A method and device for positioning the base of inverted papilloma based on paranasal sinus CT image

Abstract

The application discloses a method and device for positioning a base of inverted papilloma in a nasal cavity based on a CT image of a nasal cavity. The method for positioning the base of the inverted papilloma in the nasal cavity based on the CT image of the nasal cavity comprises the following steps: acquiring image information of a patient to be predicted; acquiring a trained positioning model; inputting the image information of the patient to be predicted into the trained positioning model, so as to acquire a segmentation mask. The positioning model provided in the application can realize high-precision segmentation and positioning of an IP base region in a CT image, improve the ability of the model in multi-scale heterogeneous structure identification, key region focusing and noise robustness, and further provide reliable technical support for preoperative evaluation, surgical path planning and postoperative efficacy monitoring.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of image processing, in particular to a method and device for positioning a root base of an inverted papilloma based on a paranasal sinus CT image. BACKGROUND

[0002] An inverted papilloma (IP) is a benign epithelial tumor originating from the nasal cavity and paranasal sinuses, which has a high local invasiveness, recurrence rate, and potential risk of malignant transformation. Since the root base of the lesion is often deep in the hidden area of the paranasal sinus, accurate positioning of the root base in preoperative image assessment is a key step to achieve complete resection and reduce the recurrence rate. Currently, paranasal sinus CT images are mainly relied on for assessment, but the manual interpretation process is limited by the experience of physicians, and there are problems such as strong subjectivity, poor recognition consistency, and unstable positioning accuracy.

[0003] In recent years, medical image segmentation methods based on deep learning have been gradually applied to the auxiliary analysis of head and neck tumors such as IP. Among them, classic models such as U-Net and its variants have shown good generality in medical image segmentation tasks and can achieve a certain degree of automatic contour recognition. However, in the identification of the IP root base, which has a complex anatomical structure, blurred boundaries, and variable volume, there are still many challenges. The traditional single-scale input strategy makes it difficult for the model to balance local details and overall morphology when facing root base regions with significant size differences, often resulting in missed detection or inaccurate edge determination. In addition, the IP root base is often located near the mucosal folds or bone structures, with complex image gray scale distribution, and the traditional convolutional network has limited attention to key areas, which is easily disturbed by background tissues, leading to false positives or positioning deviations.

[0004] To improve the model's ability to identify key lesion areas, some studies have attempted to introduce attention mechanisms (such as SE-Block, CBAM, etc.) to enhance feature channels or spatial location responses, but most of them remain at a single scale or lack optimization for small target identification, failing to effectively address practical problems such as small target class imbalance and strong noise interference. At the same time, most existing methods use simple feature concatenation or weighted averaging for feature fusion, without considering the coupling strength and information complementarity between different branch features, which can easily introduce redundant features and affect the final segmentation accuracy.

[0005] In summary, the prior art has the following shortcomings in the positioning of the base of inverted papilloma: first, the joint modeling of different scale information is lacking, making it difficult to meet the segmentation needs of different volume base; second, the attention mechanism for key areas is limited, and the model is easily affected by background interference; third, the feature fusion strategy is still rough, and the complementarity between multi-source information has not been fully tapped. Therefore, an improved segmentation method with multi-scale input, spatial-channel joint attention guidance and efficient learnable feature fusion mechanism is needed to improve the automatic positioning accuracy of IP base in CT images and its clinical application value.

[0006] Therefore, it is desirable to have a technical solution to overcome or at least alleviate at least one of the aforementioned deficiencies of the prior art.

[0007] Summary of the application

[0008] Glossary:

[0009] IP(Inverted Papilloma)

[0010] Inverted papilloma, a benign tumor originating from the nasal cavity or sinuses, has local invasive and certain risk of malignancy. The IP base, which is the location where the lesion first attaches or grows, is a key area for preoperative evaluation and path planning.

[0011] CT(Computed Tomography)

[0012] Computed tomography, a technique for obtaining multi-layer cross-sectional images of the sinuses, has important value in the discovery, positioning and boundary identification of nasal tumors.

[0013] Multi-scale input(Multi-scale Input)

[0014] The same image is input into a deep neural network with different resolutions (such as 400x400, 200x200, etc.) to obtain feature expressions under different receptive fields, improving the model's perception of small targets and global structures.

[0015] Spatial attention mechanism(SpatialAttention Mechanism)

[0016] A mechanism for guiding the network to focus on key areas in the image, combining spatial location and feature response to generate a weight map, thereby improving the model's sensitivity to specific location areas.

[0017] Channel attention mechanism(ChannelAttention Mechanism)

[0018] Used to measure and adjust the importance of each feature channel in the current task, enhance the expression ability of high correlation features, and suppress redundant or irrelevant channels.

[0019] Hadamard Product

[0020] Refers to the operation of multiplying the corresponding elements of two matrices of the same dimension, used to enhance the high confidence area in multi-path feature fusion, commonly used in feature interaction and response enhancement link.

