EnamelNet-TRiX: A Lesion-Aware Dual-Transformer With Cross-Attention for Early and Advanced Enamel Caries Diagnosis.

EnamelNet-TRiX: A Lesion-Aware Dual-Transformer With Cross-Attention for Early and Advanced Enamel Caries Diagnosis - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Int Dent J . 2026 Apr 7;76(3):109437. doi: 10.1016/j.identj.2026.109437 Search in PMC Search in PubMed View in NLM Catalog Add to search EnamelNet-TRiX: A Lesion-Aware Dual-Transformer With Cross-Attention for Early and Advanced Enamel Caries Diagnosis Sastika Balachandran Sastika Balachandran 1 School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, Tamil Nadu 600127, India Find articles by Sastika Balachandran 1 , Raja Marappan Raja Marappan 1 School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, Tamil Nadu 600127, India Find articles by Raja Marappan 1, ⁎ Author information Article notes Copyright and License information 1 School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, Tamil Nadu 600127, India ⁎ Corresponding author at : Vellore Institute of Technology, Chennai, Tamil Nadu 600127, India [email protected] Collection date 2026 Jun. © 2026 The Authors This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). PMC Copyright notice PMCID: PMC13090983  PMID: 41950800 Abstract Purpose Dental caries is one of the most prevalent oral diseases worldwide, and predicting Early Enamel Caries (EEC) and Advanced Enamel Caries (AEC) in intraoral imaging is a significant clinical research challenge. To overcome the challenges, this research presents an EnamelNet-TRiX, an automated diagnosis system for enamel caries based on a lesion-aware dual-transformer framework with across-attention guidance. Methods The proposed framework is developed using a shallow convolutional lesion-aware module (LAM) that incorporates on the lesions of the enamel combined with a Swin Transformer that zooms in on textures of the lesions, and a Vision Transformer (ViT) that captures the overall global structure to provide contextual global structural guidance followed by a multi-scale feature fusion stage that augments cross-stream attention with concatenation and unification for classification. The model is trained on the Caries-Spectra dataset, a set of internal images comprising 2000 intraoral images with three diagnostic classes (EEC, AEC, and No Enamel Caries [NEC]), which are cross-validated using an 80-20 split on a patient basis for training and testing. Results The model is externally evaluated on the DentRT-2 dataset that consists of 300 real-world intraoral images with diverse diagnostic conditions. The model achieved accuracy: 99.25%, precision: 98.94%, recall: 99.12%, and F1-score: 99.03%, respectively, for the Caries-Spectra dataset, while confirming 96.33% accuracy on DentRT-2. The simulation results show the solid domain generalisation of caries diagnosis. Conclusion The proposed lesion-aware dual-transformers with cross-attention, a multiscale fusion mechanism incorporating local, global, and lesion-prior features, and exhaustive internal and external clinical testing on a real-time dataset shed light on EnamelNet-TRiX as a trustworthy framework for the diagnosis of enamel caries. Key words: Enamel caries, Dual transformer, Lesion priors, Cross attention, Deep learning, Medical imaging Introduction Dental caries is one of the most prevalent global oral health issues, affecting every age group and significantly straining the public health system. 1 The condition begins with slow demineralisation of the enamel surface. If left untreated, it progresses to an extensive case of enamel caries, leading to cavitation and ultimately resulting in the loss of the affected tooth. 2 This is why early diagnosis is crucial, as invasive restorative techniques are often necessary for advanced carious lesions. Remineralisation therapies, on the other hand, can potentially stop the carious process or reverse early caries. Clinical diagnosis of dental caries traditionally relies on visual inspection supplemented by imaging modalities such as bitewing radiographs, panoramic radiographs, cone-beam computed tomography (CBCT), and intraoral photography. 2 , 3 , 4 , 5 However, these methods are limited by clinician subjectivity, variable imaging quality, and interobserver variability across practitioners. 6 CBCT offers volumetric visualisation and can assist in detecting subsurface lesions; however, its use is associated with higher radiation exposure and increased acquisition complexity. 7 Importantly, radiation-related concerns arise from the imaging modalities rather than from computational or deep learning (DL) algorithms, which operate exclusively on acquired images and do not introduce additional radiation burden. Several studies have demonstrated that well-optimised two-dimensional (2D) imaging approaches, including bitewing radiographs and intraoral photographs, can achieve diagnostic performance comparable to CBCT for enamel-level caries when combined with robust analytical models. 8 , 9 , 10 These issues highlight the need for systems that can provide automated, objective, and reproducible diagnostics to aid clinical judgment. 11 Figure 1A shows the global prevalence of dental caries by WHO region, illustrating its widespread burden across all age groups. Fig. 1. Open in a new tab A, Global prevalence of Dental caries by WHO region. B, Global breakdown of dental caries cases. C, Performance metrics of diagnostic tools for caries detection. Early computational approaches to dental caries detection employed traditional machine learning (ML) techniques based on handcrafted features and classical classifiers. Although these methods demonstrated initial feasibility, their performance was constrained by feature design bias, sensitivity to illumination variations, and limited adaptability to heterogeneous clinical datasets. 12 These problems are resolved using computational methods, such as diagnostic aids. The emergence of DL, particularly the application of convolutional neural networks (CNNs), can aid in identifying more specific features of enamel lesions and irregularities, as well as the boundaries of these lesions. 13 CNN-based models have consistently outperformed traditional ML approaches by learning hierarchical representations of enamel texture, lesion boundaries, and morphological variations. 6 , 9 Systematic reviews and meta-analyses further confirm that DL-based frameworks achieve superior diagnostic accuracy and robustness compared to conventional ML techniques, especially when trained on sufficiently large datasets. 14 These methods help clinicians document both specific and general anatomical information, which improves the visualisation of lesions. However, issues such as overlapping anatomical structures, varying illumination, and interindividual differences affect the accuracy and consistency of the diagnosis. 15 Figure 1B illustrates the global distribution of dental caries cases by severity and population group, highlighting the disparity in cases. Architectural advancements, such as U-Net and ResNet, have further strengthened CNN-based medical image analysis by enhancing feature propagation, localisation accuracy, and training stability in complex imaging tasks. 16 , 17 Unlike CNNs, models are inherently limited in capturing long-range spatial dependencies and global enamel context, which are critical for differentiating subtle early-stage lesions from normal enamel variations. To address these limitations, recent studies have explored transformer-based and hybrid CNN–transformer architectures that integrate local feature extraction with global contextual reasoning. 