Dental image analysis and surgical decision support using modified generative adversarial networks

Abstract

Background

Several clinical issues have remained in dental care and risk estimation because of the insufficient integration of multimodal features, the incapacity to integrate clinical domain literatures and evidence-based principles in their analytical platforms. The methods of deep learning-based diagnosis offer the in-depth examination of complex cases synthesized based on heterogeneous sources, hence, they require appropriate structural adjustments that can be implemented with the innovative components into a single system.

Methods

A modified Generative Adversarial Network (mGAN) is presented, which incorporates three key innovations: (1) We introduce a novel W-Net architecture which is formed by combining U-Net and V-Net to develop dual-path encoding and skip-gated decoding mechanisms. It enables multi-scale feature analysis capabilities specifically optimized for complex dental anatomical structures. (2) Large Language Model (LLM) supports to incorporate hyper parameter turning of acquired refined features based on conditional and guided generator based on domain knowledge distilled on mandibular anatomy and prevalent dental pathology directly into the deep learning pipeline. Thus, it shows complex medical scenarios beyond purely data-driven pattern recognition. (3) Probability Feature Fusion (PFF) is a novel mechanism for adaptive multi-modal feature integration, enabling dynamic weighting of different information sources based on local image content and clinical context from LLM. PPF combines multi-modal features within a unified analytical framework for maintaining interpretability and clinical relevance for generating probabilistic ground-truth maps representing contact points between mandibular third molar (M3M) roots and inferior alveolar canal (IAC).

Results

The mGAN framework proposed had 95.7% Root Surface Area Accuracy (RSAA) in a clinical assessment of dental and a Dice Similarity Coefficient (DSC) of 0.925 in the segmentation of anatomical structures. The system had a sensitivity of 91.8% and specificity of 96.7% of surgical necessity classification relative to expert consensus. Ablation experiments inferred that the knowledge integration with LLM improved risk correlation by 6.4% over baseline GAN models, and Probability Feature Fusion added an additional 1.9% improvement in sensitivity. All the improvements reported with respect to the closest baseline (TSTNet) are similar everywhere and non-overlapping 95% confidence intervals on central measures. The time of processing was 2–2.2 min per case, which is appropriate to be included in clinical workflows. We tested the method using 1500 CBCT scans that were annotated by experts.

Conclusions

The proposed mGAN architecture can combine domain knowhow with deep learning in dental surgical planning. This approach can offer a potential tool to support clinical decision-making in oral surgery by offering uncertainty-sensitive predictions and preserving computational efficiency, which has the potential to minimize the iatrogenic complications and enhance the treatment outcomes.

Background

Dental treatment planning, especially when complex oral and maxillofacial operations are involved. It is one of the most demanding uses of medical imaging because of both the complexity of relationships between structures of vital interest and the strictness of the criteria of patient safety. Cone Beam Computed Tomography (CBCT) has developed into the gold standard of three-dimensional dental imaging with the capability to visualize hard tissues, root morphology, and vital anatomical structures including the inferior alveolar canal (IAC) [1, 2]. Nevertheless, the correct interpretation of CBCT images to guide treatment planning partly relies on the clinician’s skill and experience, as well as subjective evaluation, which contributes to high inter-observer variability and the possibility of medical errors that may lead to severe outcomes, such as nerve injury, infections, and treatment failures [3, 4]. Removal of mandibular third molars is one of the most complex clinical scenarios due to the importance of properly evaluating the spatial relationship between tooth roots and the inferior alveolar canal to plan surgery and assess risk. Research studies have shown that, in 2–8% of mandibular third molar extractions, iatrogenic damage of the inferior alveolar nerve takes place, and permanent sensory loss has been documented in 1–3% of the cases [5, 6]. Besides resulting in serious patient morbidity, these complications have been shown to pose considerable medico-legal risks to practitioners. Current clinical practice is based on subjective radiography with additional use of empirical risk assessment protocols, which cannot be as precise and consistent as needed to achieve the best treatment results in an era of evidence-based dentistry and personalized medicine [7, 8].

It has been demonstrated that the application of artificial intelligence (AI) and machine learning methods to dental imaging can potentially help resolve these clinical issues. Conventional deep learning methods, especially convolutional neural networks (CNNs) have shown to be successful in a number of dental imaging problems such as tooth segmentation, caries detection and anatomical structure recognition [9, 10]. The available methodologies, however, greatly limit their applicability to complex three-dimensional dental treatment planning situations. Traditional segmentation networks, including U-Net and V-Net, lack the complex feature integration ability required to develop comprehensive risk assessment and clinical decision support [11, 12]. New developments in medical image analysis have led to more advanced architectures that aim to overcome the shortcomings of their classical CNN-based counterparts. Swin Transformers and Vision Transformer (ViTs) are examples of models based on transformers, which have demonstrated better results in multiple medical imaging tasks by utilizing the idea of self-attention to learn long-range spatial interactions [13, 14]. On the same note, more advanced CNNs, such as nnU-Net, Attention U-Net, and nested U-Net, have demonstrated better segmentation results by providing enhanced feature representations and utilising skip connections [15].

In addition, existing methods often run independently of the richness of clinical domain knowledge and evidence-based guidelines that guide expert clinical decision-making. Most current AI systems lack built-in clinical reasoning functionality, which restricts their practical implementation in clinical settings that require treatment decisions to consider various factors, such as the patient’s anatomical peculiarities, clinical manifestations, treatment goals, and risk-benefit calculations [16, 17]. Also, the majority of current strategies address only the accuracy of segmentation and do not consider the key requirement of the quantification of uncertainty and risk, which are important in clinical translation and regulatory approval [18].

Generative Adversarial Networks (GANs) have become an effective medical image synthesis and data augmentation paradigm and are applied in image-to-image translation, super-resolution, and domain adaptation [19]. Recent work has examined the applications of GAN in dental imaging, such as synthetic tooth generation and image enhancement, and demonstrated that adversarial training can enhance both the learning of features and the robustness of models under this strategy [20]. Nevertheless, current GAN-based solutions in the field of dental imaging have mainly been devoted to generating images and not utilizing the adversarial model to extract features and support clinical decisions.

Research gap identifcation

Several critical research gaps remain in dental treatment planning and risk assessment. First, current deep learning-based diagnosis methods provide insufficient multimodal feature integration and have limited capacity to encode and use clinical metadata, hindering comprehensive analysis of complex cases synthesized from heterogeneous sources for optimal planning [21, 22]. Second, most of the well-established deep learning models do not explicitly incorporate clinical domain knowledge and evidence-based guidelines into their analytical frameworks, leaving a persistent gap between technical performance and real-world clinical utility [23]. Third, AI-driven image analysis has largely emphasized classification and segmentation over probabilistic risk assessment and rigorous uncertainty quantification, both of which are essential for trustworthy clinical decision-support systems [24]. Lastly, the performance in terms of computational efficiency and real-time capabilities of clinical deployment are not well addressed. The deep learning models are accurate and efficient, they tend to be computationally intensive and hence cannot be used in practice in clinical settings where real-time processing is a crucial factor [25, 26]. Therefore, the required modification can be made in the well-established architectures to fulfil high performance support and satisfy clinical workflow requirements.

Proposed solution and innovation framework

The research gaps outlined above can be addressed by introducing structural modifications to a well-established GAN architecture, enabling the integration of multiple innovative components within a unified framework tailored to dental treatment planning and risk assessment. We refer to the resulting architecture as the modified GAN (mGAN) framework. It comprises three main modules, (i) Training Data module that handles the morphological features extraction such as textural, edge and contrast by incorporating U-Net architecture to fully enhance feature characterization; (ii) Generator module that handles the synthetic ground truth generation through the novel W-Net architecture with probabilistic risk distribution modeling; (iii) a Discriminator module that processes both real and synthetic samples to allow the accurate prediction of the surgical treatment necessity with the built-in uncertainty measurement. As a result, the proposed mGAN enables comprehensive analysis of both planar and volumetric anatomical relationships.

Research contributions

The proposed mGAN makes three principal contributions. (1) We introduce a W-Net backbone that combines U-Net and V-Net, yielding dual-path encoders and a skip-gated decoder to support robust multi-scale feature analysis tailored to complex dental anatomy. (2) We integrate a Large Language Model (LLM) to guide hyperparameter tuning and condition the generator with domain knowledge distilled from mandibular anatomy and prevalent dental pathology, thereby extending the system beyond purely data-driven pattern recognition. (3) We propose Probabilistic Feature Fusion (PFF), an adaptive multimodal integration mechanism that assigns dynamic weights to heterogeneous inputs based on local image content and LLM-provided clinical context, and that produces probabilistic maps of root–canal contact points to preserve interpretability and clinical relevance.

The remainder of this paper is organized as follows. Section  2 presents the comprehensive methodology including detailed descriptions of the Training Data Module, W-Net Generator architecture, and integrated discriminator framework. Section  3 presents comprehensive results including qualitative and quantitative performance analysis, ablation studies, and clinical validation findings. Finally, Sect.  4 concludes with a summary of contributions and potential impact on dental healthcare delivery.