[0021] Joint Loss Function

[0022] In the present invention, it refers to the loss function design of weighted combination of binary cross entropy (BCE) and Dice coefficient loss, which takes into account the optimization of classification accuracy and target boundary overlap.

[0023] Cosine Annealing Scheduler

[0024] A learning rate dynamic adjustment strategy, periodically reduces the learning rate as the training process progresses, prevents the model from falling into local optimum, and speeds up the convergence speed.

[0025] Gated Mechanism

[0026] In the present invention, it is used to control the contribution degree of different feature paths in the fusion process, automatically adjust the information flow intensity through a learnable weight vector, and improve the fusion flexibility and adaptability.

[0027] Image Processing System Device

[0028] Refers to the image recognition and processing platform developed in the present invention for actual deployment, including image input, preprocessing, model inference and result display module, supporting automatic recognition and visual positioning of IP root base in clinical operation process.

[0029] The present application aims to provide a method for positioning the root base of inverted papilloma based on CT images of nasal sinuses to overcome or at least alleviate at least one of the aforementioned deficiencies of the prior art.

[0030] To achieve the above-mentioned purpose, the present application provides a method for positioning the root base of inverted papilloma based on CT images of nasal sinuses, which comprises:

[0031] Obtaining patient image information to be predicted;

[0032] obtaining a trained positioning model;

[0033] inputting the to-be-predicted patient image information into the trained positioning model, so as to obtain a segmentation mask.

[0034] Optionally, the positioning model comprises:

[0035] an input preprocessing module, which comprises four parallel processing channels, and is configured to process the obtained to-be-predicted patient image information into a first independent resolution image matrix, a second independent resolution image matrix, a third independent resolution image matrix, and a fourth independent resolution image matrix;

[0036] a feature extraction module, which is configured to extract features of the first independent resolution image matrix, the second independent resolution image matrix, the third independent resolution image matrix, and the fourth independent resolution image matrix respectively, so as to obtain a first channel feature map, a second channel feature map, a third channel feature map, and a fourth channel feature map respectively;

[0037] a spatial attention enhancement module, which is configured to perform spatial attention enhancement on the first channel feature map, so as to obtain a first enhanced feature map;

[0038] a channel attention enhancement module, which is configured to perform channel enhancement on the second channel feature map, the third channel feature map, and the fourth channel feature map respectively, so as to obtain a second channel enhanced feature map, a third channel enhanced feature map, and a fourth channel enhanced feature map;

[0039] a multi-scale fusion module, which is configured to fuse the first enhanced feature map, the second channel enhanced feature map, the third channel enhanced feature map, and the fourth channel enhanced feature map, so as to obtain a fusion feature;

[0040] an output module, which is configured to generate a segmentation mask according to the fusion feature.

[0041] Optionally, the feature extraction module comprises:

[0042] a first feature extraction branch, which comprises two layers of 3×3 convolution, each layer of which maintains 32 channels, and each layer of which is followed by BN+ReLU, and is configured to generate the first independent resolution image matrix, the size of the first independent resolution image matrix being 400×400;

[0043] The second feature extraction branch includes two layers of 3*3 convolution, each layer maintains 64 channels, and each layer is followed by BN+ReLU, and the second feature extraction branch is used to generate a second independent resolution image matrix, and the size of the second independent resolution image matrix is 200*200;

[0044] The third feature extraction branch includes two layers of 3*3 convolution, each layer maintains 128 channels, and each layer is followed by BN+ReLU, and the third feature extraction branch is used to generate a third independent resolution image matrix, and the size of the third independent resolution image matrix is 100*100;

[0045] The fourth feature extraction branch includes two layers of 3*3 convolution, each layer maintains 256 channels, and each layer is followed by BN+ReLU, and the fourth feature extraction branch is used to generate a fourth independent resolution image matrix, and the size of the fourth independent resolution image matrix is 50*50.

[0046] Optionally, the spatial attention enhancement module includes:

[0047] The feature adjustment module is used to optimize the first channel feature map, so as to obtain a first channel optimized feature map;

[0048] The parallel pooling module is used to perform parallel pooling on the first channel optimized feature map, so as to respectively obtain an average pooling feature and a maximum pooling feature;

[0049] The boundary detection module is used to obtain a single-channel boundary mask according to the to-be-predicted patient image information;

[0050] The spatial attention weight generation module generates a spatial attention weight map according to the average pooling feature, the maximum pooling feature and the boundary mask;

[0051] The spatial feature fusion module is used to fuse the spatial attention weight map and the first channel optimized feature map, so as to obtain a first enhanced feature map.

[0052] Optionally, the spatial attention weight generation module generates a spatial attention weight map through the following formula:

[0053] A s (x,y)=σ(W s *[AvgPool(F),MaxPool(F),M]);

[0054] Wherein, A s(x, y) is a spatial attention weight map; σ is a Sigmoid activation function; AvgPool(F) is an average pooling feature; MaxPool(F) is a maximum pooling feature; M is a boundary mask; W s * is a convolution operation.

