An orthodontic treatment post-treatment soft and hard tissue collaborative prediction method and system based on a conditional diffusion model and a computer readable storage medium

Abstract

The application discloses a kind of orthodontic treatment after soft and hard tissue collaborative prediction method, system and computer readable storage medium based on conditional diffusion model, belong to oral orthodontic treatment technical field.For the problems that existing method lacks soft and hard tissue collaborative prediction, precision is limited, and adaptability to complex cases is poor, the method is first to the lateral view of head before and after treatment is completed registration, cutting and normalization preprocessing, constructs the P-CDGAN network of U-Net architecture, constructs overall loss function by fusing adversarial loss, content loss, perceptual loss and structural similarity loss, and completes adversarial training using progressive gaussian noise-Reverse denoising mechanism;The preoperative image to be predicted is input into the trained model, and the collaborative prediction image containing the morphology of soft and hard tissue is output.The application realizes the integrated accurate prediction of soft and hard tissue, and the overall average radial error MRE of the landmark is less than 2mm, and the prediction performance is significantly better than the existing commercial orthodontic prediction software.

Description

Technical Field

[0001] This invention relates to the field of orthodontic treatment technology, specifically to a method, system, and computer-readable storage medium for predicting the co-prediction of soft and hard tissue changes after orthodontic treatment based on a conditional diffusion model. It is suitable for accurately predicting and visually simulating changes in soft and hard tissues after orthodontic treatment. Background Technology

[0002] Improving facial aesthetics is the most common reason for seeking orthodontic treatment. Orthodontic treatment indirectly improves soft tissue facial appearance through anterior tooth retraction and alveolar bone remodeling after tooth extraction. However, the patterns of change in soft and hard tissues vary greatly from person to person and are influenced by multiple factors such as lip tension, muscle activity, ethnicity, and age, resulting in a complex nonlinear coupling characteristic in the response of soft tissues to hard tissue movement. Therefore, establishing reliable methods for predicting changes in soft and hard tissues is a key clinical problem that urgently needs to be solved in the field of orthodontics.

[0003] Traditional prediction methods are primarily based on empirical soft and hard tissue movement ratios, and based on this, visualization software for treatment targets (such as Dolphin Imaging) has been developed. However, due to differences in the correlation between soft and hard tissues among different individuals and in different locations, and because the parameters in the model are fixed, the prediction accuracy is limited. Multiple studies have shown that the prediction errors of such software in the lower lip and chin regions often exceed 2 mm. Although statistical methods such as multiple regression are superior to traditional ratio methods in terms of prediction accuracy, they still struggle to characterize the complex nonlinear spatial relationships between soft and hard tissues, especially in the lip region where the prediction error remains relatively large. While three-dimensional finite element analysis can construct personalized facial models with clinically acceptable accuracy, this method is computationally intensive, modeling is complex, and the mathematical transformation processing of cone-beam computed tomography (CBCT) data is time-consuming and prone to morphological distortion, making it difficult to widely apply in clinical practice.

[0004] In recent years, AI-driven soft tissue prediction systems have shown potential in terms of accuracy and efficiency. In 2015, artificial neural networks (ANNs) were first used to predict changes in lip curvature in cases of tooth extraction and non-extraction, achieving better prediction accuracy than stepwise regression analysis. However, they still relied on cephalometric data, which had limitations such as large annotation workload and loss of image information. In 2025, Professor Bai Yuxing's team proposed the Soft-P-CGAN model, which, based on conditional generative adversarial networks (CGAN), verified the feasibility of using image generation networks for orthodontic posterior lateral radiograph prediction. However, it was only a non-end-to-end single-tissue framework, relying on doctors to input hard tissue parameters and only outputting soft tissue contours. The single-step generation architecture could not accurately depict the nonlinear coupling relationship between hard and soft tissues. Moreover, its small sample size limited representativeness, making it difficult to generalize to complex cases (such as cases of convex facial deformities). In terms of 3D prediction, Tanikawa et al. combined geometric morphology measurement with neural networks to build a prediction model based on 3D radiographic data; Park's team used the CGAN (conditional generative adversarial network) framework with CBCT data as input, achieving a prediction accuracy of 80.8%. However, these methods rely on manual annotation and complex 3D reconstruction, which increases the clinical burden, and the radiation risks and costs of CBCT limit its clinical application.

[0005] The existing studies mentioned above only predict soft tissue changes and lack synergistic prediction of changes in both soft and hard tissues. Orthodontic treatment indirectly affects soft tissues through tooth movement and alveolar bone remodeling; predicting only soft tissue changes has limited guiding value for treatment planning. Furthermore, Cho et al. compared the predictive performance of AI (artificial intelligence) models and conventional statistical methods on 44 soft and hard tissue landmarks in a large sample (887 cases), finding that the overall average error of the conventional method (1.69 mm) was better than the AI ​​model (2.12 mm), with AI only performing better in highly variable areas. This indicates that existing AI prediction models still have the following shortcomings: First, their prediction accuracy is limited, especially in highly variable areas of complex cases (such as the lip and chin), where errors are significant, and the training set bias caused by small samples further limits generalization performance; second, the prediction information is incomplete, with existing models only outputting soft tissue contours and lacking synergistic prediction of hard tissue driving information, making it difficult to directly guide treatment planning.

[0006] Conditional diffusion models (DDPM) offer a novel technical approach for medical image generation. Through forward denoising and backward denoising, they generate high-quality images, offering advantages over generative adversarial networks (GANs) such as more stable training and lower risk of pattern collapse. Compared to GANs, DDPM progressively approximates the target distribution through a gradual denoising process, exhibiting stronger expressive power and a more stable training process in fitting high-dimensional nonlinear mappings. Theoretically, it is more suitable for handling the nonlinear coupling relationships between soft and hard tissues in complex cases such as convex facial deformities. However, research on efficiently and accurately generating predictive images of soft and hard tissue changes after orthodontic treatment based on conditional diffusion models has not yet been reported. Summary of the Invention

[0007] To address the shortcomings of the prior art, this invention provides a method, system, and computer-readable storage medium for the coordinated prediction of soft and hard tissue changes after orthodontic treatment based on a conditional diffusion model. The aim is to achieve coordinated and accurate prediction of changes in soft and hard tissues after orthodontic treatment by leveraging the stable training and high-quality generation capabilities of the conditional diffusion model, combined with a progressive denoising mechanism.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A method for predicting the combined soft and hard tissue effects after orthodontic treatment based on a conditional diffusion model includes the following steps:

[0010] S1: Constructing a conditional diffusion generative adversarial network: Using the lateral cephalometric radiograph before orthodontic treatment as the hard tissue condition input, and the diffusion time step and malocclusion category as auxiliary conditions, a conditional diffusion generative adversarial network containing a U-Net generator and a discriminator is constructed.