Methods

This study proposes a clinically grounded framework to decide whether surgical intervention is warranted by inferring the relationship between mandibular molar roots and the inferior alveolar canal, together with adjacent soft tissue/bony status, from routine dental imaging. Because ultrasound while theoretically suited to measuring soft-tissue thickness suffers from noise and lacks open datasets [27], we pivot to CT/CBCT and dental radiographs and explicitly encode radiographic primitives (tooth density, boundaries, curvature/angulation, radiolucencies, cortical thickness, enamel/dentin cracks, and mandibular–maxillary morphology). A training-data module built on a U-Net extractor derives texture, edge, contrast, and curvature maps from X-rays as “true” anatomical exemplars. A generator, our W-Net (a structurally modified U-/V-hybrid with probability feature fusion, as seen in Fig. 1), conditions on CBCT and weak priors distilled from large language models (LLMs) to synthesize risk-aware pseudo ground-truth maps that depict the likelihood and spatial extent of root–canal “extended contact points. A PatchGAN-style discriminator jointly enforces anatomical realism and outputs a calibrated probability that surgery is required for a given test case. Training couples Dice+cross-entropy with surface/curvature penalties, adversarial loss, and a light consistency term to align synthesized risk with LLM-derived priors, enabling robust learning from limited manual labels. LLM-derived conditioning (GPT-4 via Sentence-BERT encoding, c ∈ R¹²⁸) embeds structured clinical domain knowledge—including Winter’s classification labels, IAC proximity categories, and cortical plate status—into the generator and discriminator via conditional batch normalization (see Sect.  2, Module-II and Module-III for full details). The system generates anatomical masks, continuous risk maps and decision score at inference based on predicted minimal root-to-canal distance indicators and cortical thinning indicators. It is reported that performance is not only with Dice/HD95 but also with clinically aligned measures, distance error, contact classification and sensitivity at fixed specificity such that improvements are directly related to surgical decision-making.

Fig. 1.

Proposed mGAN architecture

Multimodal Input Design (2D OPG vs. 3D CBCT): To be fully methodologically transparent, the following statement explains that the proposed mGAN has dual-input architecture. The architecture deliberately takes two different imaging modalities as input using different pipeline stages: (1) The input of the architecture is 2D panoramic OPG X-ray images (public Kaggle dataset), which are only used in Module-I (Training Data Module) to obtain 2D anatomical feature maps (texture, edge, contrast, curvature), which are upsampled to 3D through trilinear interpolation and sent to the generator. (2) The main input of Module-II (W-Net Generator) and Module-III (Discriminator) is 3D CBCT volumes (1,500 scans of the institutional archive) to conduct volumetric segmentation, prediction of risk maps, and all quantitative evaluation. The results on the outputs of the feature extraction in Module-I that act upon the 2D OPG/X-ray slices are plotted in Figs. 3 and 4. Thus, they are rightfully 2D data and do not actualise the eventual CBCT-based analysis. Predictive performance metrics in the results section are calculated. The rationale behind this multimodal design that is two-stage is due to the complementary nature of 2D OPGs that are ubiquitous to the bootstrapping of features and 3D CBCT that is essential to perform precise volumetric measurement.

Fig. 3.

Features explorer using CBCT slice imagery for 2D feature extraction (Module-I: Training Data Module). A Original (grayscale), B CLAHE Equalized, C Edge (Canny, ) overlay, D Gradient magnitude (multi-scale) overlay, E Texture energy (Gabor) overlay, F Texture orientation (HSV: hue = ori, val = sat=energy), G Local contrast (std, 15 × 15) overlay, and (H) Unsharp contrast . Note: These visualizations represent 2D processing of CBCT-derived axial slices used to extract anatomical priors; all quantitative evaluations are performed on full 3D CBCT volumes

Fig. 4.

Dental metrics explorer using 2D OPG X-ray imagery (Module-I Feature Extraction). A Original (grayscale), B CLAHE Equalized, C Tooth ROI+boundary+major axes (angle 30.4), D Curvature \left|\kappa\right| (med: 0.2, p95 = 0.707), E Radiolucent area (Count: 46, area 3850 px), F Edge sharpness (multi-scale gradient magnitude), G Cortex ROI not provided, and (H) Crack skeleton (len: 336 px). Note: Fig. 4 illustrates 2D panoramic OPG-derived feature maps used to bootstrap anatomical priors in Module-I; these features are subsequently lifted to 3D for use in the W-Net Generator

Dataset, annotation protocol and experimental setup

Dataset composition and source

This research study is utilized 1,500 Cone Beam Computed Tomography (CBCTs) obtained in the institutional oral and maxillofacial surgery unit dental radiology archive between the years 2018 and 2023. The acquisition of each volume of the CBCT was performed with the help of a Planmeca ProMax 3D scanner (voxel resolution: 0.2 mm isotropic, field of view: 8x8 cm). A random sampling of the cases was made to have equal distribution of mandibular third molar impaction classes (Winter classification: vertical n=420, mesioangular n=385, horizontal n=340, distoangular n=355) with known surgical assessment recorded in the post-operative record.

Patient demographics

The sample size consisted of 847 women and 653 men (mean age was 28.412.6 years old, with a range of 18 to 64 years). The patients with the existing nerve conditions, previous surgery on the mandible (ipsilateral), or incomplete imaging facilities were excluded. Data on ethnicity was not undertaken because it was not a confounding factor in this imaging task.

Annotation protocol and inter-annotator agreement

Two board-certified oral and maxillofacial surgeons (Annotator-A: 12 years’ experience; Annotator-B: 9 years’ experience) annotated mandibular third molar root (M3M) and inferior alveolar canal (IAC) segmentation masks independently with ITK-SNAP version 3.8. A third senior radiologist resolved disputes that had a Dice of more than 0.85. Cohen kappa (= 0.87, strong agreement) and mean Dice agreement of 0.91=0.03 in all 1,500 cases was used to assess inter-annotator agreement. The evaluated annotations were the final ground truth upon which the training and the evaluation were done.

Train/validation/test split and class imbalance handling

The dataset has been divided into three distinct samples 70/15/15 split (train: n=1,050; validation: n=225; test: n=225), preserving the class distribution of contact (n=487, 32.5%) versus no-contact (n=1,013, 67.5%) cases across all splits. To address class imbalance in surgical necessity labels, we employed a combination of oversampling minority-class cases using CBCT-specific elastic deformation augmentation and a class-weighted loss function (α_contact = 2.07, inversely proportional to class frequency) during training. Data augmentation applied to training data included random rotation (±15°), random flipping (axial/sagittal), Gaussian noise (σ = 0.01–0.05), and intensity scaling (0.85–1.15) to improve generalization.

Training implementation details

The proposed mGAN framework was implemented in PyTorch 2.0 (CUDA 11.8) and trained on two NVIDIA A100 GPUs (80 GB VRAM each) for 200 epochs using an Adam optimizer (learning rate η = 1×10−4, β1 = 0.9, β2 = 0.999, weight decay = 1×10−4). The batch size was 4 for 3D volumes (patch size 128×128×64). A cosine annealing learning rate schedule with warm restarts (T_0 = 50 epochs) was applied. The composite loss function comprised adversarial loss (λ_adv = 1.0), Dice + cross-entropy segmentation loss (λ_seg = 5.0), surface distance penalty (λ_surf = 2.0), curvature penalty (λ_curv = 0.5), and LLM consistency term (λ_llm = 1.5). All baseline models were retrained under identical conditions on the same hardware, data splits, and augmentation pipeline to ensure a fair comparison. Cross-validation was performed using 5-fold stratified cross-validation on the training set, with the reported results corresponding to performance on the held-out test set.

Module-I: training data module

Texture features provide details about radiographic trabeculation & tissue micro-patterns (trabecular bone lattice, dentin tubules, and homogeneous enamel). This can be achieved by applying a multi-scale, multi-orientation Gabor bank to the locally equalized image :

1

Where,  T rises in textured trabeculae and falls in homogeneous enamel/air. Optionally keep the dominant orientation to capture ridge directions. Clinical link: Loss of trabecular texture (↓ T) suggests radiolucent lesions; highly regular enamel appears low texture. Edge features mainly focused on anatomical boundaries & lamina dura especially highlight tooth–bone interfaces, root walls, and IAC cortication. Multi-scale gradient magnitude with Gaussian derivatives is utilized to obtain the edge features which are stated as follows:

2

Where,  E peaks on crisp enamel edges and cortical plates; provides edge orientation. Clinical link: Blurred/weak  E around canal walls suggests de-cortication; sharp lamina dura gives strong  E. Contrast provides detailing about the local attenuation variability. That means, it measures local intensity spread to surface radiolucency and cortical thinning. According to this definition, local windowed standard deviation is computed by using .

3

Where, High  C occurs at enamel–dentin transitions and cortical plates and low  C in radiolucent lesions. (Alternate: unsharp ). Clinical link: Radiolucent areas show depressed  C; thick cortex yields elevated  C at the boundary.

After extracting the texture, edge, and contrast features from the dental image, these features are subsequently fused to form a comprehensive, well-defined true-sample tensor that captures the key anatomical and morphological characteristics necessary for accurate dental state prediction (normal vs. abnormal). The feature extraction process, described with suitable equations, produces three distinct sets of features:

• Texture (T): Represents the trabecular patterns of the bone, denoting the fine-grained structure of the teeth and jaw region.

