Yila
文献综述样本返回文献综述

Artificial Intelligence in Medical Imaging Diagnosis: Clinical Performance, Explainability, Bias, and Implementation Challenges

Author: yila.ai

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.

2. Clinical Performance of AI in Medical Imaging

2.1 Diagnostic Accuracy Across Modalities

The evidence base for AI-assisted medical image interpretation has expanded exponentially. Deep learning models have been applied to virtually every imaging modality, including chest radiography, computed tomography (CT), magnetic resonance imaging (MRI), ultrasound, retinal photography, dermatoscopy, and digital pathology [1,4]. Convolutional neural networks remain the dominant architectural paradigm, though vision transformers and hybrid architectures are increasingly investigated [2].

In radiology, AI systems have achieved area under the receiver operating characteristic curve (AUC) values exceeding 0.90 for tasks such as pneumonia detection on chest X-rays, intracranial hemorrhage identification on head CT, and pulmonary nodule detection on thoracic CT [4,5]. In ophthalmology, deep learning models for diabetic retinopathy screening from retinal fundus photographs have received regulatory clearance and been deployed in real-world settings, with reported sensitivities and specificities comparable to or exceeding those of trained ophthalmologists [4]. In dermatology, CNN-based classifiers have demonstrated accuracy comparable to board-certified dermatologists for melanoma detection from dermoscopic images [1].

Disease-specific systematic reviews further illustrate the breadth of applications. In oncology, AI models applied to breast cancer screening have achieved AUCs ranging from 0.65 to 0.975 across studies, with performance strongly modulated by dataset characteristics and validation methodology [28]. In glioblastoma, AI systems integrating MRI, histopathology, and molecular profiling have shown promise for tumor grading, subtype differentiation, and treatment response prediction, though heterogeneity across studies limits meta-analytic synthesis [29]. In non-small cell lung cancer (NSCLC), multimodal AI approaches combining thoracic imaging, digital pathology, and multiomics data are beginning to enable precision diagnosis and personalized prognostication [30]. In ophthalmology, generative AI techniques—including generative adversarial networks (GANs), autoencoders, and diffusion models—are being explored for image synthesis, enhancement, and anomaly detection [31]. In pediatric imaging, AI-based models for necrotizing enterocolitis (NEC) diagnosis from abdominal radiographs have been systematically reviewed, with meta-analytic evidence indicating promising but heterogeneous diagnostic performance [32]. In rheumatology, deep learning has been applied to classification, segmentation, and anomaly detection across plain radiography, ultrasound, and MRI for rheumatoid arthritis, osteoarthritis, and spondyloarthritis [33]. In ocular oncology, machine learning models for uveal melanoma diagnosis and prognosis from multimodal imaging have been meta-analyzed, highlighting persistent concerns about generalizability and clinical readiness [34]. At the point of care, AI-enabled echocardiography systems are emerging to reduce operator dependency and expand access to cardiovascular imaging in resource-limited settings [35]. Most recently, diffusion models—a class of generative models—have been systematically reviewed for medical image-based diagnosis, with applications spanning image reconstruction, super-resolution, and anomaly detection, though clinical integration remains nascent [36].

2.2 AI Versus Clinicians: Comparative Evidence

A landmark systematic review by Nagendran et al. examined studies comparing the diagnostic performance of deep learning algorithms with that of expert clinicians [5]. Of 81 identified non-randomized studies, only 10 were prospective, and the median number of patients in the validation set was 274. The review identified a high or unclear risk of bias in the majority of studies, with most using an insufficiently rigorous reference standard. Critically, few studies evaluated performance in real clinical settings with the full spectrum of disease presentations and comorbidities [5].

A related concern is the near-total absence of randomized controlled trials (RCTs) demonstrating that AI-assisted diagnosis improves patient outcomes. While AI systems frequently match or exceed clinician performance in silico, evidence that their deployment changes clinical decision-making, reduces mortality, or improves cost-effectiveness remains sparse. The CONSORT-AI extension, published in 2020, established minimum reporting standards for clinical trials evaluating AI interventions, aiming to address the transparency deficits identified in the existing literature [6].

2.3 Reporting Standards and Evidence Quality

The methodological quality of AI diagnostic studies has attracted sustained criticism. Common limitations include: (i) retrospective design with convenience sampling; (ii) absence of external validation on data from different institutions or populations; (iii) use of clean, curated datasets that do not reflect real-world image quality variability; (iv) inadequate handling of confounding variables; and (v) selective reporting of performance on the most favorable subgroups [5,6]. The CONSORT-AI and SPIRIT-AI guidelines represent an important corrective, though their adoption in the literature remains incomplete.

