A whole body SPECT lung cancer bone imaging data processing method and system

Abstract

The application discloses a kind of whole body SPECT lung cancer bone imaging data processing method and system, it is related to image processing technical field, comprising: the whole body SPECT lung cancer bone imaging data is input into the improved UNet network after training, local texture feature is extracted by profile wave decomposition module, local texture feature is fused with original image, and the image feature is extracted by inputting fusion feature into UNet encoder;Using the UNet decoder of adding profile wave feature splicing and reconstruction strategy to the extracted image feature, obtains lung cancer bone metastasis lesion area.The application can improve the feature extraction capability of bone imaging lesion segmentation method under low resolution and fuzzy boundary condition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a whole body SPECT lung cancer bone imaging data processing method and system. BACKGROUND

[0002] As a non-invasive imaging technique, single photon emission computed tomography (SPECT) bone imaging can reflect bone metabolic activity through the distribution of tracers in the skeletal system, and has become one of the main means for current bone metastasis screening. However, due to the limited imaging resolution and high background noise of SPECT images, the lesion boundary is blurred, and it is easy to be confused with benign bone lesions, which affects the recognition accuracy.

[0003] In the prior art, for the automatic segmentation of bone metastasis lesions in bone imaging, various methods based on machine learning and deep learning have been proposed. Method one proposes a method based on BP neural network. This method performs image preprocessing by least squares method, uses shallow neural network for rough classification, and combines template matching and symmetry removal strategy to improve the accuracy of segmentation. However, since this method mainly relies on hand-designed features and shallow network modeling, it has limited recognition ability for complex texture features, blurred boundaries and small lesion areas in SPECT images, and is difficult to fully adapt to low imaging resolution and heavy noise interference in actual clinical data. And other researchers have proposed methods based on CNN, semi-supervised convolutional network, Efficient-BtrflyNet butterfly structure and Transformer-convolution hybrid network (Dbahnet). These methods have achieved high segmentation accuracy on positron emission computed tomography PET / CT, bone scan, µCT (micro-CT) and other medical image data, and can capture global features and local contour information of lesions to a certain extent. However, the above methods have the shortcomings of limited feature extraction mechanism, ignoring multi-scale directional texture information and insufficient targeted optimization.

[0004] In summary, how to improve the feature extraction capability of bone imaging lesion segmentation method under low resolution and blurred boundary conditions is an important problem to be solved. SUMMARY

[0005] The embodiments of the present application provide a whole body SPECT lung cancer bone imaging data processing method and system, which can improve the feature extraction capability of bone imaging lesion segmentation method under low resolution and blurred boundary conditions.

[0006] The embodiments of the present application provide a whole body SPECT lung cancer bone imaging data processing method, which comprises the following steps:

[0007] Obtaining whole body SPECT lung cancer bone imaging data;

[0008] The whole body SPECT lung cancer bone imaging data is input into the improved UNet network trained, the local texture features are extracted through the similar profile wave decomposition module, the local texture features are fused with the original image, and the fused features are input into the UNet encoder to extract image features; the UNet decoder with the similar profile wave feature splicing and reconstruction strategy is used to the extracted image features, and the segmentation map of the whole body SPECT lung cancer bone imaging data is obtained.

[0009] Further, the local texture features are extracted through the similar profile wave decomposition module, and the specific steps include:

[0010] The input whole body SPECT lung cancer bone imaging data is subjected to multi-level Laplacian pyramid decomposition, and high-frequency residual signals of different scales are extracted; the high-frequency residual signals of each layer are subjected to spectrum direction modeling using a learnable direction filter bank (DFB), and local texture features serving as multiple direction feature subbands are extracted.

[0011] Further, the UNet decoder with the similar profile wave feature splicing and reconstruction strategy is used to the extracted image features, and the specific steps include:

[0012] In the skip connection stage of the UNet, the image features extracted by the encoder are spliced with the feature subbands of the corresponding direction; the splicing result is input into the up-sampling stage of the decoder of the UNet, the splicing and reconstruction strategy of the similar profile wave feature is used, and the segmentation map of the whole body SPECT lung cancer bone imaging data is obtained.

