Pulmonary tuberculosis CT image recognition method and system and electronic equipment

Abstract

The invention relates to a pulmonary tuberculosis CT image recognition method and system and electronic equipment. The method comprises the following steps: acquiring lung CT image data; the method comprises the following steps: preprocessing lung CT image data, inputting the preprocessed lung CT image data into a backbone network to extract multi-scale features, and obtaining three feature maps output at different depths; three feature maps output at different depths are input into a feature diffusion pyramid module, and after two-stage multi-scale feature focusing and feature diffusion are performed in sequence, corresponding diffusion fusion features, semantic enhancement features and detail enhancement features are output; and taking the diffusion fusion feature, the semantic enhancement feature and the detail enhancement feature as input of a target detection head to obtain a pulmonary tuberculosis image detection result. According to the method, features of different depths are extracted, low-level detail information and high-level semantic information are integrated through two-level multi-scale feature focusing and feature diffusion, and the target detection capability in a feature fuzzy and high-noise region is effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tuberculosis detection, and in particular to a tuberculosis CT image recognition method and system and electronic equipment. BACKGROUND

[0002] Tuberculosis, as a chronic respiratory infectious disease that seriously endangers human health, is caused by Mycobacterium tuberculosis infection and is a major challenge to global public health. According to the relevant report of the World Health Organization, there are millions of new tuberculosis cases worldwide each year, and timely and accurate diagnosis of tuberculosis is crucial for effective treatment, control of transmission and improvement of patient prognosis. At present, CT image examination plays a key role in the diagnosis of tuberculosis, which can present the fine structure and lesion characteristics of the lung, helping doctors to identify various imaging manifestations of tuberculosis and providing key evidence for disease diagnosis. However, the manifestations of tuberculosis in CT images are extremely complex and diverse, and the lesion edges are prone to be blurred, and the imaging process is easily affected by artifacts, device noise and other factors, which brings great difficulty to accurate recognition. SUMMARY

[0003] Therefore, the main purpose of the present application is to provide a tuberculosis CT image recognition method, system and electronic equipment, which extracts features of different depths, and integrates low-level detail information and high-level semantic information through two-level multi-scale feature focusing and feature diffusion, effectively improving the target detection capability in feature fuzzy and high noise regions.

[0004] In a first aspect, the embodiments of the present application provide a tuberculosis CT image recognition method, which includes the following steps: obtaining lung CT image data; after preprocessing the lung CT image data, inputting the lung CT image data to a backbone network to extract multi-scale features, and obtaining three feature maps of different depths output; inputting the three feature maps of different depths output into a feature diffusion pyramid module, sequentially passing through two-level multi-scale feature focusing and feature diffusion, and outputting corresponding diffusion fusion features, semantic enhancement features and detail enhancement features; taking the diffusion fusion features, the semantic enhancement features and the detail enhancement features as inputs of a target detection head, and obtaining a tuberculosis image detection result.

[0005] In some embodiments, after preprocessing the lung CT image data, inputting the lung CT image data to a backbone network to extract multi-scale features, and obtaining three feature maps of different depths output, includes: preprocessing the lung CT image data; The preprocessed lung CT image data is input into a backbone network for multi-scale feature extraction, and low-level features, middle-level features and high-level features are output.

[0006] In some embodiments, when the preprocessed lung CT image data is input into the backbone network for multi-scale feature extraction, a feature enhancement module is used for feature enhancement processing.

[0007] In some embodiments, the three feature maps output at different depths are input into a feature diffusion pyramid module, and after two levels of multi-scale feature focusing and feature diffusion, corresponding diffusion fusion features, semantic enhancement features and detail enhancement features are output, including: After the low-level features, middle-level features and high-level features output at different depths are input into the feature diffusion pyramid module, the intermediate fusion features are output after one call of the feature focusing module for processing; The intermediate fusion features are respectively down-sampled and up-sampled, and after being fused with the high-level features and the low-level features, deep semantic features and detail features are obtained; The feature focusing module is called twice, and the intermediate fusion features, the deep semantic features and the detail features are input to construct diffusion fusion features; The diffusion fusion features are subjected to a double-path expansion operation, and after the final semantic enhancement features and detail enhancement features are obtained, they are output together with the expansion fusion features.

