proposes the optimization of Interval Type-3 fuzzy systems for disease classification. The structure of these classifiers is optimized using the Salp Swarm Algorithm, which searches the parameters of the membership functions of each input and their corresponding fuzzy rules. These optimized classifiers are designed using three databases: Immunotherapy, Cryotherapy, and Haberman's Survival, and where the average accuracy achieved is 84.72, 89.17, and 76.64, respectively. The results accomplished are compared with Type-1 and Interval Type-2 fuzzy systems designed by the same optimization algorithm, and with fuzzy systems designed using a different optimization technique.

The management of long-term conditions like type-1 diabetes relies on accurate glucose forecasting to optimize patient care. In this paper, we present a novel ensemble neural network (NN) model for glucose prediction in type 1 diabetes patients. The proposed ensemble model consists of three different types of recurrent networks (RNN). A fuzzy weighted averaging integration method is employed to compute the final prediction. The proposed ensemble NN model accomplishes a mean root mean square error (RMSE) of 0.001921 ( ± 0.001077), based on 30 experiments done using data generated from the UVA/Padova simulator. To further enhance the model’s performance, we conducted an optimization of the model’s architecture using a genetic algorithm which then obtained an RMSE of 0.001416 ( ± 0.000545). The genetic algorithm was employed to find the best architecture for the ensemble model, including the number of layers, neurons, LSTM cells, and other parameters. The comparative analysis highlights the performance of our proposed ensemble NN and how it improves the results of existing ensemble models. These findings emphasize the potential of our model in advancing glucose prediction accuracy so that patients can make better decisions when managing their condition.

Through the use of Convolutional Neural Networks (CNN) has demonstrated great success in handling uncertainty in image classification tasks. Applying optimization algorithms to a CNN can significantly improve the model's structure, leading to enhanced performance. In this model, the Particle Swarm Optimization (PSO) algorithm was employed to refine the CNN structure. PSO was specifically used to identify optimal hyperparameters, among them the number of convolutional layers and the filters within each layer, thereby maximizing the model's capabilities. Facial Emotion Recognition (FER) has become focus of ongoing exploration among researchers due to its wide-ranging applications across various fields. To validate the results, the FER2013 dataset was utilized.

Diabetic retinopathy (DR) is one of the many consequences caused by diabetes mellitus (DM). The way to detect DR is through different distinctive features that can be observed through retinal images. With this goal, expert technicians look for the characteristic features in the images for the correct detection of the disease. However, some authors dedicated to research have used convolutional neural networks (CNNs). Due to the good results offered by this technique, this paper presents the implementation of Type-1 fuzzy logic to combine it with convolutional neural networks to increase performance of the obtained results. The implementation of fuzzy logic to adjust the hyperparameters allowed us to obtain a mean precision of 0.9273 with a standard deviation of 0.0130, offering better results than when Type-1 fuzzy logic is not implemented, where a mean precision of 0.9021 was obtained with a standard deviation of 0.1065.

Within the realm of neural networks, various regularization techniques are used to prevent overfitting. One of the most common methods is the dropout function, which randomly deactivates a fraction of neurons during training. This approach reduces dependency on specific neurons. However, standard dropouts apply a fixed probability of deactivation, which may not always be optimal for different databases. Consequently, this work aims to enhance this function by adapting the dropout rate based on a fuzzy system that estimates the optimal probability according to the network architecture. The values obtained during the training process such as validation loss and validation accuracy are those that allow obtaining the inputs of the fuzzy system to obtain the new probability value for the next training epoch. The present work was implemented in a convolutional neural network architecture replacing the dropout function and testing it using MNIST and Fashion-MNIST databases. This shows that fuzzy logic can be used in the dropout function. The proposed method allowed us to obtain 0.9208 accuracy as the best experiment in the Fashion-MNIST database, surpassing the traditional Dropout.

One problem that the medical field has faced is the early detection of various existing diseases. Patients with diabetes mellitus are prone to additional conditions, one of which is diabetic retinopathy. Due to the increasing number of people with diabetes mellitus, the number of expert technicians is insufficient to adequately treat them. To solve this problem, computer tools have been used to automate the detection of diseases. One of these tools is the use of artificial neural networks. These networks have the characteristic that they can be adapted to a specific disease, which allows for the creation of different neural network models. Each model has parameters that adjust the weights of its neurons. These parameters are assigned by the designer of network architecture before training. This requires time for testing and fine-tuning the parameters until the desired result is obtained. One of these parameters is the learning rate of the training algorithm. This value can only be modified before training, so selecting the most appropriate one may require a significant investment in time and analysis. This work proposes a method that adjusts the learning rate of the Adaptive Moment Estimation training algorithm between each epoch using interval Type-2 and generalized Type-2 fuzzy inference systems, taking as input the average training and validation loss values, as well as the epoch number. This reduces analysis time, allowing for a focus on other network parameters. The proposed method is applied to two different convolutional neural network architectures: disease detection and classifying disease severity.