[0055] Optionally, the channel attention enhancement module comprises:

[0056] a global information extraction module, configured to obtain any one of the second channel feature map, the third channel feature map, and the fourth channel feature map as a to-be-processed feature map, and generate a corresponding channel description vector according to the to-be-processed feature map;

[0057] a channel attention weight calculation module, configured to generate a channel attention weight vector according to the channel description vector;

[0058] an activation re-scaling module, configured to generate a channel attention weight with upper and lower limit constraints according to the channel attention weight vector;

[0059] a weight application module, configured to generate a channel enhancement feature map according to the to-be-processed feature map and the channel attention weight with upper and lower limit constraints;

[0060] a channel feature fusion module, configured to fuse the channel weighted feature map and the first enhancement feature map, so as to obtain a channel enhancement feature map corresponding to the to-be-processed feature map; wherein the second channel feature map corresponds to a second channel enhancement feature map; the third channel feature map corresponds to a third channel enhancement feature map; and the fourth channel feature map corresponds to a fourth channel enhancement feature map.

[0061] Optionally, the channel attention weight with upper and lower limit constraints is obtained by the following formula:

[0062] A c (i) = β + (γ - β) · σ(W2 · ReLU(W1 · GAP(F)));

[0063] wherein A c (i) is the channel attention weight with upper and lower limit constraints of the i-th channel; β is an activation re-scaling lower limit; γ is an activation re-scaling upper limit; σ is a Sigmoid activation function; W1 is a first convolution weight matrix; W2 is a second convolution weight matrix; GAP is a global average pooling; and F is a channel attention weight vector.

[0064] Optionally, the multi-scale fusion module adopts the following formula for fusion:

[0065]

[0066] wherein F fused is the fusion feature; F i is the input feature of the i-th scale; w i is the fusion coefficient; θi is a learnable gating vector; θj is the gating parameter of the j-th scale; e is the base of natural logarithm.

[0067] The application also provides a device for positioning a root base of an inverted papilloma based on a CT image of a nasal sinus, which comprises:

[0068] A to-be-predicted patient image information acquisition module is configured to acquire to-be-predicted patient image information.

[0069] A positioning model acquisition module is configured to acquire a trained positioning model.

[0070] A segmentation mask acquisition module is configured to input the to-be-predicted patient image information into the trained positioning model, so as to acquire a segmentation mask.

[0071] The positioning model provided by the application is an improved deep neural network model which fuses a multi-scale feature extraction mechanism, a space-channel joint attention mechanism and a learnable feature fusion structure, and is matched with a complete automatic positioning device system, and is aimed at realizing high-precision segmentation and positioning of an IP root base region in a CT image, improving the ability of the model in multi-scale heterogeneous structure recognition, key region focusing and noise robustness, and then providing reliable technical support for preoperative assessment, surgical path planning and postoperative efficacy monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0072] Figure 1 FIG. 1 is a flowchart of a method for positioning a root base of an inverted papilloma based on a CT image of a nasal sinus according to an embodiment of the application.

[0073] Figure 2 FIG. 3 is a structure diagram of a multi-scale attention fusion deep network provided by the application.

[0074] Figure 3 FIG. 5 is a segmentation comparison result of the model of the application in the Unclear / Clear-cut lesion case.

[0075] Figure 4 FIG. 7 is a comparison of various indicators of the model on the test set. DETAILED DESCRIPTION

[0076] For the purpose of the embodiments of the present application, the technical solutions and advantages, the technical solutions of the embodiments of the present application will be described in more detail below in combination with the drawings in the embodiments of the present application. In the drawings, the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The described embodiments are part of the embodiments of the present application, not all embodiments. The embodiments described below by reference to the drawings are exemplary and are intended to explain the present application, and cannot be understood as limiting the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application. The embodiments of the present application will be described in detail below in combination with the drawings.

[0077] In the description of the present application, it should be understood that the terms "center", "longitudinal", "transverse", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the purpose of facilitating the description of the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the scope of protection of the present application.

[0078] As shown in the method for positioning the base of inverted papilloma based on the CT image of the nasal cavity, the method comprises: Figure 1 Obtaining image information of a patient to be predicted;

[0079] Obtaining a trained positioning model;

[0080] Inputting the image information of the patient to be predicted into the trained positioning model to obtain a segmentation mask.