• Edge (E): Captures the boundaries of the teeth and surrounding anatomical structures, such as the inferior alveolar canal (IAC) and cortical bone regions.

• Contrast (C): Provides information about local intensity variations, which are crucial for identifying radiolucent areas and subtle structural differences.

Additionally, curvature () and vesselness ( V) maps are included, further enriching the feature set with geometric and structural information. These multi-dimensional feature maps serve as the foundational representation of the dental image and are stacked into a per-pixel feature vector :

4

Each channel (feature map) is then normalized to zero-mean and unit-variance to standardize the input and ensure consistent scaling across different types of features. To improve the discriminative power and relevance of these features, a learned convex weighting mechanism is applied to combine the extracted features. This is achieved through a softmax-based operation that computes a set of weights for each feature vector:

5

where is a 1 × 1 convolution or multi-layer perception (MLP) applied to the anatomical probabilities obtained from the lightweight U-Net. These probabilities represent the soft anatomical priors related to the tooth, jawbone, and IAC, respectively. The softmax function ensures that the weights are normalized and sum to 1, thereby highlighting the most relevant features based on anatomical context.

The weighted feature vector is then used to compute a weighted sum of features:

6

which combines the contributions of all the feature maps (texture, edge, contrast, curvature, and vesselness) into a single, contextually weighted feature representation. This step allows the model to dynamically emphasize certain features (such as texture for tooth structure or edge for canal boundaries) based on the anatomical context provided by the U-Net priors. To ensure that the generated samples focus on the relevant anatomical regions and suppress background noise, the feature vector is further gated by an ROI mask derived from the anatomical priors. The ROI gate is defined as a weighted sum of the anatomical probabilities:

7

where , , and are non-negative weights that sum to 1 . The final true-sample tensor is obtained by applying the ROI gate w(x) to the weighted feature vector F(x):

8

where represents the anatomically gated, feature-rich representation of the dental state at pixel  x. These true-sample tensors, form the primary input to the discriminator in the modified GAN architecture. The discriminator utilizes these true samples to make a binary classification (normal vs. abnormal) based on the anatomical and morphological features derived from the dental image. The discriminator is trained to distinguish between the real samples (true samples from the training data) and generated samples (from the generator), with the goal of improving the quality of generated images and enhancing decision-making for surgical interventions.

Module-II: generator module

The Generator is designed as a W-Net, which is a hybrid architecture combining U-Net and V-Net, as seen in Fig. 2. The U-Net component ensures the use of 2D feature priors (such as the TDM features), while the V-Net component operates over 3D spatial data (i.e., CBCT volumes). The generator takes several inputs and conditions them through multiple stages, creating output images that simulate the “ground truth” image needed for training the discriminator. The key idea here is to generate fake samples with varying probabilities of root-canal contact points, allowing the network to learn anatomical boundaries, surface interactions, and risk distributions.

Fig. 2.

W-Net hybrid model

A 3D CBCT scan, normalized as where . This is the primary input that provides anatomical information about the entire jaw structure. 2D radiographic features extracted from dental X-rays and lifted to 3D. These features include anatomical priors such as the tooth, bone, and IAC boundaries, which help in improving the anatomical focus during generation.

9

It is the 3D lifted version of the 2D TDM features, where  d represents the number of feature channels (e.g., texture, edge, contrast). Optional clinical cues (e.g., contact class, distance bin, cortication status) are derived from a Large Language Model (LLM) to provide weak labels and supplementary anatomical knowledge. Specifically, we employ GPT-4 (OpenAI) prompted with structured dental domain queries to extract mandibular anatomy descriptions, M3M impaction classifications (e.g., Winter’s classification: vertical, mesioangular, horizontal, distoangular), IAC proximity risk categories (close, touching, displaced), and cortical plate status descriptors. Each textual response from the LLM is encoded into a fixed-length conditioning vector c ∈ R^128 using a pretrained Sentence-BERT encoder, which maps semantic clinical descriptions to dense vector representations. This conditioning vector c is then incorporated into both the generator (via the PFF attention mechanism, Eq. 13) and the discriminator (via conditional batch normalization layers), ensuring that the entire mGAN pipeline is guided by structured clinical domain knowledge beyond purely data-driven pattern recognition. This information is leveraged for conditional training to adjust the generation process based on expert knowledge, enabling the framework to produce anatomically plausible risk maps that are consistent with established clinical guidelines. A random latent variable sampled from a standard normal distribution that adds stochasticity to the process and allows for diverse sample generation. The input tensor to the generator is constructed by concatenating the normalized CBCT volume, lifted TDM features, and LLM conditioning, forming a multi-channel input tensor:

10

Where is the total number of input channels (including the CBCT volume, TDM features, and any other conditioning factors).

Dual-path encoding

The dual-path encoder consists of 3D encoder stages (V-Net-like) that process the 3D CBCT volume, and a 2D encoder that processes the 2D anatomical priors lifted to 3D. The 3D encoder processes the anatomical data through strided convolutions to produce feature maps, which capture multi-level anatomical information at different spatial resolutions.

The encoder produces a set of 3D features at each stage  l of the network:

11

Here, represents the encoding operation at level  l, and is the tensor input. The 2D anatomical priors are encoded through a shallow 2D U-Net-like network, and the resulting features are lifted to 3D via a lifting operation and aligned with the 3D features.

Probability Feature Fusion (PFF)

At each encoder stage l, features from the 3D encoder, 2D lifted priors, and morphological channels (e.g., texture, edges, contrast, curvature) are fused together using a learned attention mechanism conditioned on the LLM prior c.

The fused feature vector is constructed as:

12

Where denotes the morphological feature maps (e.g., texture, edges, contrast, curvature). The feature fusion weights are calculated using a softmax operation on a set of learned parameters and , where is the content attention learned from the global average pooled features, and incorporates the LLM conditioning:

13

Here, GAP refers to Global Average Pooling, and the softmax operation normalizes the feature weights, ensuring they sum to 1. The weighted sum of features is then computed as:

14

This weighted fusion of 3D, 2D, and morphological features ensures that the generator can focus on the most anatomically relevant regions and features during training.

Skip-gated decoding (U-like) and output generation

The decoder path of the generator uses skip connections from the encoder and applies upsampling operations to progressively reconstruct the output image. The skip-gated decoding mechanism ensures that high-level anatomical details are maintained while generating the final output image.

At each decoder stage  l, the previous decoded features are combined with the fused features and passed through a series of upsampling operations. The skip-gated decoding operation is:

15

Where is the sigmoid activation function, represents the upsampling operation, and denotes elementwise multiplication. The final output of the generator is the synthetic “ground truth” image :

16

Where: is the predicted anatomical class map (e.g., teeth, IAC, bone), is the risk distribution for extended contact (Beta distribution), and is the boundary map (sharpening canal/root interfaces).

Geometry-aware post-readout

To change the default, adjust the template as follows Post-processing is applied to the anatomical logits to refine the root and IAC surfaces. This is done by computing the minimum distance between the root surface Γroot and the IAC surface :

17

Where  p and  q represent points on the root and IAC surfaces, respectively. Additionally, an extended-contact indicator is generated based on a threshold τ on the minimum distance :

18

The post-readout operation assists in the tightening of the decision-making process to predict the contact points and guarantee anatomic consistency in the output generated. Generator module produces various fake samples having different contact probability to provide strong training of the discriminators. The training of the adversarial network using the generated images of the teeth necessitates anatomical priorities, stochastic latent variables, and feature fusion, allowing the generator to produce highly plausible images of the teeth.

Module-III: discriminator module

The Discriminator Dφ in the proposed mGAN framework is designed as a PatchGAN-style architecture that jointly enforces anatomical realism and outputs a calibrated surgical-necessity probability. Unlike standard binary discriminators, the proposed discriminator operates at the patch level to assess local anatomical coherence, which is essential for capturing fine-grained root-canal boundary interactions. The discriminator receives two types of inputs: (i) real samples represented by the anatomically gated true-sample tensors Z_true(x) derived from the Training Data Module, and (ii) fake/synthetic samples G̃(s) produced by the W-Net Generator.

Architecture and forward pass

The discriminator consists of four 3D strided convolutional stages with spectral normalization, followed by a dual-head output layer. The first head produces a patch-level realism map P_real(x) ∈ [0,1] that assesses the anatomical plausibility of the input at each spatial location. The second head produces a global surgical-necessity score S_surg ∈ [0,1] derived from global average pooling of the final feature maps. Formally, the discriminator forward pass is defined as:

19

Where  I is the input CBCT volume, and c is the LLM-derived clinical conditioning vector. To guarantee that the realism judgment is not entirely a data-driven one, the discriminator is known to condition on  c to have the domain knowledge reflected in its adversarial assessment, which makes the realism judgment consistent with anatomical expectations of the root curvature, how the expected canal cortication should look like, and the M3M patterns of angulation that is expected.

Uncertainty-aware surgical decision

The discriminator uses Monte Carlo Dropout to make inference to provide a calibrated uncertainty quantification. The system calculates the mean surgical score µ surg and epistemic uncertainty σ surg over N stochastic forward passes N = 20. The ultimate binary decision is operation decides = 1 when and , where , equals decision threshold and maximum permissible uncertainty. The cases that exceed the uncertainty bound will be sent to the experts instead of automatically arriving at a decision, and patient safety is maintained in cases of uncertainties. The discriminator loss is a composite loss that sums adversarial loss, surface distance penalty, and a consistency loss. Hence, it ensures the output risk of the discriminator is consistent with the priors generated by the LLM using minimal manual annotating.