[0011] The generator includes an encoding path, a bottleneck layer and an attention module, and a decoding path. The encoding path extracts multi-scale features of the input image through a residual downsampling module. The bottleneck layer and attention module model the spatial dependencies of the features through an attention gating mechanism, enabling the model to focus on key structural regions in the image. The decoding path restores spatial resolution through upsampling and performs channel concatenation with the feature maps of the corresponding layers of the encoder through skip connections. Simultaneously, it adaptively modulates the features based on temporal embedding and malocclusion category labels.

[0012] The discriminator adopts a temporal conditional discriminator structure, taking the splicing result of the predicted image and the corresponding diffusion state image as input, and combining time step information and conditional labels to distinguish between true and false images;

[0013] S2: Design the overall loss function: An overall loss function is constructed by introducing adversarial loss, content loss, perceptual loss, and structural similarity loss. The content loss constrains the pixel-level consistency between the predicted image and the real image and applies weighted constraints to the soft tissue contour region. The specific expression of the overall loss function is as follows:

[0014] L G =L adversarial + 2L content + L perceptual + 0.1L ssim ,

[0015] In the formula, L G Let L be the overall loss function. adversarial To combat losses, L content For content loss, Lperceptual For sensing loss, L ssim For structural similarity loss;

[0016] S3: Model Training: Using the pre-processed pre-treatment lateral cephalometric radiograph as the conditional input and the corresponding post-treatment image as the target output, the conditional diffusion generative adversarial network is trained adversarially using a progressive denoising mechanism. During training, the generator predicts the reverse denoising result based on the noise state during the diffusion process, combined with the pre-treatment image, time step information, and category conditions. The discriminator distinguishes between real denoised samples and generated denoised samples under the same time step conditions. Through adversarial training, the generator learns the nonlinear coupling relationship between the pre-treatment hard tissue conditions and the post-treatment soft and hard tissue morphological changes.

[0017] S4: Soft and hard tissue co-prediction: Input the pre-treatment lateral cephalometric radiograph to be predicted into the trained conditional diffusion generative adversarial network to generate a soft and hard tissue co-prediction image, which includes the hard tissue morphology and soft tissue contour morphology after treatment.

[0018] The method also includes data acquisition and preprocessing steps: acquiring lateral cephalometric images, registering the images, extracting the lower 1 / 3 region of interest, performing image enhancement and normalization processing, and unifying the image size.

[0019] The data collection included: subjects aged 18-40 years with skeletal Class I and Class II malocclusion who had completed treatment and achieved the design goals, had good craniofacial symmetry, and had no history of orthodontic treatment, trauma, or surgery; and excluded subjects with a history of serious systemic diseases, developmental disorders, syndromes, missing teeth, craniofacial defects, and asymmetrical malocclusions.

[0020] The registration is performed using a registration system for point-to-point registration.

[0021] The lower 1 / 3 region of interest is identified and cropped using the YOLOv8 model;

[0022] The image enhancement is histogram equalization;

[0023] The unified image size means unifying the image pixel size to 256×256;

[0024] The preprocessing also includes: applying the same spatial transformation to the registered image pairs using a synchronous enhancement protocol, including random rotation, random scaling and random horizontal flipping, and then randomly dividing the dataset into training set, validation set and test set.

[0025] The residual downsampling module performs feature compression through a 3×3 convolutional layer with a stride of 2, and combines it with an FIR (Finite Impulse Response) filter kernel to complete downsampling.

[0026] The temporal conditional discriminator takes the concatenation result of the predicted image and the corresponding diffusion state image as input, and performs joint discrimination by combining time step information and conditional labels. The network consists of a 1×1 input convolution, 6 residual downsampling modules, a 3×3 convolutional layer and a global pooling layer. Each residual downsampling module adopts the LeakyReLU activation function, and completes feature alignment and residual connection through a shortcut branch, gradually compressing the feature map resolution and improving the ability to distinguish between true and false, and finally outputs the discrimination result.

[0027] The method also includes a prediction result evaluation step: using structural similarity, peak signal-to-noise ratio, and mean absolute error to evaluate the quality of the generated image; and calculating the mean radial error of soft tissue markers and hard tissue markers, as well as the success rate of detection, to evaluate the prediction accuracy.

[0028] A soft and hard tissue collaborative prediction system includes: a data acquisition module, a data preprocessing module, a model building module, a loss calculation module, a model training module, a prediction output module, and an evaluation and verification module;

[0029] The data acquisition module is used to acquire lateral cephalometric radiographs before and after orthodontic treatment;

[0030] The data preprocessing module is used to perform registration, cropping of the lower 1 / 3 region of interest, image enhancement, and normalization on the lateral cephalometric radiographs.

[0031] The model building module is used to build a conditional diffusion generative adversarial network model containing a U-Net generator and a temporal conditional discriminator, with a pre-treatment cephalometric radiograph as the hard tissue conditional input and diffusion time step and malocclusion category as auxiliary conditions.

[0032] The loss calculation module is used to construct an overall loss function by introducing adversarial loss, content loss, perceptual loss, and structural similarity loss.

[0033] The model training module is used to perform adversarial training on the model using a progressive denoising mechanism;

[0034] The prediction output module inputs the pre-treatment image to be predicted into the trained model to generate a soft and hard tissue co-prediction image, which includes the hard tissue morphology and soft tissue contour morphology after treatment.

[0035] The evaluation and verification module is used to evaluate the image quality and the accuracy of the marker prediction results.

[0036] Preferably, the evaluation of the prediction results is as follows: the mean radial error (MRE) of 11 soft tissue landmarks and the mean radial error of 6 hard tissue landmarks are calculated in skeletal Class I malocclusion and skeletal Class II malocclusion, respectively, and the success rate (SDR) is calculated within the error range of 1.0 mm, 2.0 mm, 3.0 mm and 4.0 mm.

[0037] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method.

[0038] A computer program product includes a computer program that, when executed by a processor, implements the method.

[0039] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method.

[0040] The beneficial effects obtained by this invention are:

[0041] Coordinated prediction of soft and hard tissues: This invention breaks through the limitation of existing technologies that can only predict changes in soft tissues, and realizes the coordinated prediction of changes in soft and hard tissues after orthodontic treatment. It can simultaneously output the location of hard tissue landmarks and the shape of soft tissue contours, providing a more comprehensive visual basis for treatment planning.