2.4 Foundation Models and Emerging Capabilities

A paradigm shift is underway with the emergence of foundation models—large-scale, self-supervised models trained on diverse datasets and capable of adapting to multiple downstream tasks with minimal task-specific fine-tuning. Moor et al. articulated the vision of generalist medical AI (GMAI), in which a single model flexibly interprets combinations of imaging, electronic health records, laboratory results, genomics, and clinical text [7]. Early examples include models that jointly process chest X-rays and radiology reports, and large language models (LLMs) that encode substantial clinical knowledge, as demonstrated by Med-PaLM 2 on the MultiMedQA benchmark [8]. These models raise both opportunities—unprecedented flexibility and scalability—and new challenges around validation, regulatory oversight, and the interpretability of outputs [7].

2.5 Broader Diagnostic Applications

Beyond imaging-specific applications, AI techniques have been evaluated across a broad spectrum of disease diagnosis and prediction tasks. Ghaffar Nia et al. surveyed machine learning and deep learning applications in healthcare diagnostics, noting that AI-driven analysis of medical images has reduced physician workload, decreased diagnostic errors, and improved prediction performance across multiple disease domains [9]. Dias and Torkamani reviewed AI applications spanning clinical imaging and genomic diagnostics, highlighting the convergence of computer vision and genomic data analysis as a particularly promising frontier [10].

3. Explainability in Medical AI

3.1 The Black-Box Problem in Clinical Decision-Making

Deep neural networks derive their power from learning hierarchical representations directly from raw data, but this very property renders their internal decision logic opaque. In clinical contexts, this opacity carries particular weight: a physician who cannot understand why an AI system flagged a lesion as malignant may be unwilling to act on the recommendation, and a patient who receives a diagnosis mediated by an unexplainable algorithm may rightfully question its validity [12,13]. The black-box problem is thus not merely a technical inconvenience but a fundamental barrier to trust, accountability, and clinical adoption.

3.2 Taxonomy of Explainable AI Methods

Barredo Arrieta et al. provided a foundational taxonomy that distinguishes between transparent models (inherently interpretable, such as linear regression or decision trees) and post-hoc explanation methods applied to black-box models [11]. Post-hoc methods can be further categorized as model-agnostic (applicable to any model) or model-specific (designed for particular architectures), and as producing global explanations (characterizing overall model behavior) or local explanations (explaining individual predictions). This taxonomy has been widely adopted and extended in subsequent work [14,15,16].

Tjoa and Guan specifically surveyed explainability methods for medical applications, categorizing approaches by the type of explanation produced: feature attribution (identifying which input features drove a prediction), example-based (identifying similar training instances), and concept-based (linking predictions to human-understandable concepts) [12]. Their review highlighted that most medical XAI research has focused on feature attribution methods applied to CNN-based imaging classifiers, with comparatively less attention to other model architectures and clinical tasks.

3.3 Post-Hoc Explanation Methods: SHAP, LIME, and Grad-CAM

SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) are the most widely cited post-hoc explanation methods in biomedical applications [17]. SHAP is grounded in cooperative game theory and assigns each input feature an importance value for a particular prediction, while LIME approximates a complex model locally with an interpretable surrogate. In medical imaging, gradient-based methods such as Grad-CAM generate heatmaps highlighting image regions that most influenced a CNN's classification decision [4,12].

Salih et al. critically examined the practical limitations of SHAP and LIME in biomedical contexts [17]. They demonstrated that both methods produce unstable explanations in the presence of feature collinearity—a common scenario in medical data where biomarkers, imaging features, and clinical variables are often correlated. Moreover, explanation outputs are model-dependent: different model architectures trained on the same data can produce substantially different SHAP values for identical predictions, undermining the notion that these methods reveal a ground-truth "reason" for the model's decision [17].

3.4 Clinical, Legal, and Ethical Dimensions

Amann et al. provided a multidisciplinary analysis of explainability in medical AI, examining its relevance from technological, legal, medical, and patient perspectives [13]. They argued that explainability requirements should be calibrated to the clinical context: an AI system that triages screening examinations may require less granular explanation than one that recommends invasive procedures or treatment changes. From a legal standpoint, the European Union's General Data Protection Regulation (GDPR) and the proposed AI Act create, in some interpretations, a "right to explanation" for automated decisions, though the precise scope and enforceability remain debated [13,25].