18 , 19 , 20 While these models demonstrate enhanced representational capacity, many existing approaches rely on single-branch feature extraction and lack explicit lesion-aware reasoning, limiting interpretability and clinical transparency. 21 The reported diagnostic performance in dental artificial intelligence (AI) literature varies widely, depending on the dataset composition, annotation quality, and validation strategy. CNN-based and hybrid models commonly report accuracies ranging from approximately 92% to 98% under internal validation settings; however, their performance often declines when evaluated on external or multicentre datasets. 6 , 22 Systematic reviews highlight that many studies employ single-centre datasets, image-level data splits, and limited clinical diversity, which can lead to inflated performance estimates that do not accurately reflect real-world diagnostic conditions. These findings emphasise the importance of external validation and clinically realistic evaluation protocols. Dataset characteristics further influence the reliability of dental AI systems. Publicly available datasets such as Caries-Spectra have facilitated reproducible research by providing annotated intraoral images across different stages of enamel caries. Nevertheless, many dental datasets suffer from class imbalance, limited sample sizes, and single-institution acquisition, particularly for early enamel lesions that are challenging to annotate consistently. 23 Interobserver variability among clinicians further complicates ground-truth labelling and impacts model training and evaluation. Multicentre datasets and independent testing cohorts have been shown to provide more reliable estimates of clinical performance by accounting for variability in imaging devices, patient demographics, and acquisition conditions. 24 , 25 Beyond predictive accuracy, clinical deployment of dental AI systems requires interpretability, robustness, and trustworthiness. Several studies have noted that black-box models, despite achieving high accuracy, face barriers to clinical adoption due to the lack of transparent lesion-level reasoning. Recent comprehensive reviews in dental AI emphasise that interpretability, external validation, and readiness for real-world deployment remain persistent challenges despite methodological advances. 26 , 27 Emerging clinically grounded applications, such as AI-driven CBCT-based decision-support systems, further highlight the necessity of rigorous validation and clinically meaningful evaluation when translating AI technologies into dental practice. 28 Figure 1C depicts the performance metrics of different tools for caries detection and highlights the advancements made using DL frameworks. Collectively, existing literature reveals persistent gaps in lesion-aware modelling, balanced local-global feature integration, interpretability, and robust cross-dataset validation for enamel caries diagnosis. These limitations underscore the need for a diagnostic framework that can explicitly focus on lesion-prone regions while preserving the global enamel context and maintaining stable generalisation across heterogeneous clinical datasets. In this context, the present study proposes EnamelNet-TRiX to address these challenges through lesion-aware representation learning and hybrid modelling strategies for reliable automated enamel caries diagnosis. Figure 2 illustrates the limitations of various models in dental caries studies. Fig. 2. Open in a new tab Limitations of dental caries analysis. A, DL models. B, ML models. C, Hybrid CNN–Transformer models. D, Explainable AI (XAI) models. 2 , 3 , 4 , 5 , 6 , 7 , 8 , 11 , 14 , 15 , 16 , 17 , 18 , 19 , 20 , 21 , 23 , 30 , 31 , 34 Material and methods This section outlines the comprehensive workflow of EnamelNet-TRiX, encompassing stages from dataset preparation and curation to optimisation. The Caries-Spectra and DentRT-2 datasets are collected to serve as benchmark and real-world evaluation datasets, respectively. The input images are standardised during the preprocessing and augmentation stages, and extra variability is added for robust training. EnamelNet-TRiX is designed with lesion-aware priors for feature extraction and classification, utilising dual transformer branches and multiscale fusion strategies. These strategies are streamlined with training to ensure reproducibility and clinical relevance, addressing the lack of variability in neural architecture. Dataset description The proposed EnamelNet-TRiX framework is assessed using the Caries-Spectra and DentRT-2 datasets. The Caries-Spectra is used for benchmark training and internal validation. This dataset comprises 2000 intraoral images across three diagnostic groups: AEC (800 images), EEC (800 images), and NEC (400 images). This dataset was assembled in 2021 and 2022 from multiple dental clinics in Rajshahi and Dhaka, Bangladesh, with variations in patient demographics and imaging parameters. To prevent data leakage, this dataset has been split at the patient level, with 80% allocated for training and 20% for testing. Moreover, 20% of the training set is kept aside for validation. The model kept the same patient images in a single partition to allow a realistic assessment of its diagnostic performance. Figure 3A shows representative intraoral images taken from the Caries-Spectra dataset corresponding to the visual features of the three diagnostic stages. Fig. 3. Open in a new tab Sample images illustrating caries stages. A, the Caries-Spectra dataset. B, Real-world DentRT-2 dataset. The EnamelNet-TRiX model is tested using the real-world DentRT-2 dataset, which was collected independently in 2025 from Pallavi Dental Clinic and Kovur Multispeciality Dental Clinic, and involves 300 intraoral images, with 100 images in each of the AEC, EEC, and NEC categories. DentRT-2 is used in real-life clinical settings and collected under controlled conditions, which naturally differ in light, reflections, and angles. This variability, which lacks structure, makes DentRT-2 a helpful tool in evaluating strength and clinical usefulness. Figure 3B shows sample intraoral images taken from the DentRT-2 dataset, demonstrating real-world imaging variations. Preprocessing stage The input images from the datasets are processed through a specific pipeline to reduce problems that arise from acquisition. Enhanced diagnostic enamel structures of sample images are shown in Figure 4A and B . There are several common frameworks utilised to preprocess dental images in achieving transparency and reproducibility in AI-based prediction modelling. 29 The original image from the intraoral cavity is defined as I o r i g which possesses a spatial resolution of h × w (height h , and width w ) consists of three-colour channels, c = { R , G , B } . The normalised image I n o r m is obtained from sequential processes. The intraoral image resolution is then improved by bilinear interpolation. Each image is resized to a resolution of h × w regardless of the original dimension. For the enhanced image using the interpolation feature filters, the value of the pixel is a result of a linear blend of the weights of the original pixel values, with the addition of parameters that w i j demonstrate. This transforms the parameters into a rectangular value without distorting the original dimension of the image. All datasets utilised in the DL processes require transformable dimensions. 13 , 14 Fig. 4. Open in a new tab Preprocessed image samples. A, The Caries-Spectra dataset. B, DentRT-2 dataset. C, The overall preprocessing flowchart for dental caries detection. The application of equalised histogram image processing resulted in the enhancement of the wreath-shaped points of enamel, thereby improving the sharpening of structures. Each intensity level is calculated using the equation s k , which is defined from the overall statistical collection of the approximate values of the specific images. The improved images enhance the overall sharpness of the subtle enamel lesions. 