Artificial Intelligence in Nailfold Capillaroscopy: A Scoping Review1of Validation, Reproducibility, and Clinical Translation

Technical Summary: Artificial Intelligence in Nailfold Capillaroscopy: A Scoping Review of Validation, Reproducibility, and Clinical Translation

Problem Statement
Nailfold capillaroscopy (NFC) and nailfold videocapillaroscopy (NVC) are established non-invasive techniques for visualizing peripheral microvascular architecture, particularly in systemic sclerosis (SSc) and Raynaud's phenomenon (RP). However, manual interpretation is time-consuming, observer-dependent, and highly sensitive to image acquisition conditions. While Artificial Intelligence (AI) has been increasingly applied to automate tasks such as capillary detection, segmentation, quantitative measurement, and disease classification, the field suffers from methodological heterogeneity. A critical issue identified is the risk of data leakage in repeated-measures imaging settings (where multiple images, fingers, or fields originate from the same participant). If validation strategies fail to separate data at the participant level, performance estimates may be artificially inflated. Furthermore, there is a lack of systematic evidence regarding external validation, reproducibility, open resource availability, and the readiness of these models for clinical translation.

Methodology
This study is a scoping review conducted in accordance with the PRISMA Extension for Scoping Reviews (PRISMA-ScR).

• Search Strategy: The authors searched PubMed, Ovid MEDLINE, Scopus, Web of Science Core Collection, and IEEE Xplore from inception to March 28, 2026.
• Selection Criteria: Eligible studies were peer-reviewed journal articles or full conference proceedings reporting original, extractable AI results using human NFC/NVC images or videos. Abstract-only, preprint-only, non-original, and duplicate reports were excluded.
• Data Extraction and Coding: Two reviewers independently screened titles/abstracts and assessed full texts. A standardized form extracted data on study design, clinical domains, AI task families, model architectures, and validation practices.
• Validation Indicators: Six strict indicators were coded for each study to assess translational readiness: (1) independent external clinical validation, (2) multicenter design, (3) direct comparison with experts, (4) publicly accessible code, (5) publicly accessible data, and (6) explicit participant-level data separation.
• Synthesis: Results were summarized descriptively. No pooled diagnostic accuracy estimates were calculated due to heterogeneity in tasks, units of analysis, and metrics.

Key Contributions

1. Evidence Mapping: The review maps the current landscape of AI in NFC/NVC, covering 43 primary studies (37 journal articles, 6 conference proceedings) identified from 3,059 initial records.
2. Task and Architecture Taxonomy: It categorizes AI applications into six principal task families: classification/prediction, detection/quantification, segmentation, dynamic/flow analysis, image-quality/preprocessing, and dataset/benchmark development. It documents the evolution from classical machine learning and handcrafted features to deep learning architectures (CNNs, U-Nets, YOLO, Vision Transformers, and hybrid systems).
3. Clinical Scope Expansion: While SSc/RP remains the dominant application (33%), the review identifies expanding applications in general microcirculation, diabetes, juvenile dermatomyositis (JDM), other autoimmune diseases, neonatal retinopathy, and amyloidosis.
4. Reproducibility Audit: The study provides a rigorous audit of validation and reporting practices, highlighting that 65.1% of included studies met none of the six key validation/reproducibility indicators.

Results

• Validation Deficits: Under strict coding criteria, only one study provided independent external clinical validation. Only four studies were multicenter, 11 directly compared outputs with experts, four provided public code, two provided public data, and five explicitly confirmed participant-level separation.
• Data Leakage Risks: The majority of studies did not explicitly report participant-level separation, raising concerns about data leakage where correlated observations from the same individual appeared in both training and test sets.
• Sample Size Reporting: Among studies reporting participant counts, the median cohort size was 30 participants. However, many studies reported image or frame counts without corresponding participant denominators, potentially overstating effective sample sizes.
• Device Portability: While low-cost and portable devices are emerging, device-stratified or device-external validation remains limited.
• Reporting Gaps: Critical reporting elements such as calibration, uncertainty estimates, subgroup performance (e.g., by skin phototype), and prospective clinical utility endpoints were largely absent or incompletely reported.

Significance and Claims
The paper claims that while AI-NFC has expanded technically and clinically—moving beyond simple capillary counting to complex microvascular computing—the evidence base for clinical generalisability and implementation readiness remains weak.

The authors argue that technical diversity and strong internal performance do not equate to clinical readiness. The primary barrier to deployment is not model innovation but the lack of robust, independent external validation, controlled data splitting, and open reproducibility. The review positions AI-NFC not merely as an image processing challenge but as a clinical decision-support problem requiring adherence to specific reporting frameworks (e.g., TRIPOD+AI, CLAIM, DECIDE-AI).

The study concludes that future research must prioritize:

• Locked external testing across independent centers, devices, and patient subgroups.
• Explicit reporting of split units and participant-level separation to prevent leakage.
• Development of open, reusable benchmarks with clear annotation standards.
• Prospective evaluation of clinical utility, including workflow outcomes and patient-relevant endpoints, rather than relying solely on retrospective accuracy metrics.

The review serves as a cautionary synthesis, urging the field to shift focus from headline accuracy metrics toward trustworthy implementation, equity, and regulatory readiness.