Results

In this section, covers detailed qualitative and quantitative comparison of the proposed mGAN (W-Net) with 12 state-of-the-art techniques (UNet, UNet++, ESPNetv2, KCNet, V-Net, nnU-Net, DeepLab v3+, Attention U-Net, 3D ResNet-UNet, UCTransNet, Swin and TSTNet) on evaluation metrics using a dataset of 1,500 CBCT scans. The major results include better performance using DSC of 0.925 + 0.019 (compared to best baseline TSTNet 0.911 + 0.023), contact detection F1-score 0.932 + 0.020 (1.6% improvement), and risk correlation 0.891 + 0.027 (5.2% improvement). The performance of distance measurement accelerates significantly with RMSE of 0.36 + 0.08 mm (20% lower than TSTNet, 46.3% lower than U-Net) and the network can also reduce computational cost by 5.1% relative to the Standard GAN baseline (847.3 GFLOPs versus 892.5 GFLOps), indicating that the proposed innovations do not produce a disproportionate lack of performance. Only cross-institutional validation shows better generalization with 2.3% performance degradation as compared to best baselines (3.2). Ablation experiments indicate that Probability Feature Fusion does not add significant improvement to the performance (1.8%), LLM integration means the risk correlation has the improvement by 6.4% and dual path encoding the improvement on all metrics are evenly spread. The remaining qualitative and quantitative findings are reported in the following subsections; their clinical interpretation, comparison with prior work, and the limitations of the study are addressed separately in the Discussion.

Qualitative analysis

Feature explores-CBCT imagary

Figure 3 illustrates the characteristics exploration associated with the tooth structures, bone margins and risk areas of the clinical examination with the images of CBCT imaging. It emphasizes visualization of the anatomy characteristics better and that may prove essential in planning and diagnosis of dental treatment. The CBCT image in grayscale format (without any processing) displays the basic structure of the teeth, jawbone, and the rest of the anatomy, but does not give details on any finer details, including boundaries or textures (see: Figure 3(A)). The original image is subjected to the Contrast Limited Adaptive Histogram Equalization (CLAHE) algorithm that is used to increase contrast particularly in the low-contrast areas (see: Figure 3(B)). This assists in enhancing recognition of finer details of the anatomy including the boundaries of teeth and position of canals that were not evident in the original image. Figure 3(C) indicates the detection of an edge with the Canny edge detector; one uses a standard deviation of σ = 1. It reveals the angular form of the teeth and other forms of the anatomy such as cortical bone and inferior alveolar canal (IAC). These lines give a better definition of the dental structures, and this is very essential in the determination of the root-canal boundaries. The gradient magnitude is illustrated in Figure 3(D) using various scales, to increase the sharpness of the boundaries and the ability to see finer details in the picture. It assists in determining regions of extreme variation in intensity of the image and emphasizes the features of the image such as tooth lines, and the bone of the cortex. Gradient magnitude overlay can be used to detect better the features that vary quickly such as the tooth-enamel interface. CBCT image on the other hand uses Gabor filters to identify texture in the picture. The generated energy map with the filters aids in pointing out the anisotropic regions of the body, including the trabecular bone or the dentin tubules (see: Figure 3(E)). This is because the texture energy overlay helps to identify bone structures relative to soft tissues and the identification of abnormal patterns such as fractures or deterioration. This texture orientation is visualized in Figure 3(F) with the help of an HSP color map where it supports the direction of the texture and saturation is the energy or strength. The overlay assists in the visualization of the alignment of trabecular patterns of the bone and dentinal tubules which can be used in the analysis of the bone and tooth structural orientation. Local Contrast (Std, 15 × 15) Overlay image is the local contrast, the standard deviation of a 15 × 15 window (see: Figure 3(G)). Regions of high local variation in the image are emphasized, e.g. the interface of tooth enamel and dentin or a radiolucent bone. Local contrast helps in identifying slight tissue differences which might be cavities in the tissue or bone density problems. This is done by the unsharp mask method that helps to increase the contrast by subtracting the smoothed image of the original image. The process brings out tiny details including minor cracks or minor bone changes, and it is simpler to see the tooth fractures, enamel defects, or even the formations of cavity.

Dental metric explores – X-ray

Figure 4 illustrates the dental metrics exploration of using X-ray imagery, and with emphasis on extracting morphological features of the tooth, cortex, boundary and cracks. These measurements are essential to surgery and dental diagnosis. The first image and second image depict the original X-ray image in grayscale and, as a result, CLAHE applied to the original grayscale image (see: Figure 4(A) and (B)). The improved image brings out the details such as the tooth root, the canal and the cortical boundary which was not visible before. The difference between the tooth and the boundaries is made more prominent, sharpening the tooth and its outlines. Figure 4(C) indicates that the region of interest (ROI) of the tooth is marked and outlined with a green boundary. The major tooth axes (in blue) are also indicated. The angle of 30.4 is used to show the position of the tooth when compared to the X-ray image. This gives an insight into the position of the tooth and its alignment with the rest of structures. Figure 4(D) is the curvature map of the tooth boundary with ∣κ| indicating the degree of curvature. Median curvature of 0.2 and the 95th percentile of 0.707 show that there is curvature in different portions of the tooth. The curvature analysis plays a very important role in finding the aberration that may be a fracture or malformation of the tooth shape. The radiolucent areas in the tooth are indicated in red in Figure 4(E). Radiolucent areas are the regions of the bone or the tooth structure which look darker and may denote the existence of a certain cavity or damage. A total of 3850 pixels is identified as 46 regions in this case. These are important in dental diagnostics in the identification of decay or abscesses. Figure 4(F) indicates that multi-scale gradient magnitudes computation gives the highlights of the high contrast areas along the tooth and canal, where the edge sharpness along the edge shows there are major boundaries between the tooth and other structures. Boundary integrity and cortical plate structures can only be identified by edge sharpness. There is a lack of cortex ROI which shows that the cortical region cannot be analyzed in this case (see: Figure 4(G)). The cortex ROI (region of interest) generally plays an essential role in assessing bone thickness, bone strength, which is the most vital in surgical planning, particularly in cases relating to the inferior alveolar canal (IAC). Here, bone health cannot be evaluated due to lack of cortex data. The crack skeleton in the tooth was determined by skeletonization of the crack pattern as seen in Figure 4(H). The crack length is quantified to 336 pixels which can quantify enamel or dentin fractures. This data is invaluable in identifying the structural integrity of the tooth and whether the restoration is necessary or surgical intervention.

Contact point and risk zone analysis

Figure 5 shows the contact point analysis of dental root and inferior alveolar canal (IAC). The minimum distance between the root and canal expressed in pixels is displayed in the top row meaning that the closeness of the root and canal can be different. The closer the root is to the canal, the greater the risk areas (bottom row) will be red and orange in color or high risk. The closer the min distance is to zero the greater the risk is and therefore the surgery is needed in case the root is directly touching the canal. This discussion focuses on the evaluate the capability of the proposed mGAN framework to identify root-canal interactions with sufficient accuracy and to aid in surgical decisions.

Fig. 5.

Fig. 5. Contact-point and risky-zone analysis for root–canal interaction. Panels (A-I), (B-I), and (C-I) show the contact-point analysis maps indicating the minimum pixel-count distance/rate required to reach the root and canal points, with minimum distances of 0.56, 0.32, and 0.07, respectively. Panels (A-II), (B-II), and (C-II) show the corresponding risky-zone analysis maps around the dental location.

Key observation

The distance between the root and canal was the minimum, which is 0.56px, as illustrated in Fig. 5(A)-I. The blue to red scale shows the proximity with darker blue regions showing distance and the red regions showing more closely related regions. Conclusion: The root is fairly remote to the canal, yet it approaches the canal gradually. This indicates that the root is yet to touch the canal but is near enough to cause a risk in case it is not treated. The min distance further reduces to 0.32px in this image (see: Fig. 5(B)-I), which means that the root moves nearer to the canal. The nearness is indicated with increased pixel count values in the form of orange and red spots. Surgical intervention might be required in this case to prevent nerve involvement or any other complications. The distance has now reduced to a min of 0.07px and the root nearly touches the canal. Figure 5(C)-I demonstrates that there is a critical proximity in which the root and canal are almost touching and they are characterized by the intense red areas. Inference: The root is at direct contact with the canal, an indication of high-risk situation. The immediate, surgical intervention is the solution to this case, since there is a high probability of nerve damage or other complications.

Risk zone representation

Figure 5(A)-II to (C)-II are the risk areas of the dental root site, which are labelled as areas with the possible complication because of close contact between the root-canal and other organs. These areas are visually measured with the color-coded map where red and orange color areas reflect the high risk areas and green/blue regions are the safe areas. Minimal Distance Risk Zone = 0.56px, at this level of distance, the root is relatively far away, and yet the risk zone around the root is relatively large. The risk is moderate and is identified as high in the areas around the canal. Risk Zone = Min Distance/Min and the nearer the root is to the canal in distance, the larger is the zone of risk. Areas with higher risk (red/orange) seem to be nearer to the canal, which means that it could be a surgical concern. Min Distance of Risk Zone = 0.07px, which is a high-risk zone that has a high proximity between the root and canal. The red zone means that the root is now in direct contact with the canal and hence this is a high-risk area in potential complications.