The ethical imperative for explainability is reinforced by considerations of informed consent, shared decision-making, and professional autonomy. Physicians cannot fulfill their duty of care if they cannot independently verify the basis of an AI-generated recommendation, and patients cannot provide truly informed consent to AI-mediated diagnosis without some understanding of how the system reaches its conclusions [13,24].

3.5 Limitations and Unresolved Challenges

Several persistent challenges limit the clinical utility of current XAI methods. First, the faithfulness of explanations—whether they accurately reflect the model's actual decision process—is rarely validated; many popular post-hoc methods have been shown to produce plausible but unfaithful explanations [17,30]. Second, explanation stability across similar inputs is often poor, eroding clinician confidence. Third, most XAI methods produce explanations that require substantial expertise to interpret, limiting their utility for patients and non-specialist clinicians [14,15]. Ali et al. surveyed the state of XAI assessment methods and available tools, concluding that the field lacks standardized evaluation protocols and that progress toward truly "trustworthy AI" requires closing gaps in explanation completeness, correctness, and consistency [15].

4. Bias in Medical Imaging AI

4.1 Sources of Bias

Bias in medical AI systems can originate at multiple stages of the development and deployment pipeline. Challen et al. provided an early framework categorizing bias into data bias (unrepresentative training samples, measurement error, labeling inconsistency), algorithmic bias (model inductive biases, optimization objectives that disregard fairness), and deployment bias (use in populations or settings different from those in which the system was developed) [18]. Ferrara subsequently extended this taxonomy to encompass generative AI biases and human decision biases that interact with algorithmic outputs [21].

Data bias is particularly consequential in medical imaging. Training datasets that over-represent certain demographic groups, disease presentations, or imaging equipment types produce models whose performance degrades on under-represented subgroups—a phenomenon known as distributional shift [18,21]. Even large, ostensibly diverse datasets may contain hidden biases: for instance, models may learn to detect chest drain tubes or portable X-ray markers rather than the pathological findings of interest, exploiting spurious correlations that do not generalize [5].

4.2 Empirical Evidence of Demographic Disparities

Two studies provide particularly compelling evidence of demographic bias in medical imaging AI. Larrazabal et al. demonstrated that gender imbalance in chest X-ray training datasets produced classifiers with systematically different performance for male and female patients [19]. When training data contained fewer than 25% female samples, the resulting models exhibited significantly lower diagnostic accuracy for women across multiple disease categories. This effect persisted across three different CNN architectures and two publicly available datasets, indicating that the problem is structural rather than model-specific [19].

Seyyed-Kalantari et al. examined algorithmic underdiagnosis—the failure of AI systems to correctly identify disease in patients who truly have it—across three large chest X-ray datasets and one multi-source dataset [20]. They found that state-of-the-art computer vision classifiers consistently and selectively underdiagnosed under-served patient populations: female patients, Black patients, and patients of low socioeconomic status were more likely to be erroneously labeled as healthy. These disparities persisted across multiple disease labels and could not be fully explained by differences in disease prevalence [20]. The clinical implications are grave: underdiagnosis bias means that the very populations already experiencing health disparities may be further disadvantaged by AI systems intended to improve care.

4.3 Clinical Safety Implications

The clinical safety consequences of biased AI extend beyond individual misdiagnoses. At a population level, systematic underdiagnosis of certain groups can widen existing health disparities, erode trust in healthcare institutions, and create feedback loops in which biased AI outputs influence clinical decisions that, in turn, generate data that reinforce the original bias [18,20]. Challen et al. emphasized that clinical AI systems operate within complex sociotechnical environments where bias can propagate through multiple pathways, and that conventional pre-deployment accuracy metrics may fail to capture these dynamics [18].

4.4 Mitigation Strategies

Proposed mitigation strategies span the AI development lifecycle. At the data level, intentional oversampling of under-represented groups, synthetic data generation, and rigorous dataset auditing have been advocated [19,21]. At the algorithmic level, fairness constraints can be incorporated into model training objectives, though there exists an inherent tension between fairness optimization and aggregate accuracy that must be navigated transparently [21]. At the deployment level, continuous post-market surveillance, disaggregated performance reporting by demographic subgroup, and prospective evaluation in diverse populations are essential [6,18]. Federated learning approaches, discussed further in Section 5, offer a complementary strategy by enabling models to be trained across heterogeneous, multi-institutional datasets without centralizing sensitive data, potentially improving representativeness [23].

