A COMPREHENSIVE EXPLORATION OF NEURAL NETWORKS FOR DENTAL CARIES DETECTION A PROJECT REPORT Submitted by GUNAR (212220060059) JAYASHREE R (212220060079) In partial fulfilment for the award of the degree of BACHELOR OF ENGINEERING in ELECTRONICS & COMMUNICATION ENGINEERING SAVEETHA ENGINEERING COLLEGE (AUTONMOUS) ANNA UNIVERSITY: CHENNAI 600 025 APRIL 2023 ANNA UNIVERSITY: CHENNAI 600 025BONAFIDE CERTIFICATE Certified that this report titled “A COMPREHENSIVE EXPLORATION OF NEURAL NETWORKS FOR DENTAL CARIES DETECTION” is the bona-fide work of “GUNA R (212220060059) , JAYASHREE R (212220060079” who carried out the work under my supervision. SIGNATURE SIGNATURE Dr. SRIGITHA S NATH, ME. Ph.D Dr. MAUSAM CHOUKSEY, ME. Ph.D HEAD OF THE DEPARTMENT SUPERVISOR Professor Assistant Professor Department of ECE, Department of ECE, Saveetha Engineering College, Saveetha Engineering College, Thandalam, Thandalam, Chennai - 602105 Chennai ~ 602105 Submitted for the project viva-voce examination held on INTERNAL EXAMINER EXTERNAL EXAMINERPROJECT APPROVAL SHEET The project sheet “A COMPREHENSIVE EXPLORATION OF NEURAL NETWORKS FOR DENTAL CARIES DETECTION” submitted by “GUNA R (212220060059) , JAYASHREE R (212220060079)” is approved for submission, as partial requirement for the award of the Degree of Bachelor of Engineering in Electronics and Communication, Anna University during the academic year 1222023. Submitted for the project viva-voce examination held on INTERNAL EXAMINER EXTERNAL EXAMINERACKNOWLEDGEMENT If words were considered as symbols of approval and token of acknowledgement, then let words play the heralding role of not praising and glorification but also to exhibit the deeply embedded feelings of thanks and gratefulness. We wish to express our gratitude to our Founder president, Dr. N.M. Veeraiyan, President Dr. V. Saveetha and Director Dr. S. Rajesh and for providing us with the facilities for the completion of our project. We also thank our principal Dr.N.Duraipandian , M.E-Ph.D., for his continuous support and encouragement in carrying out our project work. Also, our thanks to Dr. Srigitha S Nath M.E., Ph.D., Head of the Department, Electronics and Communication Engineering of our college for providing us facilities and permission to carry out our project work We mention our special thanks to our project coordinator Dr. K. Nirmala devi M.E., Ph.D., Assistant Professor, Department of Electronics and Communication Engineering for his wisdom, guidance, and constant encouragement. We manifest cut deep sense of gratitude to our internal guide Dr. Mausam chouksey M.E, Ph, Assistant professor, Department of Electronics and Communication Engineering for his encouragement and faith in as though this endeavor. We take this opportunity to acknowledge the efforts to staff members and friends at Saveetha Engineering College for their sincere cooperation in guiding us. Finally, we express our deepest gratitude to our family, for their constant encouragement, support and motivation throughout under graduate career and for always being there for us.CHAPTER NO. Al 31 3.2 33 34 35 41 5.1 5.2 TABLE OF CONTENTS TITLE ABSTRACT LIST OF TABLE LIST OF FIGURES LIST OF SYMBOLS INTRODUCTION MODELS AND METRICS OF DIFFERENT ALGORITHMS, ADVANCEMENTS AND CHALLENGES IN DENTAL CARE LITERATURE REVIEW INTRODUCTION PAPERS PROPOSED WORK SYSTEM ARCHITECTURE 3.1.1 UART COMMUNICATION ACTUATATION MODULES FUTURE SCOPE PROPOSED SYSTEM FLOW DIAGRAM. THE DIAGNOSIS OF DENTAL CARIES MATLAB DESIGN SOURCE CODE CONCLUSION RESULT AND GRAPHS CONCLUSION w NO. 12 13 13 14 15 16 7 19 20 30 30 35ABSTRACT Dental caries, a progressive bacterial infection, stands as a leading cause of tooth loss, primarily stemming from inadequate dental hygiene practices. This pervasive issue gives rise to various other dental afflictions. The primary objective of this project is to advance early-stage detection of dental caries through the utilization of digital color imagery, facilitating more straightforward and effective treatment protocols. The proposed classification methodology holds promise for tele-dentistry, aligning with the burgeoning field of tele-informatic oral healthcare. To achieve this, we have implemented a Convolutional Neural Network (CNN) as the deep learning model of choice. Extensive training and testing were conducted on a binary dataset encompassing images both with and without cari . Remarkably, our results show an impressive accuracy rating of 99.13%. This remarkable outcome underscores the effectiveness of the MI-DCNNE model, illustrating its potential to significantly enhance the classification of dental caries. The findings of this study represent a crucial step towards revolutionizing early-stage detection of dental caries, offering a profound impact on oral healthcare practices and outcomes. wi)TABLE NO 1d 5.1 LIST OF TABLES NAME OF THE TABLE MODELS AND METRICS OF DIFFERENT ALGORITHMS DENTAL CARIES DATASET FEATURES DENTAL CARIES DATASET SPLIT ti) PAGE NO 4 31 34LIST OF FIGURES FIGURE NO NAME OF THE FIGURE PAGE NO COMPARATIVE PERFORMANCE OF DENTAL, 21 CARIES DETECTION ALOGRITHMS 11 3.1 PROPOSED SYSTEM ARCHITECTURE 13 3.2 PROPOSED SYSTEM FLOW DIAGRAM 14 5.1 MATLAB PROCESSING PHASE 31 FINAL CLASSIFICATION OF DENTAL a CARIES By FINAL PREDICTION MASK FOR DENTAL 53 CARIES 33 TRAINING PROGRESS - LOSS AND 54 INTERSECTION OVER UNION (IOU) PLOT 34 PROPOSED DENTAL CARIES CIRCUIT DIAGRAM 5.5 34 lv]LIST OF SYMBOLS AND ABBREIVATIONS CNN = Convolution Neural Networks Al - Artificial Intelligence Leb = liquid-crystal display ML - Machine Learning bc - Deep Leaning UART ~ Universal Asynchronous