Image detection method for helicobacter pylori in digital slice
Through digital slice scanning and neural network super-resolution image processing methods, the problem of difficult detection of Helicobacter pylori in pathological sections was solved, and efficient and accurate Helicobacter pylori computer detection was achieved.
Patent Information
- Application Number
- CN202510756317.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-06
- Publication Date
- 2025-09-19
AI Technical Summary
In the existing technology, Helicobacter pylori is small in size and small in number in pathological sections, making it difficult to accurately detect through conventional microscopy and special staining methods, resulting in detection difficulties.
Image blocks are obtained by digital slice scanning, and super-resolution images are generated through a neural network-based super-resolution image processing method and a hierarchical deep learning framework for training target detection models and computer identification of Helicobacter pylori.
It improves the accuracy and efficiency of Helicobacter pylori detection, reduces the workload of manual microscopic screening, enhances the detailed features recognized by computers, and realizes comprehensive detection of Helicobacter pylori.
Smart Images
Figure CN120673405A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical image processing, and more particularly to an image detection method for Helicobacter pylori in digital slices. Background Art
[0002] Helicobacter pylori (HP) is a bacterium that thrives in the acidic environment of the stomach. HP infection in the human stomach often leads to chronic gastritis, gastric ulcers, and other diseases, and can increase the risk of gastric cancer. HP detection methods include breath tests and serum antibody tests. Visual observation of HP in pathological sections of biopsied tissue during endoscopic biopsy is a highly intuitive and accurate method. Therefore, HP detection in biopsies has important diagnostic value and clinical significance.
[0003] However, HP is approximately 0.5 to 1.0 μm wide and 2 to 5 μm long. Compared to the size of human cells (usually between 10 and 30 μm in diameter), HP is relatively small. At the same time, the number of HP in the gastric mucosa is relatively small, especially in the early stages of chronic inflammation, when only a small number of bacteria may be present, making manual observation under a conventional microscope very difficult. In addition, although special staining methods such as Giemsa staining or special silver staining can be used to find Helicobacter pylori in pathological sections, different staining methods have different visualization effects on bacteria, which may affect the interpretation of the results.
[0004] In summary, the detection of HP in pathological sections is of clinical significance. HP is small in size and small in number, making it difficult to observe and find under the microscope. Special staining is only partially helpful and is not sufficient to meet detection needs.
[0005] Therefore, an image detection method for Helicobacter pylori in digital slices is designed to overcome the above problems. Summary of the Invention
[0006] The object of the present invention is to overcome the shortcomings of the existing technology and provide an image detection method for Helicobacter pylori in digital slices. The method comprises scanning a pathological slice to obtain a digital slice, converting the digital slice into an image block as an original image, and processing the original image through the neural network-based super-resolution image processing method disclosed in the present invention. The super-resolution image thus obtained can provide more detailed features for target detection, and can also ensure that the super-resolution image obtained has the characteristic information of the original image, which is helpful for computer recognition and detection of HP. The present invention can comprehensively search for HP by detecting the complete digital slice, reducing the workload of step-by-step screening under manual microscopy.
[0007] The present invention is achieved through the following technical solution: a method for detecting Helicobacter pylori images in digital slices, comprising the following steps:
[0008] 1) Scanning the pathological sections with a digital slide scanner and converting the obtained digital slides into image blocks as original images;
[0009] 2) The original image in step 1) is processed using a neural network-based super-resolution image processing method and a hierarchical deep learning framework to obtain a super-resolution image. The super-resolution image has characteristic information of the original image and is used to provide detailed features for the trained target detection model, thereby facilitating computer recognition and detection of HP and reducing the workload of manual microscopic step-by-step screening. 3) Helicobacter pylori is detected using the target detection model trained in step 2) to obtain HP detection results.
[0010] Preferably, the digital slice in step 1) is a pathological slice of gastric mucosal biopsy tissue, which is complete and a full-scan image. The pathological slice is stained with HE staining, which facilitates image processing in step 2).
[0011] Preferably, the algorithm of the neural network-based super-resolution image processing method in step 2) specifically includes the following steps:
[0012] 1) Fourier transform is used to extract the high-frequency enhanced image of the image block, and the semantic features of the image block are extracted in parallel;
[0013] 2) extracting a pixel feature map of the high-frequency enhanced image using a convolutional neural network, and concurrently extracting a 2D morphological feature map using stacked staggered blocks;
[0014] 3) merging the semantic features in step 1) and the pixel feature map in step 2) to obtain gated pixel features, and concurrently merging the 2D morphological feature map and the semantic features to obtain gated morphological features;
[0015] 4) The gated pixel features in step 3) are combined with the gated features to obtain an embedded feature map, which is further calculated to obtain a super-resolution feature map, and finally a super-resolution image is generated.