5. Implementation Challenges

5.1 Technical Barriers to Clinical Translation

The gap between published AI performance and real-world clinical utility is substantial. Kelly et al. identified challenges intrinsic to the science of machine learning—including dataset shift, confounding, and the difficulty of specifying clinically meaningful objective functions—as primary barriers to translation [22]. Models trained on retrospectively curated datasets frequently fail when deployed prospectively because the data distribution in live clinical environments differs from the training distribution in ways that are difficult to anticipate or correct [22].

Generalizability remains perhaps the single greatest technical barrier. A model trained on images from one hospital's equipment may perform poorly on images from another vendor's scanners. A model developed in a high-resource setting may fail in a low-resource setting where disease presentations, comorbidities, and imaging protocols differ. External validation on diverse, multi-institutional data, while increasingly recommended, remains infrequently performed [5,22].

5.2 Regulatory and Legal Frameworks

The regulatory landscape for medical AI is evolving rapidly but remains fragmented. In the United States, the Food and Drug Administration (FDA) has cleared several hundred AI-enabled medical devices, predominantly under the 510(k) pathway, which requires demonstration of substantial equivalence to an existing device rather than de novo evidence of clinical benefit [4,6]. The United Kingdom's Medicines and Healthcare products Regulatory Agency (MHRA) and the European Union's Medical Device Regulation (MDR) have introduced or are developing AI-specific regulatory frameworks, but harmonization across jurisdictions is limited [25].

Legal questions surrounding liability represent a significant unresolved challenge. When an AI-assisted diagnosis leads to patient harm, responsibility may be distributed among the developer, the deploying institution, and the clinician who acted on (or overrode) the AI's recommendation [25,26]. Gerke et al. analyzed the ethical and legal dimensions of AI-driven healthcare, identifying liability, intellectual property, data protection, and informed consent as areas where existing legal frameworks are insufficient or ambiguous [25]. They argued for adaptive governance approaches that balance innovation with patient protection.

5.3 Clinical Workflow Integration

Even technically robust AI systems may fail if they disrupt clinical workflows or impose unacceptable cognitive burdens on users. Radiologists already face high workloads and alert fatigue; adding AI-generated findings without careful integration into existing picture archiving and communication systems (PACS) and reporting workflows can reduce rather than enhance efficiency [4,22]. Human-AI interaction design—determining when and how the AI presents its outputs, the level of detail provided, and the mechanism for clinician override—is a critical and understudied determinant of clinical impact [22].

The phenomenon of automation bias, in which clinicians over-rely on AI recommendations and fail to critically evaluate them, and its converse, algorithm aversion, in which clinicians dismiss correct AI outputs due to mistrust, both represent significant risks [4,22]. Achieving the optimal balance—where AI augments rather than replaces clinical judgment—requires deliberate attention to training, interface design, and organizational culture.

5.4 Data Governance, Privacy, and Federated Learning

Medical imaging data are subject to stringent privacy regulations, including the Health Insurance Portability and Accountability Act (HIPAA) in the United States and GDPR in Europe [23,25]. These regulations, while essential for patient protection, create barriers to the data aggregation that deep learning models typically require. Federated learning has emerged as a promising solution: models are trained across multiple institutions without raw data leaving each site, with only model updates (gradients) being shared and aggregated [23].

Kaissis et al. provided a comprehensive review of secure, privacy-preserving, and federated machine learning techniques for medical imaging [23]. They described a technical stack combining federated learning with differential privacy guarantees and secure multi-party computation, demonstrating feasibility across several medical imaging applications. However, they also noted significant unresolved challenges: communication overhead, heterogeneity in data formats and annotation standards across institutions, and the need for robust defenses against gradient-leakage attacks that could potentially reconstruct patient data from shared model updates [23].

5.5 Economic, Organizational, and Sociocultural Factors

The economic case for AI in medical imaging remains incompletely established. While AI may reduce radiologist workload and improve efficiency, the upfront costs of procurement, integration, training, and ongoing maintenance are substantial [22,26]. In low- and middle-income countries (LMICs), where the potential public health impact of AI-assisted imaging may be greatest given shortages of specialist radiologists, infrastructure constraints—including unreliable internet connectivity, limited computational resources, and inconsistent imaging equipment—present formidable barriers [27].

Schwalbe and Wahl examined the role of AI in global health, noting that most AI development has occurred in high-income countries and may not transfer to LMIC settings without substantial adaptation [27]. They cautioned that without deliberate attention to equity, AI could widen rather than narrow global health disparities.