[0008] In some embodiments, after the low-level features, middle-level features and high-level features output at different depths are input into the feature diffusion pyramid module, the intermediate fusion features are output after one call of the feature focusing module for processing, including: After the low-level features, middle-level features and high-level features are input into the feature diffusion pyramid module, the original fusion features are obtained after one call of the feature focusing module for feature alignment and channel dimension splicing; The original fusion features are subjected to deep convolution, and the convolution path output features are stacked with the original fusion features and summed element by element to obtain cross-scale fusion features; The cross-scale fusion features are subjected to channel reconstruction and semantic compression to obtain the to-be-enhanced features, which are connected in residual with the original fusion features to obtain the intermediate fusion features.

[0009] In some embodiments, when the feature focusing module is called once for feature alignment: The low-level features are subjected to down-sampling operation for bidirectional compression of size and channel to obtain aligned low-level features; performing an up-sampling operation on the high-level feature, and performing channel compression through convolution to obtain an aligned high-level feature; performing channel adjustment on the middle-level feature according to a compression coefficient to obtain an aligned middle-level feature.

[0010] In some embodiments, the intermediate fusion feature is subjected to down-sampling and up-sampling operations, and is fused with the high-level feature and the low-level feature to obtain deep semantic features and detail features, including: performing a convolution and a down-sampling operation on the intermediate fusion feature, and generating deep semantic features after the obtained sampling result is fused with the high-level feature; performing an up-sampling operation on the intermediate fusion feature, and generating detail features after the obtained sampling result is fused with the low-level feature.

[0011] In some embodiments, the diffusion fusion feature is subjected to a double-path expansion operation, and after the obtained final semantic enhancement features and detail enhancement features are obtained, the diffusion fusion feature is output together, including: performing a convolution and a down-sampling operation on the diffusion fusion feature, and extracting final semantic enhancement features after the obtained sampling result is fused with the deep semantic features and the down-sampling output path again; performing an up-sampling operation on the diffusion fusion feature, and extracting final detail enhancement features after the obtained sampling result is fused with the detail features and the up-sampling path; outputting the semantic enhancement features and the detail enhancement features together with the diffusion fusion feature.

[0012] In a second aspect, the embodiments of the present application provide a pulmonary tuberculosis CT image recognition system, including: a data acquisition module configured to acquire pulmonary CT image data; a preprocessing module configured to input the pulmonary CT image data after preprocessing to a backbone network to extract multi-scale features and obtain three feature maps of different depths; a feature focusing diffusion module configured to input the three feature maps of different depths to a feature diffusion pyramid module, and sequentially pass through two levels of multi-scale feature focusing and feature diffusion to output corresponding diffusion fusion features, semantic enhancement features and detail enhancement features; a result generation module configured to input the diffusion fusion features, the semantic enhancement features and the detail enhancement features as inputs of a target detection head to obtain a pulmonary tuberculosis image detection result.

[0013] In a third aspect, an electronic device is provided, which includes a processor, a storage medium, and a computer program. The computer program is stored in the storage medium and, when executed by the processor, implements the tuberculosis CT image recognition method of any of the preceding aspects.

