Multi-direction and multi-scale dental focus detection method for oral cavity panoramic film

Abstract

The invention provides a multi-direction and multi-scale dental focus detection method for an oral cavity panorama. The method comprises the following steps: firstly, obtaining an oral cavity panorama, performing screening and standardization preprocessing, completing focus bounding box labeling and dividing a training set, a verification set and a test set; performing data enhancement on the training sample, wherein the data enhancement comprises basic enhancement and focus visibility enhancement; and performing visibility enhancement on at least part of the samples according to a preset proportion, wherein the enhancement at least comprises contrast enhancement and gray inversion type transformation. The method comprises the following steps of: constructing a first-stage dental focus detection network DENTAL-NET: introducing an edge-Aware Feature Extraction (EAFE) into a backbone network, including an edge-Aware Multi-Directional Scanning (EAMDS), and used for generating a multi-directional edge response and performing adaptive fusion based on a directional weight to enhance boundary and directional structure characterization, and constructing a second-stage dental focus detection network DENTAL-NET based on the edge-Aware Multi-Directional Scanning (EAMDS); a multi-scale adaptive feature fusion module is arranged in the neck network, scale alignment is performed on different levels of features, and multi-receptive field representation is constructed to realize multi-scale fusion. And training by using the enhanced training set and determining an optimal weight through the verification set to obtain a detection model, and inputting a to-be-detected panorama to output a focus category, a confidence coefficient and a positioning frame result. The method can improve the detection performance under the background of low contrast, gray ambiguity and complex texture, reduces missing detection and false detection, and is suitable for auxiliary film reading scenes such as clinical preliminary screening and follow-up visit.

Description

Technical Field

[0001] This invention belongs to the field of intelligent medical image analysis and computer vision target detection technology, specifically relating to an automatic detection method for dental lesions based on orthopantomogram (OPG) images, and more particularly to a dental lesion detection method that combines lesion visibility enhancement, multi-directional edge scanning, and multi-scale adaptive feature fusion. Background Technology

[0002] Panoramic radiographs can simultaneously present the overall anatomical structure of the upper and lower jaws and dentition, and are clinically used for screening and assessment of oral diseases such as dental caries, periapical lesions, periodontal bone resorption, impacted teeth, and tooth defects. Traditional manual interpretation of radiographs relies heavily on the doctor's experience and is prone to omissions and misinterpretations under high-intensity work, exhibiting problems such as repetitive labor and significant subjective differences. Therefore, the use of deep learning for the automated detection of dental lesions on panoramic radiographs has significant application value.

[0003] Object detection methods can be broadly categorized into three types: two-stage detection, single-stage detection, and Transformer-based end-to-end detection. Single-stage detection, due to its relatively simple structure, fast inference speed, and ease of deployment, can balance detection accuracy and efficiency to a certain extent, making it commonly used in engineering applications. However, directly applying existing single-stage object detection networks to dental lesion detection on panoramic oral radiographs still presents certain challenges. First, panoramic oral radiographs generally exhibit imaging characteristics such as grayscale ambiguity, low contrast, blurred boundaries, and complex background textures. The grayscale of lesions is similar to that of surrounding tissues, and there is significant overlap between tooth edges and alveolar bone texture. This results in insufficient modeling ability of general feature extraction modules for edge details and directional structures, easily leading to category confusion, missed detections, and false detections, thus reducing localization accuracy. Second, dental lesions often exhibit characteristics of being predominantly small targets with a large scale span. General feature fusion strategies are insufficient for representing small targets, making it difficult to effectively retain fine-grained semantic information, further exacerbating the problem of missed detections. Third, the dental arch exhibits a significant directional distribution along the jaw arch. If the network lacks directionally sensitive edge and structural representation capabilities, the model is more prone to false detections or detection box offsets in complex texture backgrounds.