Organizational readiness and clinician trust are equally critical. Surveys consistently show that while many clinicians are optimistic about AI's potential, a substantial minority express concerns about deskilling, liability, and erosion of the doctor-patient relationship [22,26]. Building trust requires not only technical validation but also transparency about AI limitations, involvement of clinicians in development and deployment decisions, and clear institutional policies governing AI use [24,26].

Zhang and Zhang proposed an ethical governance framework for trustworthy medical AI organized around five domains: data quality, algorithmic bias, opacity, safety and security, and responsibility attribution [24]. They argued that governance must operate at multiple levels—technical standards, institutional policies, professional guidelines, and legal regulation—and that effective governance requires ongoing stakeholder engagement across developers, clinicians, patients, and regulators.

6. Evidence Synthesis

The table below synthesizes findings across the four dimensions of this review, assessing the maturity of the evidence base, the strength of consensus, and the most critical unresolved questions.

Dimension Key Findings Evidence Maturity Consensus Strength Critical Gaps
Clinical Performance DL models achieve expert-level accuracy on narrow diagnostic tasks [1,4]; few prospective or RCT-level studies exist [5]; external validation is rare; CONSORT-AI reporting standards established but adoption incomplete [6]; foundation models (GMAI) and LLMs emerging with new capabilities and validation challenges [7,8] Moderate: many retrospective studies, few prospective evaluations Moderate consensus: AI can match/exceed clinicians on specific benchmarks, but real-world generalizability uncertain Prospective RCTs; external validation across populations and settings; patient-outcome evidence; standardized benchmarks
Explainability XAI taxonomy well-established: transparent vs. post-hoc, model-agnostic vs. model-specific [11]; medical XAI surveyed across feature attribution, example-based, and concept-based methods [12]; SHAP/LIME/Grad-CAM widely used but faithfulness of post-hoc explanations rarely validated [15,17]; multidisciplinary perspectives highlight clinical, legal, and ethical dimensions [13,16] Moderate: extensive methods literature, limited clinical validation Low consensus on which methods are sufficient for clinical use; debate on whether post-hoc explanations are adequate [14,15] Explanation faithfulness validation; standardized evaluation protocols; clinically meaningful explanation formats; patient-level explainability
Bias and Fairness Clear empirical evidence of demographic underdiagnosis bias: gender imbalance in training data produces biased classifiers [19]; state-of-the-art models selectively underdiagnose under-served populations (female, Black, low-SES patients) [20]; bias sources span data, algorithm, and deployment stages [18,21]; mitigation strategies proposed but rarely evaluated prospectively Moderate: compelling case studies, limited systematic mitigation evidence High consensus that bias is real and clinically significant [18,20]; low consensus on optimal mitigation [21] Prospective evaluation of fairness interventions; bias auditing standards; intersectional bias analysis; generative AI bias
Implementation Multifactorial barriers identified: technical (generalizability, dataset shift) [22], regulatory fragmentation [25], privacy constraints addressed by federated learning [23], clinical workflow disruption and trust deficits [22,26], economic uncertainty, and LMIC infrastructure gaps [27]; governance frameworks proposed spanning data quality, algorithmic bias, opacity, safety, and responsibility [24] Low to moderate: many commentaries and frameworks, limited implementation science evidence High consensus that barriers are real and multifactorial; low consensus on which interventions are effective Implementation science studies; cost-effectiveness analyses; LMIC-specific implementation research; human-AI interaction optimization

The evidence reviewed reveals a field at a critical juncture. Technical capabilities have advanced rapidly, but the translational infrastructure—regulatory frameworks, validation standards, implementation science, and governance mechanisms—lags behind. The four dimensions are deeply interconnected: biased models cannot be clinically performant across all populations; unexplainable models undermine the trust needed for implementation; and implementation in the absence of rigorous validation risks patient harm.

7. Future Directions and Conclusion

7.1 Toward Trustworthy and Generalizable Medical AI

Several research priorities emerge from this review. First, the field urgently needs prospective, multi-institutional RCTs evaluating AI-assisted diagnosis against standard-of-care, with patient-centered outcomes as primary endpoints. The CONSORT-AI framework provides methodological guidance; what is now required is funding, regulatory support, and institutional commitment to execute such trials at scale [6].

Second, explainability research must move beyond method development toward clinical validation. The critical question is not whether an explanation can be generated, but whether it improves diagnostic accuracy, clinician trust, or patient outcomes when deployed in practice. Standardized evaluation protocols that assess explanation faithfulness, stability, and clinical actionability are essential [15,17].