[0042] Prediction accuracy is significantly improved: Through the progressive denoising mechanism of the conditional diffusion model, the model can effectively learn the inherent nonlinear coupling relationship between hard tissue changes and soft tissue morphological changes, with a comprehensive MAE of 1.07 mm (IQR: 0.63–1.62), which is significantly better than Dolphin (2.49 mm) and iOrtho (2.40 mm) (p<0.05). The MAEs for bone type I and II are 1.01 mm and 1.14 mm, respectively, and the average MAEs for soft / hard tissues are 1.26±0.42 / 0.98±0.31 mm and 1.38±0.50 / 1.12±0.35 mm, respectively.

[0043] Training stability and generation quality: Compared with traditional generative adversarial networks, the conditional diffusion model has the advantages of more stable training and lower risk of mode collapse. The generated predicted images have higher anatomical rationality and visual realism. The average accuracy of the visual Turing test is close to the random level (49.31%), indicating that experts have difficulty distinguishing between generated images and real images.

[0044] Highly clinically applicable: This invention can make predictions based on conventional lateral cephalometric radiographs (two-dimensional X-ray films), without the need for three-dimensional imaging such as CBCT, thus avoiding radiation risks and additional costs. Furthermore, it eliminates the need for complex manual annotation and three-dimensional reconstruction, lowering the threshold for clinical use and helping to improve the efficiency of doctor-patient communication, as well as assisting in clinical teaching and case database construction.

[0045] Multi-scale feature fusion: By introducing an attention mechanism, the model can capture multi-level features of the image from coarse to fine, so that the prediction results maintain high accuracy at different anatomical scales. Attached Figure Description

[0046] Figure 1 This is an overall flowchart of the method for predicting the synergistic effect of soft and hard tissues after orthodontic treatment based on the conditional diffusion model of the present invention.

[0047] Figure 2 This is a schematic diagram of the P-CDGAN conditional diffusion generative adversarial network model architecture. The model takes a pre-treatment lateral cephalometric radiograph (x), diffusion time step (t), and malocclusion category (z) as conditional inputs. The generator G (U-Net encoder-decoder structure) outputs a denoised post-treatment predicted image (y). The discriminator D, based on temporal conditions, distinguishes between the real and generated images. The generator uses adversarial loss (L... adversarial Content loss (L) content ), structural similarity loss (L ssim ) and perceptual loss (L) based on features extracted by VGG-16 (VGG16 deep convolutional pre-trained network). perceptual Perform multi-constraint joint optimization.

[0048] Figure 3 A detailed structural diagram of the P-CDGAN generator and discriminator;

[0049] Figure 4 Violin plots showing the distribution of baseline ANB angles (angles at points A-N-B, measured by cephalometric shadows of the relative positions of the maxilla and mandible) in patients with skeletal Class I and Class II malocclusions. The plots show the median (solid black line) and interquartile range (dashed black line), with α=0.001.

[0050] Figure 5 This is a schematic diagram showing the location of soft and hard tissue landmarks on the lateral side of the skull.

[0051] Figure 6 Comparison of typical sample images predicted after orthodontic treatment using different generative models;

[0052] Figure 7 Success rate (SDR) curves for different malocclusion landmarks within the range of 1.0 mm, 2.0 mm, 3.0 mm, and 4.0 mm.

[0053] Figure 8 A comparative chart showing the scores of clinical practicality questionnaires for experts with different seniority levels. Detailed Implementation

[0054] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The following embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. This study was exempted from ethics committee review, and this exemption has been approved by the Ethics Committee of the School of Stomatology, Guangxi Medical University (Approval No.: 2024KY0084). All procedures were strictly performed in accordance with relevant laws and regulations.

[0055] The overall flowchart of the method for predicting the synergistic effect of soft and hard tissues after orthodontic treatment based on the conditional diffusion model of this invention is as follows: Figure 1 As shown. This study is divided into three stages:

[0056] (1) Data annotation was completed using Labelme and X-AnyLabeling tools. Image registration was achieved through a registration system. After extracting the target region using the YOLOv8 model, the data was normalized and data augmentation was implemented.

[0057] (2) By comparing four representative generative models, the optimal BBDM conditional diffusion generative model was selected as the basic architecture. The residual U-Net generator and the temporal conditional discriminator were introduced into it to complete the model construction.

[0058] (3) The model was validated in multiple dimensions by using Likert scale scoring, visual Turing test, and comparative analysis with the prediction results of iOrtho and Dolphin software.

[0059] Example 1: Model Building and Training

[0060] 1. Data Preparation

[0061] (1) Data collection

[0062] This invention is a retrospective cohort study, including 1052 adult patients aged 18-40 who underwent orthodontic treatment at multiple campuses of the Affiliated Stomatological Hospital of Guangxi Medical University between January 2015 and December 2025. Basic information and pre- and post-treatment lateral cephalometric radiographs were collected. Inclusion criteria were:

[0063] (1) Adults aged 18-40 (inclusive);

[0064] (2) Complete orthodontic treatment (extraction orthodontic treatment) and achieve the treatment goals designed in the plan;

[0065] (3) Good craniofacial symmetry, with the anterior point of the chin in hard tissue deviating from the midline of the face by ≤1mm to the left and right;

[0066] (4) No history of previous orthodontic treatment;

[0067] (5) No history of trauma or surgery;

[0068] (6) The patient's clinical and imaging data are complete, all images are clear and complete, and the imaging meets the standards;

[0069] (7) Skeletal Class I and Skeletal Class II malocclusions.

[0070] Exclusion criteria:

[0071] (1) History of severe systemic diseases;

[0072] (2) Developmental disorders, syndromes, missing teeth, craniofacial defects, and asymmetrical deformities;

[0073] (3) Previous orthodontic treatment or orthognathic surgery history.

[0074] Lateral cephalometric radiographs were taken using Myriad Hyperion X9 (Italy's Satellite Group) and Planmecaromexis pro max (Finland) cameras. The original images were 2460×1950 or 1752×2108 pixels, with a resolution of 0.1 mm / pixel.

[0075] (2) Data preprocessing

[0076] Prior to cephalometric measurements, three orthodontists were trained by an orthodontic expert with over 15 years of experience, and 8% of the sample size was periodically calibrated. Finally, three senior orthodontists with 10 years of clinical experience used Labelme and X-AnyLabeling tools to annotate 22 anatomical landmarks (5 registration points, 6 hard tissue points, and 11 soft tissue points). Thirty samples were randomly selected and re-annotated by the aforementioned physicians four weeks after the initial annotation. Intra-class correlation coefficient (ICC) was used to assess the consistency between and within the annotators (ICC > 0.90 was considered excellent).