Receiver-Transmitter MEDCNNE = __ Multi-stage Deep Learning Convolution Neural Networks RGB: - Red,Green,Blue(color model) ToT - Internet of Things SVM - Support Vector Machine GSM - Global System for Mobile Communication Jo) - Intersection Over Union MATLAB - Matrix Laboratory wlCHAPTER 1 INTRODUCTION An estimated two billion adults and 520 million children experience dental caries in their permanent and primary teeth, respectively. Dental caries is predominantly preventable, yet if left untreated, it represents a significant concern. A significant health burden. The incidence of dental caries is on the rise in the majority of low and middle-income nations, indicating a growing trend. The issue at hand has a disproportionate impact on marginalized and underserved populations .In addition to this, dental facilities and clinicians are frequently dispersed in an uneven manner across geographic areas, which increases the number of hurdles to accessing dental treatment. There has been some discussion in over the use of artificial intelligence (AI) models to assist in the diagnosis of dental diseases. Deep learning (DL), a subfield of machine learning (ML), is the foundation for recent applications of artificial intelligence (Al) for imaging. The primary way in which DL diverged from past ML approaches was through the substitution of high-capacity neural networks trained on massive amounts of data for the human-engineered features that were previously used. This enabled automatic feature extraction. The earliest uses of DL in dentistry revolved around radiographic image analysis, such as for the detection of dental cavities. In dentistry, there have been more and more papers written about how AI models can be used. A recent systematic study looked at how DL can be used to find cavities in teeth using different types of images, like radiographs and near- infrared light transillumination. Dental diseases are common in humans and are getting worse, even though most of us know how to take care of our teeth. They are a major public health problem in many countries and affect people their whole lives, causing pain, discomfort, and deformity. They are also linked to other majornoncommunicable diseases like diabetes, heart disease, cancer, and chronic lung disease. ML methods are used to solve problems without any help from a person in order to predict diseases from oral health records. The algorithm and the amount of data used to train the model are the major things that determine how well and how efficiently the network works. Machines use both statistical and probabilistic models to make better decisions. They learn from past examples and can use this, information to make guesses, find new forms, or sort new data. In section I, contextualizing dental caries detection, emphasizing its significance in oral health. Reviewing existing research on dental caries detection methods, highlighting strengths and limitations are discussed in section II. In Section II Detailing the MI-DCNNE model and its constituent stages, including image processing, feature extraction, and classification. In section IV Presenting MI- DCNNE's 99.13% accuracy, discussing its implications in dental care. In section 5 summarizing MI-DCNNE's effectiveness and proposing future research directions for enhanced dental caries detection. 1.1 MODELS AND METRICS OF DIFFERENT ALGORITHMS When doing an analysis of research in dentistry that is based on AI, it is important to differentiate between picture classification, detection, and segmentation. The term "image classification refers to the process of providing a categorical label, typically binary, to an entire image. The term detection is used to refer to the process of determining the region of an outcome measure by using bounding boxes. Detection can also be thought of as a combination of classification and localization. The term segmentation refers to the process of establishing a specific outline for the result measure and separating that portion of the image from the rest of the image by annotating it pixel-by-pixel ., Because the included studies used such a wide variety of research approaches, it wasimpossible to aggregate their findings or directly compare their findings. However, when the results were broken down according to the different photographic modalities and formulation tasks, it became clear that AI models performed significantly better when they were given images captured by professional and intraoral cameras as opposed to photographs captured by smartphones. Oral health practitioners can considerably benefit from the detection of dental caries in oral images captured by professional and intraoral cameras. This is because the detection of dental caries in oral photographs can objectively verify the existence of caries and their severity, thereby minimizing diagnostic differences. As a consequence of this, only these studies were able to account for the possibility of confounding factors when evaluating AI model performance. The primary finding was that one-stage models, such as YOLOv, did not have a significant advantage over two-stage models like R-CNN. This was the conclusion drawn from the research. This discovery may be attributable to several different variables, such as overfitting as a result of small sample sizes, redundant complexity in the two-stage models, and data quality. However, more tests are needed to corroborate these findings before they can be considered conclusive. In terms of the different types of caries, the AT models showed significantly improved performance when detecting initial caries in