[0013] Further, the whole body SPECT lung cancer bone imaging data is obtained, and the specific steps include:

[0014] The noise in the whole body SPECT lung cancer bone imaging data is removed, and the blank area is cut off; the non-target area is identified and cut off; the image is rotated around the geometric center by a set angle, and randomly translated by a set distance in the horizontal or vertical direction.

[0015] An embodiment of the present application provides a whole body SPECT lung cancer bone imaging data processing system, which comprises:

[0016] A data acquisition module is configured to acquire whole body SPECT lung cancer bone imaging data; a model construction module is configured to input the whole body SPECT lung cancer bone imaging data into the improved UNet network trained, extract local texture features through the similar profile wave decomposition module, fuse the local texture features with the original image, and input the fused features into the UNet encoder to extract image features; and a UNet decoder with the similar profile wave feature splicing and reconstruction strategy is used to the extracted image features, and a segmentation map of the whole body SPECT lung cancer bone imaging data is obtained.

[0017] The embodiment of the present application provides a whole body SPECT lung cancer bone imaging data processing method and system, compared with the prior art, the beneficial effects are as follows:

[0018] The improved UNet network extracts local texture features through a similar contour wave decomposition module, fuses the local texture features with the original image, and inputs the fused features into a UNet encoder to extract image features; a UNet decoder with a similar contour wave feature splicing and reconstruction strategy is used to the extracted image features, so as to obtain a segmentation map of the whole body SPECT lung cancer bone imaging data; the similar contour wave decomposition module and the similar contour wave feature splicing and reconstruction strategy belong to similar contour wave knowledge guidance, which can extract the contour information and edge details of the image, thereby improving the feature extraction capability of the UNet network under the condition of low resolution and fuzzy boundary. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A whole body data acquisition flowchart is provided for the embodiment of the present application;

[0020] Figure 2 A whole body data acquisition flowchart is provided for the embodiment of the present application;

[0021] Figure 3 A similar contour wave knowledge decomposition diagram is provided for the embodiment of the present application;

[0022] Figure 4 A similar contour wave knowledge decomposition diagram is provided for the embodiment of the present application;

[0023] Figure 5 A similar contour wave knowledge decomposition diagram is provided for the embodiment of the present application; DETAILED DESCRIPTION

[0024] In order to make the above-mentioned objects, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. In the following description, a large number of specific details are set forth in order to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the scope of the present application, therefore the present application is not limited to the specific embodiments disclosed below.

[0025] Referring to Figure 5 The embodiment of the present application provides a whole body SPECT lung cancer bone imaging data processing method, comprising the following steps:

[0026] Step one: acquiring whole body SPECT lung cancer bone imaging data.

[0027] Step two: input the whole body SPECT lung cancer bone imaging data into the improved UNet network trained, extract the local texture features through the contourlet-like decomposition module, fuse the local texture features with the original image, and input the fused features into the UNet encoder to extract image features; use the UNet decoder with contourlet-like feature splicing and reconstruction strategy to extract the image features, and obtain the segmentation map of the whole body SPECT lung cancer bone imaging data.

[0028] The specific scheme is as follows:

[0029] The application introduces a contourlet-like decomposition module (Contourlet-like Decomposition Module) at the front end of the UNet backbone network. The module decomposes the original SPECT bone scan image into feature subbands with different directions and frequency components through a learnable multi-scale and multi-direction filter bank, thereby extracting rich texture details and structural boundary information.

[0030] Secondly, the application fuses the multiple feature subbands obtained by decomposition with the original image as guided features input into the UNet encoder. This fusion method explicitly enhances the perception ability of the model to different scale and direction features, improving the learning effect of the encoder on small lesion areas and fuzzy boundaries.