[0077] A registration system was used to perform point-to-point registration on lateral cephalometric radiographs before and after treatment. All registration operations were performed independently by the three senior physicians mentioned above. The registration results were cross-checked, and consensus was reached through consultation when there were disagreements. The YOLOv8 model was used to identify the lower 1 / 3 region of interest in the images, which was then cropped to obtain accurate images of the oral cavity's soft and hard tissues. Subsequently, histogram equalization and normalization were performed on the image data, and the image pixel size was uniformly adjusted to 256×256. To ensure the consistency of anatomical structures before and after treatment, a synchronous enhancement protocol was used in the data loading logic. Based on the image center as the reference point, a center rotation within ±20°, a random scaling factor of 0.5 to 1.0, and a horizontal flip with a probability of 0.3 were randomly applied to ensure that the registered image pairs underwent identical spatial transformations. The dataset was randomly divided into training, validation, and independent test sets in a training:validation:test ratio of 8:1:1.

[0078] 2. Model Architecture

[0079] Model Selection: To select the optimal generative framework suitable for image prediction after orthodontic treatment, this study used pre-treatment images as input and compared four representative models: Pix2Pix (paired image conditional generative adversarial network), CycleGAN (cyclic generative adversarial network), ResViT (CNN-Transformer fusion medical image generation network), and BBDM (conditional backdiffusion generative model). Pix2Pix represents a conditional generation framework based on paired data, CycleGAN represents an image transformation framework based on cycle consistency constraints, ResViT represents a medical image generation architecture that fuses CNN (convolutional neural network) and Transformer (transformer), and BBDM represents a conditional diffusion generative model.

[0080] Comparative analysis shows that the first three types of models still have problems such as blurred edges, local artifacts, or insufficient structure preservation in the prediction of complex soft and hard tissue structures; although BBDM has better generation stability and structure fidelity, it still has shortcomings such as complex diffusion sampling process and lack of explicit realism constraints in generation process.

[0081] Based on the advantages and disadvantages of the above models, this invention constructs the overall architecture of P-CDGAN (Conditional Diffusion Generative Adversarial Network) as follows: Figure 2 As shown, a conditional generative adversarial design is adopted: the pre-treatment cephalometric radiograph, diffusion time step, and malocclusion category are used as conditional inputs. The generator is responsible for outputting the denoised post-treatment predicted image, and the discriminator combines the temporal conditions to distinguish between real and generated images. The generation quality is gradually improved through adversarial iteration.

[0082] A detailed structural diagram of the P-CDGAN generator and discriminator is shown below. Figure 3 As shown, the generator is based on a U-Net-style encoder-decoder architecture and incorporates residual convolutional blocks and a self-attention mechanism to adapt to the image generation task in the diffusion denoising process. The specific design is as follows:

[0083] (1) Encoder path: The input is a 256×256 pixel image. After the first layer of 3×3 convolution (blue), a 128×128×64 feature map is output. Then, it is passed through the residual downsampling module (gray, ResDown Block) for multi-level downsampling. The feature map size and number of channels are 64×64×128, 32×32×128, and 16×16×256 respectively, until the bottleneck layer reaches 8×8×256. Each downsampling level is performed by a 3×3 convolution with a stride of 2, and combined with an FIR (Finite Impulse Response) filter kernel [1,3,3,1] to complete the downsampling.

[0084] (2) Bottleneck Layer and Attention Block: An attention block (red) is introduced at the 8×8×256 bottleneck layer. After attention gating, the dependencies between different spatial locations are modeled. The attention features combine global contextual information, enabling the model to pay more attention to key target regions and improve structural consistency and detail reconstruction capabilities.

[0085] (3) Decoder path: It consists of 5 upsampling layers (purple, Upsample) connected in series. It uses 2x upsampling to restore the spatial resolution of the feature map. The number of output channels is adjusted by a 3×3 convolution layer, and FIR filter kernel [1,3,3,1] is used for smooth reconstruction. After each upsampling, the feature map of the corresponding layer of the encoder is concatenated (white circle, Concatenation) to form a skip connection to retain multi-scale detail information. The final output is a generated image of the same size as the input (256×256).

[0086] (4) Residual Convolution Module: This module consists of a main branch and a shortcut branch. The main branch contains two 3×3 convolution layers, combined with group normalization, SiLU (Sigmoid-Weighted Linear Unit) activation function, time-step embedding injection, and Dropout operation. The time-step embedding t is used to represent the diffusion stage and injects intermediate features in an additive form; the condition label z is used to represent the condition information corresponding to each patient sample, and the features are adaptively modulated by AdaGN (Adaptive Group Normalization). The shortcut branch is used to preserve the original input features and uses 1×1 convolution (gray) to align them when the channel dimensions do not match, and finally adds them to the output of the main branch to form the output.

[0087] The discriminator employs a temporal conditional discriminator structure. Its input is the concatenated result of the image to be discriminated and the diffusion state image, and it performs joint discrimination by combining time step information and conditional labels. The network consists of a 1×1 input convolution (gray), six residual downsampling modules (yellow, ResBlock), a 3×3 convolutional layer (blue), and a global pooling layer (light blue). Each residual downsampling module uses the LeakyReLU activation function (with a negative slope of 0.2) and completes feature alignment and residual connection through a shortcut branch, thereby gradually compressing the feature map resolution and improving the ability to distinguish between true and false images, ultimately outputting the discrimination result.

[0088] Model training:

[0089] A diffusion time step t is randomly selected, and noise of an appropriate level is added to the input image X to obtain the noisy image X at the current time step. t Subsequently, X t The time step information t and the bone type condition z are input into the generator to obtain the predicted state X' of the previous time step. t-1 During the generator optimization phase, the network parameters of the discriminator D are fixed, and only the generator parameters are updated. The generator loss function consists of the adversarial loss L. adversarial Content loss L content Perceived loss L perceptual and structural similarity loss L ssim Together, they form a system that optimizes the generator parameters through various loss constraints. During the discriminator optimization phase, the network parameters of the generator G are fixed, and the true state X is... t-1 and the generated state X' t-1 The discriminator D is used to classify true and false states, and the discriminator parameters are updated through adversarial loss so that it can effectively distinguish between true diffusion states and generated diffusion states.

[0090] Model Inference:

[0091] During the inference phase, only the trained generator is used for image generation. The model takes the input image X, the skeletal category condition z, and the time step information t as input, and performs a four-step backdiffusion process starting from the initial noise state to gradually restore the predicted post-treatment image. The first round of denoising mainly restores the overall maxillofacial contour information; the second round of denoising further restores the main anatomical structures such as the jawbone and dentition; the third round of denoising focuses on the refinement of local boundaries of soft and hard tissues; and the fourth round of denoising completes residual noise suppression and texture detail enhancement, finally obtaining the predicted post-treatment image.