[0004] In summary, it is necessary to propose an improved method for detecting dental lesions that addresses the characteristics of panoramic oral radiographs, such as grayscale ambiguity, blurred boundaries, distinct directional structures, and small targets at multiple scales. This would enhance the accuracy and robustness of the detection, provide a more stable and reliable auxiliary reference for clinical practice, reduce the burden on doctors in interpreting images and subjective differences, and decrease missed and false detections in automated detection. Summary of the Invention

[0005] (a) Purpose of the invention

[0006] The purpose of this invention is to provide a multi-directional, multi-scale dental lesion detection method for panoramic oral radiographs, in order to solve the problems of insufficient edge direction structure modeling and insufficient multi-scale fusion of small targets caused by existing first-stage target detection networks under the conditions of low contrast, gray-level ambiguity, blurred boundaries and significant dentition directionality in panoramic oral radiographs, thereby improving the accuracy and robustness of automatic dental lesion detection while taking into account inference efficiency.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention provides a method for multi-directional, multi-scale dental lesion detection using panoramic radiographs, comprising the following steps:

[0009] Step 1: Data preparation stage. Acquire panoramic images of the oral cavity and construct a dataset; perform standardized preprocessing on the images; annotate the bounding boxes of at least one dental lesion to form labeled data; divide the samples into training set, validation set and test set.

[0010] Step 2: Data Augmentation Stage. OPG-specific data augmentation (DABL) is performed on the training set to improve model robustness. DABL includes Basic Augmentation (BAug) and Lesion Augmentation (LAug). Basic Augmentation includes, but is not limited to, one or more of random rotation, flipping, scaling, cropping, and stitching enhancements. Lesion Augmentation includes at least contrast enhancement and grayscale inversion transformations.

[0011] Step 3: Model Construction Stage. A dental lesion detection network, DENTAL-NET, is constructed. DENTAL-NET is a one-stage target detection network designed for panoramic dental radiographs, comprising a backbone network, a neck network, and a head detection head. An edge-aware feature extraction (EAFE) module is introduced into the backbone network to enhance boundary and orientation structure representation. The core of EAFE is the edge-aware multi-directional scanning unit (EAMDS). A multi-scale adaptive feature fusion module is introduced into the neck network to enhance multi-scale feature interaction and small target representation. The module positions are as follows: Figure 2 As shown.

[0012] Step 4: In the model training phase, the enhanced training set obtained in Step 2 is used to iteratively train the DENTAL-NET constructed in Step 3 to obtain the dental lesion detection model; during the training process, the optimal weight parameters are selected as the final model based on the validation set.

[0013] Step 5: Model inference stage. Input the panoramic oral radiograph to be detected into the detection model and output the category information of the dental lesion and its localization box result.

[0014] Preferably, the contrast enhancement in step 2 is achieved using Contrast-Limited Adaptive Histogram Equalization (CLAHE) to improve local contrast and enhance the visibility of lesion details. The gray-level inversion transformation in step 2 is achieved using Negative Transformation (NegTrans), which inverts the pixel gray-level values ​​according to the upper limit of gray level to highlight low-contrast structures and alleviate gray-level ambiguity. The upper limit of gray level for 8-bit images is 255, thereby enhancing the gray-level separability of the lesion area and surrounding tissues and suppressing information overload caused by low-contrast imaging. This ensures that effective detail representations that can be used for detection can still be obtained against the background of blurred boundaries and texture superposition.

[0015] Preferably, the core of EAFE in step 3 is an edge-aware multi-directional scanning unit (EAMDS), which includes multi-directional edge scanning and directional weight fusion; the preset direction set is preferably... Trainable directional weight parameters are set for the edge response features obtained in each direction, and the weights for each direction are obtained by normalization using the Softmax function. Then, the edge response features in each direction are weighted and fused according to the normalized directional weights to obtain the fused edge features, thereby enhancing the sensitivity to the dental arch structure and improving the representation of boundary details. This further suppresses false edge responses caused by texture overlay, reduces false detections in complex backgrounds, and reduces detection box offset.