[0016] Preferably, the target detection model trained by the hierarchical deep learning framework in step 2) is defined as HLIP, which extracts hierarchical features from the high-frequency enhanced image based on the intrinsic features of the pathological image, and then establishes a super-resolution image based on the local pathological image pattern.
[0017] Preferably, the specific steps of establishing the super-resolution image are as follows:
[0018] 1) First, the original image is converted into a high-frequency enhanced image to recover the high-frequency texture features that are highly relevant to pathological diagnosis, and used as the input of the HLIP model;
[0019] 2) For the input pathology image p with width W and height H, apply a two-dimensional Fourier transform to convert the pixel value p_i(x,y) of channel i in RGB space to frequency space:
[0020]
[0021] Where j is the imaginary unit and (x, y) is the coordinate of a single pixel;
[0022] 3) After frequency calibration, place the zero-frequency component at the edge of the feature map;
[0023] 4) By designing high-pass filters (HPFs) with different cutoff frequencies, the m k The mask is used to perform spectrum filtering to obtain the multi-level high-frequency characteristic spectrum G of each channel. i,k * (u,v):
[0024] G i,k * (u,v)=G i (u,v)⊙m k
[0025] Where ⊙ represents the Hadamard product of the matrix;
[0026] 5) Obtain the corresponding high-frequency spatial feature map h through inverse Fourier transform i,k (x,y):
[0027]
[0028] 6) Using the original image g i (x,y) and multiple high-frequency spatial feature maps h i,k (x,y) is superimposed to obtain the high-frequency enhanced image q:
[0029] q(x,y)=concat(g1(x,y),g2(x,y),…g c (x,y),h 1,1 (x,y),h 1,2 (x,y),…,h c,d (x,y))
[0030] Where c is the number of original channels, d is the number of high-frequency spatial feature maps;
[0031] 7) Using attention-based methods to integrate hierarchical features, including semantic features and extracting features using stacked convolutional neural networks;
[0032] 8) Using the general base model for computational pathology INI Extract the semantic features V of a single image block in the original image p ∈R W′×H′×F The model is pre-trained on the Mass-100K dataset and, in parallel, a single-layer CNN is used to extract V based on the enhanced image q. m ∈R W′×H′×F 2D morphological feature map, and further based on the pixel features, the stacked residual blocks with CNN backbone are used to extract V m ∈R W′×H′×F 2D morphological feature map;
[0033] v s =f UNI (p)
[0034] V p =CNN(p)
[0035] V m =StackedCNN(V p );
[0036] 9) Take the semantic feature as the query object and the 2D morphological feature map as the key value to obtain the gated morphological feature V m,gated , for the gated pixel feature V m,gated The same construction procedure is applied to the pixel feature map of :
[0037] V p,gated =rossAttention(query=v s ,key=V p ,value=V p )
[0038] V m,gated =CrossAttention(query=v s ,key=V m ,value=V m )
[0039] 10) Aggregate the gated morphological features and gated pixel features to obtain an embedded feature map M∈R with super resolution W×H×F :
[0040] M=V p,gated +V m,gated ;
[0041] Attention-based hierarchical feature integration is used to upsample local implicit image blocks to generate a super-resolution feature map Z∈R with a given magnification ratio δ. δW×δH×F, and obtain the final super-resolution image based on the feature map.
[0042] As a preferred embodiment, the specific steps of obtaining a super-resolution image based on the super-resolution feature map are as follows:
[0043] 1) Map M(a,b) to Z(δa,δb), where a∈{1,2,…,W} and b∈{1,2,…,H}, and resize the embedded feature map M to the size of Z, expressed as:
[0044] Z(δa,δb)=M(a,b)
[0045] 2) By taking the weighted sum of the features of the four nearest coordinates that are divisible by δ, the features corresponding to the other coordinates are generated. For example, for the feature z with coordinates (x, y) in the super-resolution feature map, x,y ∈R F , where x∈{1,2,…,δW} and y∈{1,2,…,δH}, get the coordinates of the nearest upper left corner (x * ,y * )as follows:
[0046]
[0047] 3) The four closest features: upper left M(x * ,y * )、Upper right M(x * +1,y * )、lower left M(x * ,y * +1) and the lower right M(x * +1,y * +1) as the weighted sum of spatial neighbors, the weight S is the target coordinate (x, y) and the opposite spatial neighbor coordinate (ie M(x * ,y * ) is weighted by (x,y) and (x * +1,y * +1); the super-resolution feature z x,y The calculation process is as follows:
[0048] d t =‖(x t ,y t )-(x,y)‖2
[0049]
[0050] Where (x t ,y t )∈{(x *,y * ),(x * +1,y * ),(x * ,y * +1),(x * +1,y * +1)}, f is a linear feature embedding layer.