comparison to secondary caries. This difference is probably due to the fact that secondary caries can look different in real life (for example, amalgam or composite fillings) and there are fewer secondary caries that have been marked in the dataset. Table 1 describes the models and their metrics of different algorithms.Ref. ‘Algorithm. Techniques Metrics no Sensitivity/Recall (%) ti] [| MaskR-CNN | Segmenta on 82.0% [12] | SVMDecision | Segmenta on - Tree [13] | SSD MobileNetV2] — Detec on 71.6% Faster RCNN (ResNet50) [14] | Customized CNNs | Classifica on 95.0% Us} [Mask R-CNN | Classifica on 97.7% [16] | YOLOv2 ssp Detec on 89.8% U7) [Mask R-CNN Detec on 96.0 % U8] CNN (NM) Classifica on = [19] | ResNet Shufflenet ] Classifica on 64.4 % V2 DenseNet ResNext [20] Squeeze Net Classifica on 58.2% Table 1.1 MODELS AND METRICS OF DIFFERENT ALGORITHMS 1.2 ADVANCEMENTS AND CHALLENGES IN DENTAL CARE The majority of dental problems are preventable, and when they do arise, they can typically be addressed more quickly. Oral diseases are becoming increasingly prevalent across the majority of low- and middle-income countries as a direct result of the urbanization process and associated shifts in living conditions. This is due in part to insufficient fluoride intake as well as inadequate access to dental health care services within the community. The increase in intake of substances that cause the problems that were previously discussed as well as new non-communicable illnesses has been contributed to by the advertising of sugary foods and drinks, as well as tobacco and alcohol. Thedevelopment of dental caries, which occurs when plaque accumulates on the surface of a tooth, is what triggers the breakdown of the free sugars found in food and drink into acids that eventually cause the tooth to be lost. Caries, which can cause discomfort, illness, and even tooth loss in extreme cases, can be caused by the continued consumption of free sugars at a high rate, by a lack of exposure to fluoride, and by a lack of brushing to eliminate plaque. The convolutional neural network (CNN) is an element of the deep learning algorithms that are particularly helpful for image analysis. Given that it is a fully connected network that mimics natural systems and draws inspiration from nature, this technique is prone to overfitting. CNN has a considerable advantage over other methods due to the fact that this method requires only a small amount of pre-processing. The prevailing limitation of current approaches lies in their inability to accurately gauge disease activity, a crucial factor in determining appropriate therapeutic interventions. Relying on existing tools may predispose patients to overdiagnosis and overtreatment, potentially resulting in inflated costs, compromised health, reduced healthcare acc sibility, and a widening of oral healthcare disparities. AI shows promise, it is imperative to approach its integration cautiously, given its nascent stage of development. Notwithstanding, automated caries screening holds the potential to shift from sporadic visits or reliance solely on observable disease indicators to real-time monitoring. Careful consideration of these factors will be instrumental in realizing the full potential of Al-based models in dental care. The primary objectives of this research papers are as follows: . To enhance feature extraction accuracy through the utilization of eight diverse pre-processing algorithms.+ Extracting RGB values from images, followed by conversion into Grayscale, forming a foundational step in the analysis process. + To implement advanced analytical features like Entropy, Kurtosis, Skewness, along with a sharpening filter for detailed focus on infected regions; additionally, integrate a voice note system for caries stage assessment.CHAPTER 2 LITERATURE REVIEW 2.1 INTRODUCTION: A concise and informative summary of scholarly sources related to a specific topic. It involves summarizing, synthesizing, and critiquing the literature that has been found through research. The purpose of a literature review is to provide background and context for a research project, identify developments in the field, validate the originality of the research, and evaluate methods and approaches. It helps researchers identify gaps in the existing literature and highlights strengths and weaknesses. The level of detail in a literature review depends on the purpose and audience, with more comprehensive reviews being necessary for academic works like theses and dissertations. 2.2 PAPERS: 1. Application of machine learning to stomatology In recent years, machine learning methods has been widely used in various fields, such as finance, spatial sciences, smart grid, intelligent transportation, renewable energy, agriculture, especially medicine. In the era of big medical data, the advantage of machine learning is that it can predict and diagnose through the analysis of a large number of clinical data, and its performance is very close and competitive to or even better than the performance of clinicians. This article focuses on the application of machine leaning techniques in the field of stomatology and detailedly describes application cases involving oral cancer, dental caries, periodontitis, dental pulp diseases, periapicallesions, oral implants, and orthodontics. Finally, the research obstacles and future work are discussed. 