[0016] More preferably, the EAFE has two structural forms: a directly embedded EAFE and a C3k nested EAFE. In the directly embedded EAFE, EAMDS directly serves as the bottleneck unit within the EAFE; in the C3k nested EAFE, the bottleneck unit within the EAFE is the C3k_EAMDS submodule, and the C3k_EAMDS, except that its bottleneck unit is replaced by EAMDS, has the same structure as the original C3k submodule of the YOLO11 structure; the above two structural forms are respectively as follows: Figure 4 and Figure 5As shown. Further, multiple EAFEs set in the backbone network can independently adopt direct-embedded EAFEs or C3k nested EAFEs respectively, and the two structural forms can be combined and configured in the same backbone network, so that the edge and direction structure representation capabilities can be flexibly configured according to the network hierarchy and feature scale and enhanced as needed.

[0017] Preferably, the multi-scale adaptive feature fusion module takes multi-scale feature maps as input. Taking the , , in the backbone network as an example, they respectively correspond to feature layers with different downsampling ratios: is the shallow feature with high resolution, is the medium-resolution feature, is the deep feature with low resolution; the multi-scale adaptive feature fusion module first performs scale and channel alignment on each scale feature through upsampling, downsampling and channel transformation, and completes feature fusion after alignment; then multiple groups of depth convolution branches with different convolution kernel sizes are set in parallel to construct multi-receptive field feature representations, where the convolution kernel sizes are preferably 5, 7, 9, 11 in sequence. The outputs of each branch are aggregated and then subjected to residual connection with the fusion feature to represent the semantics and detail information of targets with different scales, and enhance the network's feature extraction and representation capabilities for lesions with different scales and different morphologies, so as to take into account shallow details and deep semantics on the basis of scale alignment and enhance multi-receptive field representations; furthermore, improve the detection capabilities of small targets and lesions with large scale spans, reduce missed detections and improve the robustness of cross-scale detection.

[0018] (III) Beneficial effects

[0019] Compared with the prior art, the present invention has at least the following beneficial effects:

[0020] 1) Aiming at the problems of low contrast and strong gray-scale ambiguity commonly existing in oral panoramic films, resulting in insignificant lesion details, the present invention performs lesion visibility enhancement LAug composed of contrast enhancement and gray-scale inversion transformations on training samples according to a preset ratio, improves the separability between lesions and the background, which is beneficial for the model to learn stable discriminative features, and further reduces missed detections and class confusion in low-contrast scenarios;

[0021] 2) Aiming at the problems of superposition of tooth edge and alveolar bone texture, blurred boundary and directional distribution of dental arch, resulting in difficult modeling of direction structure, the present invention sets EAFE with EAMDS as the core in the backbone network, and through multi-directional edge response generation, direction weight adaptive fusion based on the Softmax function and residual output mechanism, explicitly enhances the representation capabilities of boundary details and direction structure, thereby improving the stability and robustness of the positioning result and reducing the false detection probability under complex texture backgrounds;

[0022] 3) To address the problem that dental lesions are mostly small targets with a large scale span, resulting in insufficient multi-scale representation, this invention introduces a multi-scale adaptive feature fusion structure. In the neck network, multi-layer features are scale-aligned and fused, and multi-receptive field representations are constructed, thereby enhancing the feature expression ability of small targets and multi-morphological lesions, and further reducing false negatives.

[0023] 4) Experimental verification shows that, under the dataset and evaluation settings described in this embodiment (as shown in Table 1), DENTAL-NET achieves superior results in detection accuracy and maintains high inference speed even with a small number of parameters and computational load. Preferably, the mean average precision at IoUthreshold 0.5 (mAP@0.5) reaches 82.8, the number of parameters is 2.6M, the computational load (GigaFloating-point Operations, GFLOPs) is 7.8, and the inference rate (Frames Per Second, FPS) is 22.8, indicating that the present invention can balance detection accuracy and inference efficiency, and is therefore suitable for rapid auxiliary detection and deployment of dental lesions on panoramic oral radiographs.

