Oral and maxillofacial segmentation method based on multi-scale kernel cross-band interactive attention fusion

Abstract

This invention belongs to the interdisciplinary field of medical image processing and artificial intelligence, and relates to a method for oral and maxillofacial bone segmentation based on multi-scale kernel cross-band interactive attention fusion. It involves acquiring hyperspectral images through a standardized hyperspectral oral and maxillofacial bone data acquisition platform and constructing a training dataset with pathological gold standard annotations. A multi-scale kernel cross-band interactive attention fusion segmentation model is constructed, integrating multi-scale feature extraction, hierarchical attention optimization, and multi-level complexity adaptation mechanisms. The model is trained based on the annotated dataset, and a converged hyperspectral oral and maxillofacial bone segmentation network is obtained through multi-scale supervision and loss function optimization. The trained model is deployed to clinical scenarios to segment hyperspectral oral and maxillofacial bone images and output results, providing support for diagnostic and treatment decisions. This invention effectively solves the problems of insufficient accuracy, limited deployment, and poor scenario adaptability in oral and maxillofacial bone segmentation, providing an efficient and reliable image analysis tool for oral and maxillofacial surgery.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of medical image processing and artificial intelligence. Specifically, it relates to a method for oral and maxillofacial bone segmentation based on multi-scale kernel cross-band interactive attention fusion. By designing a multi-scale kernel cross-band feature extraction module and a hierarchical attention fusion mechanism, it achieves efficient feature mining and accurate segmentation of hyperspectral oral and maxillofacial bone images. It can be adapted to portable hyperspectral devices during surgery and postoperative analysis workstations, providing quantitative basis for preoperative implant planning, intraoperative tumor boundary determination, and postoperative bone healing assessment in oral and maxillofacial surgery. It is especially suitable for clinical scenarios with limited resources. Background Technology

[0002] Oral and maxillofacial bone segmentation is a core supporting task for decision-making in oral and maxillofacial surgery, and its accuracy directly determines the effectiveness of postoperative assessment. In dental implant surgery, segmentation results are needed to clarify alveolar bone thickness and the trajectory of the inferior alveolar nerve canal to avoid nerve damage and determine implant placement depth. In jawbone tumor surgery, precise differentiation between tumor infiltration areas and normal bone tissue is required to achieve radical resection while maximizing the preservation of functional structures. In postoperative healing assessment, changes in bone regeneration area and density are quantified to determine whether the healing process meets clinical expectations. Therefore, precise segmentation of the oral and maxillofacial bone is not only crucial for improving treatment efficiency but also a core guarantee for reducing postoperative complications and improving patient prognosis. However, existing segmentation techniques are limited by their underlying principles and the characteristics of clinical scenarios, exhibiting multiple bottlenecks in accuracy, efficiency, and adaptability, making it difficult to meet complex clinical needs.

[0003] Currently, the mainstream oral and maxillofacial bone segmentation methods in clinical practice mainly rely on two types of techniques, but their shortcomings have gradually become apparent: one is manual segmentation, where oral and maxillofacial surgeons manually delineate the boundaries and key structures of the jawbone by combining CT / MRI images with clinical experience. This method heavily relies on the doctor's subjective judgment, which is not only inefficient but also results in poor consistency in annotation among different doctors, especially in ensuring the accuracy of identifying fine structures. The other is automatic segmentation based on traditional images, which uses CT / MRI as the data source and employs traditional algorithms such as threshold segmentation and region growing to achieve automatic segmentation. This type of method is limited by the contrast characteristics of the images themselves. Although CT can distinguish between bone tissue and soft tissue, it is difficult to distinguish the subtle differences between cortical bone and cancellous bone. MRI has high resolution for soft tissue but blurry imaging of bone tissue, resulting in segmentation results that cannot meet the needs of refined diagnosis and treatment (such as the determination of alveolar bone density stratification in dental implant surgery).