2. A new technique for diagnosis of dental caries on the children’s first permanent molar. The current researches have been shown high prevalence and incidence of children’s teeth caries, especially for the first permanent molar, which might do a lot of harm to their general health. Fortunately, early detection and protection can reduce the difficulty of treatment and protect children's oral health. However, traditional diagnostic methods such as dentist's visual inspection and radiographic imaging diagnosis are non-automatic and time-consuming. Given the COVID-19 epidemic, these methods should not be taken into consideration, since they fail to practice social distancing and further increase the risk of infection. To address these issues, in this paper we propose a novel caries detection and assessment (UCDA) framework to achieve a new technique for fully-automated diagnosis of dental caries on the children's first permanent molar. Inspired by an efficient in- network feature pyramid and anchor boxes, the proposed UCDA framework mainly contains a backbone network that is initialized with ResNet-FPN, and two parallel task-specific subnetworks for region regression and region classification. 3. Machine learning approach identified multi-platform factors for caries prediction in child-mother dyads. The oral microbiome of mother-child dyads was used in combination with demographic-environmental factors and relevant fungal information to create a multifactorial machine learning model based on the LASSO-penalized logistic regression. The most prevalent and consequential oral diseases globallyare dental caries (tooth decay), periodontal disease, tooth loss, and cancers of the lips and oral cavity. In this first of two papers in a Series on oral health, we describe the scope of the global oral disease epidemic, its origins in terms of social and commercial determinants, and its costs in terms of population wellbeing and societal impact. Although oral diseases are largely preventable, they persist with high prevalence, reflecting widespread social and economic inequalities and inadequate funding for prevention and treatment, particularly in low-income and middle-income countries (LMICs). As with most non- communicable diseases (NCDs), oral conditions are chronic and strongly socially patterned. Children living in poverty, socially marginalised groups, and older people are the most affected by oral diseases, and have poor access to dental care. In many LMICs, oral diseases remain largely untreated because the treatment costs exceed available resources. The personal consequences of chronic untreated oral diseases are often severe and can include unremitting pain, sepsis, reduced quality of life, lost school days, disruption to family life, and decreased work productivity. 4. Detection of dental caries in oral photographs. The most common chronic oral disease in the majority of adolescents and adults all over the world, and it can be divided into primary caries and secondary caries according to previous treatment experience. Primary caries is the most common type of caries in clinical treatment, while the secondary caries is caries that recurs due to the small gap between the edge of the filling and the surrounding dentin for various reasons, or the filling is performed before removal of the original lesion tissue, leaving the plaque in situ. The results of the fourth oral health epidemiological survey in China showed that the caries rate of permanent teeth in China was 89.0% in the age group of 35~44 years old, 95.6% in the age group of 55-64 years old, and 98.0% in the age group of 65~ 974 years old. We have used the images from the Kaggle dataset, and after tuning our model we were able to achieve 71.43% accuracy. 5, Dental photography using deep learning. Artificial Intelligence, has shown to be a powerful technique for computer-aided diagnostic support tasks. In machine learning, algorithms learn patterns and structures in data (training) and may then be applied to make predictions on unseen data (inference). A popular field in machine learning is deep learning, where multi-layered (deep) neural networks are used to learn nonlinear patterns in the data. For complex data structures such as imagery, so-called Convolutional Neural Networks (CNNs), are most commonly employed, learning hierarchical features such as edges, corners, shapes, and macroscopic patterns. In healthcare, and in particular for medical image analysis, deep learning using CNN has been successfully employed to detect breast cancer in mammographies skin cancer in clinical skin screenings , or diabetic retinopathy in retinographies . In dentistry, CNNs have been applied to detect periodontal bone loss on peri- apical and panoramic radiographs, and apical lesions and caries lesions on peri- apical radiographs, all with acceptable to high accuracies. Our approach gives 95% accurate results. The proposed methodology is not harmful to health, helpful for physicians, and can be used in computer-aided learning. It provides a comparative analysis of various dental caries detection algorithms, each identified by a reference number. The accuracy percentages showease the effectiveness of each algorithm in accurately identifying dental caries from images. Figure 8 shows comparative performance of dental caries detection algorithms.In algorithm [11] (Mask R-CNN) achieved an accuracy of 82.0%. Customized CNNs in [14] exhibited a higher accuracy of 95.0%. In algorithm [15] (Mask R-CNN) demonstrated a commendable accuracy of 97.1%. YOLOV2 SSD in [16] achieved an accuracy of 89.8% . In algorithm [17] 10(Mask R-CNN) displayed an accuracy of 96.0%. The Proposed Method, MI- DCNNE, outperformed all, achieving a remarkable accuracy of 99.13%. Comparative Performance of Dental Caries Detection Algorithms 100 80 [] x ° ‘Accuracy (8) 8 