[0031] Further, in the jump connection and decoder stage of the UNet, the application introduces a contourlet-like feature splicing and reconstruction strategy (Contourlet-like Feature Fusion), which integrates boundary-sensitive features and global context information at different scales through specific fusion operations (such as weighted splicing or attention mechanism fusion), enhancing the decoder's ability to restore small lesions in detail.

[0032] The overall model still adopts the encoder-decoder structure, maintaining the multi-scale feature extraction and layer-by-layer feature restoration advantages of UNet. On this basis, the contourlet-like guided decomposition and fusion module serves as an external knowledge auxiliary channel, significantly enhancing the segmentation ability of low-contrast, fuzzy boundaries, and small-scale lesions.

[0033] Through the above technical solutions, the application can effectively improve the segmentation accuracy of lung cancer bone metastasis lesions in SPECT bone imaging data, especially in early micro-lesion detection and complex background processing, showing higher robustness and accuracy, and providing reliable technical support for clinical automatic diagnosis.

[0034] 1. Data preparation.

[0035] The present application is directed to SPECT whole-body bone imaging data, and the data preprocessing steps specifically include data cleaning, data denoising, blank area cutting, and bladder area personalized cutting. See Figure 1 .

[0036] Bladder area personalized cutting (cutting of non-target areas):

[0037] Non-lesion hot spots (such as injection sites and the bladder) can significantly affect the accurate diagnosis of true lesions. The radioactivity drug concentration in the lesion area is much lower compared to non-lesion hot spots. The excessively high radiation level of non-lesion hot spots can cause the phenomenon of "large numbers covering small numbers", making it difficult to identify lesions, thus leading to missed diagnosis. To solve this problem, non-lesion hot spots are automatically removed in the study, and the processed images are included in the experimental data set. The bladder area personalized cutting process is shown in Figure 1 .

[0038] Data expansion: data rotation, data translation

[0039] Data rotation refers to the image being randomly rotated by a certain angle to the left or right around its geometric center.

[0040] Data translation refers to the image being randomly translated by +t or -t pixels in the horizontal or vertical direction.

[0041] 2. Model construction based on knowledge guidance.

[0042] 2.1 Model overview diagram.

[0043] As shown in Figure 2 , in the encoding stage, the number of channels outside the input (i.e., 256-2x1) is doubled using compound convolution (i.e., Conv3x3+Conv3x3), and the size of the feature map is halved using the pooling operation (i.e., MaxPooling). Conversely, in the decoding stage, the feature map is restored layer by layer using transposed convolution (i.e., transposed Conv).

[0044] The encoding process is used to learn image features, focusing on the target area step by step from large to small. Then the decoding process reconstructs the image to generate the final output by step-by-step restoration. This structure not only meets the requirements of image segmentation, but also effectively extracts and utilizes the multi-scale features within the image, thereby enhancing the segmentation performance. The core components of the model that enhance segmentation performance include class profile wave knowledge decomposition and fusion. These components will be described in detail in the following chapters.

[0045] 2.2 Class profile wave knowledge decomposition.

[0046] As shown in Figure 3 , the class profile wave knowledge decomposition diagram.

[0047] To enhance the modeling capability of the model on the structural details and directional texture in the lesion area, the application proposes a Contourlet-like directional frequency enhancement module, which simulates the "pyramid-directional filtering" structure in the traditional Contourlet transform. The module mainly includes two stages: high-frequency residual signal extraction (Laplacian Pyramid) and multi-directional response modeling (DFB). The first stage: Laplacian Pyramid decomposition (high-frequency residual extraction).

[0048] Given an input image , first apply a learnable Gaussian blur operator LG to obtain the low-frequency component:

[0049] .

[0050] Wherein, V l is the low-pass image of the first l layer, LG indicates a learnable Gaussian blur process, and then the high-frequency component of the layer is obtained by a residual manner W l :

[0051] .

[0052] Wherein, W l is the high-frequency component of the l layer, I l is the input image of the first l layer, V l is the low-pass image of the first l layer V l Subsequently, the low-pass image

[0053] .