[0092] To enhance the quality of model generation, in addition to the traditional adversarial loss L... adversarialIn addition, content loss L was further introduced. content Perceived loss L perceptual and structural similarity loss L ssim .

[0093] After multiple sets of comparative experiments, taking into account the realism of the generated images, pixel-level errors, visual perception quality, and structural similarity, the adversarial loss L was ultimately selected. adversarial Content loss L content Perceived loss L perceptual and structural similarity loss L ssim The weights are set to 1.0, 2.0, 1.0, and 0.1, respectively. The final generator overall loss function L... G Represented as:

[0094] L G =L adversarial + 2L content + L perceptual + 0.1L ssim

[0095] Among them, the resistance loss L adversarial Used to enhance the realism of generated images; content loss L content Used to enhance pixel-level consistency between predicted and ground images (including emphasis-weighted constraints on soft tissue contour regions); perceptual loss L perceptual The difference between generated images and real images in high-level semantic features is determined by constraining the feature space of VGG-16 (a pre-trained VGG16 deep convolutional network); structural similarity loss L ssim Used to maintain the morphological consistency of teeth and surrounding hard tissue structures.

[0096] The model was trained on an NVIDIA RTX 4090 GPU with an input image resolution of 256×256 and a diffusion time step of 4. The training batch size was 4, and the total number of training epochs was 500. The optimizer used was Adam (Adaptive Moment Estimation Optimizer), with initial learning rates of 1×10⁻⁶ for both the discriminator and generator. -4 and 2×10 -4 The generator parameters are exponentially averaged using the EMA (Exponential Moving Average) mechanism. During training, the generator is optimized by joint constraints of adversarial loss, content loss, perceptual loss and structural similarity loss, while the discriminator is optimized by adversarial loss.

[0097] 3. Verification Phase

[0098] Image Quality Evaluation: To quantitatively assess the realism and fidelity of lateral cephalometric images generated by different models, structural similarity (SSIM), peak signal-to-noise ratio (PSNR), and mean absolute error (MAE) were selected as image quality evaluation metrics to quantitatively evaluate the image synthesis performance of each network. To verify the effectiveness and performance advantages of the P-CDGAN (Conditional Diffusion Generative Adversarial Network) model, multiple comparative experiments were conducted, comparing the proposed model with three mainstream algorithms (ResViT, Pix2Pix, and CycleGAN). The image generation performance of the proposed model was intuitively verified through various image evaluation metrics. SSIM, PSNR, and MAE were calculated using the following formulas:

[0099]

[0100]

[0101]

[0102] Landmark Prediction Accuracy Evaluation: To further verify the clinical applicability and landmark prediction accuracy of the model in predicting orthodontic treatment outcomes, a registration system was used to register the actual post-treatment images (T2) with the model-predicted post-treatment images (PT2), and the landmark error was calculated. T2 and PT2 were overlaid and compared to accurately identify the deviation between the predicted results and the actual clinical results. The anatomical landmark prediction error of the model was quantified by calculating the mean radial error (MRE) of 11 soft tissue landmarks and 6 hard tissue landmarks between T2 and PT2. Simultaneously, the success rate (SDR) at multiple thresholds of 1.0 mm, 2.0 mm, 3.0 mm, and 4.0 mm was used to comprehensively evaluate the model's predictive accuracy.

[0103] Comparison with mainstream software: To further highlight the predictive advantages of the P-CDGAN model, this paper compares its prediction results with those of mainstream clinical orthodontic prediction software (Angelica iOrtho software and Dolphin software), and statistically compares the differences in marker MRE among the three prediction methods.

[0104] Statistical Analysis: All statistical analyses were performed using SPSS statistical software (version 27; IBM, Armonk, NY, USA). After determining the data distribution type using the Shapiro-Wilk normality test, paired t-tests and Wilcoxon tests were used to analyze statistical differences between groups. Pearson correlation coefficients were used to analyze the correlation between the prediction error of the landmark point and the actual movement of the landmark point along the x-axis and y-axis before and after orthodontic treatment; p < 0.05 was considered statistically significant.

[0105] Visual Turing Test: To assess the visual realism of the generated images, a visual Turing annotation test was further conducted. The test set contained 60 external images (30 real post-treatment images and 30 model-predicted images). This test was completed by three orthodontic experts with over 15 years of clinical experience through a unified interface. To control variables, all images were presented in a fixed order, and experts were not allowed to retrospectively modify the evaluated images. Experts were informed that 30 images were real and 30 were synthetic, and they judged each image. Their judgment results were used to calculate sensitivity, specificity, and accuracy, with real images defined as positive and synthetic images as negative. The core purpose of this test was to verify the realism and visual fidelity of the model's output images, rather than assessing its clinical predictive accuracy or the degree to which treatment outcomes matched real cases.

[0106] Clinical Applicability Survey of Generated Images (Likert Scale Scoring): To evaluate the clinical applicability of images generated by the P-CDGAN model in simulating treatment effects, we conducted an expert survey. The survey selected 62 sets of lateral cephalometric radiographs, each including pre-treatment images, actual post-treatment images, and model-predicted images, which were simultaneously presented to three experts. Experts were required to score the predicted images based on a pre-defined set of 8 clinical evaluation criteria (using a 1-5 point Likert scale, with higher scores indicating a better match to the expected treatment effect).

[0107] Example 2: Predictive Performance Evaluation

[0108] This embodiment uses independent test set data and follows the logical chain of "baseline balance verification → image quality verification → quantitative accuracy verification → clinical tool comparison → visual authenticity verification → clinical usability verification" to verify the system performance of the trained P-CDGAN model. All statistical tests are considered to be statistically significant with p < 0.05.

[0109] 1. Baseline Equalization Validation

[0110] This study included 1052 participants, who were divided into a skeletal Class I group (684 cases) and a skeletal Class II group (368 cases) according to the ANB angle, the gold standard in orthodontic clinical practice. To exclude confounding factors from interfering with the subsequent performance verification results, the baseline demographic characteristics of the two groups were first tested for balance. The results are shown in Table 1.

[0111] Table 1. Comparison of baseline characteristics between skeletal Class I and Class II malocclusion patients in independent testing groups.

[0112]

[0113] Note: Categorical data are expressed as a percentage; normally distributed measurement data are expressed as mean ± standard deviation (mean ± SD), and non-normally distributed measurement data are expressed as median [interquartile range, IQR].

[0114] Statistical results showed that there was no statistically significant difference in gender distribution between the two groups (p=0.563>0.05), indicating a balanced male-to-female ratio. Although there was a statistically significant difference in age between the two groups (p=0.001<0.05), the effect size was extremely small (r=0.101, far below the small effect threshold of 0.2), explaining only about 1% of the inter-group variation, which was clinically insignificant and did not constitute a confounding factor. The ANB angle showed a highly significant difference between the two groups (p<0.001, r=0.838, large effect size), confirming that the two groups accurately distinguished malocclusion types and that the grouping was effective.

