جلد 14، شماره 2 - ( 5-1404 )
جلد 14 شماره 2 صفحات 22-12 |
برگشت به فهرست نسخه ها
Download citation:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:
Vafaeian S, Hajibagheri P. Recent Advances in Deep-Learning for Dental Imaging (2022–2025): A Narrative Review. Journal title 2025; 14 (2) :12-22
URL: http://3dj.gums.ac.ir/article-1-646-fa.html
URL: http://3dj.gums.ac.ir/article-1-646-fa.html
Recent Advances in Deep-Learning for Dental Imaging (2022–2025): A Narrative Review. عنوان نشریه. 1404; 14 (2) :12-22
چکیده: (245 مشاهده)
This narrative review synthesises 2022-2025 evidence on deep learning for dental imaging is synthesised, focusing on diagnostic accuracy (precision, recall, F1), segmentation quality (Dice), and reporting of inference speed. Representative model families are also situated —YOLO one-stage detectors, attention-augmented U-Nets, 3-D CNNs, and Segment Anything Model–derived hybrids (“SAM-derived” = architectures adapted from Segment Anything for dental images)—within current clinical workflows. From database searches of PubMed, Scopus and IEEE Xplore (January 2022 - May 2025) 342 records were retrieved; after deduplication and screening, 17 primary studies were included. Detection/classification studies (n = 10) reported overall F1 values up to 0.97 (highest in peri-implantitis detection; internal test; n = 100 from an 800-image); the lowest externally evaluated detector reported ~0.53 F1 on CBCT periapicals (external site; n=195 scans). YOLOv8 achieved F1 ≈ 0.82 on bitewings on an internal test split (n = 150). Segmentation studies (n = 8) reported Dice ~0.49–0.98: Attention U-Net reached 0.963 for single-tooth CBCT on an internal test (n = 9 scans). A multi-structure Swin-UNETR reported Dice 0.936–0.965 for tooth/sinus/bone/canal on an external set (n = 55 CBCT scans). A SAM-derived model (Tooth-ASAM) achieved 0.909–0.975 Dice across mixed mixed public/private datasets. While three studies included external-site validation; none were prospective or randomised. Key priorities for clinical translation were identified: consistent speed reporting (batch-1 latency/FPS on named hardware), metric harmonisation (Dice for segmentation; F1@0.5 for detection), dockerised, inference pipelines, and multi-centre external testing followed by prospective trials to quantify clinical impact.
نوع مطالعه: Review article |
موضوع مقاله:
تخصصي
دریافت: 1404/4/15 | پذیرش: 1404/5/30 | انتشار: 1404/5/19
دریافت: 1404/4/15 | پذیرش: 1404/5/30 | انتشار: 1404/5/19
| بازنشر اطلاعات | |
 | این مقاله تحت شرایط Creative Commons Attribution-NonCommercial 4.0 International License قابل بازنشر است. |