[0081] In the present embodiment, the positioning model comprises:

[0082] An input preprocessing module, the input preprocessing module comprises four parallel processing channels, for processing the obtained image information of the patient to be predicted into a first independent resolution image matrix, a second independent resolution image matrix, a third independent resolution image matrix, and a fourth independent resolution image matrix;

[0083] A feature extraction module, the feature extraction module is used to extract the features of the first independent resolution image matrix, the second independent resolution image matrix, the third independent resolution image matrix, and the fourth independent resolution image matrix, respectively, to obtain a first channel feature map, a second channel feature map, a third channel feature map, and a fourth channel feature map, respectively;

[0084]

[0085] ​a spatial attention enhancement module, configured to perform spatial attention enhancement on the first channel feature map, so as to obtain a first enhanced feature map;

[0086] a channel attention enhancement module, configured to perform channel enhancement on the second channel feature map, the third channel feature map and the fourth channel feature map respectively, so as to obtain a second channel enhanced feature map, a third channel enhanced feature map and a fourth channel enhanced feature map;

[0087] a multi-scale fusion module, configured to fuse the first enhanced feature map, the second channel enhanced feature map, the third channel enhanced feature map and the fourth channel enhanced feature map, so as to obtain a fused feature;

[0088] an output module, configured to generate a segmentation mask according to the fused feature.

[0089] In the embodiment, the feature extraction module comprises:

[0090] a first feature extraction branch, comprising two layers of 3*3 convolution, each layer maintaining 32 channels, and each layer being followed by BN+ReLU, the first feature extraction branch being configured to generate a first independent resolution image matrix, the size of the first independent resolution image matrix being 400*400;

[0091] a second feature extraction branch, comprising two layers of 3*3 convolution, each layer maintaining 64 channels, and each layer being followed by BN+ReLU, the second feature extraction branch being configured to generate a second independent resolution image matrix, the size of the second independent resolution image matrix being 200*200;

[0092] a third feature extraction branch, comprising two layers of 3*3 convolution, each layer maintaining 128 channels, and each layer being followed by BN+ReLU, the third feature extraction branch being configured to generate a third independent resolution image matrix, the size of the third independent resolution image matrix being 100*100;

[0093] a fourth feature extraction branch, comprising two layers of 3*3 convolution, each layer maintaining 256 channels, and each layer being followed by BN+ReLU, the fourth feature extraction branch being configured to generate a fourth independent resolution image matrix, the size of the fourth independent resolution image matrix being 50*50.

[0094] In the embodiment, the spatial attention enhancement module comprises:

[0095] a feature adjustment module, configured to optimize the first channel feature map, so as to obtain a first channel optimized feature map;

[0096] a parallel pooling module configured to perform parallel pooling on the first channel-optimized feature map to obtain average-pooled features and maximum-pooled features, respectively;

[0097] a boundary detection module configured to obtain a single-channel boundary mask based on the to-be-predicted patient image information;

[0098] a spatial attention weight generation module configured to generate a spatial attention weight map based on the average-pooled features, the maximum-pooled features, and the boundary mask;

[0099] a spatial feature fusion module configured to fuse the spatial attention weight map and the first channel-optimized feature map to obtain a first enhanced feature map.

[0100] In an alternative embodiment, the spatial attention enhancement module further includes a position encoding module configured to perform position encoding on the to-be-predicted patient image information to obtain a position embedding matrix; in this alternative embodiment, the position encoding module generates the position encoding of each pixel point by the following formula, wherein each pixel point includes a sine position encoding and a cosine position encoding.

[0101] The sine position encoding formula is as follows:

[0102]

[0103] The cosine position encoding formula is as follows:

[0104]

[0105] wherein x is the horizontal coordinate of the pixel point, 2i is the even-dimensional channel index, d is the total dimension of the position encoding, y is the vertical coordinate of the pixel point, 2i+1 is the odd-dimensional channel index, and i is the frequency control factor.

[0106] In this embodiment, the spatial attention weight generation module generates the spatial attention weight map by the following formula:

[0107] A s (x,y)=σ(W s *[AvgPool(F),MaxPool(F),M]);

[0108] wherein A s(x, y) is a spatial attention weight map; σ is a Sigmoid activation function; AvgPool(F) is an average pooling feature; MaxPool(F) is a maximum pooling feature; M is a boundary mask; W s is a convolution operation.

[0109] In this embodiment, the channel attention enhancement module comprises:

[0110] a global information extraction module, configured to obtain any one of the second channel feature map, the third channel feature map and the fourth channel feature map as a to-be-processed feature map, and generate a corresponding channel description vector according to the to-be-processed feature map;

[0111] a channel attention weight calculation module, configured to generate a channel attention weight vector according to the channel description vector;

[0112] an activation re-calibration module, configured to generate a channel attention weight with upper and lower limit constraints according to the channel attention weight vector;

[0113] a weight application module, configured to generate a channel enhancement feature map according to the to-be-processed feature map and the channel attention weight with upper and lower limit constraints;

[0114] a channel feature fusion module, configured to fuse the channel weighted feature map and the first enhancement feature map, so as to obtain a channel enhancement feature map corresponding to the to-be-processed feature map; wherein the second channel feature map corresponds to a second channel enhancement feature map; the third channel feature map corresponds to a third channel enhancement feature map; and the fourth channel feature map corresponds to a fourth channel enhancement feature map.