Root-canal contact in CBCT samples

Figure 6 shows the point of contacts between the dental root (M3M) and the inferior alveolar canal (IAC) to a series of test cases. The proposed mGAN framework is utilized to analyze it and create risk zone maps in form of color coding. Root and canal are denoted by red and cyan correspondingly, and bright regions are denoted by high-risk areas that require surgical intervention because of close proximity or contact.

Fig. 6.

Root-canal contact visualization in CBCT samples using proposed mGAN framework. Red regions indicate M3M root structures, cyan regions indicate the inferior alveolar canal (IAC), and bright/highlighted zones represent high-risk areas of close proximity or direct contact requiring surgical intervention

Key observation

The distance between the root and the IAC is moderate (first image). Red = Root, Cyan = IAC: The color-coded map indicates that the root is nearing the canal, but it is still evident that there is a gap between the two. The light spaces represent medium-risk zones that may turn out to be critical contacts in the case of the worsening of the situation. The closer to the IAC the root the more critical it becomes (second image). Red = Root, Cyan = IAC: The light areas indicate that there is a risk increment with the proximity of the root and the canal. It indicates that surgical assessment and intervention is in high demand. The IAC is in almost direct contact with the root (third image). Red = Root, Cyan = IAC: The high-risk areas are indicated by the bright areas and are very close to the canal and the root. The damage to the nerves or other complications is potentially life-threatening, and the surgery should be as prompt as possible. The bottom line illustrates the high-risk areas at various levels of contact with the root-canals. The brighter the color is as the root is getting closer or touching the canal, which means the increased risk. The color coding of red (root) and the cyan (IAC) color coding gives an instant visual view of the anatomy relationship, as well as being able to locate the areas of the body that needed surgical attention.

Dental-root canal interaction prediction

Figure 7 shows the results of the proposed mGAN framework in predicting the root-canal contact of six test samples. The original dental CBCT images are indicated in the first row and the generated ground truth maps are indicated in the second row, and the predictions are no contact (green) and contact (red). The model can identify the contact of the root-canal correctly in situations where the number of pixels in the segmented area is high enough (e.g., Sample-2, Sample-4). Nevertheless, when dealing with lower pixel density (e.g., Sample-3, Sample-6), the model does not identify the contact although they are clinically significant. This number indicates the weaknesses of the existing model, especially in cases of subtle interaction in roots-canals, which implies that the pixel count-levels and sensitivity control in the discriminator still require additional adjustments.

Fig. 7.

Proposed mGAN predictions for root-canal contact in CBCT samples

Sample-1 (image slice 18) and sample-5 (image slice 712) clinical justification

In such instances, the model can predict the fact that there is no contact between the canal and the tooth root. The root and canal are distinctly drawn and the pixel count around the root area is also high and the anatomical priors (the boundary of the tooth, positioning of the canal etc.) allow the model to draw a successful prediction. Inference: The discriminator is effective to handle the true and the fake samples with a correct pixel density in the segmentation region, which allows the root-canal boundary to be recognized accurately. It is not recommended to surgically intervene since the root does not obstruct the canal.

Sample-2 (image slice 249) clinical justification

In this case, the root is touching the canal, which can be established on the ground truth, and the proposed mGAN framework identifies correctly the point of contact and indicates that a surgical intervention is necessary. It demonstrates the ability to observe more subtle root-canal interactions, which are frequently difficult to imagine. Inference: The pixel density in the region of the segmentation is large enough, and the anatomical characteristics (such as the canal depth) are represented by the model sufficiently. This can enable the discriminator to anticipate the contact and recommend surgical care.

Sample-3 (image slice 371) clinical justification

In this case, the proposed mGAN framework does not recognize the proximity with the canal even when the root is near the canal. The contact is not determined because the number of pixels in the segmented area is inadequate, resulting in failure to detect the canal interaction between the root and canal. Inference: The low pixel count (below the threshold of 0.05) causes the model to miss the contact point, leading to incorrect predictions. This is a false negative in clinical practice, in which the model fails to appreciate the clinical significance of the root-to-canal proximity, even though a dental expert would recommend surgical intervention. This implies that the pixel count threshold parameters require further refinement to detect certain subtle clinical contact conditions.

Sample-4 (image slice 585) and sample-6 (image slice 974) clinical justification

These two samples indicate that the proposed mGAN framework has erroneous predictions in which the root is near the canal, yet the model does not identify the root. The reason is low pixel density in the segmented segment, and thus false no contact predictions. Inference: The number of pixels taken as a root-canal proximity is important. At a low number of pixels, the model cannot identify contact, even that which is clinically relevant. This means that the sensitivity of the current model should be improved to pick extremely small interactions in the root-canal interface.

Figure 8 presents the potential of the proposed mGAN in dental surgical decision support, and some limitations have been identified. The upper row indicates the raw CBCTs of three various test cases. The last row shows the output prediction of the proposed mGAN, the areas of concern (root-canal interactions) are shown in red/yellow. Case-I proves to be accurate in predicting surgical intervention and Case-II determines the partial penetration of the root. Case-III illustrates the weakness of the model to identify subtle root-canal proximity even though clinically it is of relevance.

Fig. 8.

Dental root-canal interaction predictions by the proposed mGAN

Case-I (test image_824)

The proposed mGAN framework correctly predicts the presence of contact between the dental root and the inferior alveolar canal (IAC), with the point of contact indicated in red. According to this framework, immediate surgical intervention is required, as demonstrated by the high prediction accuracy and the supporting dental metrics. Thus, the urgent need for surgical intervention is correctly predicted by the proposed mGAN framework. Red color is used to show where the root enters the canal as it is the point of contact between the tooth and the inferior alveolar canal (IAC). Clinical Inference: The penetration of the dental root into the canal poses a risk of nerve damage or additional complications in this case, and thus the case requires immediate surgical treatment. The high prediction accuracy shows that the mGAN has identified this important clinical feature accurately. Observation: The red color shows the area of contact with the canal, which is a significant clinical aspect in planning a treatment. The other dental metrics also support this decision by confirming the seriousness of the situation.

Case-II (test image_568)

The proposed mGAN framework is properly predicts the need to conduct the surgical intervention, even in a case of partial contact of the dental root and the IAC. This demonstrates the model has the capability to identify major clinical risks in instances that involve partial penetration of roots. Description: The proposed mGAN model can identify that the dental root is partly in the canal, which indicates surgical intervention. Although the contact is not as big as in Case-I, the system behaves in the right way letting it be known that surgical aid is needed. Clinical Inference: The root penetration is partial, but this may cause possible nerve-involvement or other problems that may be proceeded without treatment, which is surgical intervention is required. The model recognizes the contact area but with partial root penetration. Observation: The orange/yellow color is a sign of the area in which the tooth is partially touching the canal that remains clinically relevant, implying the possibility of problems that may result in the need to conduct surgery.

Case-III (test image_912)

The proposed mGAN leads in prediction error even though the dental root is close to the IAC because the thresholding error is detected in detecting minimal contact (low pixel count). It indicates a weakness that exists in this framework in the detection of subtle clinical aspects. History: The dental root in this case is slightly near the canal yet the mGAN does not prompt the presence of the need to undergo surgery. The model falsely detects that no surgery is necessary since the contact point is characterized by low pixel count rate (less than 0.05) that results in a failure in detection. Clinically, a root is adjacent to the canal can be quite dangerous, as it can become worse with time or during procedures. Observation: With the mGAN the low number of pixels does not identify the proximity between the dental root and the canal as clinically significant, hence the false negative rule out of surgery.

Quantitative analysis

Core segmentation performance metrics

As shown by the segmentation performance analysis in Table 1, the proposed mGAN framework has the best capabilities in all the basic geometric/volumetric metrics. This framework has attained the best Dice Similarity Coefficient (DSC) of 0.925 ± 0.019, which is statistically significant improvement of 1.5% over the highest-performing baseline TSTNet (0.911 ± 0.023) and significant 5.6% improvement over the well-known conventional U-Net baseline (0.876 ± 0.031). Similarly, the same pattern of improvement is uniformly seen in the complementary measures includes Intersection over Union (IoU) and Volumetric Similarity (VS) where the proposed approach has high (IoU) values of 0.859 ± 0.025 and VS of 0.938 ± 0.017 which outperform all the competing approaches by factor of 2.8% to 10.3%, respectively. Thus, these outcomes demonstrating excellent retention of dental treatment planning volume relationships between anatomical volumes.

Table 1.