[0024] 5) The present invention outputs detection results such as lesion candidate boxes, categories and confidence levels, which can be used in clinical screening and follow-up and other image reading assistance scenarios to provide doctors with consistent computer-aided reference information, thereby reducing the burden of repetitive image reading and improving screening efficiency. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall process of the dental lesion detection method described in this invention;

[0026] Figure 2 This is a schematic diagram of the DENTAL-NET network structure described in this invention;

[0027] Figure 3 This is a schematic diagram of the data augmentation framework described in this invention;

[0028] Figure 4 This is a schematic diagram of the structure of the directly embedded EAFE described in this invention;

[0029] Figure 5 This is a schematic diagram of the C3k nested EAFE structure described in this invention;

[0030] Figure 6 This is a schematic diagram of the structure of the multi-scale adaptive feature fusion module described in this invention. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings. It should be understood that the following embodiments are only for illustrating the present invention and not for limiting the scope of protection; those skilled in the art can make equivalent substitutions or changes to the specific steps, parameter ranges, module insertion positions and quantities without departing from the spirit of the present invention, and all such substitutions or changes should fall within the scope of protection of the present invention.

[0032] Example: A method for multi-directional, multi-scale dental lesion detection using panoramic radiographs.

[0033] This embodiment takes a panoramic oral radiograph as input and outputs the type and bounding box of the dental lesion, such as... Figure 1 As shown, the specific steps include:

[0034] 1. Data Preparation Stage

[0035] Step 1_1 Obtain panoramic oral radiographs. The images can be grayscale or color images. The same image can contain multiple lesion targets. The samples can be collected clinically or from publicly available data sources.

[0036] Step 1_2: Perform quality screening on the samples, and remove image samples with severe motion artifacts, severe overexposure / underexposure, or missing key anatomical structures.

[0037] Step 1_3: Cover or crop the area containing patient identification information in the image;

[0038] Step 1_4: Adjust the image to the preset input size and perform pixel value normalization or standardization; in a preferred embodiment, the input size is 640×640, and the image is scaled proportionally and aligned to the input size using a padding method to avoid deformation;

[0039] Step 1_5: Annotate the bounding boxes of the dental lesions and generate annotation files that correspond one-to-one with the images. The annotation information includes at least the category label and the bounding box coordinates. Multiple bounding boxes are allowed in the same image, and the category set may vary depending on the data source.

[0040] Step 1_6 divides the samples into training, validation and test sets for model training, model selection and performance evaluation.

[0041] 2. Data augmentation stage (e.g.) Figure 3 (As shown)

[0042] This embodiment uses the DABL enhancement framework, which includes the basic enhancement BAug and the lesion visibility enhancement LAug.

[0043] Step 2_1 Perform BAug on the training samples to improve sample diversity. The basic augmentation includes, but is not limited to, one or more of random rotation, horizontal flipping, random scaling, random cropping, and splicing augmentation.

[0044] Step 2_2 Select a portion of the training samples according to a preset ratio to perform LAug. The ratio can be set according to the data size and is preferably 0.05 to 0.30.

[0045] Step 2_3 Perform contrast enhancement processing on the selected samples, preferably using contrast-limited adaptive histogram equalization (CLAHE).

[0046] Step 2_4 After completing CLAHE, perform a grayscale inversion-type transformation to highlight low-contrast details and alleviate grayscale ambiguity. Preferably, this includes a negative transformation (NegTrans), as shown in the following formula:

[0047]

[0048] in, For pixels Pixel value at that location, These are the transformed pixel values. This is a constant representing the upper limit of grayscale when the image is an 8-bit image. ;

[0049] Step 2_5: Combine the samples obtained by BAug and LAug processing with the original samples to form an enhanced training set for subsequent model training.