[0115] violin diagrams showing the baseline ANB angle distribution in skeletal Class I and Class II malocclusion patients. Figure 4 As shown in the figure, the solid black line represents the median and the dashed black line represents the interquartile range. The difference in the angular distribution of ANB between the two groups is statistically significant (p<0.001), confirming that the grouping of skeletal type I and type II is accurate and effective.

[0116] In summary, apart from the expected significant difference in grouping based on the ANB angle, there were no clinically significant differences in other demographic characteristics between the two groups. The baselines were balanced and comparable, and the subsequent model performance validation results were not affected by confounding factors, demonstrating scientific validity and reliability.

[0117] 2. Image quality verification

[0118] On the independent test set, after image preprocessing by the registration system, using pre-treatment lateral cephalometric radiographs (T1) as input and paired real post-treatment images (T2) as the gold standard, P-CDGAN was compared with three mainstream image generation control models (ResViT, Pix2Pix, and CycleGAN) using mean absolute error (MAE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM) as evaluation metrics. MAE measures pixel-level average deviation, with smaller values ​​indicating smaller errors; PSNR measures overall image fidelity, with larger values ​​indicating smaller distortions; SSIM measures anatomical structural consistency, with values ​​closer to 1 indicating better structural reproduction. The results are shown in Table 2.

[0119] Table 2. Image quality evaluation metrics comparing images generated by different models.

[0120]

[0121] The results show that P-CDGAN achieves the best performance in all three image quality metrics: compared with the second-best control model ResViT, MAE is reduced by 21.1%, PSNR is improved by 1.15dB, and SSIM is improved by 8.3%, significantly outperforming the Pix2Pix and CycleGAN models.

[0122] Figure 6 This paper presents a comparison of typical visual effects of different models in the image prediction task after orthodontic treatment, using real post-treatment images as the gold standard. From left to right: T1 (original input image before treatment), T2 (real gold standard image after treatment), the result generated by the P-CDGAN method of this invention, the result generated by ResViT, the result generated by Pix2Pix, and the result generated by CycleGAN; each row corresponds to one patient sample, showing a total of 3 typical cases. Visual comparison shows that: the image generated by P-CDGAN has the highest consistency with the real post-treatment image in terms of incisor position, jawbone contour, and soft tissue profile, with complete edge details and natural anatomical structure; the overall structure of the ResViT result is basically reasonable, but local edges are blurred, and details such as teeth and soft tissues are significantly lost; the Pix2Pix image can roughly reflect the contour changes after treatment, but the overall contrast is low, and the boundaries between soft and hard tissues are not clear enough; the CycleGAN result has the largest deviation, with significant differences between the jawbone morphology and tooth position and the real result, and obvious artifacts and structural distortions.

[0123] The combined quantitative indicators and visual sample results confirm that the predicted images generated by P-CDGAN after orthodontic treatment have the smallest pixel-level error, the highest image fidelity, and the best ability to restore the dentofacial anatomy, providing a reliable basis for the accuracy of subsequent landmark positioning.

[0124] 3. Quantitative accuracy verification

[0125] (1) Prediction accuracy analysis of landmarks: To quantitatively evaluate the prediction accuracy of P-CDGAN landmarks, this study selected 17 standard cephalometric landmarks as observation objects, and their positions on the lateral cephalometric radiographs are as follows: Figure 5 As shown, it includes 11 soft tissue landmarks and 6 hard tissue landmarks, covering the core clinical observation areas such as the nose, lips, chin, maxilla, mandible, and incisors. Using the actual post-treatment image (T2) as the gold standard, the mean absolute error (MAE) of each landmark in the horizontal (X-axis) and vertical (Y-axis) directions between the model-predicted image (PT2) and the gold standard was calculated. The statistical results of the axis errors are shown in Table 3.

[0126] Table 3. Comparison of landmark distance errors between skeletal Class I and Class II malocclusions.

[0127]

[0128] A: Point A (maxillary alveolar seat); B: Point B (mandibular alveolar seat); U1-R: Apical point of maxillary central incisor; U1: Incisional point of maxillary central incisor; L1: Incisional point of mandibular central incisor; L1-R: Apical point of mandibular central incisor; Prn: Nasal apex; Sn: Nasal floor point; A': Soft tissue point A; Ls: Upper lip border point; Stms: Upper oral fissure point; Stmi: Lower oral fissure point; Li: Lower lip border point; B': Soft tissue point B; Pog': Soft tissue anterior chin point; Me': Soft tissue submental point; Gn': Soft tissue chin apex. Hard tissue landmarks are: A, B, U1, U1-R, L1, L1-R; the rest are soft tissue landmarks.

[0129] Visual assessment showed that the generated images were highly consistent with the real images in terms of anatomical morphology, with an overall MRE < 2 mm, which is within the clinically acceptable range. There were no statistically significant differences between the predicted and actual values ​​of any landmarks in the two groups (p > 0.0029).

[0130] Patients with skeletal class II had significantly higher overall mean refractive error (MRE) than those with skeletal class I (p < 0.05), primarily due to differences in mean amplitude of action (MAE) along the X-axis (p < 0.05), while there was no significant difference in the mean amplitude of action (MAE) along the Y-axis (p > 0.05). The rank mean of hard tissue error was higher in class II than in class I (54.48 vs. 42.67, p < 0.05), while there was no difference in soft tissue error between the two groups, suggesting that the increased difficulty in predicting class II may stem from the unpredictability of hard tissue changes. Both groups showed that soft tissue prediction was more challenging than hard tissue prediction: in class I, the MAE of soft tissue was significantly greater than that of hard tissue (1.26 ± 0.42 vs. 0.98 ± 0.31 mm, p < 0.001), and the trend was consistent in class II (1.38 ± 0.50 vs. 1.12 ± 0.35 mm, p < 0.0125). Errors were concentrated at points U1 and L1 in hard tissue and points Ls, Li, B', and Gn' in soft tissue, indicating that these anatomical regions are challenging for prediction.

[0131] In summary, the overall prediction error of P-CDGAN is within the clinically acceptable range, but the prediction accuracy for Class II (especially hard tissue) and soft tissue landmarks still needs to be improved.

