وجهة نظر جديدة تعتمد على التعلم العميق لرقم الأسنان واكتشاف التسوس
A novel deep learning-based perspective for tooth numbering and caries detection

المجلة: Clinical Oral Investigations، المجلد: 28، العدد: 3
DOI: https://doi.org/10.1007/s00784-024-05566-w
PMID: https://pubmed.ncbi.nlm.nih.gov/38411726
تاريخ النشر: 2024-02-27
المؤلف: Baturalp Ayhan وآخرون
الموضوع الرئيسي: الأشعة السينية السنية والتصوير

نظرة عامة

هدفت الدراسة إلى أتمتة الكشف عن الأسنان وترقيمها في الأشعة الرقمية للعضة بينما يتم تقييم الكفاءة التشخيصية للأسنان المتسوسة باستخدام خوارزميات التعلم العميق. تم استخدام مجموعة بيانات تتكون من 1,170 صورة أشعة مجهولة الهوية، والتي خضعت لعملية تسمية قبل تقسيمها إلى مجموعات تدريب واختبار. شملت عملية العمل المقترحة ثلاث مراحل: أولاً، تم استخدام شبكة عصبية تلافيفية مسبقة التدريب (CNN) لتحديد جانب صور العضّة؛ ثانياً، تم استخدام نموذج YOLOv7 المحسن لترقيم الأسنان وكشف التسوس؛ وأخيراً، قامت الخوارزمية بتقييم وجود التسوس من خلال مقارنة التسوس المكتشف مع الأسنان المرقمة باستخدام مقياس التقاطع على الاتحاد.

أشارت النتائج إلى أداء عالٍ للنموذج، مع قيم الاسترجاع والدقة ودرجة F1 لكشف الأسنان عند 0.994 و0.987 و0.990 على التوالي، ولترقيم الأسنان عند 0.974 و0.985 و0.979. حقق كشف التسوس قيم استرجاع ودقة ودرجة F1 تبلغ 0.833 و0.866 و0.822. أظهرت أداء المطابقة لترقيم الأسنان وكشف التسوس دقة واسترجاع وخصوصية ودقة ودرجة F1 بقيم 0.934 و0.834 و0.961 و0.851 و0.842 على التوالي. تؤكد النتائج على إمكانية استخدام الشبكات العصبية التلافيفية لدعم الأطباء من خلال أتمتة هذه العمليات، مما يعزز كفاءة التقييم ويوفر الوقت في البيئات السريرية.

مقدمة

تتناول مقدمة ورقة البحث أهمية الكشف المبكر وعلاج تسوس الأسنان، وهو مرض معدي شائع وقابل للتجنب يمكن أن يؤدي إلى مشاكل صحية خطيرة إذا لم يتم علاجه. انتقلت طب الأسنان الحديثة نحو أساليب وقائية وأقل تدخلاً، مع التأكيد على أهمية التشخيص الدقيق من خلال طرق متنوعة، بما في ذلك الفحوصات البصرية اللمسية والتصوير الشعاعي. بينما تحتوي الطرق التقليدية على قيود، خاصة في الكشف عن الآفات التسوسية المبكرة، فإن التقدم في تصوير الأشعة الرقمية للعضة قد حسن من دقة وكفاءة التشخيص.

تسلط الورقة الضوء على إمكانيات الذكاء الاصطناعي (AI)، وبشكل خاص تقنيات التعلم العميق (DL) مثل الشبكات العصبية التلافيفية (CNNs)، في تعزيز عملية التشخيص في طب الأسنان. أظهرت الدراسات السابقة فعالية الشبكات العصبية التلافيفية في تشخيص التسوس من الأشعة الرقمية للعضة. ومع ذلك، يشير المؤلفون إلى وجود فجوة في الأدبيات بشأن الكشف المتزامن عن الأسنان وترقيمها وتحديد التسوس باستخدام نهج الشبكات العصبية التلافيفية. تهدف هذه الدراسة إلى سد تلك الفجوة من خلال اقتراح بنية أنبوبية جديدة تعتمد على التعلم العميق تدمج وحدات انتباه الكتل التلافيفية (CBAMs) مع نموذج YOLOv7. تم تصميم هذه البنية لتحسين كفاءة ترقيم الأسنان وكشف التسوس، مما يقلل في النهاية من وقت الجلوس ويعزز فعالية العلاج.