Third, fairness must be treated as a first-order design requirement rather than a post-hoc correction. This requires: (a) mandatory disaggregated performance reporting by demographic subgroup in all AI diagnostic studies; (b) development and prospective validation of bias mitigation strategies; and (c) regulatory frameworks that condition market approval on evidence of equitable performance across populations [19,20,21].

Fourth, implementation science methodologies must be adapted to AI in healthcare. The barriers identified—technical, regulatory, workflow, economic, and sociocultural—are each amenable to systematic study, but such research requires interdisciplinary collaboration that has been uncommon to date [22,24,27].

7.2 Research Gaps

The synthesis table in Section 6 enumerates specific gaps. Several warrant emphasis. The evidence base for AI diagnostic performance is dominated by single-institution, retrospective studies in high-income settings; evidence from LMICs, from pediatric populations, and from real-world deployment is sparse [27,32]. The interaction between multiple simultaneous biases—for example, the intersection of race, gender, and socioeconomic status—is almost entirely unexamined [20]. And while foundation models and large language models are attracting enormous investment and attention, their clinical safety, bias properties, and explainability characteristics are poorly characterized relative to the narrower AI systems that preceded them [7,8].

7.3 Conclusion

Artificial intelligence in medical imaging diagnosis stands at a crossroads between technological promise and translational reality. Deep learning models have demonstrated diagnostic accuracy that in many domains rivals or exceeds that of expert clinicians, and emerging foundation models promise to extend these capabilities across modalities and tasks. Yet the evidence that AI improves patient outcomes, that it performs equitably across all populations, and that it can be safely and effectively integrated into diverse clinical environments remains incomplete.

Explainability methods have proliferated but have not yet delivered the transparent, trustworthy decision support that clinicians and patients need. Bias in AI systems is real, clinically significant, and inadequately addressed by current development practices. And the infrastructure for implementation—regulatory, technical, organizational, and economic—is immature relative to the pace of algorithmic innovation.

Closing these gaps will require a deliberate shift from technology-driven development to evidence-driven translation, in which rigorous clinical evaluation, fairness assurance, and implementation science are accorded the same priority as algorithmic performance. The goal is not merely more accurate AI, but AI that is trustworthy, equitable, and genuinely useful in the service of patient care.

证据综合表

| Dimension | Key Findings | Evidence Maturity | Consensus Strength | Critical Gaps |
|---|---|---|---|---|---|
| Clinical Performance | DL models achieve expert-level accuracy on narrow diagnostic tasks [1,4]; few prospective or RCT-level studies exist [5]; external validation is rare; CONSORT-AI reporting standards established but adoption incomplete [6]; foundation models (GMAI) and LLMs emerging with new capabilities and validation challenges [7,8] | Moderate: many retrospective studies, few prospective evaluations | Moderate consensus: AI can match/exceed clinicians on specific benchmarks, but real-world generalizability uncertain | Prospective RCTs; external validation across populations and settings; patient-outcome evidence; standardized benchmarks |
| Explainability | XAI taxonomy well-established: transparent vs. post-hoc, model-agnostic vs. model-specific [11]; medical XAI surveyed across feature attribution, example-based, and concept-based methods [12]; SHAP/LIME/Grad-CAM widely used but faithfulness of post-hoc explanations rarely validated [15,17]; multidisciplinary perspectives highlight clinical, legal, and ethical dimensions [13,16] | Moderate: extensive methods literature, limited clinical validation | Low consensus on which methods are sufficient for clinical use; debate on whether post-hoc explanations are adequate [14,15] | Explanation faithfulness validation; standardized evaluation protocols; clinically meaningful explanation formats; patient-level explainability |
| Bias and Fairness | Clear empirical evidence of demographic underdiagnosis bias: gender imbalance in training data produces biased classifiers [19]; state-of-the-art models selectively underdiagnose under-served populations (female, Black, low-SES patients) [20]; bias sources span data, algorithm, and deployment stages [18,21]; mitigation strategies proposed but rarely evaluated prospectively | Moderate: compelling case studies, limited systematic mitigation evidence | High consensus that bias is real and clinically significant [18,20]; low consensus on optimal mitigation [21] | Prospective evaluation of fairness interventions; bias auditing standards; intersectional bias analysis; generative AI bias |
| Implementation | Multifactorial barriers identified: technical (generalizability, dataset shift) [22], regulatory fragmentation [25], privacy constraints addressed by federated learning [23], clinical workflow disruption and trust deficits [22,26], economic uncertainty, and LMIC infrastructure gaps [27]; governance frameworks proposed spanning data quality, algorithmic bias, opacity, safety, and responsibility [24] | Low to moderate: many commentaries and frameworks, limited implementation science evidence | High consensus that barriers are real and multifactorial; low consensus on which interventions are effective | Implementation science studies; cost-effectiveness analyses; LMIC-specific implementation research; human-AI interaction optimization |