[0115] In this embodiment, the channel attention weight with upper and lower limit constraints is obtained by the following formula:

[0116] A c (i) = β + (γ - β) · σ(W2 · ReLU(W1 · GAP(F)));

[0117] wherein, A c (i) is the channel attention weight with upper and lower limit constraints of the i-th channel; β is the activation re-calibration lower limit; γ is the activation re-calibration upper limit; σ is the Sigmoid activation function; W1 is the first convolution weight matrix; W2 is the second convolution weight matrix; GAP is the global average pooling; and F is the channel attention weight vector.

[0118] In this embodiment, the multi-scale fusion module is fused by the following formula:

[0119]

[0120] where F fused is the fusion feature; F i is the input feature of the i-th scale; w i is the fusion coefficient; θi is the learnable gating vector; θj is the gating parameter of the j-th scale; and e is the base of the natural logarithm.

[0121] The present application will be further described in detail below by way of example, and it can be understood that the example does not constitute any limitation on the present application.

[0122] The embodiment discloses a deep neural network model with a fusion improved multi-scale feature extraction structure, an enhanced space-channel joint attention mechanism and a learnable gating fusion strategy, aiming to improve the precise segmentation and stable positioning ability of inverted papilloma (IP) root base in CT images. This method is particularly aimed at the difficulties such as small tumor root, fuzzy edge, large anatomical variability and the like in clinical practice, and realizes sensitive identification and boundary enhancement of key areas through systematic structural innovation, and is suitable for preoperative auxiliary diagnosis, path planning and postoperative monitoring and the like scenes.

[0123] The overall architecture of the model is shown in Figure 2 The input part adopts a four-resolution parallel branch structure, and inputs image matrices of 400x400, 200x200, 100x100 and 50x50 respectively to construct a multi-scale feature perception path. In each scale branch, a three-layer convolution stacking module is adopted, each layer is respectively set to have a channel number of 32, 64 and 128, a kernel size of 3x3 and a stride of 1, and the feature distribution is normalized through BN and ReLU activation function. This structure can effectively extract hierarchical representations from edge details to structural semantics, and further realize information backflow through a cross-scale skip connection mechanism to alleviate the influence of detail loss caused by low-scale input.

[0124] To significantly improve the attention ability of the model to the spatial region of the IP root base, an improved spatial attention module is introduced in each scale branch. The module introduces a position embedding item (PositionEncoding) based on the traditional AvgPool and MaxPool, and its two-dimensional form is defined as:

[0125]

[0126] The coordinate position is taken as a structural guide to enhance the position perception ability of the network to the boundary region. The attention weight generation formula is as follows:

[0127] A s (x,y)=σ(W s *[AvgPool(F),MaxPool(F),M])

[0128] where MMM is the boundary prior mask, which is generated by Canny algorithm preliminarily, σ is a Sigmoid function, * represents convolution operation, and the weight map As is multiplied with the original feature map element by element and then enters the subsequent fusion structure.

[0129] To enhance the adaptive modeling of inter-channel weights, an improved channel attention module is embedded in the decoding path, and an activation rescaling mechanism with upper and lower limits is adopted, and the formula is as follows:

[0130] A c (i) = β + (γ - β) · σ(W2 · ReLU(W1 · GAP(F)))

[0131] where β = 0.5 and γ = 1.5 are the channel weight limit interval, The compression ratio r = 16. The structure can prevent attention saturation and ensure stable weight distribution.

[0132] The feature fusion part between the multi-scale paths is one of the core innovations of the application. The element-by-element Hadamard product strategy with a gating weight is adopted, instead of simple concatenation, and a weight learning module is set to generate a fusion coefficient w i :

[0133]

[0134] where θi is a learnable gating vector that automatically adjusts the contribution strength of different scale paths. This strategy can improve the robustness and segmentation consistency of the model in complex structure regions and reduce false positives caused by differences in multi-path prediction.

[0135] In the output stage, two layers of upsampling and convolution operations (channels 128→64→1) are used, the output size is consistent with the input, and the output mask is obtained by Sigmoid activation to get the pixel-level probability map. In the difficult area of boundary discrimination, gradient-guided loss is used to enhance the edge fitting ability, and the total loss function is as follows:

[0136]

[0137] where λ1 = 1.0, λ2 = 1.0, and λ3 = 0.5, and the third term is the boundary gradient loss, which is used to constrain the edge similarity between the predicted graph and the real label.

[0138] In this embodiment, the training data used for training the positioning model of the application comes from actual cases (n = 186) approved by ethics, among which the IP root base is verified by surgery and pathology, and the training / validation / testing is divided according to 7:1:2. The image preprocessing adopts a multi-mode enhancement strategy, including rotation (±15°), Gaussian noise (σ = 0.02), affine stretching (±10%), etc., to simulate the real interference in the CT acquisition process.