5 , 9 After enhancing contrast, brightness was normalised due to differences in illumination of intraoral images. Linear modification is done using a scale factor α and a shift term β , thus producing the brightness-adjusted image I b r i g h t , where α is responsible for the brightness scaling, while β adjusts global intensity offsets. These changes diminished the impact of lighting differences, which are prevalent in clinical situations when intraoral images are taken using nonstandardised methods. 7 , 8 Lastly, normalisation is done on the individual colour channels of each image, c ∈ { R , G , B } . On each channel, the values of the mean intensity ( μ c ) , and the standard deviation ( σ c ) are used to standardise pixel values in producing the final normalised image I n o r m . This procedure enhances the intensity distribution of the images due to pretrained vision models, thus improving convergence stability and generalisation in subsequent training. 14 , 19 The intraoral images are finally normalised as I = { I 1 , I 2 , I 3 … I n } . These are provided as input to produce the augmented images, which are standardised, enhanced for lesion visibility, adjusted for brightness, and further optimised for model processing. Figure 4C sketches the stages of the dental caries image preprocessing workflow defined in Algorithm 1 and Table 1 presents the parameters used during the preprocessing stage. Algorithm 1. Preprocessing of intraoral images for EnamelNet-TRiX. Input: Raw intraoral images I o r i g . Output: Preprocessed normalised images I n o r m . // Resize image 1: Set the target height and width to h × w = 224 × 224 . 2: Apply bilinear interpolation using I r e s ( x , y ) = ∑ i = x x + 1 ∑ j = y y + 1 w i j ( x , y ) I o r i g ( i , j ) // Apply contrast enhancement (histogram equalisation) 3: Compute the probability distribution of intensity levels, p ( r k ) . 4: Calculate the cumulative distribution function using s k = ( L − 1 ) ∑ j = 0 k p ( r j ) // Adjust the brightness 5: Determine scale factor α and shift β from the image histogram. 6: Update the pixel brightness using I b r i g h t ( x , y ) = α · I r e s ( x , y ) + β //Normalisation on individual colour channels 7: Normalise for each colour channel c = { R , G , B } using I n o r m ( c ) = I b r i g h t ( c ) − μ c σ c 8: Return the preprocessed, normalised images I n o r m = { I 1 , I 2 , I 3 … I n } . Open in a new tab Table 1. Parameters used in Algorithm 1. Parameter Value h × w 224 × 224 p ( r k ) [ 0 , 1 ] s k [ 0 , 255 ] α [ 0.8 , 1.2 ] β [ ( − 30 ) , ( + 30 ) ] μ c [ 0.485 , 0.456 , 0.406 ] σ c [ 0.229 , 0.224 , 0.225 ] Open in a new tab Augmentation stage As part of generalisation and to mitigate the effects of class imbalance, an all-inclusive augmentation approach is employed for the Caries-Spectra dataset. In contrast, the DentRT-2 dataset has not been augmented to maintain its real-world diagnostic integrity. Each preprocessed I n o r m image is transformed into several versions using geometric, photometric, and distortion manipulation, resulting in a balanced dataset of I a u g with equal representation across all diagnostic classes. The augmentation process starts with applying geometric transformations. Horizontal flipping is done with probability, p f l i p , so that the model can perform mirroring on anatomical structures that are frequently present in intraoral images. Then the random rotations are performed within θ defined boundaries to simulate the horizontal or vertical misplacements due to the rater or intraoral camera position. These geometric transformations enhanced the ability to perform cross-dental imaging. 6 , 13 To imitate alterations in sharpness, noise, and defocusing, the Gaussian blur with a kernel size k g and standard deviation is σ g are applied along with the brightness and contrast by scaling factor α and shift term β . These techniques enhanced the range of illumination conditions and improved the stability of device-specific lighting. 7 , 8 Then, the local contrast improvements are performed by contrast limited adaptive histogram equalisation (CLAHE) with the clipping threshold c l i p l i m i t . This technique is necessary to improve the contrast on fine details, enhance poorly defined boundaries of lesions without saturating noise, and distribute pixel intensities. The elastic transformations are applied through displacement fields δ x and δ y , and grid distortion, governed by grid size n × n , is employed to simulate diversity in morphology. Elastic transformations are used to shear images, and grid distortion is applied to introduce systematic changes in the alignment of enamel structures. 10 , 16 To simulate lens-associated imprecisions, optical distortion is synthesised by means of the distortion centre ( x c , y c ) , radius of distortion r , and the distortion parameter k . Thus, the realistic distortions associated with intraoral imaging devices are formed. Along with this, motion blur is applied with the kernel size. k m o t i o n to model artefacts resulting from patient movement and hand-held device instability. 30 , 31 Lastly, increases in tonal diversity are achieved by implementing colour-based transformations. To enhance the dynamic range of pixel intensity distributions, colour inversion is applied in conjunction with gamma correction, utilising a parameter γ to control the nonlinear intensity mapping of the colour range. These transformations are essential for the models to adapt to differences in the devices used in acquisition and the quality of the photographs taken. 14 , 19 The pipeline created a clinically varied dataset I a u g through geometric, blurring, distortion, and colour-based modifications. This method equally distributed the classes for the range of intraoral imaging conditions, which aided the model in classifying more heterogeneous clinical scenarios. The representative images demonstrating the difference in augmentation pipeline defined in Algorithm 2 are shown in Figure 5 and Table 2 shows the simulation parameters used in the augmentation stage. Algorithm 2. Image augmentation pipeline for Caries-Spectra dataset. Input: Preprocessed, normalised images I n o r m . Output: Augmented and balanced dataset I a u g . 1: For each input image I n o r m , perform steps (2) to (13). 2: Apply a horizontal flip along the vertical axis with probability p f l i p . 3: Apply random rotation within the defined angle range θ . 4: Apply a Gaussian blur with a kernel size k g and standard deviation σ g . 5: Adjust brightness and contrast by a scaling factor α and shift term β . 6: Apply CLAHE with the clipping threshold c l i p l i m i t . 7: Perform elastic deformation using displacement fields δ x , δ y . 8: Apply grid distortion of control grid size n × n . 9: Apply optical distortion with the distortion centre ( x c , y c ) , radial distance r , and the distortion coefficient k . 10: Apply motion blur with kernel size k m o t i o n . 11: Invert the colours using I ′ ( x , y ) = 255 − I ( x , y ) . 12: Apply gamma correction with parameter γ . 13: Store all augmented variants of the images into the dataset I a u g . 14: Return the dataset with the equal class distribution. Open in a new tab Fig. 5. Open in a new tab Caries-Spectra dataset preprocessed sample images after augmentation. Table 2. Parameters used in Algorithm 2. Parameter Value p f l i p 0.5 θ ±15° k g [ 3 , 5 ] σ g [ 0.1 , 2.0 ] α [ 0.8 , 1.2 ] β [ − 30 , + 30 ] c l i p l i m i t [ 2.0 , 4.0 ] δ x , δ y [ − 10 , 10 ] n [ 4 , 8 ] r 0 – max k [ − 0.3 , + 0.3 ] k m o t i o n [ 3 , 7 ] γ [ 0.8 , 1.5 ] Open in a new tab Model architecture AI has been applied in dental image analysis, caries detection and its schematic representation for prediction using different stages is analysed to develop the new model. 