[0004] The rise of hyperspectral imaging technology has provided a new technical approach for the segmentation of the oral and maxillofacial bones. Compared to CT / MRI, hyperspectral imaging can capture spectral information in 30-64 bands, reflecting differences in the chemical composition and microstructure of tissues. For example, the spectral reflectance of normal cortical bone is significantly higher than that of cancellous bone in the 650nm band, and tumor infiltration areas show characteristic spectral shifts (decreased reflectance in the 620-680nm band) due to tissue necrosis. This "spectral fingerprint" provides a key basis for the accurate differentiation of different tissues in the jawbone. However, the application of hyperspectral technology in oral and maxillofacial segmentation still faces three major challenges: First, the disaster of data dimensionality. The high-band characteristics (dimensions 30-64) of hyperspectral data lead to a surge in data volume, making it difficult for traditional processing algorithms to effectively reduce dimensionality, resulting in overfitting and heavy computational burden (processing time for a single 512×512 image often exceeds 5 seconds). Second, poor scene adaptability. The oral surgical environment presents problems such as tissue displacement, instrument interference (e.g., surgical forceps, suction devices), and illumination fluctuations, which can lead to distortion of spectral features. Existing algorithms lack anti-interference mechanisms. Third, band redundancy and information waste. There is a strong correlation between hyperspectral bands (e.g., the spectral information overlap of adjacent bands exceeds 60%), which not only increases the computational load but also introduces noise, affecting segmentation accuracy.

[0005] From the perspective of algorithm development, current hyperspectral segmentation technology for the oral and maxillofacial bones can be mainly divided into two categories, but their limitations are quite obvious:

[0006] One category is traditional feature extraction and machine learning methods. These methods rely on manually designed spectral features (such as spectral mean and band variance) or spatial features (such as edge gradient and texture entropy), which cannot capture the joint "spatial-spectral" features of hyperspectral data. They are extremely poorly adapted to complex oral structures (such as the curved shape of the mandibular angle and the tortuous course of the nerve canal). At the same time, the expressive power of hand-designed features is limited, making it difficult to distinguish tissues with highly similar spectral features (such as early tumor areas and inflammatory areas). The segmentation accuracy is generally lower than 0.8, and the "curse of dimensionality" is prone to occur when processing high-dimensional data, which cannot meet the accuracy requirements of clinical practice.

[0007] Another category is deep learning methods, represented by UNet and its improved models (such as UNet++ and MobileUNet). Although traditional UNet can capture multi-scale features through the encoder-decoder structure, the number of parameters is as high as 5 million or more, and the computational cost exceeds 10 GFLOPs, making it impossible to deploy in portable intraoperative devices (such as NVIDIA Jetson embedded terminals and mobile snapshot hyperspectral imagers). To achieve lightweighting, models such as MobileUNet use depthwise separable convolutions to simplify the structure, but excessive compression leads to a decrease in feature extraction capabilities, especially for fine structures (such as the alveolar nerve canal with a diameter of only 2-3 mm), where the segmentation accuracy drops sharply (Dice coefficient is below 0.85), and the edge error exceeds 3 pixels. In addition, the attention mechanisms of existing deep learning models (such as CBAM and SE-Net) have obvious defects. They either rely on fully connected layers to compress channels, resulting in redundant parameters and an increase in computational cost of more than 30%, or they use 3×3 small convolutional kernels to capture spatial features, which cannot adapt to the irregular curved surface morphology of the oral and maxillofacial bones, resulting in insufficient spatial positioning accuracy and difficulty in accurately segmenting the boundary between tumor boundaries and normal bone tissue.