[0139] The model training platform is an NVIDIA RTX 4080 Super GPU + Intel i9-14900 CPU, the training period is 40 epochs, the batch size is 8, the optimizer is Momentum-SGD (momentum factor 0.9), the initial learning rate is 1e-3, and cosine annealing scheduling is used to the lowest 1e-5. After each round of training, the IoU index is monitored for early stopping to prevent overfitting.

[0140] Referring to Figure 3 The experimental comparison results show that the model has high precision and high stability in "fuzzy lesion boundary" and "clear boundary" cases, with an average IoU improvement of 10.2% and a boundary F1 improvement of 12.6%, which has a significant advantage compared with traditional U-Net and DeepLabV3 (for details Figure 4 , the red part is the index of the present application).

[0141] The present application proposes a method for positioning the root base of inverted papilloma based on CT images of nasal sinuses, which is specifically used for precise positioning of the root base of inverted papilloma in CT images of nasal sinuses, and has significant structural innovation and clinical adaptability advantages.

[0142] In terms of image modeling, the present application designs four image input paths with different resolutions to process original structures, lesion adjacent regions, overall morphology and coarse-grained feature information respectively, covering full-scale features from high resolution to low resolution. This multi-scale input strategy can effectively improve the modeling ability of the model for small volume and significantly structural variation root base regions, and is particularly suitable for complex scenarios with fuzzy root base edges and various lesion morphologies. Compared with traditional single-scale convolution models, the present application has made obvious breakthroughs in target contour retention ability and spatial distribution understanding.

[0143] In terms of key region recognition ability, the present application proposes a space-channel bidirectional attention mechanism. The spatial attention module adaptively adjusts the response strength of each position on the feature map to guide the model to focus on the high-confidence lesion region, effectively suppressing background interference. The channel attention mechanism reweights each channel to enhance the attention to the feature dimensions highly relevant to root base recognition, thereby improving the discriminability of overall feature expression and the ability to compress irrelevant features. Spatial and channel attention work together to make the model have stronger local perception ability and semantic recognition ability.

[0144] In terms of feature fusion strategy, the application innovatively proposes a learnable weighted fusion mechanism to replace the traditional splicing or average fusion method. The fusion module automatically learns the importance of each scale feature at different spatial positions during model training, and uses this to guide the feature weighting process, thereby achieving optimal combination while maintaining the difference between each branch expression. This fusion method can strengthen the complementary cooperation between multi-scale features, improve the consistency of the model in identifying the root base area boundary, and reduce the risk of deviation or missed detection caused by single scale misjudgment.

[0145] In terms of training stability and performance optimization, the application uses a combination of loss functions designed for medical image characteristics, including Dice loss sensitive to boundary contours and cross-entropy loss stable for classification. Combined with cosine annealing scheduling and momentum acceleration strategy, the model has improved convergence speed, volatility control and final performance compared to existing solutions. In addition, various geometric disturbances and noise simulations are added to the data augmentation part, further enhancing the model's adaptability under different clinical acquisition conditions.

[0146] In experimental verification, the model showed stable high precision in multiple real case test sets, especially in images with blurred root base edges or obvious structure overlap, its intersection over union and error rate were significantly better than current mainstream segmentation frameworks, and had good clinical promotion potential.

[0147] The application focuses on the accurate positioning of the root base of inverted papilloma (IP) in nasal sinus CT images, and designs a deep learning model with fine structure perception ability, attention guiding mechanism and feature interaction enhancement ability. An image processing system device that can be clinically deployed is also developed. The application not only integrates multiple optimization strategies in the structure design, but also refines and improves the core computing path, fusion mechanism and information flow method, forming an integrated solution with advanced technology and conversion potential.

[0148] In terms of feature extraction path structure setting, the application uses a parallel multi-scale input strategy to simultaneously scale the same CT image to four typical sizes (400x400, 200x200, 100x100, 50x50) to simulate the hierarchical perception process of doctors when observing lesions from macroscopic overall appearance to detailed boundary. Each path is configured with fixed convolution layer depth and progressive channel number to ensure that the receptive field expands layer by layer while maintaining feature consistency. Cross-scale paths are connected through a jump connection and a reverse feature channel sharing mechanism to establish a bidirectional bridge for semantic backflow and detail compensation, thereby ensuring that the model still has stable recognition ability even in cases where the root base volume is small or the boundary is blurred. This resolution difference driven information fusion design constitutes the first technical core element of the application.

[0149] In terms of region perception and feature discrimination, the application introduces a spatial and channel joint attention mechanism, breaking the limitations of traditional single attention in expression dimension. The spatial attention module adds a position embedding mechanism when generating the weight map, enabling the model to adjust the perception center based on structural position coding in the spatial dimension, thereby actively focusing on the area where the lesion may occur. The channel attention module combines the statistical information of global average pooling and max pooling, introduces an activation re-labeling strategy, dynamically adjusts the feature importance in the channel dimension, effectively suppresses channel redundancy and background interference. The combined effect of the two makes the model have dual sensitivity to "where is the most important" and "which dimensions are the most critical" in the feature expression process, constituting the second technical key of the application.