26 The EnamelNet-TRiX model is based on a hybrid framework that aims to enhance robust enamel caries diagnosis by integrating lesion-aware priors, dual transformer encoders, and multiscale fusion. The overall workflow of the EnamelNet-TRiX model is sketched in Figure 6A . The architecture is composed of four core components: the lesion-aware prior module (LAM), one branch in the Vision Transformer (ViT), the Swin Transformers, and a classification head with scales of cross-attention, multiscale feature fusion (MFF), and loss functions that account for uncertainty. 35 The internal architecture of the LAM module, including the convolutional stem, decoder, and mask generation, is sketched in Figure 6B . Each augmented input image I a u g is first passed through a shallow convolutional stem to produce low-level spatial embeddings, denoted as F s t e m . Then, a lightweight decoder generates a lesion attention mask M , which emphasises enamel regions with potential demineralisation. The lesion-aware prior representation I p r i o r is obtained by element-wise modulation of F s t e m and M . This approach guarantees that early enamel irregularities are selective and easily extractable. 5 , 31 Fig. 6. Open in a new tab EnamelNet-TRiX framework and model components. A, The overall workflow. B, The lesion-aware module (LAM) architecture. C, The Vision Transformer. D, The Swin Transformer. E, The classification head. The ViT branch, which emphasises its tokenisation, positional encoding, and attention process for global feature extraction, is illustrated in Figure 6C . The ViT branch captures global enamel structure by dividing the input I a u g into n p non-overlapping patches of size p × p . Each patch x p ϵ R p 2 C is flattened and linearly projected into an embedding space. To maintain spatial order, positional encodings E p o s are added, which help to maintain spatial relationships, and the input sequence Z 0 is formed. In each transformer layer, self-attention is performed with a learnable projection W Q , W K and W V , where the multiple heads capture the concatenation to produce the output embedding. The resulting global representation is denoted F V i T ϵ R D 1 , in which D 1 is the embedding dimension. 10 , 18 The Swin transformer architecture with the shifted window scheme over the four stages is sketched in Figure 6D . In parallel, the Swin Transformer branch divides the input into smaller portions containing local windows W j ∈ R M × M , where M is the size of the window. Self-attention is performed in each window with the addition of relative position bias terms B . Through merging hierarchical representations across multiple stages, the Swin branch generates the abstract and feature representations denoted F s w i n ∈ ℝ D 2 , where D 2 is the dimension of the features. This pathway captures finer details of enamel, especially the borders of lesions and the local textures. 16 , 19 To integrate complementary information, cross-attention is conducted over the global ViT features F V i T and the local Swin features F S w i n producing output features F g l o b a l and F l o c a l . These features are concatenated with the lesion-aware prior I p r i o r to generate a global feature vector F c o n c a t . Then, a projection function is applied to obtain the final fused embedding feature F f i n a l . This step ensures that the lesion contours, global structure, and local texture are maintained in the joint feature embedding for classification. 11 , 17 , 21 The final fused embedding F f i n a l is now mapped to class logits for the parameters W c and b c to obtain the predictions of AEC, EEC, and NEC classifications. The final output probabilities are calculated through the softmax layer and are denoted as y ^ . The categorical cross-entropy loss is used to supervise the model. In contrast, focal loss is applied to capture the minority classes, and aleatoric uncertainty regularisation is applied to handle ambiguous situations. Dropout with a rate of ρ is used to increase generalisation, and the other layers of the network are kept fixed. By using the AdamW optimiser, the model is trained with a certain learning rate η , with the weight decay and neural network structure defined by λ w d . This decay is scheduled with the cosine transformation. 18 , 32 The classification head component is sketched in Figure 6E , which includes the fully connected layers, a Softmax output, and the loss functions. Through this design, EnamelNet-TRiX effectively integrates lesion-aware priors, global context, and local lesion cues into a unified diagnostic representation, achieving high accuracy, sensitivity, and robustness in both benchmark and real-world datasets is defined in Algorithm 3 . Table 3 presents the simulation parameters used in the EnamelNet-TRiX model architecture. Algorithm 3. EnamelNet-TRiX model architecture. Input: Intraoral images I = { I 1 , I 2 , I 3 . . . I n } . Output: Caries class y ^ ϵ ℝ 3 = { A E C , E E C , N E C } . 1: Apply Algorithm 1 on each image I to get I n o r m . 2: Augment I using Algorithm 2 to get I a u g . // LAM 3: For each image I a u g , perform shallow feature extraction using a convolution stem: F s t e m = C o n v S t e m ( I a u g ) 4: Generate a lesion attention mask M from F s t e m via a light decoder: M = D e c o d e r ( F s t e m ) 5: Apply the mask to the stem features (element-wise multiply). I p r i o r = F s t e m ⊙ M // ViT branch 6: Segment I a u g into n p = h w P 2 nonoverlapping patches of size p × p : each patch x p ϵ R p 2 C , where h × w is the image resolution. 7: Each patch is flattened, linearly projected, and position encoding is added to form tokens. Z 0 = X p E + E p o s , E ϵ ℝ ( p 2 C ) × D t , E p o s ϵ ℝ n p × D t 8: For each layer l of the transformer, form queries, keys, and values. Q = Z ( l − 1 ) W k , V = Z ( l − 1 ) W v 9: Computes scaled dot-product attention for each head h : A t t n ( Q , K , V ) = s o f t m a x ( Q K T d k ) V 10: Multihead self-attention: M H S A ( Q , K , V ) = C o n c a t ( h e a d 1 , h e a d 2 , h e a d 3 . . . h e a d h ) W o 11: Obtain the ViT feature after the last block (CLS or pooled): F V i T ϵ R D 1 (or R n p × D 1 before pooling) // Swin Transformer branch 12: Divide I a u g into M × M windows W j ϵ R M × M using the shifted -window scheme over four stages. 13: Calculate Q , K , V per window and per stage. Q , K , V = P r o j ( W j ) 14: Find the attention to a window attributable to a relative Bias B . A t t n ( Q , K , V ) = s o f t m a x ( Q K T d k + B ) V 15: Aggregate over stages to obtain Swin features F s w i n ϵ R D 2 (or window-pooled). // Cross-Attention, MFF 16: Compute cross-attention of global (ViT) and local (Swin). F g l o b a l , F l o c a l = C r o s s A t t e n t i o n ( F V i T , F S w i n ) 17: Concatenate fused global, fused local, lesion-aware prior. F c o n c a t = c o n c a t ( F g l o b a l , F l o c a l , I p r i o r ) 18: Apply nonlinearity fusion projection. F f i n a l = R e L U ( W f F c o n c a t + b f ) // Classification head. 19: Apply linear mapping to logits. Z = W c F f i n a l + b c 20: Calculate SoftMax probabilities. y i = e Z i ∑ j = 1 3 e z j , i ∈ { 1 , 2 , 3 } 21: Classify enamel caries using y ^ = arg max i ∈ { 1 , 2 , 3 } y i = { A E C i f y 1 = max ( y 1 , y 2 , y 3 ) E E C i f y 2 = max ( y 1 , y 2 , y 3 ) N E C i f y 3 = max ( y 1 , y 2 , y 3 ) // Loss 22: Calculate categorical cross-entropy. L = − ∑ i = 1 3 y i t r u e log ( y i ) // Optimisation 23: Optimise