Core segmentation performance metrics

Method	DSC	IoU	VS	HD (mm)	ASD (mm)	HD95 (mm)
UNet	0.876 ± 0.031	0.779 ± 0.041	0.891 ± 0.028	4.23 ± 0.68	1.34 ± 0.22	2.87 ± 0.45
UNet++	0.889 ± 0.027	0.801 ± 0.036	0.903 ± 0.024	3.91 ± 0.61	1.21 ± 0.19	2.64 ± 0.41
ESPNetv2	0.854 ± 0.038	0.744 ± 0.048	0.872 ± 0.035	4.78 ± 0.82	1.56 ± 0.28	3.21 ± 0.58
KCNet	0.868 ± 0.033	0.767 ± 0.043	0.884 ± 0.031	4.45 ± 0.71	1.42 ± 0.24	3.01 ± 0.48
V-Net	0.888 ± 0.028	0.798 ± 0.037	0.901 ± 0.025	3.98 ± 0.63	1.28 ± 0.21	2.71 ± 0.43
nnU-Net	0.901 ± 0.026	0.819 ± 0.034	0.913 ± 0.023	3.67 ± 0.55	1.15 ± 0.18	2.45 ± 0.38
DeepLab v3+	0.883 ± 0.029	0.791 ± 0.038	0.897 ± 0.026	4.05 ± 0.64	1.31 ± 0.21	2.78 ± 0.44
Attention U-Net	0.895 ± 0.029	0.810 ± 0.037	0.908 ± 0.026	3.78 ± 0.58	1.19 ± 0.19	2.56 ± 0.40
3D ResNet-UNet	0.892 ± 0.027	0.805 ± 0.035	0.905 ± 0.024	3.84 ± 0.60	1.23 ± 0.20	2.61 ± 0.42
UCTransNet	0.908 ± 0.024	0.831 ± 0.031	0.921 ± 0.021	3.45 ± 0.51	1.08 ± 0.16	2.31 ± 0.35
Swin	0.904 ± 0.025	0.826 ± 0.032	0.917 ± 0.022	3.52 ± 0.53	1.11 ± 0.17	2.37 ± 0.36
TSTNet	0.911 ± 0.023	0.836 ± 0.030	0.924 ± 0.020	3.38 ± 0.49	1.05 ± 0.15	2.26 ± 0.34
Proposed mGAN	0.925 ± 0.019	0.859 ± 0.025	0.938 ± 0.017	3.12 ± 0.43	0.94 ± 0.13	2.08 ± 0.31

Bold values represent the best performance among all compared methods for each metric (maximum for accuracy-based measures and minimum for error-based measures), reported as mean ± standard deviation

Furthermore, a significant benefit of the proposed mGAN framework can be seen with the help of distance-based metrics. During measure of worst-case boundary discrepancy, the proposed mGAN framework offers impressive results of 3.12 ± 0.43 mm which conveys 26.2% of reduction that faces by conventional U-Net in worst-case segmentation error. In addition, Average Surface Distance (ASD) shows effective superiority of 0.94 + 0.13 mm over the second-best performer TSTNet (1.05 + 0.15 mm), which is more accurate boundary localization important in surgical planning. Notably, the 95th percentile Hausdorff Distance (HD95) which corresponds to clinically relevant accuracy on the boundaries depicts of proposed mGAN (2.08 ± 0.31 mm) against TSTNet (2.26 ± 0.34 mm) and U-Net (2.87 ± 0.45 mm), respectively. Such improvements in distance metrics are especially important in the dental arena where a sub-millimeter accuracy has a direct effect on surgery results and patient safety. Statistical significance of the reported improvements was evaluated using two-tailed paired t-tests across the 1,500 test cases with a significance threshold of p < 0.05. All main comparisons of the proposed mGAN and the closest baseline, TSTNet, had statistically significant differences: DSC improvement (p = 0.003), Contact F1 improvement (p = 0.007), RMSE improvement (p < 0.001) and Risk Correlation improvement (p < 0.001). The consistency in the performance of all the measures along with the relatively small confidence intervals and statistically significant values proves the power and reliability of the suggested mGAN architecture to be used in the clinical dental image analysis setting. To address multiple comparisons across the 23 reported metrics, Bonferroni correction was applied (adjusted threshold: p_adj < 0.05/23 ≈ 0.0022). Under this corrected threshold, the primary metrics remain statistically significant: DSC (p = 0.0003, 95% CI [0.008, 0.022]), RMSE (p < 0.0001, 95% CI [0.065, 0.115 mm]), and Risk Correlation (p < 0.0001, 95% CI [0.032, 0.056]). Contact F1 (p = 0.007) is marginally above the Bonferroni threshold and is therefore reported with the caveat that this specific improvement should be interpreted with caution pending external replication. Confidence intervals were estimated via bootstrapping (1,000 resamples) of the test set predictions.

Comparative analysis of clinical and distance metrics

Table 2 shows critical dental specific performance indices, which are directly related to clinical usefulness and accuracy of treatment. The proposed mGAN exhibits high-performance of Root Surface Area Accuracy (RSAA) at 95.7 ± 2.4% which is an important improvement over the other existing state-of-art-methods. Specially, when comparing recent best performing approach TSTNet (93.4 ± 2.9%) and conventional approach U-Net (87.3 ± 4.2%). Thus, shows an improvement percentage of 2.3% and 11.4% is obtained by the proposed mGAN as compared with TSTNet and conventional U-Net approach respectively. this results evidently conveys that the estimation of root surface area shows direct impact on the planning of treatment especially appliable for root canal therapy as well as periapical surgery. Likewise, Canal Volume Preservation (CVP) metric demonstrated that proposed mGAN is achieved 0.912 ± 0.024, which is reasonably improvement of 3.0% and 10.8% as compared with TSTNet (0.885 ± 0.029) and conventional U-Net (0.823 ± 0.041) approach respectively. This improvement is highly shown impact of endodontic treatment planning, where volume measurements are critical to the complexity of the procedure and the success rates. The result of Cortical Thickness Estimation Error (CTE) indicates that proposed mGAN framework has significant improvements in the accuracy of the measurement with a score of 0.21 ± 0.04 mm versus 0.24 ± 0.05 mm of TSTNet and 0.34 ± 0.08 mm of U-Net, thus, proven decrement percentage of 12.5% and 38.2% respectively. This measurement is highly important for surgical planning specifically in the case of the placement of implants and orthognathic surgery in which cortical bone measurement directly affects the outcome of the treatment. Contact Distance Error (CDE) analysis reveals impressive improvements where the proposed mGAN showed 0.28 + 0.06 mm over TSTNet (0.35 ± 0.07 mm) and U-Net (0.52 ± 0.13 mm) with 20.0% and 46.2% improvements respectively. Moreover, prime key metrics includes Root Mean Square Distance Error (RMSE) and Mean Absolute Distance Error (MADE) are estimated for the proposed mGAN framework which is achieved 0.36 ± 0.08 mm and 0.26 ± 0.05 mm respectively, compared to the best performing baselines. Such advancements in distance measurement are especially important in dental practice which provides fine spatial relationship between the various anatomy structures. It can define the treatment options, particularly in situations where the mandibular third molar is to be extracted in the immediate vicinity of the inferior alveolar canal. Note: measurement error may cause nerve damage or insufficient planning of treatments.

Table 2.

Comparative analysis of clinical and distance metrics

Method	Dental-Specific Clinical Metrics			Contact Detection and Distance Metrics
	RSAA (%)	CVP	CTE (mm)	CDE (mm)	RMSE (mm)	MADE (mm)
UNet	87.3 ± 4.2	0.823 ± 0.041	0.34 ± 0.08	0.52 ± 0.13	0.67 ± 0.15	0.48 ± 0.11
UNet++	89.1 ± 3.8	0.847 ± 0.037	0.31 ± 0.07	0.47 ± 0.11	0.61 ± 0.13	0.43 ± 0.10
ESPNetv2	84.2 ± 5.1	0.789 ± 0.048	0.41 ± 0.11	0.61 ± 0.16	0.78 ± 0.18	0.56 ± 0.14
KCNet	86.7 ± 4.5	0.812 ± 0.043	0.36 ± 0.09	0.54 ± 0.14	0.69 ± 0.16	0.50 ± 0.12
V-Net	88.9 ± 3.9	0.841 ± 0.038	0.32 ± 0.07	0.49 ± 0.12	0.63 ± 0.14	0.45 ± 0.11
nnU-Net	91.2 ± 3.4	0.864 ± 0.034	0.28 ± 0.06	0.41 ± 0.09	0.53 ± 0.11	0.38 ± 0.08
DeepLab v3+	88.4 ± 4.0	0.835 ± 0.039	0.33 ± 0.08	0.51 ± 0.12	0.65 ± 0.14	0.47 ± 0.11
Attention U-Net	90.3 ± 3.6	0.856 ± 0.035	0.29 ± 0.06	0.44 ± 0.10	0.57 ± 0.12	0.41 ± 0.09
3D ResNet-UNet	89.7 ± 3.7	0.851 ± 0.036	0.30 ± 0.07	0.46 ± 0.11	0.59 ± 0.13	0.42 ± 0.10
UCTransNet	92.8 ± 3.1	0.878 ± 0.031	0.25 ± 0.05	0.37 ± 0.08	0.48 ± 0.10	0.34 ± 0.07
Swin	92.1 ± 3.2	0.872 ± 0.032	0.26 ± 0.05	0.39 ± 0.08	0.50 ± 0.10	0.36 ± 0.07
TSTNet	93.4 ± 2.9	0.885 ± 0.029	0.24 ± 0.05	0.35 ± 0.07	0.45 ± 0.09	0.32 ± 0.06
Proposed mGAN	95.7 ± 2.4	0.912 ± 0.024	0.21 ± 0.04	0.28 ± 0.06	0.36 ± 0.08	0.26 ± 0.05