[0014] Technical effects of the present application: The present application extracts features of different depths, and integrates low-level detail information and high-level semantic information through two-level multi-scale feature focusing and feature diffusion, effectively improving the target detection capability in feature fuzzy and high noise areas. Specifically, the lung CT image data is obtained; after the lung CT image data is preprocessed, it is input into the backbone network to extract multi-scale features, and three feature maps of different depths are obtained; the three feature maps of different depths are input into the feature diffusion pyramid module, and after two-level multi-scale feature focusing and feature diffusion, the corresponding diffusion fusion features, semantic enhancement features and detail enhancement features are output; the diffusion fusion features, semantic enhancement features and detail enhancement features are taken as the input of the target detection head, and the tuberculosis image detection result is obtained. Therefore, through two-level multi-scale feature focusing and feature diffusion, not only the efficient interaction of shallow spatial information and deep semantic information is promoted, but also the spatial correlation between different scales is implicitly modeled, so that the diffusion fusion features, semantic enhancement features and detail enhancement features are input into the detection head, realizing the diffusion fusion of semantic information and the sufficient interaction of scale information, having stronger feature perception ability when processing fuzzy structure, and effectively improving the detection result precision.

[0015] Other features and aspects of the present disclosure will become apparent from the following detailed description of exemplary embodiments, taken in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0016] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate exemplary embodiments, features, and aspects of the present disclosure and serve to explain the principles of the present disclosure.

[0017] Figure 1 A flowchart of the tuberculosis CT image recognition method according to an embodiment of the present application is shown; Figure 2 A structure diagram of the feature diffusion pyramid module in the tuberculosis CT image recognition method according to an embodiment of the present application is shown; Figure 3 A structure block diagram of the tuberculosis CT image recognition system according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0018] Various exemplary embodiments, features, and aspects of this disclosure will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0019] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0020] Furthermore, to better illustrate this disclosure, numerous specific details are set forth in the following detailed description. Those skilled in the art will understand that this disclosure can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art have not been described in detail in order to highlight the main points of this disclosure.

[0021] like Figure 1 As shown, in a first aspect, embodiments of this application provide a method for recognizing CT images of pulmonary tuberculosis, comprising the following steps: S100, Acquire lung CT image data; In this step, lung CT image data refers to medical image data generated after a CT scan during tuberculosis detection, which includes density information of lung tissue and is stored in DICOM format.

[0022] S200. After preprocessing the lung CT image data, it is input into the backbone network to extract multi-scale features, resulting in three feature maps with different depths. In this step, the preprocessed lung CT image data is processed using a backbone network to extract multi-scale features, resulting in feature maps at three key scales. These include low-level feature P3, mid-level feature P4, and high-level feature P5. Low-level feature P3 contains the richest edge and texture details and has the highest resolution. High-level feature P5 has the strongest representational ability and the highest level of semantic abstraction, but has the lowest resolution. Mid-level feature P4 falls between the two, serving as an intermediate scale.

[0023] S300: Input the three feature maps output at different depths into the feature diffusion pyramid module. After passing through two levels of multi-scale feature focusing and feature diffusion, the corresponding diffusion fusion features, semantic enhancement features and detail enhancement features are output. In this step, a multi-scale feature diffusion pyramid module FFM is added between the backbone network and the detection head. By inputting the low-level feature P3, the middle-level feature P4 and the high-level feature P5 into the feature diffusion pyramid module FFM, the multi-scale of different semantic levels in the backbone network is fully fused, and the conflict between semantic information and spatial details is alleviated. Through two-level multi-scale feature focusing and feature diffusion, the corresponding diffusion fusion feature F4, the semantic enhancement feature F5 and the detail enhancement feature F6 are obtained, which not only promotes the efficient interaction of shallow spatial information and deep semantic information, but also implicitly models the spatial correlation between different scales.

[0024] S400, the diffusion fusion feature, the semantic enhancement feature and the detail enhancement feature are taken as inputs of the target detection head to obtain a pulmonary tuberculosis image detection result.

[0025] In this step, by taking the diffusion fusion feature F4, the semantic enhancement feature F5 and the detail enhancement feature F6 as inputs of the target detection head, diffusion fusion of semantic information and sufficient interaction of scale information are realized, which has stronger feature perception ability when processing boundary fuzzy structures and effectively improves the detection result precision.

[0026] In some embodiments, after the lung CT image data is preprocessed, it is input into the backbone network to extract multi-scale features, and three feature maps of different depths are obtained, including: The lung CT image data is preprocessed. The preprocessed lung CT image data is input into the backbone network for multi-scale feature extraction, and low-level features, middle-level features and high-level features are output. The backbone network is a YOLO v8 network.