[0008] In summary, while oral and maxillofacial segmentation technology has evolved from manual to automated segmentation, existing methods still suffer from three major shortcomings: First, it is difficult to balance lightweight design with accuracy; traditional deep learning models cannot adapt to portable intraoperative devices, while lightweight models lack sufficient accuracy. Second, hyperspectral spatial-spectral features are not fully utilized; neither traditional methods nor existing deep learning models have designed dedicated feature fusion modules for the dimensional characteristics of hyperspectral data, making it impossible to effectively mine the correspondence between "spectral fingerprints" and anatomical structures. Third, clinical scenario adaptability is poor; it cannot be compatible with the single / multi-channel input formats of different hyperspectral devices and lacks mechanisms to resist intraoperative environmental interference. Therefore, there is an urgent need for a hyperspectral oral and maxillofacial segmentation technology that balances lightweight design, high accuracy, and strong scenario adaptability to overcome the bottlenecks in the clinical deployment of existing methods and provide an efficient and reliable image analysis tool for oral and maxillofacial surgery. Summary of the Invention

[0009] This invention addresses the shortcomings of existing technologies in oral and maxillary bone segmentation, such as the difficulty in balancing lightweight design and precision, insufficient utilization of hyperspectral spatial-spectral features, and poor adaptability to clinical scenarios. It proposes an oral and maxillary bone segmentation method based on multi-scale kernel cross-band interactive attention fusion, which solves the problem that existing technologies cannot meet the needs of portable device deployment and refined diagnosis and treatment.

[0010] To achieve the above-mentioned objectives, the technical solution adopted by the present invention is as follows:

[0011] The oral and maxillary bone segmentation method based on multi-scale kernel cross-band interactive attention fusion includes the following steps:

[0012] Step 1: Acquire hyperspectral images using a standardized hyperspectral oral and maxillofacial bone data acquisition platform and construct a training dataset with pathological gold standard annotations;

[0013] Furthermore, step 1 includes the following sub-steps:

[0014] Step 1.1: Set up a hyperspectral oral and maxillofacial bone data acquisition platform (spectral range 610-730nm, 9 bands, spatial resolution 1020×1020), and use a dental fixation frame to acquire images of the jawbone in vitro or during surgery. The image size is uniformly 256×256~512×512 pixels.

[0015] Step 1.2: Two or more associate chief physicians, in conjunction with CT / MRI and pathological slides, manually labeled the necrotic bone, normal bone, and background areas in the hyperspectral images. The consistency of the labeling was verified by the Kappa coefficient (Kappa≥0.85).

[0016] Step 1.3: Using fixed random seed sampling, the labeled dataset is divided into training set, validation set and test set in a ratio of 8:1:1. The experiment is repeated 3 times to verify the model's generalization ability.

[0017] Step 1.4: Preprocess the hyperspectral data, including standard reflector spectral correction and pixel normalization, to form a standardized training dataset.

[0018] Step 2: Construct a multi-scale kernel cross-band interactive attention fusion segmentation model, integrating multi-scale feature extraction, hierarchical attention optimization, and multi-level complexity adaptation mechanisms;

[0019] Furthermore, step 2 includes the following sub-steps:

[0020] Step 2.1: Design the core modules of the model, including: multi-scale kernel depth separable inverted residual blocks (integrating 1×1 / 3×3 / 5×5 convolutional kernels, reducing computation through "spectral channel expansion - spectral feature extraction - spectral channel compression", supporting summation / splitting fusion); cross-band grouped interactive attention gates (convolutions are designed according to groups, and features are dynamically selected across bands, reducing computation by 50%); hierarchical channel-spatial attention (CA-SA, CA adjusts the compression ratio according to the hierarchy, SA uses 7×7 convolutional kernels, and implements "spectral selection - spatial enhancement" in sequence); and multiple model families (-T / S / M / L, adapted to different hardware through channel number gradient).

[0021] Step 2.2: Construct the overall model architecture, including a data input layer, a multi-scale encoder (5 stages, including "multi-scale kernel residual block + 2×2 max pooling", which preserves intermediate features), an attention fusion layer (4 GAGs + 5 sets of CA-SA), a multi-scale decoder (5 stages, including "multi-scale kernel residual block + 2×2 bilinear interpolation"), and a multi-stage output layer (4 1×1 convolutional branches, out1~out3 for supervision, out4 for output results).