Hybrid intelligent systems allow the development of robust solutions to help solve complex problems in different application areas, such as pattern recognition or control problems. Human recognition allows us to determine who a person is or if a person is who it claims to be. Using multimodal biometry allows us to be more confident about the identification. In this work, a genetic algorithm is developed to design Interval Type-3 fuzzy systems to integrate modular neural network responses, where each modular neural network recognizes using a specific biometric measure. The design includes the number, type, parameters of membership functions, and fuzzy rules. The effectiveness of the Type-3 fuzzy systems is proved by combining modular neural network responses using images with and without noise, and the results are compared with Interval and General Type-2 fuzzy systems. The results achieved show statistically better results than interval Type-2 fuzzy systems with images without and with noise, whereas, compared with General Type-2 fuzzy systems, the advantages of the interval Type-3 fuzzy systems are observed only when images with noise are simulated in the modular neural networks. Regarding execution time, the proposed method has an average time lower than the Interval and General Type-2 fuzzy systems because it needs fewer fuzzy if–then rules to achieve a better average recognition rate.

In this paper, the latest global COVID-19 pandemic prediction is addressed. Each country worldwide has faced this pandemic differently, reflected in its statistical number of confirmed and death cases. Predicting the number of confirmed and death cases could allow us to know the future number of cases and provide each country with the necessary information to make decisions based on the predictions. Recent works are focused only on confirmed COVID-19 cases or a specific country. In this work, the firefly algorithm designs an ensemble neural network architecture for each one of 26 countries. In this work, we propose the firefly algorithm for ensemble neural network optimization applied to COVID-19 time series prediction with type-2 fuzzy logic in a weighted average integration method. The proposed method finds the number of artificial neural networks needed to form an ensemble neural network and their architecture using a type-2 fuzzy inference system to combine the responses of individual artificial neural networks to perform a final prediction. The advantages of the type-2 fuzzy weighted average integration (FWA) method over the conventional average method and type-1 fuzzy weighted average integration are shown.

Vision problems are common in patients with diabetes mellitus (DM) because they may suffer from diabetic retinopathy (DR). Because the symptoms of this condition are not easy to detect without the intervention of an expert technician, the use of convolutional neural networks (CNN) has been implemented to speed up the process of analyzing retina images. Due to the good results of this technology, efforts have been made to combine it with other technologies. In this paper, we present the use of an intelligent hybrid system that uses CNNs and Fuzzy Logic with the aim of improving the accuracy obtained. The implementation of fuzzy logic to adjust the hyperparameters of the network allowed us to obtain a mean of 0.9526 with a standard deviation of 0.008521158 in the binary case study, while in the multiclass case study we obtained a mean of 0.7299 and a standard deviation of 0.015614013, offering better results when fuzzy logic is combined compared to when not.

An essential aspect of healthcare is receiving an appropriate and opportune disease diagnosis. In recent years, there has been enormous progress in combining artificial intelligence to help professionals perform these tasks. The design of interval Type-3 fuzzy inference systems (IT3FIS) for medical classification is proposed in this work. This work proposed a genetic algorithm (GA) for the IT3FIS design where the fuzzy inputs correspond to attributes relational to a particular disease. This optimization allows us to find some main fuzzy inference systems (FIS) parameters, such as membership function (MF) parameters and the fuzzy if-then rules. As a comparison against the proposed method, the results achieved in this work are compared with Type-1 fuzzy inference systems (T1FIS), Interval Type-2 fuzzy inference systems (IT2FIS), and General Type-2 fuzzy inference systems (GT2FIS) using medical datasets such as Haberman’s Survival, Cryotherapy, Immunotherapy, PIMA Indian Diabetes, Indian Liver, and Breast Cancer Coimbra dataset, which achieved 75.30, 87.13, 82.04, 77.76, 71.86, and 71.06, respectively. Also, cross-validation tests were performed. Instances established as design sets are used to design the fuzzy inference systems, the optimization technique seeks to reduce the classification error using this set, and finally, the testing set allows the validation of the real performance of the FIS.