[0027] In this embodiment, after the lung CT image data is obtained, the original DICOM format file is preprocessed, including image size adjustment, standardization, and de-noising and enhancement. Then it is input into the backbone network for multi-scale feature extraction. After a series of convolution operations, three feature maps of different depths are obtained, which are low-level feature P3, middle-level feature P4 and high-level feature P5.

[0028] It should be noted that the backbone network of this embodiment is based on YOLOv8 architecture. In order to improve the diversity of feature representation and the context modeling ability, a multi-scale feature diffusion pyramid model FFM is introduced between the YOLOv8 backbone network and the detection head. Specifically, the neck part of YOLOv8 is improved. Through the diffusion mechanism, the model can propagate features with rich context information from one scale to another scale, capture the spatial dependency between different structures, and further improve the understanding of the positional relationship between structures while maintaining spatial precision.

[0029] In some embodiments, the feature enhancement module is used for feature enhancement processing when the preprocessed lung CT image data is input into the backbone network for multi-scale feature extraction.

[0030] In this embodiment, the feature enhancement module is used for feature enhancement processing during the extraction of multi-scale features. Specifically, the feature enhancement module captures the semantic dependency and local detail changes of the target structure through a multi-scale local-global channel attention mechanism. On the basis of the original C2f of the YOLOv8 network, an improved module CAC2f that integrates a multi-scale channel attention mechanism is added. The MLCABottleneck structure is introduced to embed the multi-scale local-global attention mechanism MLCA into each Bottleneck branch to enhance the model's ability to perceive the blurred boundaries of weak targets.

[0031] It should be noted that the front part of the CAC2f module is consistent with the original C2f module. A 1 × 1 convolution is used to map the input channel c1 to an intermediate feature of 2c, which is then split into two branches. One branch is directly used as a residual path, and the other branch is sequentially transmitted to several MLCABottleneck structures with attention mechanisms. The internal structure of each MLCABottleneck maintains a standard residual connection, and an attention module MLCA is added after the main branch. Finally, the outputs of all branches are spliced, and a 1 × 1 convolution is used to restore the output channel number c2.

[0032] It should be noted that the MLCA module of this embodiment can achieve multi-scale top-down modeling, dual-channel attention paths, and fusion and re-scaling. For multi-scale top-down modeling, MLCA uses global average pooling to extract channel responses in the entire image range, models global semantic dependency, and uses local average pooling to extract local area context with a fixed window size to enhance the model's sensitivity to local structures. Through the dual-channel attention path, local attention weights and global attention weights are generated. Finally, the local attention and global attention are fused in a adjustable weighting manner, the channel attention weight after fusion is applied to the original feature map, the channel re-scaling and enhancement are realized, the discriminability of feature expression is improved, the response strength of the target region is enhanced, the model pays more attention to the channel information related to the target in the lung CT, and irrelevant information such as background noise is suppressed, so as to improve the ability to capture details and semantics.

[0033] In some embodiments, the three feature maps output at different depths are input into the feature diffusion pyramid module, and after two levels of multi-scale feature focusing and feature diffusion, the corresponding diffusion fusion features, semantic enhancement features, and detail enhancement features are output, including: The low-level features, the middle-level features and the high-level features output at different depths are input to the feature diffusion pyramid module, and after one call of the feature focusing module, intermediate fusion features are output; The intermediate fusion features are respectively down-sampled and up-sampled, and after being fused with the high-level features and the low-level features, deep semantic features and detail features are obtained; The feature focusing module is called twice, and the intermediate fusion features, the deep semantic features and the detail features are input to construct diffusion fusion features; After performing a double-path expansion operation on the diffusion fusion features, the final semantic enhancement features and the detail enhancement features are obtained, and the diffusion fusion features are output together.