[0022] Step 2.3: Configure the basic parameters of the model. Set the input image size to 256×256 and the number of output categories to 3 (select three categories as needed).

[0023] Step 3: Train the model based on the labeled dataset, and obtain a converged hyperspectral oral and maxillofacial segmentation network through multi-scale supervision and loss function optimization;

[0024] Furthermore, step 3 includes the following sub-steps:

[0025] Step 3.1: Configure training parameters. The optimizer is AdamW (initial learning rate 1e-4, weight decay 1e-5). The loss function is a weighted sum of Dice loss and cross-entropy loss (1:1). The training epochs are set to 200 epochs. The batch size is adjusted according to the hardware (workstation Batch=16).

[0026] Step 3.2: Perform model training. The training set is augmented with random flipping, rotation, scaling, and Gaussian noise. The early stopping strategy is enabled (the Dice coefficient on the validation set stops if it does not improve after 5 rounds). Save the optimal weight file.

[0027] Step 3.3: Evaluate model performance using Dice and IoU as metrics. Ensure that Dice ≥ 92% and IoU ≥ 85% on the test set. If the criteria are not met, adjust the hyperparameters and retrain.

[0028] Step 4: Deploy the trained model to a clinical setting, segment hyperspectral oral and maxillofacial images and output the results to support diagnostic and treatment decisions.

[0029] Furthermore, step 4 includes the following sub-steps:

[0030] Step 4.1: Deploy the model to the clinical scenario, with embedded device deployment - T / S (TensorRT quantization acceleration) and workstation deployment - M / L (supporting high resolution / batch processing).

[0031] Step 4.2: Perform segmentation inference, acquire the image to be detected and automatically preprocess it, and zero-fill to standard size if the size is insufficient;

[0032] Step 4.3: Output the segmentation results and generate a color-coded map (differentiated colors distinguish tissues);

[0033] Step 4.4: Provide diagnostic and treatment support, including preoperative planning of the implantation path, intraoperative indication of tumor boundaries, and postoperative assessment of bone healing, forming a closed-loop support system throughout the entire cycle.

[0034] Compared with the prior art, the advantages of the present invention are as follows:

[0035] This invention combines hyperspectral spatial-spectral characteristics with clinical needs, achieving a synergy of lightweight design and high precision through multi-scale kernels and hierarchical attention:

[0036] (1) Extremely lightweight and multi-device adaptability: Model parameters are 86,000 to 1,200,000 (1 / 50 to 1 / 4 of traditional UNet), computation is 1.2 to 8 GFLOPs, and inference time on embedded devices is ≤0.5s / frame, solving the bottleneck of deployment during operation;

[0037] (2) In-depth utilization of hyperspectral spatial and spectral features: multi-scale kernel captures spectral differences, CA-SA optimizes the joint spatial and spectral information, and fine structure segmentation Dice≥85%, with an accuracy improvement of 12%~15% compared with existing lightweight models;

[0038] (3) Full clinical cycle adaptation: Supports single / multi-channel input, covering preoperative / intraoperative / postoperative scenarios, automatically outputs quantitative reports, and reduces subjective errors.

[0039] This invention can effectively solve the problems of "insufficient precision, limited deployment, and poor scene adaptability" in the existing oral and maxillofacial bone segmentation technology, and provide an efficient and reliable image analysis tool for oral and maxillofacial surgery. It provides doctors with objective evidence in dental implant surgery, jawbone tumor resection and other procedures, and ultimately improves patient prognosis. Attached Figure Description

[0040] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0041] Figure 1 Construct a flowchart for the dataset;

[0042] Figure 2 Flowchart for constructing a lightweight network based on multi-scale kernel cross-band interactive attention fusion;

[0043] Figure 3 A flowchart for the training and testing process;

[0044] Figure 4 This is a diagram of a lightweight network structure based on multi-scale kernel cross-band interactive attention fusion.