طرق البحث

في هذه الدراسة، تم إجراء جميع التجارب باستخدام إعداد حوسبة عالية الأداء يتكون من وحدة معالجة الرسوميات NVIDIA 1080 Ti بسعة 12 جيجابايت من الذاكرة الفيديوية ومعالجات Intel Xeon CPU المزدوجة المزودة بسعة 64 جيجابايت من الذاكرة العشوائية، تعمل على منصة Ubuntu. استخدم الإطار التجريبي برمجة بلغة بايثون، مستفيدًا من مكتبات التعلم العميق مثل PyTorch وKeras، إلى جانب مكتبة OpenCV للرؤية الحاسوبية لمهام معالجة الصور. سهلت هذه التهيئة تنفيذ وتقييم المنهجيات المقترحة بشكل فعال.

النتائج

في مجموعة بيانات الاختبار التي تتكون من 170 صورة أشعة رقمية للعضة، كشفت التحليلات عن إجمالي 1679 سنًا و2842 سطحًا تقاربيًا، مع انتشار الآفات بنسبة 21.18%. أظهرت التقييمات الأولية دون التمييز بين الجانبين الأيمن والأيسر قيم استرجاع ودقة ودرجة F1 لترقيم الأسنان بلغت 0.409 و0.236 و0.299، ولكشف التسوس 0.676 و0.707 و0.691 على التوالي.

عند تقسيم مجموعة البيانات إلى الجانبين الأيمن والأيسر، تحسنت مقاييس الأداء بشكل كبير: بلغت قيم الاسترجاع والدقة ودرجة F1 لكشف الأسنان 0.994 و0.987 و0.990؛ لترقيم الأسنان كانت 0.974 و0.985 و0.979؛ ولكشف التسوس 0.833 و0.866 و0.822 على التوالي. بالإضافة إلى ذلك، حقق أداء المطابقة لترقيم الأسنان وكشف التسوس قيم دقة واسترجاع وخصوصية ودقة ودرجة F1 بلغت 0.934 و0.834 و0.961 و0.851 و0.842 على التوالي. تشير النتائج إلى أن نموذج YOLOv7-AP-CBAM يظهر قدرات قوية في كشف وترقيم الأسنان، وكذلك في كشف التسوس على الأشعة الرقمية للعضة.

المناقشة

في هذه الدراسة، طور المؤلفون نموذج تعلم عميق (DL)، وهو نموذج YOLOv7 المحسن المدمج مع وحدة انتباه الكتل التلافيفية (CBAM)، لتعزيز دقة ترقيم الأسنان وكشف التسوس في الأشعة الرقمية للعضة. استخدمت عملية تحديد حجم العينة اختبار فيشر الدقيق وحساب تأثير التصميم (DE)، مما أدى إلى مجموعة بيانات مطلوبة تتكون من 2136 سطحًا، بينما كانت مجموعة البيانات الفعلية تتكون من 170 صورة أشعة للعضة. تم تدريب النموذج على 1000 صورة، محققًا دقة مذهلة بنسبة 100% في تصنيف الجانبين الأيسر والأيمن من الصور، وهو أمر حاسم لترقيم الأسنان بشكل فعال.

تتكون البنية المقترحة من ثلاثة مكونات رئيسية: الكشف عن الجانب باستخدام نموذج DenseNet-121، وترقيم الأسنان، وكشف التسوس من خلال إطار عمل YOLOv7. سمح دمج CBAM للنموذج بالتركيز على الميزات البارزة داخل الصور، مما يعزز أداء الكشف. أفادت الدراسة بمعدلات استرجاع ودقة عالية لكل من ترقيم الأسنان وكشف التسوس، حيث حقق النموذج استرجاعًا بنسبة 99.4% ودقة بنسبة 98.75%. تشير النتائج إلى أن نموذج YOLOv7-AP-CBAM المحسن يمكن أن يساعد بشكل كبير ممارسي طب الأسنان من خلال توفير أداة موثوقة وفعالة لتشخيص تسوس الأسنان، مما يعالج التحديات المرتبطة بالأخطاء البشرية في البيئات السريرية.

Journal: Clinical Oral Investigations, Volume: 28, Issue: 3
DOI: https://doi.org/10.1007/s00784-024-05566-w
PMID: https://pubmed.ncbi.nlm.nih.gov/38411726
Publication Date: 2024-02-27
Author(s): Baturalp Ayhan et al.
Primary Topic: Dental Radiography and Imaging

Overview

The study aimed to automate the detection and numbering of teeth in digital bitewing radiographs while evaluating the diagnostic efficiency of decayed teeth using deep learning algorithms. A dataset of 1,170 anonymized radiographs was utilized, which underwent a labeling process before being divided into training and test sets. The proposed end-to-end pipeline included three stages: first, a pre-trained convolutional neural network (CNN) identified the side of the bitewing images; second, an enhanced YOLOv7 model was employed for tooth numbering and caries detection; and finally, the algorithm assessed caries presence by comparing detected caries with numbered teeth using the intersection over union metric.

