Glaucoma is one of the leading causes of blindness worldwide and is characterized by progressive visual field loss due to optic nerve damage. Early detection is fundamental, yet it is often hindered by the asymptomatic nature of the disease in its initial stages. In response to this challenge, advanced techniques such as Deep Learning (DL) and computer vision are emerging as potential tools to revolutionize glaucoma diagnosis. This article aims to systematically evaluate the current state of artificial intelligence approaches for fundus image-based glaucoma detection and to identify trends, challenges, and opportunities. Following the PRISMA methodology, we conducted a comprehensive systematic review examining a total of 63 publications available in Scopus, ScienceDirect, IEEE, and Web of Science databases, available in English and published during the study period between 2020 and 2024. The review revealed key techniques in the critical stages of automated glaucoma detection. Convolutional neural networks dominated recent literature, with ResNet architectures achieving optimal performance (accuracy range: 82.37%-98.48%). For localization and segmentation, U-Net variants, attention-guided networks, and advanced ensemble approaches were prominent. In feature extraction, methods exploiting structural and textural metrics, wavelet-based transformations, and attention mechanisms showed considerable potential. Classification tasks benefited from Convolutional Neural Network (CNN) optimizations, attention-based architectures, hybrid models, and transformer frameworks, demonstrating high accuracy for both binary and multiclass glaucoma detection. Deep learning (DL) approaches have demonstrated significant potential for both binary and multiclass glaucoma classification from color fundus images. Key findings from this study include: i) attention mechanisms and transformer architectures show superior performance in capturing subtle disease features, with accuracies exceeding 95%; ii) hybrid approaches combining multiple techniques achieve better generalization across datasets; iii) curriculum learning strategies improve multiclass severity grading accuracy; and iv) challenges persist in standardizing evaluation metrics and managing variations in data quality. Future research should focus on developing more robust architectures that can handle diverse image qualities and incorporate clinical knowledge into the learning process. Additionally, systematic analysis revealed critical implementation barriers: only 30%-40% of studies included external validation, with significant performance degradation on independent datasets; approximately 60%-70% relied on structure-only approaches, excluding essential visual field correlation for clinical decision-making.

Objectives: This paper focuses on the reduction of harmonic content in the voltage output waveform of a 31 level asymmetric monophasic cascaded multilevel inverter using the Simulated Annealing Optimization (SAO) algorithm. Method: The SAO algorithm was used to find the appropriate switching angles for the stages of the converter using MATLAB. First, the technique employed was simulated and evaluated; then, a prototype of the multilevel inverter was developed for validating the results obtained in simulation. Findings: The output waveform, the harmonic profile graphic, and THD evolution were obtained both the simulation and the prototype. The SAO algorithm allowed to calculate the appropriate switching angles for the asymmetric multilevel inverter, permitting to obtain a low THD (less than 5%), using a relatively low number of semiconductors compared with the symmetric topology. The SAO algorithm also allowed finding the solution of transcendental equations system through a heuristic approach, avoiding the local minima problem presented in traditional methods. Novelty /Improvement: The proposed method (SAO) avoids the potential issues with local minima present in traditional methods for the switching angles computation in cascaded multilevel inverters.