[0054] Wherein, I l+1 is the input image of the next layer, DS(·) indicates a spatial down-sampling operation, V l indicates a low-pass image.

[0055] The above process can be recursively performed multiple times to form a multi-level Laplacian pyramid structure. With 4-level decomposition, the input image with original size of 888x256 is progressively processed to obtain multi-scale high-frequency residual signals with sizes of 444x128, 222x64, 111x32, and 56x16, respectively W 0 , W 1 , W 2 , W 3 .

[0056] The second stage: DFB direction filtering (frequency direction perception modeling).

[0057] For each layer of high-frequency residual signal W l , further spectrum direction modeling is performed through a learnable directional filter bank (DFB, Directional Filter Bank) to extract local texture features in each direction. Specifically, the directional filtering operation is defined as follows:

[0058] .

[0059] wherein represents the response result of the l th direction channel in the k th layer. Wherein F k is the learnable 3x3 convolution kernel of the k th direction.

[0060] Through the above DFB direction modeling, the model can identify the repeated structure, texture directionality, organization discontinuous edge and other complex local high-frequency details in the image, providing structural prior support for subsequent feature fusion.

[0061] 2.3 Contourlet knowledge guided module.

[0062] To effectively introduce the direction frequency information in the image and guide the network structure perception, a contourlet knowledge guided module is proposed, as shown in Figure 4 .

[0063] To enhance the texture and structure modeling capability of the encoder features, a direction perception feature fusion module is introduced in the shallow stage (enoder1, enoder2, enoder3). The specific process is as follows:

[0064] First, the directional response features are adjusted to the same spatial size as the encoder features E i through bilinear interpolation:

[0065] .

[0066] where, is the i-th layer encoder feature, is its corresponding directional response, H i ,W i is the feature map size, and Interp(·) denotes the bilinear interpolation operation.

[0067] .

[0068] .

[0069] where, view(·) denotes the dimension reorganization operation, B is the batch size, K is the number of directions, H i ,W i is the feature map size, is the number of channels allocated to each group.

[0070] Subsequently, the feature modulation is performed in a group-wise weighted modulation (group-wise multiplication fusion) manner:

[0071] .

[0072] where, is the learnable weight, which is used to adaptively adjust the importance of each direction.

[0073] The fused feature is restored to the original shape:

[0074] .

[0075] The effects of the present application are summarized as follows:

[0076] The present application is dedicated to developing a deep learning segmentation model based on profile-like wave knowledge guidance, which is used for automatic identification and segmentation of lesion regions in lung cancer bone scan (SPECT) images. In view of the problems of low resolution, fuzzy lesion boundary and high structural heterogeneity of SPECT images, the present application introduces a profile-like wave decomposition and fusion strategy, which significantly improves the learning ability of the model for fine-grained structural features and directional texture information.

[0077] Compared with the traditional convolutional neural network method directly based on the original map input, the feature guiding mechanism proposed in the application can effectively make up for the insufficient feature extraction caused by the limited imaging quality, and enhance the distinguishability of the lesion boundary and the detection accuracy of the small lesion area. The method not only improves the recognition accuracy and model robustness of lung cancer bone metastasis lesions, but also has good generalization potential and is suitable for other types of low-resolution medical image segmentation tasks. Overall, the segmentation framework proposed in the application has the following advantages:

[0078] (1) The class profile wave decomposition module is used to extract the structure and texture detail information in the input image in multiple scales and multiple directions, improving the perception ability of the model to the small lesion area.

[0079] (2) Through the class profile wave feature fusion strategy, the boundary details and global context information are strengthened at different levels, and the lesion reconstruction effect of the decoder stage is optimized.

[0080] (3) While keeping the simplicity of the basic UNet architecture, the segmentation performance and stability of the model under low contrast and complex background conditions are improved.

[0081] (4) The method has good scalability and can be applied to other low-resolution medical image (such as brain SPECT, liver SPECT, etc.) lesion segmentation tasks.