Mask RONN Customized CNNs m Mask R-CNN, =YOLOV2SSD Mask RCN mMLOCNNE Fig 2.1 Comparative Performance Of Dental Caries Detection Algorithms nuCHAPTER 3 PROPOSED WORK The proposed method employed eight distinct pre-proc sing algorithms to evaluate the dental caries. Initially, the RGB values of the images were extracted before being transformed into Grayscale representations. This crucial step laid the foundation for subsequent analysis. Convolutional Neural Networks (CNNs) serve as computational systems finely tuned for pattern recognition. Their versatility has ushered them into various domains, including healthcare, where they play a pivotal role in the early-stage diagnosis of diseases through image analysis. Further refining our approach, a sharpening filter was strategically applied to the Grayscale image. This process amplifies the intricate details of the affected regions, aiding in more precise identification. In a bid to augment our analytical capacity, advanced features such as Entropy, Kurtosis, Skewness, and others will be incorporated. These additions promise to furnish a more comprehensive and nuanced understanding of the underlying data. To implement a voice, note system that corresponds to different stages of caries. This innovative integration will not only facilitate more accessible and personalized communication about oral health but also mark a significant stride towards more inclusive and effective healthcare practices. This multifaceted approach stands poised to revolutionize how we understand, diagnose, and address dental caries.3.1 SYSTEM ARCHITECTURE It shows the proposed system architecture for finding the dental caries. The proposed block diagram outlines a comprehensive system for dental caries detection and notification. It leverages various components including a dental caries dataset, MATLAB for image processing, UART communication, an Arduino microcontroller, an LCD display, a buzzer, a GSM module, and an loT module. POWER SUPPLY ia Lm) MATLAB Arduino +) lor mopute Dental Caries DATASETS: } > GSM + AFFECTED f U __J Leo Fig. 3.1 Proposed System Architecture 3.1.1 UART COMMUNICATION Universal Asynchronous Receiver-Transmitter (UART) facilitates serial communication between the MATLAB environment and the Arduinomicrocontroller. It ensures seamless data exchange and control. The Arduino board acts as the central processing unit, It receives processed data from MATLAB through UART and executes subsequent actions based on the received information. 3.2 ACTUATATION MODULES The buzzer is employed for auditory feedback. It emits alerts or notifications in response to specific conditions, providing an additional layer of user interaction. The Global System for Mobile Communications (GSM) module enables wireless communication. In case of significant findings, the system can transmit notifications via text messages to predefined contacts. The Internet of Things (oT) module establishes a connection to the internet. This enables remote monitoring and control of the system, allowing for real-time updates and data accessibility through a web-based interface or a mobile application. DENTAL IMAGE es ENCHANCEMENT : = = ae CONTOUR) GLCMIFEATURE gj CLASSIFICATION(CNN) das conve: een FEATURES CARIES STAGES Voice OUTPUT Fig. 3.2 Proposed System Flow Diagram uBy integrating these components, the system creates a sophisticated pipeline for dental caries detection and notification. It combines image processing techniques with hardware implementation, enabling efficient and effective dental healthcare support. This system is poised to revolutionize early-stage detection and remote monitoring of dental caries, ultimately contributing to improved oral health outcomes. 3.3 FUTURE SCOPE To comprehensively explore neural networks for dental caries detection, consider these avenues: Data Augmentatio Enhance your dataset with varied examples through techniques like rotation, flipping, and scaling to improve model generalization. Transfer Learning: Leverage pre-trained models on large datasets and fine-tune them for dental caries detection, adapting the knowledge from broader medical imaging. Architectural Exploration: Experiment with different neural network architectures such as CNNs, RNNs, or hybrid models to identify the most effective structure for dental caries detection. Attention Mechanisms: Implement attention mechanisms to focus on relevant regions witl dental images, aiding in better feature extraction. 1sEnsemble Methods: Combine predictions from multiple neural networks to enhance overall model performance and robustness. Explainability: Integrate interpretability techniques to understand and visualize the decision-making process of your neural network, crucial for gaining trust in medical applications. Data Privacy and Security: Investigate methods to ensure patient data privacy and security, especially when dealing with sensitive medical information. Clinical Validation: Collaborate with dental professionals to validate the model's performance on diverse patient populations and real-world clinical scenarios. 