References

[1] Barragán-Montero AM, Javaid U, Valdés G, et al. Artificial intelligence and machine learning for medical imaging: A technology review. Physica Medica. 2021. doi:10.1016/j.ejmp.2021.04.016

[2] Alzubaidi L, Zhang J, Humaidi AJ, et al. Review of deep learning: concepts, CNN architectures, challenges, applications, future directions. Journal of Big Data. 2021. doi:10.1186/s40537-021-00444-8

[3] Al Kuwaiti A, Nazer K, Alreedy AH, et al. A Review of the Role of Artificial Intelligence in Healthcare. Journal of Personalized Medicine. 2023. doi:10.3390/jpm13060951

[4] Najjar R. Redefining Radiology: A Review of Artificial Intelligence Integration in Medical Imaging. Diagnostics. 2023. doi:10.3390/diagnostics13172760

[5] Nagendran M, Chen Y, Lovejoy CA, et al. Artificial intelligence versus clinicians: systematic review of design, reporting standards, and claims of deep learning studies. BMJ. 2020. doi:10.1136/bmj.m689

[6] Liu X, Cruz Rivera S, Moher D, Calvert M, Denniston AK; SPIRIT-AI and CONSORT-AI Working Group. Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. Nature Medicine. 2020. doi:10.1038/s41591-020-1034-x

[7] Moor M, Banerjee O, Abad ZSH, et al. Foundation models for generalist medical artificial intelligence. Nature. 2023. doi:10.1038/s41586-023-05881-4

[8] Singhal K, Azizi S, Tu T, et al. Large language models encode clinical knowledge. Nature. 2023. doi:10.1038/s41586-023-06291-2

[9] Ghaffar Nia N, Kaplanoğlu E, Nasab A. Evaluation of artificial intelligence techniques in disease diagnosis and prediction. Discover Artificial Intelligence. 2023. doi:10.1007/s44163-023-00049-5

[10] Dias R, Torkamani A. Artificial intelligence in clinical and genomic diagnostics. Genome Medicine. 2019. doi:10.1186/s13073-019-0689-8

[11] Barredo Arrieta A, Díaz-Rodríguez N, Del Ser J, et al. Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI. Information Fusion. 2019. doi:10.1016/j.inffus.2019.12.012

[12] Tjoa E, Guan C. A Survey on Explainable Artificial Intelligence (XAI): Toward Medical XAI. IEEE Transactions on Neural Networks and Learning Systems. 2020. doi:10.1109/tnnls.2020.3027314

[13] Amann J, Blasimme A, Vayena E, Frey D, Madai VI. Explainability for artificial intelligence in healthcare: a multidisciplinary perspective. BMC Medical Informatics and Decision Making. 2020. doi:10.1186/s12911-020-01332-6

[14] Hassija V, Chamola V, Mahapatra A, et al. Interpreting Black-Box Models: A Review on Explainable Artificial Intelligence. Cognitive Computation. 2023. doi:10.1007/s12559-023-10179-8

[15] Ali S, Abuhmed T, El-Sappagh S, et al. Explainable Artificial Intelligence (XAI): What we know and what is left to attain Trustworthy Artificial Intelligence. Information Fusion. 2023. doi:10.1016/j.inffus.2023.101805

[16] Angelov P, Soares E, Jiang R, Arnold NI, Atkinson PM. Explainable artificial intelligence: an analytical review. WIREs Data Mining and Knowledge Discovery. 2021. doi:10.1002/widm.1424

[17] Salih A, Raisi-Estabragh Z, Boscolo Galazzo I, et al. A Perspective on Explainable Artificial Intelligence Methods: SHAP and LIME. Advanced Intelligent Systems. 2024. doi:10.1002/aisy.202400304

[18] Challen R, Denny JC, Pitt M, Gompels L, Edwards T, Tsaneva-Atanasova K. Artificial intelligence, bias and clinical safety. BMJ Quality & Safety. 2019. doi:10.1136/bmjqs-2018-008370

