Deep Learning Models for the Early and Effective Detection of Diabetes Through Foot Images: A Systematic Review

Resumen

Abstract– This Systematic Literature Review (SLR), conducted under the PRISMA methodology, aims to identify recently proposed artificial intelligence models for the early identification of diabetes patients through foot images. It analyzes the challenges these models address, the Deep Learning (DL) techniques utilized, and the performance metrics applied. Additionally, the results obtained are examined, highlighting innovations compared to classical methods. Out of 237 articles identified in SCOPUS using the PICO methodology, 23 were selected through PRISMA. These studies explore deep learning approaches for detecting diabetic neuropathy and early signs of diabetes through foot images. The most commonly used models include convolutional neural networks (CNN) and deep neural networks (DNN), which stand out for addressing challenges such as image variability and quality. However, the lack of homogeneous databases for meaningful comparisons was identified. The approaches combine CNNs with optimization algorithms and hybrid methods, achieving accuracies ranging from 81.18%, as achieved by Inception-ResNet-v2, to 99.4%, attained by EfficientNet technology. Key innovations include advanced preprocessing techniques and the integration of diverse datasets to improve generalization. Finally, recommendations are proposed to optimize these models, such as developing homogeneous and standardized databases, implementing modern architectures like EfficientNet and Inception-ResNet-v2, and exploring hybrid approaches that integrate RGB, thermal, and spectroscopic images. These measures aim to enhance diagnostic capabilities and overcome current limitations, facilitating their clinical application. Keywords-- Deep learning; diabetic foot; diabetes; early detection; foot images.