3.4 PROPOSED SYSTEM FLOW DIAGRAM It illustrates the proposed system flow diagram. The proposed flow diagram for dental caries detection encompasses several critical stages, each contributing to accurate identification and assessment. The process initiates with a comprehensive dental caries database containing a diverse set of images, including both positive (with caries) and negative (without caries) cases. This dataset serves as the foundation for training and validation. Prior to analysis, the images undergo a crucial enhancement phase. This step aims to improve image quality, contrast, and overall visual clarity. It involves techniques like histogramequalization, noise reduction, and contrast stretching. Next, the enhanced images undergo segmentation, a pivotal step in isolating the regions of interest. In this context, segmentation focuses on delineating the dental structures from the background and isolating areas potentially affected by caries. CNN is employed to extract intricate features from the segmented images. The CNN architecture is adept at recognizing patterns and structures, allowing it to discern subtle details indicative of dental caries. The extracted features are fed into a classification CNN model. This model is trained to distinguish between images with and without caries. Through iterative learning, CNN refines its ability to accurately classify cases. In tandem with CNN-based features, the system employs Gray- Level Co-occurrence Matrix (GLCM) feature extraction. This technique quantifies the spatial relationships of pixel intensity levels, providing supplementary insights into texture properties associated with caries. The amalgamation of CNN-based features and GLCM analys s equips the system to discern nuanced patterns indicative of dental caries. Through a rigorous training process, the classification CNN refines its ability to accurately identify and categorize images into caries-positive and caries-negative cases. This integrated approach ensures a robust and reliable assessment of dental health. By leveraging the strengths of each component, the system achieves a high level of accuracy, setting the stage for improved early-stage detection and management of dental caries. 3.5 THE DIAGNOSIS OF DENTAL CARIES The proposed system architecture for finding the dental caries. In this research project, the goal is to create a mobile application utilizing deep learning for the diagnosis of dental caries in X-ray images. The application comprises three layers. The user layer allows users to upload or capture dental X-ray ”images, the algorithm layer identifies whether the tooth is healthy or affected by caries, guiding subsequent fertilization in dental care practices. The final layer application interface, displaying the tooth status through the application of the RCNN technique in deep leaming. The context of detecting dental caries in X-ray images, the Fast RCNN architecture comprises two independent modules: the Region Proposal Network (RPN) and the Fast R-CNN. Notably, these modules can be trained separately. However, a significant evolution is observed in Faster R-CNN, where a unified network is established. This unified network simultaneously trains both the RPN and Fast R-CNN, introducing a core concept known as layer sharing or feature sharing.CHAPTER 4 MATLAB DESIGN This project paper delves into the groundbreaking realm of dental caries detection, employing advanced Neural Network approaches. Leveraging MATLAB as a robust computational tool, our study aims to revolutionize traditional methodologies in identifying and diagnosing dental caries. The paper presents a comprehensive exploration of various neural network architectures, their implementation, and the subsequent impact on accuracy in caries detection. This segment elaborates on the various Neural Network models employed, such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). Comparative analyses are conducted to showcase the strengths and weaknesses of each architecture. This section critically reviews prior research on dental caries detection, highlighting gaps and limitations. It establishes the foundation for the necessity of Neural Network solutions and contextualizes the significance of MATLAB in this context. Acknowledging challenges faced during the project, this section suggests potential improvements and avenues for future research. It encourages the ongoing evolution of neural network applications in dental diagnostics. Summarizing key findings and contributions, the underscores the transformative potential of Neural Network approaches in revolutionizing dental caries detection. It emphasizes the practical implications for the field and the positive impact on patient care. A comprehensive list of references is provided, citing relevant literature, research papers, and MATLAB documentation, ensuring the academic integrity of the project paper.4.1 SOURCE CODE cle clear all close all warning off [filename,pathname]=uigetfile("*.*','Pick a MATLAB code file’); filename=streat(pathname, filename); I=imread(filename); imshow(I); Istrech = imadjust(I,stretchlim(1)); figure(2),imshow(Istrech) title(‘Contrast stretched image’) %K = medfilt2(Istrech); % figure(3),imshow(K) %% Convert RGB image to gray % I = rgb2gray(Istrech); % figure(3),imshow(I1,[]) title(RGB to gray (contrast stretched) ') I = imresize(I,[200,200]); gaussianFilter = ',20, 10); img_filted = fspecial(‘ga imfilter(1, gaussianFilter, symmetric’); figure imshow(img_filted); title('gaussianFilter Filted Image’);filted_edges = edge(img_filted, Canny’); figure(): subplot(121); imshow(filted_edges); title(‘Edges found in filted image'); img_edges = edge(Z, ‘Canny’); subplot(122); imshow(img_edges); %% Apply median filter to smoothen the image K = medfilt2(1); figure(4),imshow(K) title(median filter’) % MSE and PSNR measurement [row, col] = size(I); mse = sum(sum((I(1,1) - K(1,1)).42)) / (row * col); psnr = 10 * log0(255 * 255 / mse); disp(‘<- ---- Median filter ----. disp(‘Mean Square Error '); disp(mse); disp(Peak Signal to Noise Ratio’); disp(psnr); disp('<-- imgID = 2; L = kmean(K); % Here we initialize some parameters disp(‘Segmentation.’); %% Him = adapthisteq(K); % Contrast Limited Adaptive Histogram Equalization figure,imshow(HIm); title(HISTOGRAM ADAPTIVE); se = strel(‘ball',8,8); % Structuring Element gopen = imopen(HIm,se); % Morphological Open godisk = HIm - gopen; %Remove Optic Disk medfilt = medfilt2(godisk); % 2D Median Filter background=imopen(medfilt,strel(‘disk’,100)); %imopen function 12 = medfilt - background; % Remove Background GC = imadjust(12); figure, imshow(GC); title(ADJUST IMAGES); IM=GC; IM=double(IM); [maxX,maxY ]=size(IM); IMM=cat(3,[M,IM);ccl=; ec2=2 50; ttFem =0; while(t tFem< 10) ttFem=ttFem+1; sttPem=(['ttFem =" num2str(ttFem)]); cl=repmat(ccl maxX,maxY); c2=repmat(cc2,maxX,maxY); test2=c2; end c=cat(3,cl,c2); ree=repmat(0.000001,maxX,maxY); reel=cat(3,ree,ree); distance=IMM-c; distance=distance.