[19] Larrazabal AJ, Nieto N, Peterson V, Milone DH, Ferrante E. Gender imbalance in medical imaging datasets produces biased classifiers for computer-aided diagnosis. Proceedings of the National Academy of Sciences. 2020. doi:10.1073/pnas.1919012117

[20] Seyyed-Kalantari L, Zhang H, McDermott MBA, Chen IY, Ghassemi M. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nature Medicine. 2021. doi:10.1038/s41591-021-01595-0

[21] Ferrara E. Fairness and Bias in Artificial Intelligence: A Brief Survey of Sources, Impacts, and Mitigation Strategies. Sci. 2023. doi:10.3390/sci6010003

[22] Kelly C, Karthikesalingam A, Suleyman M, Corrado GS, King D. Key challenges for delivering clinical impact with artificial intelligence. BMC Medicine. 2019. doi:10.1186/s12916-019-1426-2

[23] Kaissis G, Makowski MR, Rückert D, Braren R. Secure, privacy-preserving and federated machine learning in medical imaging. Nature Machine Intelligence. 2020. doi:10.1038/s42256-020-0186-1

[24] Zhang J, Zhang Z. Ethics and governance of trustworthy medical artificial intelligence. BMC Medical Informatics and Decision Making. 2023. doi:10.1186/s12911-023-02103-9

[25] Gerke S, Minssen T, Cohen G. Ethical and legal challenges of artificial intelligence-driven healthcare. In: Artificial Intelligence in Healthcare. Elsevier; 2020. doi:10.1016/b978-0-12-818438-7.00012-5

[26] Dwivedi YK, Hughes L, Ismagilova E, et al. Artificial Intelligence (AI): Multidisciplinary perspectives on emerging challenges, opportunities, and agenda for research, practice and policy. International Journal of Information Management. 2019. doi:10.1016/j.ijinfomgt.2019.08.002

[27] Schwalbe N, Wahl B. Artificial intelligence and the future of global health. The Lancet. 2020. doi:10.1016/s0140-6736(20)30226-9

[28] Desai M, Desai B. Evolution and integration of artificial intelligence across the cancer continuum in women: advances in risk assessment, prevention, and early detection. Cancer Causes & Control. 2025. doi:10.1007/s10552-025-02048-6

[29] Ahangari G, Norioun H, Ghaemi S, Zali A. Artificial intelligence in Glioblastoma Diagnostics: Integrating MRI, histopathology, and molecular profiling. Cancer Treatment and Research Communications. 2025. doi:10.1016/j.ctarc.2025.101040

[30] Chang L, Li H, Wu W, et al. Applications of artificial intelligence in non-small cell lung cancer: from precision diagnosis to personalized prognosis and therapy. Journal of Translational Medicine. 2026. doi:10.1186/s12967-025-07591-z

[31] Phipps B, Hadoux X, Sheng B, et al. AI image generation technology in ophthalmology: Use, misuse and future applications. Progress in Retinal and Eye Research. 2025. doi:10.1016/j.preteyeres.2025.101353

[32] Mahapatra V, Chandra A, Rehman T, et al. Artificial intelligence in the diagnosis of necrotising enterocolitis from abdominal radiographs: a systematic review and meta-analysis. Frontiers in Pediatrics. 2026. doi:10.3389/fped.2026.1852254

[33] Moon J, Jadhav P, Choi S. Deep learning analysis for rheumatologic imaging: current trends, future directions, and the role of human. Journal of Rheumatic Diseases. 2025. doi:10.4078/jrd.2024.0128

[34] Bravo-Gonzalez A, Dominguez-Ruiz P, Buitrago-Gonzalez MJ, et al. Imaging-based machine learning for the diagnosis and prognosis of uveal melanoma: a systematic review and meta-analysis. Graefe's Archive for Clinical and Experimental Ophthalmology. 2026. doi:10.1007/s00417-026-07230-3

[35] East SA, Wang Y, Yanamala N, Maganti K, Sengupta PP. Artificial Intelligence-Enabled Point-of-Care Echocardiography: Bringing Precision Imaging to the Bedside. Current Atherosclerosis Reports. 2025. doi:10.1007/s11883-025-01316-9

[36] Azad M, Fahad NM, Raiaan MAK, et al. A Systematic Review of Diffusion Models for Medical Image-Based Diagnosis: Methods, Taxonomies, Clinical Integration, Explainability, and Future Directions. Diagnostics. 2026. doi:10.3390/diagnostics16020211