[0034] In this embodiment, as shown in Figure 2 For the low-level features P3, the middle-level features P4 and the high-level features P5 input to the backbone network, the feature diffusion pyramid module FFM first calls the feature focusing module MSFM to fuse the three to form the first intermediate fusion features F1. Then the feature diffusion pyramid module FFM respectively performs down-sampling and up-sampling operations on the intermediate fusion features F1 to realize semantic enhancement and spatial completion, and after the down-sampling and up-sampling operations, the intermediate fusion features F1 are fused with the high-level features P5 and the low-level features P3, respectively, to obtain the deep semantic features F2 and the detail features F3. Further, the feature diffusion pyramid module FFM re-enters the deep semantic features F2, the intermediate fusion features F1 and the detail features F3 into the feature focusing module MSFM, i.e., the MSFM module is called twice, to construct diffusion fusion features F4 of a unified scale to enhance the feature bidirectional compensation and context reconstruction between different scales. Finally, based on the diffusion fusion features F4, a double-path expansion operation is performed to obtain the semantic enhancement features F5 and the detail enhancement features F6, which are output together with the diffusion fusion features F4 as the input of the target detection head to realize diffusion fusion of semantic information and sufficient interaction of scale information.

[0035] It should be noted that the feature diffusion pyramid module FFM of this embodiment aims to fully fuse multi-scale features of different semantic levels in the backbone network, alleviate the conflict between semantic information and spatial details, and improve the perception and reasoning ability of the model for complex spatial relationships through multi-scale diffusion, deep and shallow feature bidirectional compensation and context reconstruction.

[0036] In some embodiments, after the low-level features, the middle-level features and the high-level features output at different depths are input to the feature diffusion pyramid module, and after one call of the feature focusing module, the intermediate fusion features are output, including: After the low-level features, the middle-level features and the high-level features are input to the feature diffusion pyramid module, the feature focusing module is called once for feature alignment, and the original fusion features are obtained by concatenating in the channel dimension; The original fusion features are subjected to deep convolution, and the output features of each convolution path are stacked with the original fusion features and summed element by element to obtain cross-scale fusion features; The cross-scale fusion features are subjected to channel reconstruction and semantic compression to obtain to-be-enhanced features, which are connected in residual with the original fusion features to obtain intermediate fusion features.

[0037] In this embodiment, when the feature diffusion pyramid module FFM calls the feature aggregation module MSFM once, the efficient fusion and deep expression of multi-scale features are realized through the feature aggregation module MSFM, and the fine-grained local structure and high-level abstract semantics in different semantic levels are fully mined. Specifically, it includes feature alignment, feature fusion and feature enhancement. In the feature alignment stage, the feature focusing module MSFM receives three feature maps output by the backbone network at different depths, i.e. low-level features P3, middle-level features P4 and high-level features P5, and performs consistent alignment of spatial scales and channel numbers. Then the three aligned features are spliced in the channel dimension to construct the original fusion features with unified scale and dimension.

[0038] Further, in order to capture cross-scale context information, multi-scale deep separable dilated convolution is introduced, wherein each convolution path shares input features but independently models different spatial dependencies between features to enhance the perception ability of the model to different scale targets. Then the output features of each convolution path are stacked with the original fusion features and summed element by element to realize weighted fusion of cross-scale features, so as to enhance the diversity and robustness of feature expression.

[0039] Further, point-wise convolution is introduced to reconstruct and compress the semantics of cross-scale fusion features to enhance the dependency modeling ability between channels, and to-be-enhanced features are obtained. Moreover, in order to maintain information consistency, a residual connection mechanism is also introduced to add the original fusion features and the to-be-enhanced features element by element to obtain intermediate fusion features, realize lossless transmission of information and effective propagation of gradient flow, and further improve the generalization ability of the model.

[0040] In some embodiments, when the feature focusing module is called once for feature alignment: The low-level features are subjected to downsampling operation for bidirectional compression of size and channel to obtain aligned low-level features; The high-level features are subjected to upsampling operation and channel compression through convolution to obtain aligned high-level features; The middle-level features are subjected to channel adjustment according to the compression coefficient to obtain aligned middle-level features.