[0082] The application proposes a new automatic diagnosis model based on class profile wave guided feature learning for low-resolution SPECT bone scan images, effectively improves the segmentation accuracy of small-scale lesion areas and fuzzy boundaries, has strong robustness and wide adaptability, and can enhance the capture ability of structure information and fine-grained texture features without greatly increasing the complexity of the model.

[0083] The embodiment of the application provides a whole body SPECT lung cancer bone imaging data processing system, which comprises:

[0084] A data acquisition module is configured to acquire whole body SPECT lung cancer bone imaging data.

[0085] A model construction module is configured to input the whole body SPECT lung cancer bone imaging data into the improved UNet network trained, extract local texture features through the class profile wave decomposition module, fuse the local texture features with the original image, and input the fused features into the UNet encoder to extract image features; and use the UNet decoder with the class profile wave feature splicing and reconstruction strategy to obtain the segmentation map of the whole body SPECT lung cancer bone imaging data.

[0086] A specific implementation is as follows:

[0087] This embodiment discloses a method for processing whole-body SPECT bone scintigraphy data for lung cancer, the specific steps of which are as follows:

[0088] S1. Obtain whole-body SPECT bone scintigraphy data for lung cancer.

[0089] S2. Input the whole-body SPECT lung cancer bone scintigraphy data into the trained improved UNet network, extract local texture features through the contour-like decomposition module, fuse the local texture features with the original image, and input the fused features into the UNet encoder to extract image features; use the UNet decoder with the added contour-like feature splicing and reconstruction strategy on the extracted image features to obtain the lung cancer bone metastasis lesion area.

[0090] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for processing whole-body SPECT bone scintigraphy data for lung cancer, characterized in that, Includes the following steps: Acquire whole-body SPECT bone scintigraphy data for lung cancer; Whole-body SPECT lung cancer bone scintigraphy data are input into a trained improved UNet network. Local texture features are extracted through a contour-like decomposition module. These local texture features are then fused with the original image, and the fused features are input into the UNet encoder to extract image features. In the skip connection stage of UNet, the image features extracted by the encoder are concatenated with feature subbands in the corresponding directions, and the concatenation results are weighted and modulated group by group. The modulated features are then input into the upsampling stage of the UNet decoder to obtain the segmentation map of the whole-body SPECT lung cancer bone scintigraphy data.

2. The method for processing whole-body SPECT bone scintigraphy data for lung cancer as described in claim 1, characterized in that, The step of extracting local texture features through a contour wave decomposition module includes: Multi-level Laplacian pyramid decomposition was performed on the input whole-body SPECT lung cancer bone scintigraphy data to extract high-frequency residual signals at different scales. For each layer of high-frequency residual signal, a learnable directional filter bank (DFB) is used to model the spectral direction, and local texture features are extracted as multiple directional feature sub-bands.

3. The method for processing whole-body SPECT bone scintigraphy data for lung cancer as described in claim 1, characterized in that, The specific steps for acquiring whole-body SPECT bone scintigraphy data for lung cancer include: Noise was removed from the whole-body SPECT bone scintigraphy data for lung cancer, and blank areas were excised. Identify and cut off non-target areas; Rotate the image around the geometric center at a set angle, and then randomly translate it a set distance in the horizontal or vertical direction.

4. A system for processing whole-body SPECT bone scintigraphy data for lung cancer, characterized in that, include: The data acquisition module is used to acquire whole-body SPECT bone scintigraphy data for lung cancer. The model building module is used to input whole-body SPECT lung cancer bone scintigraphy data into the trained improved UNet network. Local texture features are extracted through a contour-like decomposition module, and the local texture features are fused with the original image. The fused features are then input into the UNet encoder to extract image features. In the skip connection stage of UNet, the image features extracted by the encoder are concatenated with the feature subbands in the corresponding direction, and the concatenation results are weighted and modulated group by group. The modulated features are then input into the upsampling stage of the UNet decoder to obtain the segmentation map of the whole-body SPECT lung cancer bone scintigraphy data.

Images