Artificial intelligence (AI) models for diagnostic imaging face reproducibility challenges due to inconsistent reporting. Existing guidelines also lack specificity for imaging-based AI diagnostics, particularly regarding clinical usability and technical transparency. To address these gaps, the Completeness, Learnability, Applicability, Interpretability, Reproducibility, and Evaluation (CLAIRE) framework was developed as a practical reporting aid by a multidisciplinary team of clinicians and AI experts. CLAIRE was retrospectively validated on a subset of 10 imaging studies selected by theoretical saturation in medical and dental imaging. Internal validation demonstrated high reliability, with inter-rater agreement improving from Cohen’s κ 0.286 to 0.987 (p < 0.01) after calibration, alongside a mean intra-rater reliability of 0.997 after a six-month washout period. This process yielded a 15-item structured checklist for standardising AI reporting, supported by an objective scoring system for quality categorisation and an editorial reference guide to facilitate systematic appraisal by reviewers and editors. CLAIRE aims to enhance clinician accessibility through plain-language technical summaries and assessments of real-world applicability. This proposal provides a unified and practical structure that improves reporting consistency, supports systematic assessment, and strengthens both reproducibility and clinical translation of AI-based imaging models.
14 May 2026 in Oral Surgery, Oral Medicine, Oral Pathology and Oral Radiology
To evaluate an annotation-free deep learning (DL) approach for the image-level detection of periapical bone rarefactions (PBRs) on panoramic radiographs (PRs). A retrospective dataset of 2,527 PRs was classified at the image level based on expert consensus using internationally accepted radiographic criteria. Multiple pretrained convolutional neural network (CNN) and vision transformer (ViT) architectures were evaluated under different modeling strategies, including fine-tuning and attention mechanisms. Model performance was assessed using accuracy, recall, specificity, and F1 score. Inferential comparisons were performed using nonparametric statistical tests. CNN-based models consistently outperformed ViT architectures in binary classification of PBRs. The fine-tuned Xception model achieved the highest overall performance, with an F1-score of 64.7% and low variability across repeated runs. Analyses were restricted to Classes A and B due to class imbalance constraints, then multiclass evaluation was not performed at this stage. Annotation-free DL demonstrated feasibility for binary image-level detection of PBRs on PRs, reducing dependence on labor-intensive spatial annotations while maintaining clinically interpretable behavior. Multiclass classification remains a target for investigation with balanced datasets. The approach supports scalable image-level analysis under clinically realistic conditions, although external validation is still required to confirm generalizability.
28 September 2026 in International Journal of Imaging Systems and Technology
This study aimed to compare an annotation‐free deep learning (DL) model with human evaluators for detecting periapical bone rarefactions (PBRs) on panoramic radiographs (PRs) and to assess the clinical acceptability of AI‐supported dental imaging. A mixed‐methods diagnostic validation was conducted using 238 PRs, equally distributed between images with ( n = 119) and without ( n = 119) PBRs. Model performance was compared with that of four calibrated dentists: two oral and maxillofacial radiologists and two endodontists. The reference standard was based on internationally accepted radiographic criteria applied by calibrated researchers under expert supervision, with final classifications verified by the principal investigator. Diagnostic performance was assessed using standard diagnostic metrics, chance‐corrected agreement statistics and receiver operating characteristic analysis, while clinical acceptability was explored through questionnaires and thematically analysed semi‐structured interviews. At the sensitivity‐oriented decision threshold of 0.27, the DL model achieved a sensitivity of 0.924, a specificity of 0.109, an accuracy of 0.517, a precision of 0.509 and an F 1‐score of 0.657. Threshold‐independent binary discrimination was moderate (ROC‐AUC = 0.726). Human evaluators showed more balanced profiles, with accuracy ranging from 0.655 to 0.752 and F 1‐scores from 0.654 to 0.779. Professional perception was generally favourable, although responses varied among evaluators. The annotation‐free DL model showed high sensitivity and moderate discrimination for PBR detection. Although low specificity limits autonomous use, the findings support annotation‐free learning as a potentially scalable strategy, warranting further training with larger and more heterogeneous datasets.