[0050] 3. Model building phase (e.g.) Figure 2 (As shown)

[0051] This embodiment constructs a dental lesion detection network, DENTAL-NET, which is a one-stage target detection network based on YOLO11. It includes a backbone network, a neck network, and a head detection head. In the backbone network, four C3k2 modules in the original YOLO11 are replaced by EAFE modules. In the neck network, two multi-scale adaptive feature fusion modules are fixed to enhance multi-scale information flow and small target detection capabilities.

[0052] Step 3_1 Backbone Network Construction and EAFE Replacement of C3k2

[0053] The overall feature extraction process of the YOLO11 backbone network is used as the basic structure. The C3k2 feature extraction module in the original YOLO11 backbone network is replaced with the EAFE module. In a preferred embodiment, the total number of replacements is four, and the output of the replaced EAFE maintains scale and channel compatibility with the original C3k2 output. The output of the replaced backbone network... , , Multi-layer feature maps are used for subsequent multi-scale fusion and head regression in the neck network.

[0054] Step 3_2 EAFE module and edge-aware multi-directional scanning mechanism (e.g.) Figure 4 , Figure 5 (As shown)

[0055] The EAFE module enhances explicit modeling of tooth structure and lesion boundaries. Its core is the Edge-Aware Multi-Directional Scanning Unit (EAMDS), which extracts features in conjunction with conventional convolutional branches. Let the input feature map be... Preset angle set ; with horizontal Sobel kernel With vertical Sobel kernel As the base core, the structural angle is Multidirectional edge kernel The formula is as follows:

[0056]

[0057] in The input feature map is compared using the edge kernels corresponding to each angle in the preset angle set. Perform channel-wise convolution (achieved by depthwise convolution or group convolution) to obtain edge response features in each direction. The edge response in each direction can be expressed as:

[0058]

[0059] Introducing trainable directional weight parameters The orientation weights are obtained by normalization using the Softmax function. The formula is as follows:

[0060]

[0061] in The number of directions and in this embodiment The edge response features in each direction are weighted and fused to obtain the fused edge features. The formula is as follows:

[0062]

[0063] Parallel configuration of regular convolution branches for input features Extracting contextual features The preferred method is to obtain the result through a 3×3 convolution, as shown in the following formula:

[0064]

[0065] Merge edge features With context features The splicing feature is obtained by splicing along the channel dimension. The formula is as follows:

[0066]

[0067] right Perform channel compression and feature fusion, and then perform pointwise convolution to obtain the fused features. The formula is as follows:

[0068]

[0069] Then Input features By adding the residual connections, pointwise convolution is performed to integrate the features, and the output is... To enhance the representation of boundary details and directional structure while maintaining the stability of the main feature distribution, the formula is as follows:

[0070]

[0071] EAFE has two structural forms: directly embedded EAFE and C3k nested EAFE. In directly embedded EAFE, EAMDS directly serves as the bottleneck unit within the EAFE (e.g., Figure 4 As shown); in the C3k nested EAFE, the bottleneck unit within the EAFE is the C3k_EAMDS submodule, and the C3k_EAMDS, except that its bottleneck unit is replaced by EAMDS, has the same structure as the original YOLO11 structure's C3k submodule (as shown). Figure 5 (As shown). In this embodiment, among the four EAFE replacement locations in the backbone network, the first two EAFEs adopt a directly embedded structure, and the latter two EAFEs adopt a C3k nested structure; the locations and structures of each EAFE are as follows. Figure 2 As shown.

[0072] Step 3_3: Neck network construction and multi-scale adaptive feature fusion module fix two locations (e.g.) Figure 2 , Figure 6 (As shown)

[0073] The neck network is used to fuse features from different levels to balance shallow details and deep semantics. In this embodiment, two multi-scale adaptive feature fusion modules are fixedly set in the neck network to enhance multi-scale information flow and small target detection capabilities; the first module selects features from the backbone network output. , , Three layers of feature maps are used as input: respectively for... Execute the adaptive downsampling branch ADown, for Execute the channel transformation branch Conv, for Perform channel compression and upsampling (UP) to achieve scale alignment and obtain aligned features. , , The formula is as follows:

[0074]