[0132] (2) Detection rate (SDR) analysis of each marker within different error ranges

[0133] Figure 7Success rate (SDR) curves for landmarks of different malocclusions within the ranges of 1.0mm, 2.0mm, 3.0mm, and 4.0mm are presented as heatmaps to visually demonstrate the success rate (SDR, i.e., the proportion of landmarks with prediction errors less than the corresponding threshold; a higher SDR indicates better prediction accuracy for that landmark) for both skeletal Class I and Class II patients. The results show three clear patterns:

[0134] ① Both sets of SDRs increased with the increase of the error threshold. When the threshold reached 3mm, the SDRs tended to saturate, indicating that the prediction error of the model for most marker points was controlled within 3mm, with no extremely large error results, and the overall prediction stability was good.

[0135] ② In most landmark locations, the SDR (Special Diagnosis Rate) of bony Class II is worse than that of bony Class I, and the difference is more pronounced within a 2mm range. At the 2mm threshold, the SDR of bony Class I at U1 and L1 points is 88.89% and 82.22%, respectively, while that of bony Class II is lower, at 67.31% and 65.38%, further supporting the conclusion that bony Class II deformities (especially in hard tissue areas) are the main challenging scenario for current model prediction.

[0136] ③ Within an error range of 2-4 mm, the SDR of soft tissue markers (such as Ls and Li) is lower than that of the corresponding hard tissues (such as U1 and L1) in the same region, reflecting the inherent complexity of soft tissue morphology prediction.

[0137] The heatmap results presented the error distribution characteristics of the model from multiple threshold and marker dimensions, providing a quantitative basis for the model's adaptation to clinical scenarios and subsequent iterative optimization.

[0138] 4. Performance comparison with existing clinical prediction software

[0139] To further verify the clinical application advantages of P-CDGAN, this study conducted a head-to-head comparison with two mainstream orthodontic prediction software programs (Dolphin Imaging and Angelalign iOrtho). The differences in MRE at various landmarks in skeletal Class I and Class II patients were statistically analyzed, and the results are shown in Tables 4-7.

[0140] Table 4. Comparison of prediction performance of P-CDGAN with other software (Skeletal Class I)

[0141]

[0142] Note: Data are expressed as median [IQR]. a represents P-CDGAN vs. Dolphin, b represents P-CDGAN vs. SiOrtho, c represents Dolphin vs. iOrtho, α=0.05.

[0143] Table 5. Comparison of prediction performance of P-CDGAN with other software (Skeletal Class II)

[0144]

[0145] Note: Data are expressed as median [IQR]. a represents P-CDGAN vs. Dolphin, b represents P-CDGAN vs. SiOrtho, c represents Dolphin vs. iOrtho, α=0.05.

[0146] Table 6 Comparison of prediction performance between P-CDGAN and iOrtho

[0147]

[0148] Note: Data are expressed as mean ± standard deviation. α = 0.0083.

[0149] Table 7 Comparison of prediction performance between P-CDGAN and iOrtho

[0150]

[0151] Note: Data are expressed as mean ± standard deviation. α = 0.0083.

[0152] In terms of soft tissue, for bony Class I convex facial deformities, the MRE of the P-CDGAN model at most landmarks was lower than that of Dolphin and iOrtho. Compared with Dolphin, except for the Li point, and compared with iOrtho, except for the Sn and A' points, the differences were statistically significant (p<0.05) (Table 4). In bony Class II, the MRE at all landmarks was lower than that of both software programs (p<0.05) (Table 5). This suggests that the P-CDGAN model has better predictive ability for soft tissue than the other two. In terms of hard tissue, for both bony Class I and II, the MRE of the P-CDGAN model at all hard tissue landmarks was lower than that of iOrtho (p<0.0083), indicating that its predictive ability for hard tissue is also better than that of iOrtho (Tables 6 and 7).

[0153] 5. Verification of the visual realism of generated images (Visual Turing Test)

[0154] To evaluate the visual realism of the P-CDGAN-generated images, 12 orthodontists were invited to participate in the Turing test (ideal accuracy 50%). The test set contained 60 external images (30 real post-treatment images and 30 P-CDGAN-generated images). Reality and falsity discrimination were performed in a blind state. All images were presented in a fixed order, and experts were not allowed to backtrack or modify them. The discrimination accuracy, sensitivity, and specificity were calculated, and the results are shown in Table 8.

[0155] Table 8. Comparison of the accuracy, sensitivity, and specificity of the Turing test for physicians of different seniority groups.

[0156]

[0157] Note: Data are expressed as mean ± standard deviation. α = 0.05.

[0158] The results showed that the overall discrimination accuracy was close to the random guessing level of 50%. There was no statistically significant difference in discrimination accuracy among physicians of different experience levels (p=0.447), indicating that regardless of clinical experience, experts could not effectively distinguish between images generated by P-CDGAN and real post-treatment images, and the visual realism of the generated images met clinical application requirements. In summary, P-CDGAN-generated images have high visual realism and are difficult for physicians of different experience levels to effectively distinguish, meeting clinical visual realism expectations. Unlike the Turing test used for architecture selection in Example 1, this test focuses on the performance of the final trained model, aiming to quantitatively verify whether it meets the visual realism expectations for clinical deployment.

[0159] 6. Expert evaluation of clinical applicability

[0160] To evaluate the clinical application potential of the P-CDGAN model in predicting lateral cephalometric radiographs, we conducted a questionnaire survey covering its practicality in treatment outcome prediction and doctor-patient communication. Sixty-two sets of lateral cephalometric radiographs (each set including pre-treatment images, actual post-treatment images, and P-CDGAN predicted images) were selected. A 1-5 Likert scale (1 = completely disagree, 5 = completely agree) was used to score the images across eight dimensions. The questionnaire covered three aspects: anatomical accuracy, clinical predictive value, and doctor-patient communication value. Specific questions included:

[0161] Q1: Do you think the position of the upper and lower incisors in the picture is close to the expected treatment?

[0162] Q2: Do you think the protrusion of the upper and lower lips in the picture is close to the expected treatment?

[0163] Q3: Do you think the improvement in the lateral facial prominence shown in the image is close to the expected treatment outcome?

[0164] Q4: Do you think the appearance of the chin in the picture is close to the expected treatment outcome?

[0165] Q5: Do you think this graph can help predict soft tissue changes in patients after treatment?

[0166] Q6: Do you think this graph can help predict changes in hard tissue after treatment?

[0167] Q7: Do you think showing this diagram to patients helps them explain and understand the treatment plan?

[0168] Q8: Do you think this diagram can help patients decide whether to receive orthodontic treatment?

[0169] The results are shown in Table 9. Overall, the scores for each issue were at a moderately high level. The mean scores showed a slight decreasing trend, with the highest in the junior group (3.68–3.90), the middle in the middle in the middle group (3.68–3.74), and the lowest in the senior group (3.37–3.70). In summary, the scores for all experience levels were close to good, indicating that the P-CDGAN model has gained consistent clinical acceptance among orthodontists of different experience levels and has good visual equivalence and application potential across experience levels.