with Adam (learning rate schedule), apply normalisation and dropout during training. Open in a new tab Table 3. Parameters used in Algorithm 3. Parameter Value/Range h × w 224 × 224 P 16 [ 8 , 32 ] n p = H W P 2 196 ( f o r 224 × 224 , P = 16 ) C 3 D t [ 384 , 768 ] d k D t h t D 1 [ 384 , 768 ] M 7 [ 6 , 12 ] D 2 [ 256 , 512 ] D f [ 256 , 512 ] ρ [ 0.1 , 0.5 ] η [ 1 × 10 − 5 , 3 × 10 − 4 ] λ w d [ 1 × 10 − 6 , 1 × 10 − 2 ] Open in a new tab Implementation details The experiments involved the use of PyTorch (v2.2) with a workstation running on an Intel Core i9 (13th Gen) with CUDA enabled, 16GB of RAM, and an NVIDIA GeForce RTX 4070 GPU. The use of Automatic Mixed Precision (AMP) helped with faster training and lower memory consumption. For reproducibility, the same deterministic random seed was used when splitting the dataset, initialising weights, and performing augmentation. Training the model was with the AdamW optimiser with an initial learning rate of η = 5 × 10 − 4 and weight decay λ w d = 5 × 10 − 4 . These values were effective by avoiding overfitting while achieving stable convergence across the dual-transformer layers. For smooth optimisation, a cosine decay learning rate scheduler with a warm-up period and the ReduceLROnPlateau method were used. The training objective was a combination of categorical cross-entropy, focal loss, which addresses class imbalance, and regularisation of aleatoric uncertainty, aiming to reduce prediction variance on ambiguous lesions. The use of a dropout rate ( ρ = 0.3 ) was implemented on the classification head to improve generalisation. For the model, the upper limit was set to 20 epochs with a batch size ( B = 32 ) , and early stopping was performed based on validation loss. The model’s best checkpoint was selected based on maximum validation F1-score and evaluated on the DentRT-2 dataset without test-time augmentation. These simulation parameters correctly classify the input images with good performance are shown in Table 4 . Table 4. Training parameters and their values. Parameter Value Learning rate 5 × 10 − 4 Weight decay 5 × 10 − 4 Dropout rate 0.3 Batch size 32 Epochs 20 Optimiser AdamW Scheduler ReduceLROnPlateau Loss function CE + Focal + Aleatoric Open in a new tab Results The EnamelNet-TRiX's performance is evaluated on the Caries-Spectra dataset (internal validation) and the DentRT-2 dataset (external testing). The evaluation utilises standard metrics, including accuracy, precision, recall, F1 score, and the area under the ROC curve (AUC-ROC). The corresponding confusion matrices provide an aggregate view of the diagnostic performance. Internal validation on the Caries-Spectra dataset evaluates the model’s ability to learn discriminative enamel features under controlled acquisition conditions. In contrast, external testing on the DentRT-2 dataset assesses generalisation under real-world variability. Reporting both results enables a transparent evaluation of robustness and mitigates concerns about dataset-specific overfitting. Evaluation metrics Accuracy: Accuracy is the ratio of the number of correctly classified samples to the total number of cases. It is expressed as a c c u r a c y = T P + T N T P + T N + F P + F N (1) where T P , T N , F P , and F N refer to the true positives, true negatives, false positives, and false negatives. TP represents the model correctly predicting an image as belonging to its actual class. TN defines an image that is correctly identified as not belonging to the actual class. FP defines the image as incorrectly predicted as belonging to a particular class when it actually belongs to another. FN defines the image that actually belongs to a particular class but is incorrectly predicted as another. EnamelNet-TRiX attained 99.25% accuracy on Caries-Spectra and 96.33% on DentRT-2. The observed reduction of approximately 2.9% between internal and external evaluation reflects differences in acquisition conditions, illumination, and patient-level variability between the datasets. The internal accuracy indicates that EnamelNet-TRiX has learned to identify features unique to enamel, and the external accuracy demonstrates that it has learned to generalise to diverse real-world conditions. Precision: Precision determines the positivity of a prediction and is defined as p r e c i s i o n = T P T P + F P (2) The proposed model achieved a precision score of 98.94% on Caries-Spectra and 95.7% on DentRT-2. With high precision achieved, this indicates that the model rarely misclassifies healthy enamel as caries, thereby avoiding clinical interventions associated with false positives. Recall (Sensitivity): Recall calculates the portion of caries that were captured. It is defined as r e c a l l = T P T P + F N (3) The EnamelNet-TRiX recorded a 99.12% recall on Caries-Spectra and 96.00% recall on DentRT-2. These figures confirm that the system is highly sensitive to the presence of both early and advanced enamel lesions, thereby minimising the chances of missing lesions in a clinical scenario. The closer measures of the datasets demonstrate that EnamelNet-TRiX calibrates well to maintain equilibrium between FP and FN. F1-Score: The F1 score is the weighted average of precision and recall, calculated as follows: F 1 s c o r e = 2 × p r e c i s i o n × r e c a l l p r e c i s i o n + r e c a l l (4) The model achieved an F1 score of 99.03% and 95.85% on Caries-Spectra and DentRT-2, respectively. This robustness suggests that EnamelNet-TRiX is counterbalanced in both positive and negative predictions, which circumvents the presence of overwhelming false negatives or false positives in the diagnostics. Confusion Matrix: To examine the results in more detail, the features across the three classes of AEC, EEC, and NEC are classified. Figure 7A shows that EnamelNet-TRiX achieved nearly perfect separation across all three classes, with very few misclassifications on both early and advanced lesions in the Caries-Spectra dataset. For DentRT-2, Figure 7B illustrates the improved performance of the proposed model, with some variations in the confusion matrix between AEC and EEC, which reflects the clinical challenge of differentiating between subtle stages of progression in real-world imaging. The model, however, can differentiate between caries-positive and caries-negative cases, which demonstrates its significant clinical value in automated enamel diagnosis. These misclassifications predominantly occur at transitional boundaries between early and advanced enamel caries, reflecting clinically recognised ambiguity rather than systematic prediction errors. Fig. 7. Open in a new tab Performance evaluation of the EnamelNet-TRiX model. A, Confusion matrix for the Caries-Spectra dataset. B, Confusion matrix for the DentRT-2 dataset. C, ROC–AUC curve for the Caries-Spectra dataset. D, ROC–AUC curve for the DentRT-2 dataset. AUC-ROC Curve: The ROC curve is a graphical representation that displays the balance between sensitivity and specificity of a test at different threshold levels. The area under the curve (AUC) is defined as follows: A U C = ∫ 0 1 T P R × d ( F P R ) (5) where TPR ( True Positive Rate ) = T P T P + F N (6) FPR ( False Positive Rate ) = F P F P + T N (7) EnamelNet-TRiX achieved an AUC of 99.78% on Caries-Spectra and 98.92% on DentRT-2, which are illustrated in Figure 7C and D . These scores reinforce the model’s impressive discriminative ability, even in diverse acquisition settings. These measures closer to 1 indicate the model’s robustness in both benchmark and clinical applications. These measures also indicate minimal overlap and strong confidence calibration in the classification. The consistently high AUC values across both datasets indicate stable class separability