[0075]

[0076]

[0077] The scale-aligned three-layer features are concatenated along the channel dimension to obtain the fused feature. The formula is as follows:

[0078]

[0079] Fusion features Multiple sets of depthwise separable convolution (DWConv) branches with different kernel sizes are executed in parallel to construct diverse receptive fields. The set of kernel sizes is as follows: The four dimensions in the equation are 5, 7, 9, and 11 respectively; the kernel size is denoted as... The output of the depthwise convolution branch is ,but:

[0080]

[0081] Connect the outputs of each branch with... Aggregation is performed to obtain aggregated features The preferred method is element-wise summation, and its aggregation formula is as follows:

[0082]

[0083] Aggregation features Performing pointwise convolution to integrate features yields ; and then with Perform residual summation output To enhance the characterization ability of lesions of different scales and morphologies, the formula is as follows:

[0084]

[0085]

[0086] The second multi-scale adaptive feature fusion module selects the output feature maps of the previous module and the two C3k2 structures in the neck network as multi-scale inputs to enhance multi-scale feature interaction.

[0087] Step 3_4 Head detection head

[0088] The detection head receives multi-scale feature maps output by the neck network, classifies candidate targets and performs bounding box regression, outputs lesion category probabilities and bounding box parameters, and can combine confidence threshold filtering and non-maximum suppression (NMS) to obtain the final set of detection boxes.

[0089] 4. Model Training Phase

[0090] The enhanced training set obtained in step 2 was used to iteratively train DENTAL-NET. During training, the input size was 640×640 (automatically processed by the YOLO training framework), the number of training epochs was 300, the batch size was 16, and the learning rate and other optimization parameters were set by default in the YOLO training framework. During training, the optimal weight parameters were selected as the final model based on the validation set.

[0091] To verify the effectiveness and deployability of the method of this invention, this embodiment selects several mainstream object detection models as comparison methods, including two-stage detection models, classic one-stage YOLO series models, and Transformer-based detection models. Detection accuracy and inference efficiency are compared and evaluated under the same input size, hardware platform, and evaluation settings. The preferred accuracy evaluation metric is mAP@0.5, and the preferred efficiency evaluation metrics are parameter quantity (Params), computational cost (GFLOPs), and inference speed (FPS), where FPS is the average frame rate obtained statistically from forward inference on the validation set. The dataset used is the publicly available Dental OPG XRAY Dataset, collected from three well-known private clinics in Dhaka, Bangladesh. The dataset contains 232 adult panoramic X-ray images (OPG), labeled with six categories: dental caries (DC), infection (IF), impacted teeth (IT), fractured teeth (FT), and crown / root damage (BDC). The comparison results are shown in Table 1.

[0092] As shown in Table 1, under the validation set and evaluation settings of this embodiment, the DENTAL-NET of the present invention achieves superior results on the mAP@0.5 metric, indicating that it has good lesion detection accuracy and localization effect under the stated dataset conditions. Further, in conjunction with the technical solution of the present invention, the lesion visibility enhancement, EAFE multi-directional edge scanning, and multi-scale adaptive feature fusion module help enhance boundary details and multi-scale representation, thereby reducing false negatives and false positives to a certain extent and improving the stability and usability of the detection results. At the same time, the model of the present invention still has a high FPS while maintaining a small number of parameters and computational load, indicating that the present invention has advantages in balancing detection performance and inference efficiency, and can be used in computer-aided screening applications of panoramic oral radiographs. It should be noted that the results shown in Table 1 are used to illustrate the technical effects of the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0093] Table 1

[0094] Model mAP@0.5 Params(M) GFLOPs FPS Faster-RCNN 78.9 28.6 102 8.6 YOLOv5 72.7 2.5 7.1 21.6 YOLOv8 69.7 3.0 8.1 22.7 YOLO11 75.6 2.5 6.3 24.3 DEIM-D-FINE 77.4 4.0 7.0 20.3 DEIM-RTDETRv2 81.6 20 60 13.8 DENTAL-NET 82.8 2.6 7.8 22.8

[0095] 5. Model Inference Stage

[0096] After preprocessing the panoramic oral radiographs to be tested according to the method in step 1, they are input into the detection model obtained in step 4. The detection model outputs the category information of dental lesions and their localization box results, and can output confidence scores for clinical reference. The candidate boxes are filtered based on the confidence threshold and NMS to obtain the final detection results, which are used for clinical initial screening, follow-up and auxiliary screening reference, thereby reducing the burden of repetitive image reading and improving screening efficiency.