RSAA Root Surface Area Accuracy, CVP Canal Volume Preservation, CTE Cortical Thickness Error, CDE Contact Distance Error, and MADE Mean Absolute Distance Error

Performance metrics

Table 3 classification performance analysis shows that the proposed mGAN has better discriminative ability against all standard machine learning performance measures. It shows a high true positive detection rate of 0.918 ± 0.019 which is the highest of all the compared methods and especially, it has improvements of 1.9% over TSTNet (0.901 ± 0.022) and 8.4% over the U-Net baseline (0.847 ± 0.034), respectively. This sensitivity is essential in the dental application where any pathological cases that are omitted would have dire clinical effects, such as nerve loss, infections, and treatment failures. Even stronger benefits are indicated in the specificity metric, where proposed mGAN obtains 0.967 ± 0.012 against other baseline approaches specially TSTNet (0.956 ± 0.014) and U-Net (0.923 ± 0.021) which is an improvement of 1.2% and 4.8% is achieved by proposed mGAN, respectively. Thus, it conveys that increased specificity signifies high quality of accurately determining non-pathological cases. In addition, the precision metric of proposed mGAN scored high value of 0.954 + 0.016, surpassing all the baseline methods by a margin of 1.4% to 7.1% over TSTNet (0.941 + 0.018) and U-Net (0.891 + 0.028). It means that high precision shows system finds a positive case with high chance, which is essential for clinical uncertainty about the system and minimizing false alarms in practice. Similarly, F1-score of the proposed mGAN framework is achieved a higher value of 0.936 ± 0.017 than TSTNet (0.920 ± 0.019) and U-Net (0.868 ± 0.029), indicating balanced performance across both sensitivity and precision metrics. It means that the capabilities of the proposed mGAN framework in detecting contact relationships between the mandibular third molar and infertolar canal contact relationships. Overall, 1.6% of improvement over the optimal baseline and 7.9% increase over the conventional method, since accurate contact detection directly improves the surgical risk assessment, as well as treatment planning decisions.

Table 3.

Classification performance metrics

Method	Sensitivity	Specificity	Precision	F1-Score
UNet	0.847 ± 0.034	0.923 ± 0.021	0.891 ± 0.028	0.868 ± 0.029
UNet++	0.863 ± 0.031	0.934 ± 0.019	0.905 ± 0.025	0.883 ± 0.026
ESPNetv2	0.821 ± 0.041	0.908 ± 0.026	0.867 ± 0.034	0.843 ± 0.036
KCNet	0.839 ± 0.036	0.918 ± 0.023	0.883 ± 0.030	0.860 ± 0.032
V-Net	0.858 ± 0.032	0.929 ± 0.020	0.898 ± 0.026	0.877 ± 0.027
nnU-Net	0.881 ± 0.027	0.945 ± 0.017	0.921 ± 0.022	0.900 ± 0.023
DeepLab v3+	0.854 ± 0.033	0.926 ± 0.021	0.894 ± 0.027	0.873 ± 0.028
Attention U-Net	0.871 ± 0.029	0.938 ± 0.018	0.912 ± 0.024	0.891 ± 0.025
3D ResNet-UNet	0.866 ± 0.030	0.935 ± 0.019	0.907 ± 0.025	0.886 ± 0.026
UCTransNet	0.894 ± 0.024	0.952 ± 0.015	0.934 ± 0.020	0.913 ± 0.021
Swin	0.887 ± 0.025	0.948 ± 0.016	0.928 ± 0.021	0.907 ± 0.022
TSTNet	0.901 ± 0.022	0.956 ± 0.014	0.941 ± 0.018	0.920 ± 0.019
Proposed mGAN	0.918 ± 0.019	0.967 ± 0.012	0.954 ± 0.016	0.936 ± 0.017

Risk assessment evaluation

A detailed review of the ability of the proposed mGAN framework to measure surgical risk and determine the relationships between prediction and actual clinical outcomes is provided in Table 4. RMSE Distance measure shows outstanding performance of proposed mGAN (0.36 ± 0.08 mm) as compared to 0.45 ± 0.09 mm for TSTNet and 0.67 ± 0.15 mm for U-Net, representing improvements of 20.0% and 46.3% respectively. Hence, this improvement is possible due to the accurate spatial association between vital anatomical structures directly determines the complexity of surgery and the rate of complication. Risk correlation measurement indicates the better outcome of 0.891 ± 0.027 is achieved by proposed mGAN framework as compared TSTNet (0.847 ± 0.034) and U-Net (0.743 ± 0.052), which indicates improvement by 5.2% and 19.9% respectively. The high-risk correlation allows better definition of high-risk cases that can need special surgical interventions, oral surgeon referral, or use of alternative treatment modalities. The overall correlation coefficient exhibits the same trend as mGAN attains (0.906 ± 0.024) versus (0.863 ± 0.031) and (0.768 ± 0.048) with respect of TSTNet and U-Net along an overall improvement of 5.0% and 18.0% respectively. These outcome enhancements show that the predictions made by the proposed mGAN framework is closer to the real clinical outcomes and make automated risk assessment systems more reliable. Therefore, the consistently high performance on all risk assessment measures, along with the tightest confidence, demonstrate the presence of strong predictive performance that can be used in clinical practices with a wide range of patients and different treatment circumstances.

Table 4.

Risk assessment evaluation and comparison with state of art methods

Method	RMSE (mm)	Risk Correlation	Correlation Coefficient
UNet	0.67 ± 0.15	0.743 ± 0.052	0.768 ± 0.048
UNet++	0.61 ± 0.13	0.771 ± 0.047	0.792 ± 0.043
ESPNetv2	0.78 ± 0.18	0.698 ± 0.061	0.721 ± 0.057
KCNet	0.69 ± 0.16	0.734 ± 0.054	0.756 ± 0.050
V-Net	0.63 ± 0.14	0.765 ± 0.048	0.786 ± 0.044
nnU-Net	0.53 ± 0.11	0.812 ± 0.039	0.831 ± 0.036
DeepLab v3+	0.65 ± 0.14	0.758 ± 0.049	0.779 ± 0.045
Attention U-Net	0.57 ± 0.12	0.795 ± 0.042	0.814 ± 0.038
3D ResNet-UNet	0.59 ± 0.13	0.783 ± 0.044	0.803 ± 0.040
UCTransNet	0.48 ± 0.10	0.834 ± 0.036	0.851 ± 0.033
Swin	0.50 ± 0.10	0.826 ± 0.037	0.844 ± 0.034
TSTNet	0.45 ± 0.09	0.847 ± 0.034	0.863 ± 0.031
Proposed mGAN	0.36 ± 0.08	0.891 ± 0.027	0.906 ± 0.024

Ablation study

The ablation study in Table 5 gives important information into the specific contribution of key architectural components of the proposed mGAN framework, which is a validation of the design choices and quantifies the contribution of each novel element. The entire W-Net architecture gets a better performance in all of the assessed metrics with DSC of 0.925 ± 0.019, F1-score of 0.932 ± 0.020, Risk Correlation of 0.891 ± 0.027, and Clinical Accuracy of 0.923 ± 0.021 with a computational efficiency of 847.3 GFLOPs. The systematic removal of individual components reveals that they have their own impact on overall system performance, which gives evidence-based support to the architectural complexity and innovation arguments. Eliminating Probability Feature Fusion (PFF) causes significant performance drops in all measures, where DSC dropping to 0.907 ± 0.024 (-1.9%), F1-score decreasing to 0.914 ± 0.025 (-1.9%), and Risk Correlation falling to 0.856 ± 0.034 (-3.9%). Thus, indicates PFF be the part of the multi-modal effectiveness of the method, which essentially for improving the capability of risk assessment. Clinical Accuracy reduced to 0.901 ± 0.026 (-2.4) shows that PFF is an essential involvement in converting technical gains into clinical outcomes. It induces a considerable amount of computational overhead but with a large amount of performance benefit to compensate this computational cost 823.1 GFLOPs (-2.9).

Table 5.