despite differences in data distribution. Performance comparison with state-of-the-art methods We further evaluated EnamelNet-TRiX against more recent state-of-the-art methods 4 , 6 , 18 , 19 , 31 , 33 in medical image classification and caries detection. Although these methods yielded positive results in radiographic and CBCT-based caries detection, some methods have chronic issues with external validation or lesion-aware priors. Table 5 shows the comparative performance of EnamelNet-TRiX on the Caries-Spectra dataset, achieving an accuracy of 99.25% with a 95% confidence interval (99.03%–99.47%), which indicates minimal variance and high reliability in the classification prediction. The model outperformed all prior methods by 1% to 3% in F1-score while maintaining an external accuracy of 96.33% with a 95% confidence interval (96.02%–96.64%) on DentRT-2, a robust and highly reliable dataset for external validation. In contrast, several existing methods reported in the literature exhibit notable performance degradation under external validation settings, highlighting sensitivity to dataset-specific characteristics. The addition of LAM with cross-attention fusion provided a specific advantage, particularly in external validation, which emphasises clinical robustness. Table 5. Comparison with state-of-the-art methods. Metrics State-of-the-art models 4 , 6 , 18 , 19 , 31 , 33 Proposed model CNN-based multi-input XAI DL model ResNet + Grad-CAM EFFResNet-ViT (Hybrid) Cross ViT with ECAP ExplainableDental Net ResNet50-SE Caries-Spectra DentRT-2 Accuracy (%) 95.8 ± 0.45 96.0 ± 0.42 95.4 ± 0.48 97.2 ± 0.36 97.5 ± 0.33 96.66 ± 0.90 98.30 ± 0.65 99.25 ± 0.22 96.33 ± 0.31 Precision 94.30 93.55 94.90 95.05 94.25 96.70 98.40 98.94 95.70 Recall 92.95 93.40 92.85 94.60 94.95 96.70 98.30 99.12 96.00 F1-Score 93.10 93.50 92.88 94.70 95.10 96.70 98.30 99.03 95.85 AUC-ROC 95.85 94.20 96.05 95.65 95.90 90.90 98.20 99.78 98.92 Open in a new tab Execution time and other resource comparison Along with the accuracy, computational efficiency is evaluated. As shown in Table 6 , EnamelNet-TRiX achieved faster inference than transformer-only models on an NVIDIA RTX 4070, which, on average, achieved an inference time of 23 ms per image. This figure is practical for real-time chairside use and results in less than 5GB of memory usage, permitting utilisation in mid-range GPUs. This confirms that the proposed model has optimal performance and efficiency and can be readily integrated and streamlined into existing clinical workflows. Table 6. Execution time and resource utilisation. Model Inference time (ms/img) GPU memory (GB) ResNet-50 15 3.2 ViT-only 35 6.8 Swin-only 28 5.7 EnamelNet-TRiX 23 4.9 Open in a new tab Ablation study To understand the contribution of each module, ablation experiments are done. The results are summarised in Table 7 . The module with LAM exhibits pronounced sensitivity to early lesions. Without cross-attention, the ability to align global and local information is weakened, and without multiscale fusion, balance is poorly maintained. In each case, the contribution is measurable, and the fusion and lesion-aware prior mechanisms are the most crucial in achieving the highest level of diagnostic accuracy. Table 7. Ablation study results. Configuration Accuracy (%) Precision (%) Recall (%) F1-score (%) Full EnamelNet-TRiX 99.25 98.94 99.12 99.03 – without LAM 97.80 96.50 97.20 96.85 – without cross-attention 97.95 96.80 97.40 97.10 – without multiscale fusion 98.10 97.00 97.65 97.32 Open in a new tab Discussion Diagnostic performance interpretation EnamelNet-TRiX participated in several distinguished diagnostic applications, resulting in the provision of competing diagnostic tools and technologies. Integration in the LAM enables the model to focus on enamel areas that are clinically prone to demineralisation, thereby reducing the influence of background artefacts and illumination variability. More focus on this effect is essential in the case of early enamel caries, where the lesions are small, and their boundaries are indistinguishable from those of normal enamel texture. By integrating a focus on both local and global components, the dual-transformer backbone enhances this functionality even further. Convolutional layers and Swin Transformer components focus on texture and boundary detail, while the Vision Transformer subnetwork surveys distant surface regions in search of longer-range spatial patterns. Validation of the model across different caries stages aligned these representations, resulting in more consistent classification. Clinical ambiguity, rather than architectural collapse, explains the borderline misclassification between early and advanced enamel caries, as clinicians may encounter challenges with transitional lesion stages. Generalisation and external validation Dental AI studies exhibit exceptionally high internal performance metrics; however, such studies are typically restricted to controlled benchmark datasets. Nonetheless, these results are not a sufficient condition to prove clinical performance. External validation on the DentRT-2 dataset in this study showed a slight decline in performance compared to the internal testing. This is to be expected and is a result of the variation in protocols for image acquisition, lighting conditions, and enamel unique to the patient. Most importantly, the relatively small degree of performance decline is a clear indication that EnamelNet-TRiX does not depend on dataset-related artefacts and is likely to perform the same diagnostic evaluations even when a domain shift is introduced. This is in contrast to the many DL models that exhibit a significant drop in accuracy when evaluated on alternative datasets. The patient-independent data splits, along with the absence of overfitting issues and external testing, demonstrate the generalisability of the suggested framework. Comparison with existing methods In EnamelNet-TRiX, compared to available CNN and transformer-based methods, several persistent issues related to dental caries detection have been adequately addressed. Traditional CNN architectures are effective in capturing local texture, but they have also been shown to be unable to incorporate a broader contextual view of the enamel. 4 In contrast, the transformer-only models are likely to underrepresent subtle features at the lesion level, which are necessary for the early detection of the lesions. Many prior hybrid approaches lack explicit lesion-focused reasoning or rely on single-branch fusion strategies, which can limit interpretability and robustness. 19 The proposed architecture addresses these limitations by focusing on lesion-aware priors and a cross-attention mechanism sharing fusion of local and global features. 18 EnamelNet-TRiX performance is compared with the recent explainable DL models in which no external validation is performed. 