[0150] The feature fusion mechanism, as the third technical focus of the application, innovatively uses a deep fusion method based on Hadamard product, replacing the conventional splicing or linear weighting method. This method takes element-by-element multiplication operation as the core, reconstructs the significance of each spatial position of the feature maps of multiple scales through learnable gating weights, strengthens the expression of regions where multiple scales produce high confidence responses in the same part, and suppresses the interference features of redundant paths. This mechanism not only improves the model's ability to capture complex targets such as fuzzy boundaries and small targets, but also effectively reduces the pseudo-differences caused by scale prediction differences, improving segmentation consistency and the accuracy of the final output.

[0151] In addition to network structure improvement, the application designs a combined loss function that fuses Dice and BCE in view of the characteristics of medical images, such as the scarcity of root base classes and the fuzziness of outlines. An edge guiding term is introduced to improve the model's attention to boundary positions. In addition, cosine annealing learning rate scheduling and momentum SGD optimizer are combined to strengthen the stability and convergence rate of model training. All structural innovations are systematically trained and verified on real IP case data confirmed by pathology, achieving significant improvements in terms of intersection over union, boundary F1, and loss convergence speed compared to traditional models.

[0152] Furthermore, to ensure the method's application in clinical processes, the application simultaneously designs an image processing device that can interface with existing PACS systems, integrating image acquisition, preprocessing, model inference, and three-dimensional visualization display modules. This supports doctors in preoperative path planning and postoperative evaluation by allowing them to call the model to complete automatic identification and positioning of IP root bases in real time, significantly improving operational efficiency and safety.

[0153] The application also provides a device for positioning the root base of inverted papilloma based on a CT image of a nasal sinus, which comprises a patient image information acquisition module, a positioning model acquisition module, and a segmentation mask acquisition module.

[0154] The to-be-predicted patient image information acquisition module is configured to acquire to-be-predicted patient image information;

[0155] The positioning model acquisition module is configured to acquire the trained positioning model;

[0156] The segmentation mask acquisition module is configured to input the to-be-predicted patient image information into the trained positioning model, so as to acquire a segmentation mask.

[0157] Finally, it should be pointed out that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for positioning the base of an inverted papilloma based on a CT image of the paranasal sinuses, characterized by, The method for positioning the base of inverted papilloma in the nose based on a CT image of a nasal sinus comprises: obtaining image information of a patient to be predicted; obtaining a trained positioning model, wherein the positioning model comprises: an input preprocessing module, which comprises four parallel processing channels, and is configured to process the obtained image information of the patient to be predicted into a first image matrix of an independent resolution, a second image matrix of an independent resolution, a third image matrix of an independent resolution, and a fourth image matrix of an independent resolution; a feature extraction module, which is configured to extract features of the first image matrix of the independent resolution, the second image matrix of the independent resolution, the third image matrix of the independent resolution, and the fourth image matrix of the independent resolution respectively, so as to obtain a first channel feature map, a second channel feature map, a third channel feature map, and a fourth channel feature map respectively; a spatial attention enhancement module, which is configured to perform spatial attention enhancement on the first channel feature map, so as to obtain a first enhanced feature map; a channel attention enhancement module, which is configured to perform channel enhancement on the second channel feature map, the third channel feature map, and the fourth channel feature map respectively, so as to obtain a second channel enhanced feature map, a third channel enhanced feature map, and a fourth channel enhanced feature map; a multi-scale fusion module, which is configured to fuse the first enhanced feature map, the second channel enhanced feature map, the third channel enhanced feature map, and the fourth channel enhanced feature map, so as to obtain a fusion feature; an output module, which is configured to generate a segmentation mask according to the fusion feature; inputting the image information of the patient to be predicted into the trained positioning model, so as to obtain the segmentation mask.

2. The method for positioning the base of inverted papilloma based on sinus CT images as claimed in claim 1, wherein, The feature extraction module comprises: a first feature extraction branch, which comprises two layers of 3x3 convolution, each layer of which maintains 32 channels, and each layer of which is followed by BN+ReLU, and is configured to generate the first image matrix of the independent resolution, wherein the size of the first image matrix of the independent resolution is 400x400; a second feature extraction branch, which comprises two layers of 3x3 convolution, each layer of which maintains 64 channels, and each layer of which is followed by BN+ReLU, and is configured to generate the second image matrix of the independent resolution, wherein the size of the second image matrix of the independent resolution is 200x200; a third feature extraction branch, which comprises two layers of 3x3 convolution, each layer of which maintains 128 channels, and each layer of which is followed by BN+ReLU, and is configured to generate the third image matrix of the independent resolution, wherein the size of the third image matrix of the independent resolution is 100x100; a fourth feature extraction branch, which comprises two layers of 3x3 convolution, each layer of which maintains 256 channels, and each layer of which is followed by BN+ReLU, and is configured to generate the fourth image matrix of the independent resolution, wherein the size of the fourth image matrix of the independent resolution is 50x50.