[0045] Figure 5 Here is a diagram of the encoder structure;

[0046] Figure 6 Diagram of multi-scale kernel cross-band interactive attention fusion structure;

[0047] Figure 7 Example of a test image;

[0048] Figure 8 Here is an example of the experimental results. Detailed Implementation

[0049] The present invention will be further described below with reference to examples and accompanying drawings. It should be noted that the embodiments do not limit the scope of protection claimed by the present invention.

[0050] like Figure 1 As shown, 100 oral and maxillofacial bone specimens (including normal jawbone, necrotic bone, and background) were selected. After medical ethics review, the specimen tissue blocks were fixed in a special oral bio-tray to ensure no tissue displacement or obstruction. A snapshot-type hyperspectral imaging platform (spectral range 610-730nm, 9 bands, spatial resolution 1020×1020) was used to acquire images. Three sets of duplicate images were acquired for each specimen to reduce random errors. Two oral and maxillofacial surgeons with the rank of associate chief physician or above manually labeled the regions in the hyperspectral images, combining CT (0.5mm slice thickness) and pathological section results. The consistency of the labeling was verified by the Kappa coefficient (Kappa≥0.88). A stratified sampling method with a fixed random seed (seed=42) was used to divide the dataset into a training set (80 cases), a validation set (10 cases), and a test set (10 cases) in an 8:1:1 ratio. The training set was used for model parameter learning, the validation set for hyperparameter tuning, and the test set for generalization ability evaluation.

[0051] like Figure 2 As shown, a lightweight hyperspectral oral and maxillofacial bone segmentation network was constructed. The entire network consists of 5 parts, as follows: Figure 4 The first part consists of the input layer and the dynamic band attention grouping module, used for effective screening and dynamic grouping of hyperspectral bands, solving the problems of spectral correlation breakage and redundant calculation caused by traditional fixed 3-channel grouping. First, the input 9-band hyperspectral image is preprocessed, and then core band extraction and adaptive grouping are achieved through dynamic band attention grouping. The core of band screening is calculating the information entropy of a single band (a measure of the effective feature quantity of a band), which can be expressed by the formula: ;in, Let the information entropy of the b-th band be... The pixel value within the band. This represents the probability distribution of pixel values ​​in this band (calculated using pixel histogram normalization). By setting a threshold H ≥ 0.6, the core bands with the Top-K information entropy (K = 3 / 6 / 9, adjusted according to hardware computing power) are selected. Subsequently, the Pearson correlation coefficient (measuring spectral correlation) between the core bands is calculated to avoid effective band splitting. The formula is: in, These are the i-th and j-th core bands, respectively. The covariance of the two bands, , ρ represents the standard deviation of the two bands. Bands with ρ≥0.8 are grouped into one group, and each group learns one band attention weight through a 1×1 convolution. After weighted summation, the weighted sums are compressed into a 3-channel feature map, and the final output is a 3-channel compatible grouped feature map, which provides high-quality input for the subsequent encoder.

[0052] The second part consists of the encoder's spectral-spatial dual-branch multi-core fusion block and adaptive band noise suppression module, used for deep extraction of hyperspectral spatial-spectral features and noise reduction. The encoder comprises five stages, such as... Figure 5 Each stage consists of an adaptive band noise suppression module, a spectral-spatial dual-branch multi-kernel fusion block, and a 2×2 max pooling module. The adaptive band noise suppression module estimates noise based on pixel variance and generates suppression weights using the following formula: ;in, The noise suppression weight for the c-th channel is... For learnable parameters, Let c be the pixel variance of the c-th channel (the larger the variance, the more noise). Multiplying the feature map channel by channel achieves "noise channel weakening and effective channel enhancement". The spectral-spatial dual-branch multi-kernel fusion block adopts a dual-branch structure: the spatial branch extracts spatial features through depthwise separable convolutions with 1×1, 3×3, and 5×5 kernels; the spectral branch captures spectral features through 1×1×3 and 1×1×5 kernels. The dual-branch fusion formula is as follows: ;middle, , Spatial and spectral weights learned for 1×1 convolution ( + =1), ensuring dynamic adaptation of spatial spectral features. Finally, 2×2 max pooling downsampling is performed, with the formula: ;in, After pooling The location pixel values ​​are pooled to reduce computation and focus on high-representation features. Encoders 1 to 5 output feature maps with sizes of 256×256, 128×128, 64×64, 32×32, and 16×16 respectively. At the same time, the features of encoders 1 to 4 are saved for subsequent fusion.

