Explainable Artificial Intelligence Techniques Based on Visual Representations and Natural Language for Medical Image-Based Diagnosis: A Systematic Review

Acosta Acosta C.A.; Mejia Cabrera H.I. · Information (Switzerland) (2026) · DOI: 10.3390/info17090896

Resumen

Explainable artificial intelligence (XAI) is increasingly used to improve the transparency of deep learning models for medical image-based diagnosis; however, the reliability and clinical value of the generated explanations remain uncertain. This systematic review followed PRISMA 2020 and examined peer-reviewed studies published from 2022 to 11 June 2026. Following the revised literature search, eligibility assessment, independent methodological appraisal by two reviewers, and consensus resolution of disagreements, 57 primary studies were included in the final evidence synthesis. Of these, 37/57 (64.9%) were published in 2025 or during the eligible portion of 2026, reflecting the recent expansion of research in this field. CNN/convolutional architectures were identified in 56/57 studies (98.2%), while 55/57 (96.5%) produced visual or attribution-based explanations without a textual reporting component; only 2/57 (3.5%) incorporated a textual component. Predictive performance was generally evaluated extensively; however, explanation-quality assessment was less consistently developed: 27/57 studies (47.4%) fully satisfied the review criterion for evaluating explanation quality, whereas 30/57 (52.6%) provided only partial evaluation, frequently relying on qualitative or otherwise incomplete assessment. Recurrent limitations included insufficiently diverse datasets, limited external and multicenter validation, scarce clinician involvement, and the absence of standardized XAI evaluation frameworks. Overall, explanation generation has advanced faster than explanation validation, and clinically dependable XAI requires architecture-aware assessment, external validation, and prospective human-centered evaluation.

Instituciones: Universidad Señor de Sipán