3. The method for positioning the base of inverted papilloma based on sinus CT images as claimed in claim 2, wherein, The spatial attention enhancement module comprises: The feature adjustment module is configured to optimize the first channel feature map to obtain a first channel optimized feature map. The parallel pooling module is configured to perform parallel pooling on the first channel optimized feature map to obtain an average-pooled feature and a maximum-pooled feature, respectively. The boundary detection module is configured to obtain a single-channel boundary mask based on the to-be-predicted patient image information. The spatial attention weight generation module is configured to generate a spatial attention weight map based on the average-pooled feature, the maximum-pooled feature, and the boundary mask. The spatial feature fusion module is configured to fuse the spatial attention weight map and the first channel optimized feature map to obtain a first enhanced feature map.

4. The method for positioning the base of inverted papilloma based on sinus CT images according to claim 3, wherein, The spatial attention weight generation module generates the spatial attention weight map based on the following formula: ； wherein, is a spatial attention weight map; The channel attention enhancement module includes: is a sigmoid activation function; AvgPool(F) is an average pooling feature; MaxPool(F) is a max pooling feature; M is a boundary mask; is a convolution operation.

5. The method for positioning the base of inverted papilloma based on sinus CT images as claimed in claim 4, wherein, The global information extraction module is configured to obtain any one of the second channel feature map, the third channel feature map, and the fourth channel feature map as a to-be-processed feature map, and generate a corresponding channel description vector based on the to-be-processed feature map. The channel attention weight calculation module is configured to generate a channel attention weight vector based on the channel description vector. The activation re-scaling module is configured to generate a channel attention weight with upper and lower limit constraints based on the channel attention weight vector. The weight application module is configured to generate a channel enhanced feature map based on the to-be-processed feature map and the channel attention weight with upper and lower limit constraints. The channel feature fusion module is configured to fuse the channel weighted feature map and the first enhanced feature map to obtain a channel enhanced feature map corresponding to the to-be-processed feature map; the second channel feature map corresponds to a second channel enhanced feature map; the third channel feature map corresponds to a third channel enhanced feature map; and the fourth channel feature map corresponds to a fourth channel enhanced feature map. The channel attention weight with upper and lower limit constraints is obtained based on the following formula:

6. The method for positioning the base of inverted papilloma based on sinus CT images as claimed in claim 5, wherein, The multi-scale fusion module is fused based on the following formula: ； wherein, is the channel attention weight with upper and lower bound constraints for the i-th channel; β is the activation re-scaling lower bound; The device for positioning the base of an inverted papilloma based on a CT image of a nasal sinus includes: is the activation re-scaling upper bound; and σ is the Sigmoid activation function; is the first convolution weight matrix; is the second convolution weight matrix; GAP is the global average pooling; and F is the channel attention weight vector.

7. The method for positioning the base of inverted papilloma based on sinus CT images as claimed in claim 6 wherein, A to-be-predicted patient image information acquisition module is configured to obtain to-be-predicted patient image information. ； wherein, is a fused feature; is an input feature of the i-th scale; is a fusion coefficient; A positioning model acquisition module is configured to obtain a trained positioning model; the positioning model includes: i is a learnable gating vector; An input preprocessing module includes four parallel processing channels and is configured to process the obtained to-be-predicted patient image information into a first independent resolution image matrix, a second independent resolution image matrix, a third independent resolution image matrix, and a fourth independent resolution image matrix. j is a gating parameter of the j-th scale; e is the base of the natural logarithm.

8. A device for locating the base of an inverted nasal papilloma based on sinus CT images, characterized in that, ​ ​ ​ ​ The feature extraction module is configured to extract features of the image matrix of the first independent resolution, the image matrix of the second independent resolution, the image matrix of the third independent resolution, and the image matrix of the fourth independent resolution respectively, so as to obtain the first channel feature map, the second channel feature map, the third channel feature map, and the fourth channel feature map respectively. The spatial attention enhancement module is configured to perform spatial attention enhancement on the first channel feature map, so as to obtain the first enhanced feature map. The channel attention enhancement module is configured to perform channel enhancement on the second channel feature map, the third channel feature map, and the fourth channel feature map respectively, so as to obtain the second channel enhanced feature map, the third channel enhanced feature map, and the fourth channel enhanced feature map. The multi-scale fusion module is configured to fuse the first enhanced feature map, the second channel enhanced feature map, the third channel enhanced feature map, and the fourth channel enhanced feature map, so as to obtain the fusion feature. The output module is configured to generate a segmentation mask according to the fusion feature. The segmentation mask acquisition module is configured to input the patient image information to be predicted into the trained positioning model, so as to obtain the segmentation mask.

Images