[0053] The third part consists of an attention fusion layer and cross-band interactive attention gates and channel-spatial attention, used for precise correlation and enhancement of encoder and decoder features, such as... Figure 6 This addresses the issue of traditional attention methods neglecting hyperspectral cross-band collaboration. First, cross-band association is constructed through a cross-band interactive attention gate: the channel cosine similarity matrix of encoder features is calculated using the following formula: ;in, Let be the similarity between channels i and j. For vector dot product, The L2 norm is used; subsequently, the decoder guiding features are processed by 3×3 grouped convolutions to obtain... encoder features Processed as The final result is obtained by multiplying the result by S. The attention weight formula is: ;

[0054] Next, key spectral channels are selected through layered enhancement: channel attention is used, with the following formula: ;

[0055] By employing dual attention synergy, a dual optimization of "cross-band correlation + spatial-spectral feature enhancement" is achieved.

[0056] The fourth part consists of the decoder and the multi-band, multi-scale adaptive output layer, used for feature map size restoration and multi-scale result adaptation to address the distortion problem caused by heterogeneous resolution in the hyperspectral bands. The decoder comprises five stages (decoder 1 to decoder 5). Each stage first restores the size using 2×2 bilinear interpolation, with the following formula: .

[0057] The fifth part is the segmentation output and clinical parameter quantification layer, which generates segmentation results and clinically usable quantitative indicators to provide a basis for diagnosis and treatment decisions. First, the segmentation output and clinical parameter quantification layer will be... Input a 1×1 convolution (output channels=3, corresponding to three types of regions: normal bone, necrotic bone, and background), activate it with softmax to obtain a category probability map, determine the pixel category according to the maximum probability, and generate a color-annotated map.

[0058] like Figure 3 The process of training the network model includes the following steps: First, the hyperspectral oral and maxillary bone dataset is input into the network, and the AdamW optimizer is used (initial learning rate 1e-4, weight decay 1e-5, cosine annealing decay). The loss function is a weighted sum of Dice loss and cross-entropy loss, and the formula is: ;in, For the prediction graph, It is the gold standard.

[0059] The process of training the network model includes the following steps: First, hyperspectral images of the oral and maxillofacial bones are input into the model, and a standardized training dataset (containing 80 specimens) is used for model training. The model is trained by calculating a weighted sum of Dice loss and cross-entropy loss as the loss function, performing gradient backpropagation and iterative optimization (using the AdamW optimizer, setting appropriate learning rates and training epochs) to continuously improve the model's segmentation accuracy and robustness. Finally, the parameters and weights that performed best on the validation set during training are saved. Next, the trained model is used to... Figure 7 The model was validated using a test dataset (containing data from 10 specimens). The Dice coefficient and mean boundary distance were evaluated in different regions (e.g., cortical bone, tumor region) to verify the model's generalization ability. Results are as follows: Figure 8 Finally, the trained models are deployed to embedded devices (such as NVIDIA Jetson Xavier NX) for use in real-world clinical scenarios such as preoperative planning and intraoperative navigation for oral and maxillofacial surgery.

[0060] The above process is merely a preferred embodiment of the present invention, and its practical application is not limited to this. The scope of protection of the present invention is not limited to the above examples; any technical solutions and modifications within the scope of the principles of the present invention should be included within the scope of protection of the present invention. For those skilled in the art, any reasonable improvements and adjustments made without departing from the core ideas of the present invention should also be considered as protected content of the present invention.