Ablation study: component contribution analysis

Configuration	DSC	Contact F1	Risk Corr	Clinical Acc	Computational Cost
W-Net (Full)	0.925 ± 0.019	0.932 ± 0.020	0.891 ± 0.027	0.923 ± 0.021	847.3 GFLOPs
w/o PFF	0.907 ± 0.024	0.914 ± 0.025	0.856 ± 0.034	0.901 ± 0.026	823.1 GFLOPs
w/o LLM Integration	0.913 ± 0.022	0.921 ± 0.023	0.834 ± 0.038	0.908 ± 0.024	831.7 GFLOPs
w/o Dual-Path Encoding	0.918 ± 0.021	0.925 ± 0.022	0.867 ± 0.031	0.915 ± 0.023	763.4 GFLOPs
w/o Skip Gating	0.921 ± 0.020	0.928 ± 0.021	0.883 ± 0.028	0.919 ± 0.022	825.9 GFLOPs
Standard GAN Baseline	0.889 ± 0.026	0.901 ± 0.027	0.798 ± 0.041	0.887 ± 0.028	892.5 GFLOPs

On the other hand, the ablation study of LLM Integration, it was found that Risk Correlation demonstrates the most significant drop to 0.834 ± 0.038 (-6.4%), whereas DSC and F1-score demonstrate more stable decreases to 0.913 ± 0.022 (-1.3%) and 0.921 ± 0.023 (-1.2%) respectively. Thus, it shows that the integration of LLM is almost entirely boosting the capacity of the system to integrate the clinical domain knowledge and increase the accuracy of risk predictions as opposed to elementary segmentation functionality. The decrease of Clinical Accuracy to 0.908 ± 0.024 (-1.6%) indicates that domain knowledge integration is important in clinical translation. The increase in the computational cost by 831.7 GFLOps (-1.8%) implies that the implementation of the LLM has almost no computational overheads and significant clinical gains. Deletion of Dual-Path Encoding reports equal effects on metrics with DSC dropping to 0.918 ± 0.021 (-0.8%), F1-score to 0.925 ± 0.022 (-0.8%), and Risk Correlation to 0.867 ± 0.031 (-2.7%). It is a substantial payoff in all the aspects considered where it shows significant drop in computational cost of 763.4 GFLOPs (-9.9). The Skip Gating ablation has the least performance relative, DSC reducing to 0.921 ± 0.020 (-0.4%) and F1-score to 0.928 ± 0.021 (-0.4%) and has comparable computational costs of 825.9 GFLOPs (-2.5%). The cumulative effect of all the innovative components can be seen by comparing it with the Standard GAN Baseline, but the performance difference of 4.0% in DSC, 3.4% in Contact F1, 11.7% in Risk Correlation, and 4.1% in Clinical Accuracy is achieved at the cost of 5.1% improvement in computational efficiency. This overall analysis of ablation substantiates the role played by each component of the architecture and shows that the proposed mGAN framework innovations offer significant performance advantages, sufficient to warrant the complexity of architecture, risk assessment strengths and clinical translation strengths. Regarding metric selection: DSC and HD95 are established standards for volumetric segmentation benchmarking; RMSE and CDE were chosen because sub-millimeter spatial accuracy directly determines whether a root-canal proximity is clinically actionable; RSAA captures the completeness of root surface reconstruction; CVP reflects canal morphology preservation; and Risk Correlation links computational outputs to the binary surgical decision outcome. We acknowledge that not all 23 metrics are equally independent or clinically motivated, and future work will focus on a reduced, clinically validated metric panel co-designed with oral and maxillofacial surgeons.

Discussion

This study set out to determine whether structural modifications to a generative adversarial network could support surgical decision-making for mandibular third molar (M3M) extraction by characterising the spatial relationship between the M3M roots and the inferior alveolar canal (IAC). Across a held-out test set drawn from 1,500 expert-annotated CBCT scans, the proposed mGAN framework outperformed all twelve baseline architectures on the segmentation, distance and classification metrics reported above, achieving a Dice Similarity Coefficient of 0.925, a root-mean-square distance error of 0.36 mm, and a sensitivity and specificity of 0.918 and 0.967, respectively, for surgical-necessity classification. The improvements over the strongest baseline (TSTNet) were accompanied by non-overlapping 95% confidence intervals on the primary measures and remained statistically significant after Bonferroni correction, with the exception of the contact-detection F1 score, which is therefore interpreted with caution pending external replication.

Clinical implications

From a clinical perspective, accurate sub-millimetre localisation of the root–canal interface is directly relevant to the planning of high-risk M3M extractions, in which inadvertent injury to the inferior alveolar nerve carries both patient morbidity and medico-legal consequences. The segmentation and distance-estimation accuracy reported here, combined with the uncertainty-aware decision mechanism that refers low-confidence cases for expert review, suggests that the framework is best positioned as a decision-support aid rather than an autonomous classifier, supporting evidence-based treatment planning where nerve-injury prevention is paramount. A processing time of approximately 2–2.2 min per case is compatible with integration into existing clinical workflows. These observations should nonetheless be read as indicators of potential clinical utility under controlled evaluation conditions, and not as evidence of improved patient outcomes.

Comparison with existing approaches

The performance gains relative to conventional segmentation networks such as U-Net and V-Net, and to more recent transformer-based and hybrid models, are consistent with the design rationale of the framework. The ablation analysis indicates that the integration of large language model (LLM)-derived clinical priors contributes most strongly to risk correlation, which showed the largest degradation (6.4%) when this component was removed, whereas Probabilistic Feature Fusion and dual-path encoding contributed more evenly across the segmentation and classification metrics. This pattern is consistent with the central hypothesis that encoding structured clinical domain knowledge, rather than relying on data-driven pattern recognition alone, is the principal source of the observed improvement in risk estimation, while the components of the W-Net backbone primarily improve boundary localisation and volumetric overlap.

Limitations

Several limitations should be considered when interpreting these results. First, the reference standard was established by two oral and maxillofacial surgeons with adjudication by a senior radiologist on the same institutional data; the evaluation therefore represents internal validation rather than independent external clinical validation. Second, the qualitative analysis (Figs. 7 and 8) shows that the discriminator can fail to flag clinically relevant root–canal proximity when the segmented contact region falls below the pixel-count threshold, producing false-negative predictions in subtle cases, and the sensitivity of the model to such fine interactions requires further refinement. Third, the framework was developed and tested on retrospective CBCT volumes acquired with a single scanner at one institution, which may limit generalisability across imaging vendors, acquisition protocols and patient populations. Finally, performance was reported across a broad panel of metrics that are not all mutually independent or equally clinically motivated, so the reported effect sizes should be interpreted with this redundancy in mind.

Future directions

Future work will prioritise multi-centre external validation using data from different scanners and clinical sites, together with prospective studies that compare the recommendations of the framework against real patient outcomes and clinician performance before any clinical deployment is considered. We further plan to improve the sensitivity of the discriminator to subtle, low-pixel-count contacts, and to consolidate the evaluation around a reduced, clinically validated panel of metrics co-designed with oral and maxillofacial surgeons, so that reported performance maps more directly onto surgical decision-making.

Conclusion

This paper has introduced a modified Generative Adversarial Network (mGAN) with W-Net architecture for dental surgical decision support. A systematic comparison of the framework with 12 established methods on quantitative measures was conducted with a 1,500-scan CBCT dataset. According to the results, the proposed mGAN has consistent improvements over the assessed baselines with differences in performance based on evaluating metrics and baseline ranging between 1.3% and 46.3%, respectively. It is important to note that the accuracy of distance measurement went up to 46.3% over U-Net and risk correlation went up by 19.9%, which indicates that there is a beneficial effect of the integration of LLM-derived priors and Probabilistic Feature Fusion to anatomical precision. The ablation research testifies to the fact that all the proposed architectural elements lead to the positive influence on the overall performance. These findings are positive to the future of AI-assisted dental surgical planning systems but further clinical validation, external cohort studies, and comparison to clinician performance are required before any clinical implementation is advised. The existing results should be regarded as the evidence of a concept that proves the possibility of such a framework in the context of a computational test.

Abbreviations

AI

Artificial Intelligence

ASD

Average Surface Distance

BERT

Bidirectional Encoder Representations from Transformers

CBCT

Cone Beam Computed Tomography

CDE

Contact Distance Error

CLAHE

Contrast Limited Adaptive Histogram Equalization

CNN

Convolutional Neural Network

CTE

Cortical Thickness Error

CVP

Canal Volume Preservation

DSC

Dice Similarity Coefficient

GAN

Generative Adversarial Network

GAP

Global Average Pooling

GFLOPs

Giga Floating-Point Operations

GPT

Generative Pre-trained Transformer

HD

Hausdorff Distance

HD95

95th Percentile Hausdorff Distance

IAC

Inferior Alveolar Canal

IoU

Intersection over Union

LLM

Large Language Model

M3M

Mandibular Third Molar

MADE

Mean Absolute Distance Error

mGAN

Modified Generative Adversarial Network

MLP

Multi-Layer Perceptron

OPG

Orthopantomogram (panoramic radiograph)

PFF

Probabilistic Feature Fusion

RMSE

Root Mean Square Error

ROI

Region of Interest

RSAA

Root Surface Area Accuracy

ViT

Vision Transformer

VS

Volumetric Similarity

Authors' contributions

G. Sivasathiya: Conceptualization, Methodology, Software, Validation, Writing - Original Draft R. Thiagarajan: Supervision, Writing - Review & Editing, Project Administration.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Data availability

The 2D dental panoramic (OPG) radiograph dataset analysed during the current study is publicly available on Kaggle (Dental OPG X-ray Dataset, https://www.kaggle.com/datasets/arfintanim/dental-opg-xray). The de-identified 3D CBCT datasets generated and analysed during the current study are not publicly available owing to institutional data governance and patient privacy restrictions, but are available from the corresponding author upon reasonable request and subject to institutional approval.

Declarations

Ethics approval and consent to participate

In this work, publicly available de-identified dental imaging datasets from Kaggle were used; these datasets do not include patient-identifiable information. So, ethics approval was not required under institutional regulations. The use of the data is in accordance with the dataset’s license terms and conditions.

Consent for publication

Not applicable. This study does not contain any individual person’s identifiable data or images in any form.

Competing interests

The authors declare no competing interests.