Abstract
Artificial intelligence (AI) has emerged as one of the most transformative technologies in medical imaging over the past decade. Deep learning (DL), particularly convolutional neural networks (CNNs), has demonstrated remarkable capabilities in image classification, segmentation, detection, and reconstruction across radiology, pathology, ophthalmology, dermatology, and numerous other imaging-intensive specialties [1,2]. The convergence of large-scale annotated datasets, advances in graphics processing unit (GPU) hardware, and algorithmic innovations has propelled AI from experimental laboratories toward clinical deployment [3,4]. Despite this momentum, the translation of AI-based diagnostic tools from research benchmarks to routine clinical practice remains uneven and contested. Several interconnected challenges define the current landscape: (i) the robustness and generalizability of reported diagnostic performance metrics; (ii) the opacity of deep learning models, colloquially termed the "black-box" problem, and the corresponding need for explainability; (iii) systematic biases embedded in training data and algorithms that may perpetuate or amplify health disparities; and (iv) a host of implementation barriers spanning regulatory, technical, organizational, and sociocultural domains. This review examines each of these four dimensions, synthesizing evidence from high-impact primary research, systematic reviews, and policy analyses published between 2019 and 2025. The scope is deliberately broad, reflecting the interconnected nature of these challenges: diagnostic performance cannot be evaluated without considering the populations on which models were trained and tested; explainability methods cannot be assessed in isolation from their clinical, legal, and ethical contexts; and implementation strategies must account for both technical infrastructure and the human factors that determine adoption. By mapping the evidence across these dimensions, this review aims to provide clinicians, researchers, and policymakers with a structured understanding of where the field stands and what remains to be done.
1. Introduction
Artificial intelligence (AI) has emerged as one of the most transformative technologies in medical imaging over the past decade. Deep learning (DL), particularly convolutional neural networks (CNNs), has demonstrated remarkable capabilities in image classification, segmentation, detection, and reconstruction across radiology, pathology, ophthalmology, dermatology, and numerous other imaging-intensive specialties [1,2]. The convergence of large-scale annotated datasets, advances in graphics processing unit (GPU) hardware, and algorithmic innovations has propelled AI from experimental laboratories toward clinical deployment [3,4]. Despite this momentum, the translation of AI-based diagnostic tools from research benchmarks to routine clinical practice remains uneven and contested. Several interconnected challenges define the current landscape: (i) the robustness and generalizability of reported diagnostic performance metrics; (ii) the opacity of deep learning models, colloquially termed the "black-box" problem, and the corresponding need for explainability; (iii) systematic biases embedded in training data and algorithms that may perpetuate or amplify health disparities; and (iv) a host of implementation barriers spanning regulatory, technical, organizational, and sociocultural domains. This review examines each of these four dimensions, synthesizing evidence from high-impact primary research, systematic reviews, and policy analyses published between 2019 and 2025. The scope is deliberately broad, reflecting the interconnected nature of these challenges: diagnostic performance cannot be evaluated without considering the populations on which models were trained and tested; explainability methods cannot be assessed in isolation from their clinical, legal, and ethical contexts; and implementation strategies must account for both technical infrastructure and the human factors that determine adoption. By mapping the evidence across these dimensions, this review aims to provide clinicians, researchers, and policymakers with a structured understanding of where the field stands and what remains to be done.