[0097] Optional implementation methods and equivalent replacements:

[0098] Without altering the core idea of ​​this invention, the following equivalent substitutions or adjustments can be made:

[0099] (1) The LAug ratio can be adjusted within the range of 0.05 to 0.30;

[0100] (2) The preset angle set can be expanded to include a set containing more than two directions;

[0101] (3) The number of EAFE replacements of C3k2 can be expanded to multiple locations on the basis of "at least one location". In this embodiment, four locations are preferred. In a preferred embodiment, the multi-scale adaptive feature fusion module in the neck network can be set to two locations.

[0102] (4) Multiple EAFEs in the backbone network can be independently adopted as directly embedded EAFEs or C3k nested EAFEs, and the two structural forms can be combined and configured in the same backbone network; in a preferred embodiment, some EAFEs adopt the directly embedded structure to enhance shallow edge details, and some EAFEs adopt the C3k nested structure to enhance deep directional structural characterization.

[0103] The above embodiments are preferred embodiments of the present invention. Any equivalent transformations or substitutions made based on the technical solutions of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for multi-directional, multi-scale dental lesion detection using panoramic oral radiographs, characterized in that, The method includes the following steps: Step 1: Data preparation stage. Obtain panoramic oral radiographs, filter data and construct datasets, perform standardized preprocessing on panoramic oral radiographs, annotate dental lesions with bounding boxes to form labeled data, and divide the samples into training set, validation set and test set. Step 2: Data augmentation stage. Data augmentation is performed on the training set. The data augmentation includes basic augmentation and lesion visibility enhancement. Specifically, a portion of the training samples are selected from the training samples according to a preset ratio to perform lesion visibility enhancement processing. The lesion visibility enhancement processing includes contrast enhancement processing and grayscale inversion transformation. Step 3: Model Construction Stage. A dental lesion detection network, DENTAL-NET, is constructed. DENTAL-NET is a one-stage target detection network designed for dental lesion detection on panoramic radiographs. An Edge-Aware Feature Extraction (EAFE) module is set in the backbone network of DENTAL-NET. The core of EAFE is an Edge-Aware Multi-Directional Scanning (EAMDS) unit. EAMDS is used to generate multi-directional edge responses and perform directional weight fusion to enhance the directional structural representation of tooth and lesion boundaries. Furthermore, a multi-scale adaptive feature fusion module is introduced into the neck network of DENTAL-NET to enhance multi-scale feature interaction and the representation ability of lesions at different scales. Step 4: Model training phase. The DENTAL-NET constructed in step 3 is trained using the training set obtained in step 1 and the enhanced training samples obtained in step 2 to obtain the dental lesion detection model. The optimal weights are selected as the final model based on the validation set. Step 5: Model inference stage. Input the panoramic oral radiograph to be detected into the detection model obtained in step 4, and output the category information of the dental lesion and its localization box result.

2. The method according to claim 1, characterized in that, The standardization preprocessing in step 1 includes the following steps: Step 1_1 Adjust the panoramic dental radiograph to the preset input size; Step 1_2: Normalize or standardize the image pixel values.

3. The method according to claim 1, characterized in that, In step 2, the preset ratio for selecting a portion of the training samples from the training samples to perform lesion visibility enhancement processing is 0.05 to 0.

30.

4. The method according to claim 1, characterized in that, The contrast enhancement process in step 2 is achieved using Contrast Limited Adaptive Histogram Equalization (CLAHE).