Claims

1. A method for oral and maxillary bone segmentation based on multi-scale kernel cross-band interactive attention fusion, characterized in that, Includes the following steps: Step 1: Acquire hyperspectral images using a standardized hyperspectral oral and maxillofacial bone data acquisition platform and construct a training dataset with pathological gold standard annotations; Step 2: Construct a multi-scale kernel cross-band interactive attention fusion segmentation model, integrating multi-scale feature extraction, hierarchical attention optimization, and multi-level complexity adaptation mechanisms; Step 3: Train the model based on the labeled dataset, and obtain a converged hyperspectral oral and maxillofacial segmentation network through multi-scale supervision and loss function optimization; Step 4: Deploy the trained model to a clinical setting, segment hyperspectral oral and maxillofacial images and output the results to support diagnostic and treatment decisions.

2. The oral and maxillary bone segmentation method based on multi-scale kernel cross-band interactive attention fusion according to claim 1, characterized in that, Step 1 includes the following sub-steps: Step 1.1: Build a hyperspectral oral and maxillofacial bone data acquisition platform with a spectral range of 610-730nm, 9 bands, and a spatial resolution of 1020×1020. Acquire images of the jawbone in vitro or during surgery using a dental fixation device. The image size is uniformly 256×256~512×512 pixels. Step 1.2: Two or more associate chief physicians manually annotate the necrotic bone, normal bone, and background areas in the hyperspectral images by combining CT / MRI and pathological slides. The consistency of the annotation is verified by the Kappa coefficient, Kappa≥0.85; Step 1.3: Using fixed random seed sampling, the labeled dataset is divided into training set, validation set and test set in a ratio of 8:1:

1. The experiment is repeated 3 times to verify the model's generalization ability. Step 1.4: Preprocess the hyperspectral data, including standard reflector spectral correction and pixel normalization, to form a standardized training dataset.

3. The oral and maxillary bone segmentation method based on multi-scale kernel cross-band interactive attention fusion according to claim 1, characterized in that, Step 2 includes the following sub-steps: Step 2.1: Design the core modules of the model, including: multi-scale kernel depth separable inverted residual blocks; cross-band grouped interactive attention gates; hierarchical channel-spatial attention; and a multi-level model family; Step 2.2: Construct the overall model architecture, including the data input layer, multi-scale encoder, attention fusion layer, multi-scale decoder, and multi-stage output layer; Step 2.3: Configure the basic parameters of the model. Set the input image size to 256×256 and the number of output categories to 3.

4. The oral and maxillary bone segmentation method based on multi-scale kernel cross-band interactive attention fusion according to claim 1, characterized in that, Step 3 includes the following sub-steps: Step 3.1: Configure training parameters. Select AdamW as the optimizer, use the weighted sum of Dice loss and cross-entropy loss as the loss function, set the training epochs to 200, and adjust the batch size according to the hardware. Step 3.2: Perform model training. The training set is augmented with random flipping, rotation, scaling, and Gaussian noise. The early stopping strategy is enabled, and the optimal weight file is saved. Step 3.3: Evaluate model performance using Dice and IoU as metrics. Ensure that Dice ≥ 92% and IoU ≥ 85% on the test set. If the criteria are not met, adjust the hyperparameters and retrain.

5. The oral and maxillary bone segmentation method based on multi-scale kernel cross-band interactive attention fusion according to claim 1, characterized in that, Step 4 includes the following sub-steps: Step 4.1: Deploy the model to the clinical setting, for embedded devices - T / S, for workstations - M / L; Step 4.2: Perform segmentation inference, acquire the image to be detected and automatically preprocess it, and zero-fill to standard size if the size is insufficient; Step 4.3: Output the segmentation results and generate a color-annotated map; Step 4.4: Provide diagnostic and treatment support, including preoperative planning of the implantation path, intraoperative indication of tumor boundaries, and postoperative assessment of bone healing, forming a closed-loop support system throughout the entire cycle.

Images