The preventive measures taken to curb the spread of COVID-19 have emphasized the importance of wearing face masks to prevent potential infection with serious diseases during daily activities or for medical professionals working in hospitals. Due to the mandatory use of face masks, various methods employing artificial intelligence and deep learning have emerged to detect whether individuals are wearing masks. In this paper, we utilized convolutional neural networks (CNNs) to classify the use of face masks into three categories: no mask, incorrect mask, and proper mask. Establishing the appropriate CNN architecture can be a demanding task. This study compares four swarm intelligent metaheuristics: particle swarm optimization (PSO), grey wolf optimizer (GWO), bat algorithm (BA), and whale optimization algorithm (WOA). The CNN architecture design involves determining the essential hyperparameters of the CNNs. The results indicate the effectiveness of the PSO and BA in achieving an accuracy of 100% when using 10% of the images for testing. Meanwhile, when 90% of the images were used for testing, the results were as follows: PSO 97.15%, WOA 97.14%, BA 97.23%, and GWO 97.18%. These statistically significant differences demonstrate that the BA allows better results than the other metaheuristics analyzed in this study.

The world has been greatly affected by the COVID-19 pandemic, causing people to remain isolated and decreasing the interaction between people. Accordingly, various measures have been taken to continue with a new normal way of life, which is why there is a need to implement the use of technologies and systems to decrease the spread of the virus. This research proposes a real-time system to identify the region of the face using preprocessing techniques and then classify the people who are using the mask, through a new convolutional neural network (CNN) model. The approach considers three different classes, assigning a different color to identify the corresponding class: green for persons using the mask correctly, yellow when used incorrectly, and red when people do not have a mask. This study validates that CNN models can be very effective in carrying out these types of tasks, identifying faces, and classifying them according to the class. The real-time system is developed using a Raspberry Pi 4, which can be used for the monitoring and alarm of humans who do not use the mask. This study mainly benefits society by decreasing the spread of the virus between people. The proposed model achieves 99.69% accuracy with the MaskedFace-Net dataset, which is very good when compared to other works in the current literature.

This work proposes convolutional neural networks (CNNs) and particle swarm optimization (PSO) for diabetic retinopathy classification. Particle swarm optimization seeks to minimize the classification error, designing the convolutional neural network using different preprocessing methods to compare results. The parameters optimized to design the CNN are the number of convolutional layers, filters with their filters size, pool size, algorithm for the learning process, number of fully connecter layers with their number of neurons, batch size, and finally the number of epochs. Among the preprocessing applied are: extraction of the retina and applying a histogram equalization to the red, green, and blue channels. The database used to test the proposed method is APTOS 2019, where the best result achieved is 96.59%, with an average of 95.33%.

One of the worse conditions caused by diabetes mellitus (DM) is diabetic retinopathy (DR) and it can be irreversible if it is not treated in time. The patient with this condition can be completely blind because DR does not have symptoms until advanced stages. Because of this, some authors have been searching for a solution to an early detection of DR. One of the most used technologies for the detection of DR is the neural networks called: Convolutional neural networks (CNN). But design a CNN model from beginning could be slow. Along this work, we proposed the design of a hierarchical genetic algorithm (HGA) to find the best hyperparameters for a CNN model for the detection of DR. Before designing the hierarchical genetic algorithm, we applied pre-processing to the APTOS 2019 database. Then we executed 30 times the hierarchical genetic algorithm and achieved 0.9650 of accuracy mean and 0.007665 of standard deviation. The best CNN model got an accuracy of 0.9781 for DR detection.

In recent years, due to the COVID-19 pandemic, there have been a large number of infections among humans, causing the virus to spread around the world. According to recent studies, the use of masks has helped to prevent the spread of the virus, so it is very important to use them correctly. Using masks in public places has become a common practice these days and if it is not used correctly the virus will continue to be transmitted. The contribution of this work is the development of a convolutional neural network model to detect and classify the correct use of face masks. Deep learning methods are the most effective method to detect whether a person is using a mask properly. The proposed model was trained using the MaskedFace-Net dataset and evaluated with different images of it. The Caffe model is used for face detection, after which the image is preprocessed to extract features. These images are the input of the new convolutional neural network model, where it is classified among incorrect mask, non-mask, and mask. The proposed model achieves an accuracy rate of 99.69

In this work, the combination of convolutional neural networks (CNNs) and particle swarm optimization techniques to automatically design CNN architectures is proposed. The proposed particle swarm optimization aims at finding CNN parameters: the number of fully connected layers with their number of neurons, the number of convolutional layers with the number and filters size, batch size, and the epochs number. This method is applied and tested with face recognition datasets. The particular databases used in this work are the ORL and Yale, where the principal goal of this work is to reduce the recognition error. Experiments are performed using different images for the training phase to observe how much information is needed to obtain a good recognition. The simulation results show a face recognition rate of 100% with 7 (ORL) and 4 (Yale) images for the training phase, respectively. The obtained results are contrasted with other works to observe the behavior of the proposed method.