*distance+ree; daoShu=l /distance; daoShu2=daoShu(:,:,1)+daoShu(:,:,2); 1).*daoShu2; ul=l /distancel; distance2=distance(:,:,2).*daoShu2;u2=I /distance2; cccl=sum(sum(ul.*ul.4TM))/sum(sum(ul.*ul)); ccc2=sum(sum(u2.*u2.4IM))/sum(sum(u2.*u2));— tmpMatrix=[abs(ccl - ccel Jee! abs(ce2-cee2)/ee2]:; pp=cat(3,u1,u2); forisl:maxX for jel:maxY if max(pp(ig.:))==Ul Gj) 1X2(i,j)=1; else IX20,j)=2: end end end if max(tmpMatrix)<0.0001 break: else ccl=cce: cc2=cce; end for IMMM(i,j)=254; else IMMMG,j)=8; end endend background=imopen(IMMM. strel(‘disk',45)); 14=IMMM-background; 14=bwareaopen(14,30); figure, imshow(14); title(MIFNET SEGMENTATION); for i=I:ma xX for jel:ma xY if TX20,j IMMM(ij)=200; else IMMMG,j)=1; end end end ffcm1=(['The Ist Cluster =" num2str(cce1)]); ffem2=(("The 2nd Cluster =" num2str(ecc2)}); [m,n]=size(14); Tn=0; Tp=0; Fp=0; Fn=0; for i=1:mTn=Tn+l; elseif MG,j==1 && Maj Tp=Tp+l; elseif 14G,j)==1 && Mij)== Fp=Fp+l; elseif MG,j)==0 && 1ij)= Fn=Fn+1; end end end auce=(Tp+Tn)/(Tp+Tn+Fp+Fn); % Accuracy sensitivity=Tp/(Tp+Fn); % True Positive Rate specificity=Tn/(Tn+Fp); 9% True Negative Rate fpr=1-specificity; % False Positive Rate ppv=Tp/(Tp+Fp); disp(‘True Positive disp(Tp); disp('True Negative =');disp(Tn); disp(False Positiv disp(Fp): disp(False Negative = '); disp(Fn); disp(‘False Positive Rate = '); disp(fpr); disp(‘Sensitivity = disp(sensitivity); disp(‘Specificity = disp(specificity); disp(‘Accuracy ='); disp(auce); disp( Positive Predictive Value ="); disp(ppv); end Img=double(I epsilon = 1; switch imgID case 1 num_it =1000;rad = 8; alpha = 0.3;% coefficient of the length term mask_init = zeros(size(Img(:,:.1))); mask_init(15:78,32:95) = 1; seg= local_AC_MS(Img,mask_init,rad,alpha,num_it,epsilon); —case2 num_it =800; rad = 9; alpha = 0.003:% coefficient of the length term mask_init = zeros(size(Img(:,:,1)));_ mask_init(53:77,56:70) = 1; seg=local_AC_UM(Img,mask_ ad,alpha,num_it,epsilon); case 3 num_it = 1500; rad alpha = 0.001% coefficient of the length term mask_init = zeros(size(Img(:,:,1)));_ mask_init(47:80,86:99) = 1; seg=local_AC_UM(Img,mask_init,rad,alpha,num_it,epsilon); end [labeledImage, numBlobs] = bwlabel(Img); props = regionprops(labeledImage, 'Centroid’, ‘Area’); allAreas = [props.Area] blobsToKeep = true(numBlobs, 1); % Get new cariesMask Img = ismember(labeledImage, find(blobsToKeep)); % Extract blobs only that are bigger than 100 pixels. Img = bwareaopen(Img, 100);% Dilate back out to original size. Img = imdilate(Img, true(5)): % Display image. subplot(2, 4, 6); imshow(img, [)): impixelinfo; % title('Caries’, ‘FontSize’, fomtSize, ‘Interpreter’, ‘None’); impixelinfo; % Get boundaries boundaries = bwboundaries(Img); % Display over original image. subplot(2, 4, 7); imshow(Img, (1); % — title(Original Grayscale ~— Image’, _—_‘FontSize’, ‘Interpreter’, 'None’); impixelinfo; hold on; fork=1: length(boundaries) thisBoundary = boundaries{k}; x= thisBoundary(:. 2); y= thisBoundary(:, 1); plot(x, y, 'r’, 'LineWidth’, 3); end fontSize,CHAPTER 5 CONCLUSION 5.1 RESULT AND GRAPHS: The dental caries detection system demonstrated remarkable accuracy in identifying early-stage dental caries. The implementation of the MI- DCNNE (Multimodal Integration with Deep Convolutional Neural Network Ensemble) approach yielded highly promising outcomes in dental caries detection. The integration of multiple modalities, including image enhancement, segmentation, CNN-based feature extraction, GLCM feature analysis, and classification CNN, collectively contributed to the system exceptional performance. Furthermore, the integration of GLCM features complemented the CNN-based analysis, offering additional texture-based insights. The combination of these features significantly contributed to the system's overall accuracy, achieving a comprehensive evaluation of dental health. Table 2 describes the dental caries dataset, indicating the total count along with the segmentation into training, testing, and validation sets. The dataset comprises a total of 1000 dental images. Among these, 75% (750 images) are allocated for training the MI- DCNNE model, ensuring a robust learning process. Subsequently, 15% (150 images) are set aside for testing, evaluating the model's performance on unseen data. An additional 10% (100 images) form the validation set, contributing to the fine-tuning and optimization of the model's parameters, This balanced distribution facilitates a comprehensive evaluation of the model's accuracy and generalization capabilities.Total Training Testing Images Images Images TABLE 5.1 Dental Caries Dataset features The initial MATLAB processing phase s played a crucial role in enhancing the quality and preparatory analysis of dental images. This preliminary step set the stage for subsequent feature extraction