33 Compared to these models, the EnamelNet-TRiX obtained better performance in internal and external validation. Besides enhancing performance in terms of metrics, this design leads to more stable behaviour across different datasets and clinically relevant feature alignment. These architectural benefits account for the increased generalisability observed in external assessments, without relying on extensive data augmentation or dataset-specific fine-tuning. Clinical significance and practical implications From a clinical perspective, EnamelNet-TRiX acts as a decision-support system, as it is not an independent diagnostic system. The system can highlight lesion-prone enamel while maintaining a good degree of sensitivity and specificity, which aids practitioners in detecting early enamel demineralisation that is often missed in standard examinations. The benefits of early detection are particularly relevant to the field of preventive dentistry. It allows the implementation of noninvasive techniques such as remineralisation therapy. It may also reduce the need for restorative treatments. The model's moderate computational requirements and ability to perform inference in real-time strengthen the case for integrating it into chair-side workflows, intraoral imaging devices, and mobile diagnostic systems. Error analysis Even with strong performance, some instances of misclassification are present, particularly with EEC and AEC on DentRT-2. Such errors can be attributed to the subtle progression of lesions and inconsistent lighting conditions in real-world settings. Studying samples that had been misclassified showed that lesions that are borderline stage transitioning are the most difficult; this aligns with the problems documented with clinical staging. 8 , 32 Although the overall classification remained robust, the fine-grained differentiation between early and advanced stages remains a bottleneck. This could be improved by using temporal progression data or multimodal imaging. Limitations and deployment risks Some shortcomings need to be addressed, despite EnamelNet-TRiX achieving reliable results. Firstly, additional optimisation strategies may be necessary for the large dataset to further improve performance measures. Secondly, although DentRT-2 has been externally validated, additional validation across varying populations and imaging devices is essential for broader generalisability. Lastly, low-resourced clinical settings still face some restrictions; however, moderate computational requirements could enable systems to be deployed. The datasets used to train the model may introduce device bias, as the images are acquired under specific clinical conditions, and variations in image quality can impact model performance. The large-scale clinical deployment of the model may involve risks, including misdiagnosis liability, privacy concerns, hardware integration constraints, model drift, and potential overreliance by clinicians. Conclusion The proposed EnamelNet-TRiX model employs a lesion-aware dual-transformer framework for automated diagnosis of EEC and AEC. The performance of the model is improved through the use of a lesion-aware prior, global Vision, local Swin Transformer, and multiscale feature fusion in the Caries-Spectra benchmark and the DentRT-2 datasets. The experiments and external validation in the DentRT-2 dataset have confirmed that the model achieved high accuracy, with 99.25% and 96.33% for internal and external validation, respectively, along with high sensitivity and positive predictive values. The EnamelNet-TRiX model outperformed the state-of-the-art methods, particularly in terms of its robustness to internal and external validation and clinical imaging heterogeneity, which highlights its clinical relevance. The model's performance also offers some reasonably lesion-aware attention maps for interpretation, which adds to clinical relevance and model trustworthiness. The model is reliable in distinguishing between caries and noncaries, as well as differentiating between early and advanced borderline lesions. The expected absolute deviation between Caries-Spectra and DentRT-2 is only 2.9% ± 0.9%, indicating that less than 3% performance drift occurred under cross-domain conditions and demonstrating the generalisation of EnamelNet-TRiX, along with the stability of LAM and cross-attention fusion. EnamelNet-TRiX provides dependable, explainable, and generalisable computer-aided diagnosis of enamel caries, potentially increasing detection rates, reducing the need for invasive procedures, and ultimately improving the outcomes of oral healthcare. Future direction Future research can address the development of EnamelNet-TRiX optimised for real-time deployment on portable devices, incorporating radiographs and fluorescence imaging, as well as multiload alongside intraoral photographs, and expanding training on comprehensive, multi-institutional datasets. The fragmentation of lesion progression will be enhanced by integrating longitudinal datasets, and combining model predictions with patient-specific risk factors can assist in extending the incorporated framework toward personalised preventive dentistry. Informed consent All participants were informed about the purpose of the study, and written informed consent was obtained before image collection. The images were captured by trained personnel using standard clinical procedures. All collected images were anonymised before analysis to ensure patient privacy and confidentiality. No identifiable personal information is included in the dataset. Author contributions Contributed to conceptualisation, methodology, software, validation, writing – original draft, and visualisation : Balachandran. Contributed to formal analysis, writing – review and editing, and visualisation : Marappan. Ethics statement This study, involving the collection and use of clinical dental images, was conducted in accordance with the ethical standards of the institutional research committee and the 1964 Helsinki Declaration and its subsequent amendments. Ethical approval for the project was obtained from Pallavi Dental Clinic, Trichy (approval no.: IEC/2025/03) & Kovur Multispeciality Dental Clinic, Chennai (approval no.: IEC/2025/15). Funding This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Conflict of interest None disclosed. Acknowledgement The authors would like to thank Vellore Institute of Technology, Chennai, for providing the necessary infrastructure to carry out this research. The authors also extend their sincere thanks to Dr. Sanjukta Mohapatra, Pallavi Dental Clinic, Trichy, and Dr. P. Manodh, Kovur Multispeciality Dental Clinic, Chennai, for providing the real-world intraoral images used in the DentRT-2 dataset and for validating the model’s results. REFERENCES 1. Shahriar H.G.M., Islam M.M., Hannan UH. Caries-Spectra: a dataset of enamel caries. Mendeley Data. 2023;V2 doi: 10.17632/9jnf2jvghy.2. [ DOI ] [ Google Scholar ] 2. Chen I.D.S., Yang C.M., Chen M.J., Chen M.C., Weng R.M., Yeh CH. Deep learning-based recognition of periodontitis and dental caries in dental X-ray images. Bioengineering. 2023;10(8):911. doi: 10.3390/bioengineering10080911. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Esmaeilyfard R., Bonyadifard H., Paknahad M. Dental caries detection and classification in CBCT images using deep learning. Int Dent J. 2024;74(2):328–334. doi: 10.1016/j.identj.2023.10.003. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 4. Oztekin F., Katar O., Sadak F., Yildirim M., Cakar H., Aydogan M., Ozpolat Z., Talo Yildirim T., Yildirim O., Faust O., Acharya U.R. An explainable deep learning model to prediction dental caries using panoramic radiograph images. Diagnostics. 2023;13(2):226. doi: 10.3390/diagnostics13020226. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Karakuş R., Öziç MÜ, Tassoker M. AI-assisted detection of interproximal, occlusal, and secondary caries on bite-wing radiographs: a single-shot deep learning approach. J Digit Imaging. 2024;37(6):3146–3159. doi: 10.1007/s10278-024-01113-x. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Imak A., Celebi A., Siddique K., Turkoglu M., Sengur A., Salam I. Dental caries detection using score-based multi-input deep convolutional neural network. IEEE Access. 2022;10:18320–18329. doi: 10.1109/ACCESS.2022.3150358. [ DOI ] [ Google Scholar ] 7. Amasya H., et al. Evaluation of a decision support system developed with deep learning approach for detecting dental caries with cone-beam computed tomography imaging. Diagnostics. 2023;13(22):3471. doi: 10.3390/diagnostics13223471. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Szabó V., Szabó B.T., Orhan K., Veres D.S., Manulis D., Ezhov M., Sanders A. Validation of artificial intelligence application for dental caries diagnosis on intraoral bitewing and periapical radiographs. J Dent. 2024 Aug;147:105105. doi: 10.1016/j.jdent.2024.105105. [ DOI ] [ PubMed ] [ Google Scholar ] 9. Lee S., Oh S.-il., Jo J., Kang S., Shin Y., Park J.-won. Deep learning for early dental caries detection in bitewing radiographs. Sci Rep. 2021;11(1):16807. doi: 10.1038/s41598-021-96368-7. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Abdinian M., Keshani F., Soltani P., Spagnuolo G., Rengo C. Evaluation of the effects of postprocessing settings in digital bitewing radiographs on proximal caries detection. Clin Exp Dent Res. 2024;10(3):e889. doi: 10.1002/cre2.889. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Bayati M., Savareh B.A., Ahmadinejad H., Mosavat F. Advanced AI-driven detection of interproximal caries in bitewing radiographs using YOLOv8. Sci Rep. 2025;15:4641. doi: 10.1038/s41598-025-11849-3. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Dhake T., Ansari N. A survey on dental disease detection based on deep learning algorithm performance using various radiographs. Proc Int Conf Adv Sci Technol (ICAST) 2022:291–296. doi: 10.1109/ICAST56720.2022.10045236. [ DOI ] [ Google Scholar ] 13. Lian L., Zhu T., Zhu F., Zhu H. Deep learning for caries detection and classification. Diagnostics. 2021;11(9):1672. doi: 10.3390/diagnostics11091672. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Abbott L.P., Saikia A., Anthonappa R.P. Artificial intelligence platforms in dental caries detection: A systematic review and meta-analysis. J Evid Based Dent Pract. 2025 Mar;25(1):102077. doi: 10.1016/j.jebdp.2024.102077. [ DOI ] [ PubMed ] [ Google Scholar ] 15. Zanini L.G.K., Rubira-Bullen I.R.F., Nunes FLS. Enhancing dental caries classification in CBCT images by using image processing and self-supervised learning. Comput Biol Med. 2024;183 doi: 10.1016/j.compbiomed.2024.109221. [ DOI ] [ PubMed ] [ Google Scholar ] 16. Asci E., Kilic M., Celik O., et al. Automatic tooth caries segmentation in panoramic radiographs of children in primary, mixed, and permanent dentition. Children. 2024;11(6):690. doi: 10.3390/children11060690. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Neeraja R., Anbarasi LJ. DentC_BSFL Net: boundary refined and spatial adaptive feature learning network for enhanced dental caries segmentation. IEEE Access. 2025;13:38945–38956. doi: 10.1109/ACCESS.2025.3456239. [ DOI ] [ Google Scholar ] 18. Panyarak W., Suttapak W., Mahasantipiya P., et al. CrossViT with ECAP: enhanced deep learning for jaw lesion classification. Int J Med Inform. 2025;193 doi: 10.1016/j.ijmedinf.2024.105666. [ DOI ] [ PubMed ] [ Google Scholar ] 19. Hussain T., et al. EFFResNet-ViT: a fusion-based convolutional and vision transformer model for explainable medical image classification. IEEE Access. 2025;13:24612–24624. doi: 10.1109/ACCESS.2025.3459899. [ DOI ] [ Google Scholar ] 20. Song B., Kc D.R., Yang R.Y., Li S., Zhang C., Liang R. Classification of mobile-based oral cancer images using the vision transformer and the Swin transformer. Cancers. 2024;16(5):987. doi: 10.3390/cancers16050987. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. Zhang X., et al. Enhancing furcation involvement classification on panoramic radiographs with vision transformers. BMC Oral Health. 2025;25:153. doi: 10.1186/s12903-025-04111-9. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Zhang L., Zhang H., Niu F., Yu Z., Zhou Z., Liu H. MPGLNet: multi-scale progressive global–local fusion dental point cloud complementation network. Biomed Sign Proc Control. 2026;111 doi: 10.1016/j.bspc.2025.108394. [ DOI ] [ Google Scholar ] 23. Li X., Ma X., Zhao Y., Hu J., Liu J., Yang Z., Han F., Zhang J., Liu W., Zhou Z. A Multi-center dental panoramic radiography image dataset for impacted teeth, periodontitis, and dental caries: Benchmarking segmentation and classification tasks. J Imaging Inform Med. 2024;37(2):831–41. 10.1007/s10278-024-00972-8. [ DOI ] [ PMC free article ] [ PubMed ] 24. Bilal A., Khan A.H., Almohammadi K., Al Ghamdi S.A., Long H., Malik H. PDCNET: deep convolutional neural network for classification of periodontal disease using dental radiographs. IEEE Access. 2024;12:14201–14212. doi: 10.1109/ACCESS.2024.3345678. [ DOI ] [ Google Scholar ] 25. Wang Y., Ye F., Chen Y., et al. A multi-modal dental dataset for semi-supervised deep learning image segmentation. Sci Data. 2025;12:117. doi: 10.1038/s41597-024-04306-9. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Samaranayake L., Tuygunov N., Schwendicke F., et al. The transformative role of artificial intelligence in dentistry: a comprehensive overview. Part 1: fundamentals of AI and its contemporary applications in dentistry. Int Dent J. 2025;75(2):383–396. doi: 10.1016/j.identj.2025.02.005. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Tuygunov N., Samaranayake L., Khurshid Z., Rewthamrongsris P., Schwendicke F., Osathanon T., Yahya N.A., The transformative role of artificial intelligence in dentistry: A comprehensive overview Part 2: The promise and perils, and the international dental federation communique, Int Dent J, 75, 2, 2025, 397-404, ISSN 0020-6539, 10.1016/j.identj.2025.02.006. [ DOI ] [ PMC free article ] [ PubMed ] 28. Deng Y., He Y., Liu C., et al. AI-driven CBCT analysis for surgical decision-making and mucosal damage prediction in sinus lift surgery for patients with low RBH. Int Dent J. 2025;75(6) doi: 10.1016/j.identj.2025.103931. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Khurshid Z., Osathanon T., Shire M.A., Schwendicke F., Samaranayake L. Artificial intelligence in dentistry: a concise review of reporting checklists and guidelines. Int Dent J. 2025;76(1) doi: 10.1016/j.identj.2025.109322. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 30. Farook T.H., Ahmed S., Rashid F., Sifat F.A., Sidhu P., Patil P., Zabin Eusufzai S., Jamayet N.B., Dudley J., Daood U. Application of 3D neural networks and explainable AI to classify ICDAS detection system on mandibular molars. J Prosthet Dent. 2025;133(5):1333–1341. doi: 10.1016/j.prosdent.2024.09.014. [ DOI ] [ PubMed ] [ Google Scholar ] 31. Inani H., Mehta V., Bhavsar D., Gupta R.K., Jain A., Akhtar Z. AI-enabled dental caries detection using transfer learning and gradient-based class activation mapping. J Ambient Intell Human Comput. 2024;15(7):3009–3033. doi: 10.1007/s12652-023-05519-5. [ DOI ] [ Google Scholar ] 32. Park E.Y., Cho H., Kang S., Jeong S., Kim EK. Caries detection with tooth surface segmentation on intraoral photographic images. BMC Oral Health. 2022;22:573. doi: 10.1186/s12903-022-02534-7. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 33. Asiri F.Y.I. Automated classification of enamel caries from intraoral images using deep learning models: a diagnostic study. J Clin Med. 2025;14(24):8959. doi: 10.3390/jcm14248959. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 34. Hasan F., Tantawi M.E., Haque F., Foláyan M.O., Virtanen JI. Early childhood caries risk prediction using machine learning approaches in Bangladesh. BMC Oral Health. 2025;25:49. doi: 10.1186/s12903-024-04091-7. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 35. Gao N., Li Y., Chen P., Tang J., Liu T. Gfdet: multi-level feature fusion network for caries detection using dental endoscope images. Big Data Min Anal. 2024;7(4):1362–1374. doi: 10.26599/BDMA.2024.9020056. [ DOI ] [ Google Scholar ] Articles from International Dental Journal are provided here courtesy of Elsevier ACTIONS View on publisher site PDF (2.8 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top