Clinical Operating-Point Evaluation of a CLAHE-Enhanced Tiled YOLO11s Framework for Proximal Caries Detection on Bitewing Radiographs
Clinical Operating-Point Evaluation of a CLAHE-Enhanced Tiled YOLO11s Framework for Proximal Caries Detection on Bitewing Radiographs
Sana Esmaeili1 Parnian Alizadeh Oskoee2 Tahmineh Razi3 Asiyeh Dadghar4 Kasra Rahimipour5 Ata Jodeiri6
1) Faculty of Advanced Medical Sciences, Tabriz University of Medical Sciences, Tabriz, Iran,
2) Faculty of Dentistry, Tabriz University of Medical Sciences, Tabriz, Iran,
3) Faculty of Dentistry, Tabriz University of Medical Sciences, Tabriz, Iran,
4) Faculty of Dentistry, Tabriz University of Medical Sciences, Tabriz, Iran,
5) Faculty of Dentistry, Tabriz University of Medical Sciences, Tabriz, Iran,
6) Faculty of Advanced Medical Sciences, Tabriz University of Medical Sciences, Tabriz, Iran,
IOI Code
:
AAHD-ABKDZB
Publication :
National Conference on Information Technology, Nanotechnology, Artificial Intelligence and Technological Futures Studies(itnaf.ir)
Abstract :
Interproximal caries are difficult to identify on bitewing radiographs because lesions are small, low contrast, and located at contact surfaces where anatomical overlap may obscure early changes. Rather than proposing a new neural-network architecture, this study evaluated a deployment-oriented framework for image-level proximal caries detection using contrast-limited adaptive histogram equalization (CLAHE), overlapping tiles, and YOLO11s. The dataset contained 1,084 bitewing radiographs from 306 patients, including 585 caries-positive and 499 normal images with 1,915 expert-annotated lesions. Preliminary comparison of full-image YOLO11s, clean tiled YOLO11s, and CLAHE-enhanced tiled YOLO11s supported selection of the CLAHE-tiled configuration for final evaluation. Five-fold patient-wise cross-validation with an inner patient-wise validation split ensured that every radiograph was evaluated once in a held-out test fold. Training retained all eligible negative tiles, whereas validation and test radiographs were processed exhaustively without annotation-guided region selection. Inspired by clinical operating-point evaluation in chest radiography, three prespecified validation-derived modes represented sensitivity-prioritized screening, balanced classification, and specificity-prioritized false-positive control. The sensitivity-prioritized mode achieved 91.94% ± 4.88% sensitivity and 43.18% ± 21.59% specificity; the balanced mode achieved 87.98% ± 4.93% sensitivity and 58.11% ± 19.76% specificity, with the highest F1-score of 78.70% ± 4.59%; and the specificity-prioritized mode achieved 69.62% ± 17.56% sensitivity and 77.78% ± 10.22% specificity. The principal contribution is methodological and evaluation-oriented, combining patient-wise exhaustive inference, validation-only operating-point selection, and explicit assessment of sensitivity–specificity trade-offs. The patient-wise exhaustive evaluation provides internal evidence for this framework, while between-fold variability and the single-center design indicate that external validation and prospective workflow evaluation are required before clinical deployment.
Keywords : bitewing radiography proximal caries yolo11s clahe clinical operating points exhaustive evaluation diagnostic bias