[0051] 4) Use another linear layer with input and output dimensions of F and 3 respectively to convert the feature map Z into RGB pixels and obtain the super-resolution image p s ;
[0052] 5) During the training process, an adversarial strategy is used to carry out HLIP, by introducing a discriminator D θ (·), the discriminator consists of 9 convolutional neural networks, which can obtain the probability that each image belongs to the generated image. For the generated super-resolution image p s , discriminator loss L d It can be calculated as follows:
[0053] L d =-log(D θ (p s ))
[0054] Among them L d is the loss function used for discriminator optimization.
[0055] As an example, the target detection model uses an overall loss L consisting of pixel loss and content loss. g Training is performed with pixel loss L pixel Expressed as the l_1-distance between each pixel of the super-resolution image and the high-resolution image, in addition, in order to obtain a semantically indistinguishable super-resolution image, the latent features of the super-resolution and high-resolution images derived by HLIP are minimized. Here, feature extraction is performed by using ResNet-50 pre-trained on ImageNet (denoted as E(·)) and a content loss L is adopted. content , specifically derived from the Jensen-Shannon (JS) divergence between potential features, the overall loss function for calculating HLIP is expressed as:
[0056] L pixel =‖p s -p h ‖1
[0057]
[0058] L g =L pixel +γL content
[0059] Where p s and p h are super-resolution and high-resolution images respectively, E(p s )∈R 2048 and E(p h )∈R 2048 are the latent features extracted by ResNet-50.
[0060] Preferably, the target detection model is a YOLOv5 model.
[0061] The beneficial effects of the present invention are as follows: the image detection method of Helicobacter pylori in digital slices designed by the present invention obtains digital slices by scanning pathological slices, converts the digital slices into image blocks as original images, and processes the original images through the neural network-based super-resolution image processing method disclosed in this patent. The super-resolution image thus obtained can provide more detailed features for target detection, and can also ensure that the super-resolution image obtained has the characteristic information of the original image, which is helpful for computer recognition and detection of HP. The present invention can comprehensively search for HP by detecting complete digital slices, reducing the workload of step-by-step screening under manual microscopy. BRIEF DESCRIPTION OF THE DRAWINGS
[0062] Figure 1 Schematic diagram of the image detection method of Helicobacter pylori in digital slices of the present invention;
[0063] Figure 2 Flowchart for establishing super-resolution images in the present invention;
[0064] Figure 3 This figure shows the comparison between the HP detection effect after image processing by the embodiment of the present invention and the existing method;
[0065] Figure 4 This is a comparison result diagram of the HP detection effect and pathologist annotations of an embodiment of the present invention;
[0066] Figure 5 This is a graph showing the HP detection results according to an embodiment of the present invention, which contain HP annotation results that were omitted by pathologists. DETAILED DESCRIPTION
[0067] In order to make those skilled in the art more clearly understand the purpose, technical solutions and advantages of the present invention, the present invention is further elaborated below in conjunction with the accompanying drawings and embodiments. In the description of the present invention, it should be understood that the orientation or position relationship indicated by the terms such as "upper", "lower", "left", "right", "inside", "outside", "horizontal", and "vertical" is based on the orientation or position relationship shown in the accompanying drawings, which is only for the convenience of describing the present invention, and does not indicate or imply that the device or original referred to must have a specific orientation, such as the singular forms "one" and "the" include plural referents, unless otherwise clearly provided in the text. It should be noted that the term "or" is generally used in its meaning including "or / and", unless otherwise clearly provided in the text, and therefore cannot be understood as a limitation on the present invention.
[0068] The present invention will be described in detail below with reference to the accompanying drawings: Figure 1 As shown, a method for detecting Helicobacter pylori in digital slices includes the following steps:
[0069] 1) Scanning the pathological sections with a digital slide scanner having a scanning magnification of 40X, and converting the obtained digital slides into image blocks as original images; the size of the image blocks is 640×640 pixels.
[0070] 2) The original image in step 1) is processed using a neural network-based super-resolution image processing method and a hierarchical deep learning framework to obtain a super-resolution image. The super-resolution image has characteristic information of the original image and is used to provide detailed features for the trained target detection model, thereby facilitating computer recognition and detection of HP and reducing the workload of manual microscopic step-by-step screening. 3) Helicobacter pylori is detected using the target detection model trained in step 2) to obtain HP detection results.