and classification in the MIDCNNE model. = ni ipee ier Seite panes TTY Fig. 5.1 MATLAB Processing Phase The initial MATLAB processing phase provided a robust foundation for subsequent stages in the MI-DCNNE model. The enhanced images and precisely segmented regions ensured that the subsequent CNN and GLCM analyses were 31conducted on high-quality data. Figure 4 shows the classification model of proposed dental caries significantly contributed to the model's remarkable accuracy of 99.13% in dental caries detection. Histogram equalization, noise reduction, and contrast stretching techniques were employed to enhance image quality. As a result, the visual clarity and distinctiveness of dental structures were significantly improved. The segmentation process successfully delineated dental structures from the background, providing clear delineation of regions potentially affected by caries. This step was pivotal in isolating relevant features for subsequent analysis. The extracted features from the segmented images laid the foundation for subsequent CNN analysis. These features encapsulated critical information about textural patterns and structures indicative of carious lesions. The GLCM analysis further enriched the feature set by quantifying spatial relationships among pixel intensity levels. This texture analysis provided supplementary insights into caries-related properties. fae ate panacea pontine se sooed: G0Gjeo : by ee i a Fig. 5.2 Final Classification of Dental Caries 32Fig. 5.3 Final Prediction mask for dental caries It presents the conclusive prediction mask generated by the MI- DCNNE model for dental caries detection. The mask highlights regions identified as potentially affected by caries based on the model's analysis of the input dental image. Darker regions on the mask indicate areas predicted as positive for caries, providing a visual representation of the model's inference. This output is crucial for clinicians in ass ng and localizing potential carious lesions, contributing to more accurate diagnosis and targeted treatment planning 33a ‘l 0 no ” 9 . fb on » 0 OD fpoch Fig, 5.4 Training Progress - Loss and Intersection over union (foU) Plot Fig. 5.5 Proposed Dental Caries Circuit diagram Ref. no ‘Algorithm ‘Accuracy (%) ay Mask CNN woe TH Cason 350% CNN Tr ‘Mik R-CNN wm 7 Mask CNN 95.0% Proposed Metbod | _ MHDCNNE 9.13% TABLE 5.2 DENTAL CARIES DATASET SPLIT5.2 CONCLUSION The implementation of the MI-DCNNE model, integrating various modalities including image enhancement, segmentation, CNN-based feature extraction, GLCM feature analysis, and classification CNN, has demonstrated exceptional promise in dental caries detection. The ensemble of techniques yielded a high classification accuracy of 99.13%. This outcome underscores the effectiveness and reliability of the proposed approach. The success of the MIDCNNE model can be attributed to its comprehensive evaluation framework, capturing both structural and textural nuances associated with caries. The integration of CNN and GLCM features proved to be particularly powerful, showeasing the model's discriminatory prowess. In future work to further enhance and extend the capabilities of the MI-DCNNE model. Incorporating a larger and more diverse dataset would bolster the model's generalizability and robustness in real-world clinical settings. 35REFERENCES * Peres, M. A., Macpherson, L. M. D., Weyant, R. J., Daly, B., Venturelli, R., Mathur, M. R., Listl, S., Celeste, R. K., Guarnizo-Herrefto, C. C., & Kearns, C. (2019). Oral diseases: A global public health challenge. The Lancet, 394(10194), 249-260 * Morid, M. A., Borjali, A., & del Fiol, G. (2021). A scoping review of transfer learning research on medical image analysis using ImageNet. Computers in Biology and Medicine, 128, 104115. © Lee, J-H., Kim, D.-H., Jeong, S.-N., & Choi, S.-H. (2018a). Detection and diagnosis of dental caries using a deep learning-based convolutional neural network algorithm. Journal of Dentistry, 77, 106-11 *® Mohammad-Rahimi, H., Motamedian, S. R., Rohban, M. H., Krois, J., Uribe, S., Nia, E. M., Rokhshad, R., Nadimi, M., & Schwendicke, F. (2022). Deep learning for caries detection: A systematic review. Journal of Dentistry, 106, 104115 © Political declaration of the high-level meeting of the general assembly on the prevention and control of non-communicable diseases: draft resolution, 2011, 13 p. « Y. W. Tang, Molecular Medical Microbiology. New York, NY, USA: Academic, 2014. © W. W. Johnson, “The history of prosthetic dentistry,” J. Prosthetic Dentistry, vol. 9, no. 5, pp. 841-846, Sep/Oct. 1959, doi: 10.1016/0022- 3913(59)90049-6. e S.C. White and M. J. Pharoah, Oral Radiology-E-Book: Principles and Interpretation. Amsterdam, The Netherlands: Elsevier, 2014. 36Hochreiter S, Schmidhuber J. Long short-term memory. Neural Comput 1997;9:1735-80. V. Geetha, K. S. Aprameya, and D. M. Hinduja, “‘Dental caries diagnosis in digital radiographs using back-propagation neural network,” Health Inf. Sci. Syst., vol. 8, no. 1, pp. 1-14, Jan. 2020. Ren S, He K, Girshick R, Sun J. Faster R-CNN: towards real-time object detection with region proposal networks. IEEE Trans PatternAnal Mach Intell 2017;39:1137-49. LeCun Y, Bottou L, Bengio Y, Haffner, P. Gradient-based learning applied to document recognition. Proc IEEE 1998:86:2278-323 Fukushima K. Neocognitron: a self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position. Biol Cybern 1980;36:193-202. Krizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep convolutional neural networks. In: Advances in neural information processing systems. 2012. p.1097-105. 37