Results indicated high performance of the model, with recall, precision, and F1-score values for teeth detection at 0.994, 0.987, and 0.990, respectively, and for teeth numbering at 0.974, 0.985, and 0.979. Caries detection yielded recall, precision, and F1-score values of 0.833, 0.866, and 0.822. The matching performance for teeth numbering and caries detection showed accuracy, recall, specificity, precision, and F1-score values of 0.934, 0.834, 0.961, 0.851, and 0.842, respectively. The findings underscore the potential of CNNs to support clinicians by automating these processes, thereby enhancing assessment efficiency and saving time in clinical settings.

Introduction

The introduction of the research paper addresses the significance of early detection and treatment of dental caries, a prevalent and preventable infectious disease that can lead to severe health issues if untreated. Modern dentistry has shifted towards preventive and minimally invasive approaches, emphasizing the importance of accurate diagnosis through various methods, including visual-tactile examinations and radiological imaging. While traditional methods have limitations, particularly in detecting early carious lesions, advancements in digital bitewing radiography have improved diagnostic accuracy and efficiency.

The paper highlights the potential of artificial intelligence (AI), specifically deep learning (DL) techniques such as convolutional neural networks (CNNs), in enhancing the diagnostic process in dentistry. Previous studies have demonstrated the effectiveness of CNNs in diagnosing caries from bitewing radiographs. However, the authors note a gap in the literature regarding simultaneous teeth detection, numbering, and caries identification using a CNN approach. This study aims to fill that gap by proposing a novel DL-based pipeline architecture that integrates convolutional block attention modules (CBAMs) with the YOLOv7 model. This architecture is designed to improve the efficiency of teeth numbering and decay detection, ultimately reducing chair time and enhancing treatment effectiveness.

Methods

In this study, all experiments were conducted using a high-performance computing setup comprising an NVIDIA 1080 Ti GPU with 12 GB of video memory and dual Intel Xeon CPU processors equipped with 64 GB of RAM, operating on the Ubuntu platform. The experimental framework utilized programming in Python, leveraging deep learning libraries such as PyTorch and Keras, alongside the OpenCV computer vision library for image processing tasks. This configuration facilitated the implementation and evaluation of the proposed methodologies effectively.

Results

In the testing dataset comprising 170 digital bitewing radiographs, the analysis revealed a total of 1679 teeth and 2842 approximal surfaces, with a lesion prevalence of 21.18%. Initial evaluations without distinguishing between right and left sides yielded Recall, Precision, and F1-score values of 0.409, 0.236, and 0.299 for teeth numbering, and 0.676, 0.707, and 0.691 for caries detection, respectively.

Upon splitting the dataset into right and left sides, the performance metrics significantly improved: Recall, Precision, and F1-score for teeth detection reached 0.994, 0.987, and 0.990; for teeth numbering, they were 0.974, 0.985, and 0.979; and for caries detection, 0.833, 0.866, and 0.822, respectively. Additionally, the matching performance for teeth numbering and caries detection yielded Accuracy, Recall, Specificity, Precision, and F1-score values of 0.934, 0.834, 0.961, 0.851, and 0.842, respectively. The results indicate that the YOLOv7-AP-CBAM model demonstrates robust capabilities in detecting and numbering teeth, as well as in caries detection on digital bitewing radiographs.

Discussion

In this study, the authors developed a deep learning (DL) model, specifically an improved YOLOv7 architecture integrated with a Convolutional Block Attention Module (CBAM), to enhance the accuracy of tooth numbering and caries detection in digital bitewing radiographs. The sample size determination utilized Fisher’s exact test and a design effect (DE) calculation, leading to a required dataset of 2136 surfaces, while the actual dataset comprised 170 bitewing radiographs. The model was trained on 1000 images, achieving a remarkable 100% accuracy in classifying the left and right sides of the images, which was crucial for effective tooth numbering.

The proposed architecture consists of three main components: side detection using a DenseNet-121 model, teeth numbering, and caries detection through the YOLOv7 framework. The integration of CBAM allowed the model to focus on salient features within the images, enhancing detection performance. The study reported high recall and precision rates for both teeth numbering and caries detection, with the model achieving a recall of 99.4% and precision of 98.75%. The findings suggest that the improved YOLOv7-AP-CBAM model can significantly aid dental practitioners by providing a reliable and efficient tool for diagnosing dental caries, thereby addressing challenges associated with human error in clinical settings.