[0071] The digital slice in step 1) is a pathological slice of gastric mucosal biopsy tissue. The pathological slice is complete and a full scan image. The pathological slice is stained with HE staining to facilitate image processing in step 2).
[0072] like Figure 2 As shown, the algorithm of the neural network-based super-resolution image processing method in step 2) specifically includes the following steps:
[0073] 1) Fourier transform is used to extract the high-frequency enhanced image of the image block, and the semantic features of the image block are extracted in parallel;
[0074] 2) extracting a pixel feature map of the high-frequency enhanced image using a convolutional neural network, and concurrently extracting a 2D morphological feature map using stacked staggered blocks;
[0075] 3) merging the semantic features in step 1) and the pixel feature map in step 2) to obtain gated pixel features, and concurrently merging the 2D morphological feature map and the semantic features to obtain gated morphological features;
[0076] 4) The gated pixel features in step 3) are combined with the gated features to obtain an embedded feature map, which is further calculated to obtain a super-resolution feature map, and finally a super-resolution image is generated.
[0077] The target detection model trained by the hierarchical deep learning framework in step 2) is defined as HLIP, which extracts hierarchical features from the high-frequency enhanced image based on the intrinsic features of the pathological image, and then establishes a super-resolution image based on the local pathological image pattern.
[0078] The specific steps of establishing the super-resolution image are as follows:
[0079] 1) First, the original image is converted into a high-frequency enhanced image to recover the high-frequency texture features that are highly relevant to pathological diagnosis, and used as the input of the HLIP model;
[0080] 2) For the input pathology image p with width W and height H, apply a two-dimensional Fourier transform to convert the pixel value p_i(x,y) of channel i in RGB space to frequency space:
[0081]
[0082] Where j is the imaginary unit and (x, y) is the coordinate of a single pixel;
[0083] 3) After frequency calibration, place the zero-frequency component at the edge of the feature map;
[0084] 4) By designing high-pass filters (HPFs) with different cutoff frequencies, the m k The mask is used to perform spectrum filtering to obtain the multi-level high-frequency characteristic spectrum G of each channel. i,k * (u,v):
[0085] G i,k * (u,v)=G i (u,v)⊙m k
[0086] Where ⊙ represents the Hadamard product of the matrix;
[0087] 5) Obtain the corresponding high-frequency spatial feature map h through inverse Fourier transform i,k (x,y):
[0088]
[0089] 6) Using the original image g i (x,y) and multiple high-frequency spatial feature maps h i,k (x,y) is superimposed to obtain the high-frequency enhanced image q:
[0090] q(x,y)=concat(g1(x,y),g2(x,y),…g c (x,y),h 1,1 (x,y),h 1,2 (x,y),…,h c,d (x,y))
[0091] Where c is the number of original channels, d is the number of high-frequency spatial feature maps; in the formula, c and d are set to 3 and 2 respectively, and the cutoff frequency is set to 5 and 10;
[0092] 7) Using attention-based methods to integrate hierarchical features, including semantic features and extracting features using stacked convolutional neural networks (CNNs);
[0093] 8) Using the general base model for computational pathology UNI Extract the semantic features V of a single image block in the original image p ∈R W′×H′×F The model is pre-trained on the Mass-100K dataset and, in parallel, a single-layer CNN is used to extract V based on the enhanced image q. m ∈R W′×H′×F 2D morphological feature map, and further based on the pixel features, the stacked residual block (StackedCNN) with CNN backbone is used to extract V m ∈R W′×H′×F 2D morphological feature map;
[0094] v s =f UNI (p)
[0095] V p =CNN(p)
[0096] V m =StackedCNN(V p );
[0097] 9) Take the semantic feature as the query object and the 2D morphological feature map as the key value to obtain the gated morphological feature V m,gated , for the gated pixel feature V m,gated The same construction procedure is applied to the pixel feature map of :
[0098] V p,gated =CrossAttention(query=v s ,key=Vp ,value=V p )
[0099] V m,gated =CrossAttention(query=v s ,key=V m ,valuer=V m )
[0100] 10) Aggregate the gated morphological features and gated pixel features to obtain an embedded feature map M∈R with super resolution W×H×F :
[0101] M=V p,gated +V m,gated ;
[0102] Attention-based hierarchical feature integration is used to upsample local implicit image blocks to generate a super-resolution feature map Z∈R with a given magnification ratio δ. δW×δH×F , and obtain the final super-resolution image based on the feature map.