[0041] In this embodiment, the feature focusing module MSFM, in the feature alignment stage, in order to realize the consistent alignment of the spatial size and the number of channels, firstly restores the spatial resolution through upsampling for the high-level feature P5, and performs channel compression through a layer of 1x1 convolution; for the low-level feature P3, performs bidirectional compression of size and channel through downsampling; for the middle-level feature P4, performs channel adjustment according to the compression coefficient e∈(0,1], wherein the compression coefficient e of this embodiment is set to 1, keeping the feature unchanged, thereby obtaining three aligned features, and constructing the original fusion feature with unified scale and dimension according to the three aligned features.

[0042] In some embodiments, the intermediate fusion feature is subjected to downsampling and upsampling operations respectively, and after being fused with the high-level feature and the low-level feature, deep semantic features and detail features are obtained, including: The intermediate fusion feature is subjected to one convolution and downsampling operation, and the obtained sampling result is fused with the high-level feature to generate deep semantic features; The intermediate fusion feature is subjected to upsampling operation, and the obtained sampling result is fused with the low-level feature to generate detail features.

[0043] In this embodiment, after the first intermediate fusion feature F1 is formed, on the one hand, the intermediate fusion feature F1 generates deeper semantic features through one convolution and downsampling, and is fused with the high-level feature P5 output by the backbone to obtain deep semantic features F2 through the C2f structure; on the other hand, the intermediate fusion feature F1 is upsampled and fused with the low-level feature P3 output by the backbone, and further processed through the C2f structure to generate detail features F3. That is, the multi-scale features of different semantic levels in the backbone network are fully fused to alleviate the conflict between semantic information and spatial details.

[0044] In some embodiments, the diffusion fusion feature is subjected to double-path expansion operation, and after the final semantic enhancement features and detail enhancement features are obtained, they are output together with the expansion fusion feature, including: After the diffusion fusion feature is subjected to convolution and downsampling operation, the obtained sampling result is fused again with the deep semantic features and the downsampling output path, and the final semantic enhancement features are extracted; After the diffusion fusion feature is subjected to upsampling operation, the obtained sampling result is fused with the detail features and the upsampling path, and the final detail enhancement features are extracted and generated; The semantic enhancement features and the detail enhancement features are output together with the expansion fusion feature.

[0045] In this embodiment, in order to further enrich the semantic expression, a double-channel expansion operation is performed on the basis of the diffusion fusion feature F4, wherein one path generates new features through convolution and down-sampling, and the deep semantic features F2 and the down-sampling path output are fused again, and the final semantic enhancement features F5 are extracted through C2f. The other path performs up-sampling on the diffusion fusion feature F4, and after the fusion of the detail features F3 and the up-sampling path, the detail enhancement features F6 are extracted and generated through C2f. That is, through the diffusion and bidirectional fusion mechanism described above, the feature diffusion pyramid module FFM of the present embodiment not only promotes the efficient interaction of shallow spatial information and deep semantic information, but also implicitly models the spatial correlation between features of different scales. The diffusion fusion feature F4, the semantic enhancement feature F5 and the detail enhancement feature F6 constitute an enhanced feature pyramid, which is used as input to the detection head. After the feature maps of each scale are processed by the detection head, the final detection result is output.

[0046] In summary, in the method of the present application, the feature enhancement module captures the semantic dependence and local detail changes of the target structure through a multi-scale local-global channel attention mechanism, and the feature focusing module receives multi-scale input and captures cross-scale information through parallel deep convolution, so that the model can consider the relationship between different scales and effectively fuse low-level detail information and high-level semantic information, thereby enhancing the detail analysis capability of the image, especially in the case of noise and boundary blur. The feature diffusion pyramid enables the model to propagate features with rich context information from one scale to another, while maintaining spatial precision and improving the understanding of the positional relationship between structures.