[0170] Table 9 Questionnaire Scoring Results

[0171]

[0172] Visualization of scoring results from experts with different seniority, for example Figure 8 As shown in the figure, the horizontal axis represents the evaluation items Q1-Q8 of the questionnaire, and the vertical axis represents the Likert scale scores of 1-5 points. The figure is categorized into three groups of experts: junior, intermediate, and senior, and displays the average scores of the eight clinical evaluation dimensions. The eight questions cover three aspects: anatomical accuracy, predictive value, and doctor-patient communication value.

[0173] The results showed that the overall scores of the three expert groups were close to 4 points (good level); the scores decreased slightly with the increase of the experts' seniority, but all dimensions reached the clinical usability level, which confirms that the predictive images generated by P-CDGAN have good recognition and practical value among clinicians of different seniority.

Claims

1. A method for predicting the coordinated soft and hard tissues after orthodontic treatment based on a conditional diffusion model, characterized in that, Includes the following steps: S1: Constructing a conditional diffusion generative adversarial network: Using the lateral cephalometric radiograph before orthodontic treatment as the hard tissue condition input, and the diffusion time step and malocclusion category as auxiliary conditions, a conditional diffusion generative adversarial network containing a U-Net generator and a discriminator is constructed. The generator includes an encoding path, a bottleneck layer and an attention module, and a decoding path; the encoding path extracts multi-scale features of the input image through a residual downsampling module. The bottleneck layer and attention module model the spatial dependency of features through an attention gating mechanism, enabling the model to focus on key structural regions in the image; the decoding path restores spatial resolution through upsampling and performs channel splicing with the feature maps of the corresponding layers of the decoding path through skip connections, while adaptively modulating features based on temporal embedding and malocclusion category labels. The discriminator adopts a temporal conditional discriminator structure, taking the splicing result of the predicted image and the corresponding diffusion state image as input, and combining time step information and conditional labels to distinguish between true and false images; S2: Design the overall loss function: An overall loss function is constructed by introducing adversarial loss, content loss, perceptual loss, and structural similarity loss. The content loss constrains the pixel-level consistency between the predicted image and the real image and applies weighted constraints to the soft tissue contour region. The specific expression of the overall loss function is as follows: L G =L adversarial + 2L content + L perceptual + 0.1L ssim , In the formula, L G Let L be the overall loss function. adversarial To combat losses, L content For content loss, L perceptual For sensing loss, L ssim For structural similarity loss; S3: Model Training: Using the pre-processed pre-treatment lateral cephalometric radiograph as the conditional input and the corresponding post-treatment image as the target output, the conditional diffusion generative adversarial network is trained adversarially using a progressive denoising mechanism. During training, the generator predicts the reverse denoising result based on the noise state during the diffusion process, combined with the pre-treatment image, time step information, and category conditions; the discriminator distinguishes between real denoised samples and generated denoised samples under the same time step conditions, and through adversarial training, the generator learns the nonlinear coupling relationship between the pre-treatment hard tissue conditions and the post-treatment soft and hard tissue morphological changes. S4: Soft and hard tissue co-prediction: Input the pre-treatment lateral cephalometric radiograph to be predicted into the trained conditional diffusion generative adversarial network to generate a soft and hard tissue co-prediction image, which includes the hard tissue morphology and soft tissue contour morphology after treatment.

2. The method according to claim 1, characterized in that, The method also includes data acquisition and preprocessing steps: acquiring lateral cephalometric images, registering the images, extracting the lower 1 / 3 region of interest, performing image enhancement and normalization processing, and unifying the image size.

3. The method according to claim 2, characterized in that, The data collection included: subjects aged 18-40 years with skeletal Class I and Class II malocclusion who had completed treatment and achieved the design goals, had good craniofacial symmetry, and had no history of orthodontic treatment, trauma, or surgery; and excluded subjects with a history of serious systemic diseases, developmental disorders, syndromes, missing teeth, craniofacial defects, and asymmetrical malocclusions. The registration is performed using a registration system for point-to-point registration. The lower 1 / 3 region of interest is identified and cropped using the YOLOv8 model; The image enhancement is histogram equalization; The unified image size means unifying the image pixel size to 256×256; The preprocessing also includes: applying the same spatial transformation to the registered image pairs using a synchronous enhancement protocol, including random rotation, random scaling and random horizontal flipping, and then randomly dividing the dataset into training set, validation set and test set.

4. The method according to claim 1, characterized in that, The residual downsampling module performs feature compression through a 3×3 convolutional layer with a stride of 2, and combines it with an FIR filter kernel to complete downsampling.

5. The method according to claim 1, characterized in that, The temporal conditional discriminator consists of a 1×1 input convolution, six residual downsampling modules, a 3×3 convolutional layer, and a global pooling layer. Each residual downsampling module uses the LeakyReLU activation function and completes feature alignment and residual connection through a side branch, gradually compressing the feature map resolution and improving the ability to distinguish between true and false data, and finally outputting the discrimination result.

6. The method according to claim 1, characterized in that, The method also includes a prediction result evaluation step: using structural similarity, peak signal-to-noise ratio, and mean absolute error to evaluate the quality of the generated image; and calculating the mean radial error of soft tissue markers and hard tissue markers, as well as the success rate of detection, to evaluate the prediction accuracy.

7. A soft and hard tissue collaborative prediction system, used to perform the method according to any one of claims 1-6, characterized in that, include: The system includes a data acquisition module, a data preprocessing module, a model building module, a loss calculation module, a model training module, a prediction output module, and an evaluation and validation module. The data acquisition module is used to acquire lateral cephalometric radiographs before and after orthodontic treatment; The data preprocessing module is used to perform registration, cropping of the lower 1 / 3 region of interest, image enhancement, and normalization on the lateral cephalometric radiographs. The model building module is used to build a conditional diffusion generative adversarial network model containing a U-Net generator and a temporal conditional discriminator, with a pre-treatment cephalometric radiograph as the hard tissue conditional input and diffusion time step and malocclusion category as auxiliary conditions. The loss calculation module is used to construct an overall loss function by introducing adversarial loss, content loss, perceptual loss, and structural similarity loss. The model training module is used to perform adversarial training on the model using a progressive denoising mechanism; The prediction output module inputs the pre-treatment image to be predicted into the trained model to generate a soft and hard tissue co-prediction image, which includes the hard tissue morphology and soft tissue contour morphology after treatment. The evaluation and verification module is used to evaluate the image quality and the accuracy of the marker prediction results.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-6.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-6.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method according to any one of claims 1-6.

Images