5. The method according to claim 1, characterized in that, The grayscale inversion type transformation in step 2 includes negative transformation (NegTrans), as shown in the following formula: in, For pixels Pixel value at that location, These are the transformed pixel values. This is a constant representing the upper limit of grayscale; when the image is an 8-bit image. .

6. The method according to claim 1, characterized in that, The preset angle set in step 3 is as follows: The multi-directional edge operator group consists of multi-directional edge operators that correspond one-to-one with the preset angle set.

7. The method according to claim 6, characterized in that, The multi-directional edge operator consists of a horizontal Sobel kernel. With vertical Sobel kernel The structure is obtained by combining elements based on angular rotation, as shown in the following formula: in, For the preset angle set The angle in the middle, For angle Corresponding multi-directional edge kernels; and employing For the input feature map Perform channel-wise convolution to obtain edge response features at the corresponding angles. The formula is as follows:

8. The method according to claim 1, characterized in that, The EAMDS includes the following steps: Step 8_1: Convolve the input feature map using the multi-directional edge operator group to obtain the edge response features in each direction; Step 8_2 Stack the edge response features in each direction according to the directional dimension to form a multi-directional edge feature tensor; Step 8_3 Introduce trainable orientation weight parameters and normalize the orientation weight parameters to obtain the orientation weights; Step 8_4: Weighted fusion of edge response features in each direction based on weights in each direction to obtain fused edge features; concatenating the fused edge features with the context features extracted by the convolutional branch in the channel dimension to obtain concatenated features; performing pointwise convolution on the concatenated features to obtain fused features, and then adding the residuals of the fused features and the input features before integrating and outputting them through pointwise convolution.

9. The method according to claim 1, characterized in that, The DENTAL-NET replaces one or more C3k2 feature extraction modules in the YOLO11 backbone network with an edge-aware feature extraction module EAFE; the EAFE includes the EAMDS; wherein the EAFE has one of the following two structural forms: In the first structural configuration, the EAMDS directly serves as the bottleneck unit within the EAFE; In the second structural form, the bottleneck unit within the EAFE is the C3k_EAMDS submodule, and the C3k_EAMDS has the same structure as the original C3k submodule in YOLO11, except that its bottleneck unit is replaced by the EAMDS.

10. The method according to claim 1, characterized in that, The multi-scale adaptive feature fusion module is fixedly installed in two locations in the neck network, namely a first multi-scale adaptive feature fusion module and a second multi-scale adaptive feature fusion module. The first multi-scale adaptive feature fusion module is located in the top-down upsampling feature fusion path of the neck network, and the second multi-scale adaptive feature fusion module is located in the bottom-up downsampling feature fusion path of the neck network. Furthermore, the first multi-scale adaptive feature fusion module includes: Step 10_1 Select from the backbone network output , , Three-layer feature maps are used as multi-scale inputs; Step 10_2 , , The three feature maps are upsampled or downsampled and then subjected to channel transformation to achieve scale alignment. Step 10_3: Concatenate the scale-aligned multi-scale features to obtain the fused features. ; Step 10_4: For the fusion features... Multiple depthwise convolutional branches with different kernel sizes are set up in parallel, with kernel sizes of 5, 7, 9, and 11 respectively; the outputs of each branch are then fused with the fused feature. After aggregation, feature integration is performed through pointwise convolution, and then combined with the fused features. Perform residual summation and output; The second multi-scale adaptive feature fusion module uses the output feature map of the first multi-scale adaptive feature fusion module and the output feature maps of the two C3k2 structures in the neck network as multi-scale inputs, and performs scale alignment, splicing and convolutional branch aggregation processing on the multi-scale inputs in the manner described in steps 10_2 to 10_4 to obtain the output.

11. The method according to claim 9, characterized in that, The backbone network is provided with multiple EAFEs, and the multiple EAFEs include EAFEs adopting the first structural form and EAFEs adopting the second structural form.

Images