[0047] As shown in Figure 3 In a second aspect, the embodiments of the present application provide a pulmonary tuberculosis CT image recognition system 10 for executing the above method, which comprises: A data acquisition module 100 configured to acquire pulmonary CT image data; A preprocessing module 200 configured to preprocess the pulmonary CT image data and input the preprocessed data to a backbone network to extract multi-scale features and obtain three feature maps output at different depths; A feature focusing diffusion module 300 configured to input the three feature maps output at different depths to a feature diffusion pyramid module, sequentially pass through two levels of multi-scale feature focusing and feature diffusion, and output corresponding diffusion fusion features, semantic enhancement features and detail enhancement features; A result generation module 400 configured to input the diffusion fusion features, the semantic enhancement features and the detail enhancement features as input of a target detection head, and obtain a pulmonary tuberculosis image detection result.

[0048] In a third aspect, the embodiments of the present application provide an electronic device, which comprises a processor, a storage medium and a computer program. The computer program is stored in the storage medium, and the computer program is executed by the processor to implement any of the above lung tuberculosis CT image recognition methods. For details, refer to the examples described in the above embodiments and optional implementation manners, which will not be repeated here.

[0049] In the description of the present application, the description of the terms "one embodiment", "an example", "a specific example" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the description of the present application, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0050] Unless otherwise defined, the technical terms or scientific terms involved in the present application should be understood as the general meaning of the technical terms or scientific terms understood by those skilled in the art to which the present application belongs. The terms "one", "a", "the" and the like similar words in the present application do not represent quantity limitation, but can represent singular or plural. The terms "include", "contain", "have" and any variations thereof involved in the present application are intended to cover non-exclusive inclusion. The terms "connect", "connect", "couple" and the like similar words involved in the present application are not limited to physical or mechanical connection, but include electrical connection, whether direct or indirect. The term "multiple" in the present application means two or more than two, and "and / or" describes the association between the associated objects, which means that there can be three kinds of relationships. The character " / " generally represents that the front and rear associated objects have an "or" relationship. The terms "first", "second", "third" and the like involved in the present application are only to distinguish similar objects, and do not represent the specific order of the objects.

[0051] The above embodiments only express several embodiments of the present application, which are described in detail and specifically, but cannot be understood as the limitation of the patent scope of the present application. It should be noted that for those skilled in the art, without departing from the concept of the present application, several modifications or improvements can be made, which are all within the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method for recognizing CT images of pulmonary tuberculosis, characterized in that, Includes the following steps: Acquire lung CT imaging data; After preprocessing the lung CT image data, it is input into the backbone network to extract multi-scale features, resulting in three feature maps with different depths. The three feature maps output at different depths are input into the feature diffusion pyramid module. After passing through two levels of multi-scale feature focusing and feature diffusion, the corresponding diffusion fusion features, semantic enhancement features, and detail enhancement features are output. The diffusion fusion feature, the semantic enhancement feature, and the detail enhancement feature are used as inputs to the target detection head to obtain the tuberculosis image detection result.

2. The method for recognizing CT images of pulmonary tuberculosis according to claim 1, characterized in that, The lung CT image data is preprocessed and then input into the backbone network to extract multi-scale features, resulting in three feature maps with different depths, including: The lung CT image data is preprocessed; The preprocessed lung CT image data is input into the backbone network for multi-scale feature extraction, and low-level features, mid-level features and high-level features are output; wherein, the backbone network is a YOLO v8 network.

3. The method for recognizing CT images of pulmonary tuberculosis according to claim 2, characterized in that, When the preprocessed lung CT image data is input into the backbone network for multi-scale feature processing, feature enhancement is performed using a feature enhancement module.

