Learning Architectures for AI-Assisted Medical Image Diagnosis: A Systematic Literature Review (2020–2025)

Resumen

Deep learning (DL) techniques have substantially transformed medical imaging diagnostics by improving diagnostic accuracy, clinical efficiency, and decision support. This article presents a systematic literature review (SLR) that aims to analyze the evolution, contributions, limitations, and clinical benefits of DL architectures applied to medical imaging diagnostics during the period 2020–2025. The review was conducted following the PRISMA framework guidelines. The literature search in the Scopus database yielded 10,623 results on artificial intelligence, deep learning, and medical imaging. After applying automated filters and a thorough manual review, 38 scientific articles were selected that met rigorous criteria for clinical validation, use of real medical images, and diagnostic utility. The results indicate that CNNs and hybrid CNN-SVM models are the most widely used architectures, accounting for more than 57% of the articles reviewed. In addition, there is a growing adoption of modern architectures such as U-Net, ResNet, DenseNet, and emerging models based on Transformers. The greatest contributions are categorized into three areas: computational (42.1%), methodological (34.2%), and clinical (23.7%). In the clinical world, DL systems consistently improved diagnostic accuracy, reduced interobserver variability, improved workflows, and aided decision-making in imaging such as MRI, CT, ultrasound, and mammography. However, challenges remain in terms of data heterogeneity, model generalization, algorithmic interpretability, regulatory frameworks, and institutional readiness. In summary, this review provides a structured, evidence-based synthesis, bringing algorithmic research closer to its practical application in the clinic.