4. The method for recognizing CT images of pulmonary tuberculosis according to any one of claims 1-3, characterized in that, The three feature maps output at different depths are input into the feature diffusion pyramid module, and after passing through two levels of multi-scale feature focusing and feature diffusion, the corresponding diffusion fusion features, semantic enhancement features, and detail enhancement features are output, including: After inputting the low-level, mid-level, and high-level features output at different depths into the feature diffusion pyramid module, the feature focusing module is called once for processing, and the intermediate fused features are output. The intermediate fusion features are downsampled and upsampled respectively, and then fused with the high-level features and the low-level features to obtain deep semantic features and detail features. The feature focusing module is called a second time, and the intermediate fusion feature, the deep semantic feature, and the detail feature are input to construct the diffusion fusion feature; A dual-path expansion operation is performed on the diffusion fusion feature to obtain the final semantic enhancement feature and detail enhancement feature, which are then output together with the expanded fusion feature.

5. The method for recognizing CT images of pulmonary tuberculosis according to claim 4, characterized in that, After inputting low-level, mid-level, and high-level features output at different depths into the feature diffusion pyramid module, and processing them once by calling the feature focusing module, the intermediate fused features are output, including: After inputting the low-level features, the middle-level features, and the high-level features into the feature diffusion pyramid module, the feature focusing module is called once to perform feature alignment, and then the features are stitched together in the channel dimension to obtain the original fused features. The original fused features are subjected to deep convolution, and the output features of each convolution path are stacked with the original fused features. The summation is then performed element by element to obtain the cross-scale fused features. The cross-scale fusion features are reconstructed and semantically compressed to obtain the features to be enhanced. These features are then residually connected with the original fusion features to obtain intermediate fusion features.

6. The method for recognizing CT images of pulmonary tuberculosis according to claim 5, characterized in that, During a single call to the feature focusing module for feature alignment: A downsampling operation is performed on the low-level features to perform bidirectional compression of size and channels, resulting in aligned low-level features; An upsampling operation is performed on the high-level features, followed by channel compression via convolution to obtain aligned high-level features; The channels of the middle layer feature are adjusted according to the compression coefficient to obtain the aligned middle layer feature.

7. The method for recognizing CT images of pulmonary tuberculosis according to claim 4, characterized in that, The process involves performing downsampling and upsampling operations on the intermediate fusion features, and then fusing them with the high-level and low-level features to obtain deep semantic features and detail features, including: Perform a convolution and downsampling operation on the intermediate fusion feature, and the corresponding sampling result is cascaded and fused with the high-level feature to generate deep semantic features; An upsampling operation is performed on the intermediate fusion features, and the resulting sampling results are cascaded and fused with the low-level features to generate detailed features.

8. The method for recognizing CT images of pulmonary tuberculosis according to claim 4, characterized in that, The process of performing a dual-path expansion operation on the diffusion-fused features to obtain the final semantic enhancement features and detail enhancement features, and then outputting them together with the expanded-fused features, includes: After performing convolution and downsampling operations on the diffusion fusion features, the corresponding sampling results are fused again with the deep semantic features and the downsampling output path, and the final semantic enhancement features are extracted. After performing an upsampling operation on the diffusion fusion feature, the corresponding sampling result is fused with the detail feature and the upsampling path, and the final detail enhancement feature is extracted and generated. The semantic enhancement features, the detail enhancement features, and the extended fusion features are output together.

9. A CT image recognition system for pulmonary tuberculosis, characterized in that, include: The data acquisition module is configured to acquire lung CT image data; The preprocessing module is configured to preprocess the lung CT image data and then input it into the backbone network to extract multi-scale features, resulting in three feature maps with different depths. The feature focusing and diffusion module is configured to input the three feature maps output at different depths into the feature diffusion pyramid module, and after passing through two levels of multi-scale feature focusing and feature diffusion, output the corresponding diffusion fusion features, semantic enhancement features and detail enhancement features. The result generation module is configured to take the diffusion fusion feature, the semantic enhancement feature and the detail enhancement feature as input to the target detection head to obtain the tuberculosis image detection result.

10. An electronic device comprising a processor, a storage medium, and a computer program, wherein the computer program is stored in the storage medium, characterized in that, When the computer program is executed by the processor, it implements the CT image recognition method for pulmonary tuberculosis as described in any one of claims 1 to 8.

Images