[0103] The specific steps to obtain a super-resolution image based on the super-resolution feature map are as follows:
[0104] 1) Map M(a,b) to Z(δa,δb), where a∈{1,2,…,W} and b∈{1,2,…,H}, and resize the embedded feature map M to the size of Z, expressed as:
[0105] Z(δa,δb)=M(a,b)
[0106] 2) By taking the weighted sum of the features of the four nearest coordinates that are divisible by δ, the features corresponding to the other coordinates are generated. For example, for the feature z with coordinates (x, y) in the super-resolution feature map, x,y ∈R F , where x∈{1,2,…,δW} and y∈{1,2,…,δH}, get the coordinates of the nearest upper left corner (x * ,y * )as follows:
[0107]
[0108] 3) The four closest features: upper left M(x * ,y * )、Upper right M(x * +1,y * )、lower left M(x * ,y * +1) and the lower right M(x* +1,y * +1) as the weighted sum of spatial neighbors, the weight S is the target coordinate (x, y) and the opposite spatial neighbor coordinate (ie M(x * ,y * ) is weighted by (x,y) and (x * +1,y * +1); the super-resolution feature z x,y The calculation process is as follows:
[0109] d t =‖(x t ,y t )-(x,y)‖2
[0110]
[0111]
[0112] Where (x t ,y t )∈{(x * ,y * ),(x * +1,y * ),(x * ,y * +1),(x * +1,y * +1)}, f is a linear feature embedding layer.
[0113] 4) Use another linear layer with input and output dimensions of F and 3 respectively to convert the feature map Z into RGB pixels and obtain the super-resolution image p s ;
[0114] 5) During the training process, an adversarial strategy is used to carry out HLIP, by introducing a discriminator D θ (·), the discriminator consists of 9 convolutional neural networks, which can obtain the probability that each image belongs to the generated image. For the generated super-resolution image p s , discriminator loss L d It can be calculated as follows:
[0115] L d =-log(D θ (p s ))
[0116] Among them L d is the loss function used for discriminator optimization.
[0117] The target detection model is constructed by an overall loss L consisting of pixel loss and content loss.g Training is performed with pixel loss L pixel Expressed as the l_1-distance between each pixel of the super-resolution image and the high-resolution image, in addition, in order to obtain a semantically indistinguishable super-resolution image, the latent features of the super-resolution and high-resolution images derived by HLIP are minimized. Here, feature extraction is performed by using ResNet-50 pre-trained on ImageNet (denoted as E(·)) and a content loss L is adopted. content , specifically derived from the Jensen-Shannon (JS) divergence between potential features, the overall loss function for calculating HLIP is expressed as:
[0118] L pixel =‖p s -p h ‖1
[0119]
[0120] L g =L pixel +γL content
[0121] Where p s and p h are super-resolution and high-resolution images respectively, E(p s )∈R 2048 and E(p h )∈R 2048 is the potential feature extracted by ResNet-50. Here γ is set to 0.5. Adam is used as the optimizer during training, and the learning rate is set to 1×10 -4 The target detection model is the YOLOv5 model.
[0122] For ease of understanding, the Figure 3-Figure 5 ,in Figure 3 This figure shows the comparison between the HP detection effect after image processing by the embodiment of the present invention and the existing method; Figure 4 This is a comparison result diagram of the HP detection effect and pathologist annotations of an embodiment of the present invention; Figure 5 This is a graph showing the HP detection results according to an embodiment of the present invention, which contain HP annotation results that were omitted by pathologists.
[0123] In the present invention, pathological sections of gastric mucosal biopsy tissue were selected and stained with hematoxylin and eosin. The sections were placed in an AperioLeica GT450 digital slide scanner for scanning. The "multi-slice scanning" mode was selected to ensure that different focal plane information of the sections was obtained. The resolution was 40X (0.25μm / pixel) and the scanning speed was 20mm / s. After the section scanning was completed, a digital slide file in the svs format was obtained. The file size was 555733156 bytes and the image size was 95989×48812 pixels. It was converted into 11550 image blocks with a resolution of 640×640. Each image block was subsequently processed and analyzed one by one.
[0124] The target detection model trained by the hierarchical deep learning framework in this invention is defined as HLIP. HLIP extracts hierarchical features from high-frequency enhanced images based on the intrinsic characteristics of pathological images, and then performs super-resolution reconstruction based on local pathological image patterns. The super-resolution image establishment process is as follows: Figure 2 As shown in Figure 3, the original image is first converted into a high-frequency enhanced image to recover the high-frequency texture features that are highly relevant to pathological diagnosis, and then used as the input of the HLIP model.
[0125] Example 1
[0126] After the original image is super-resolution processed using the method disclosed in the present invention, it is detected using the fully supervised HP target detection model based on YOLO-v5s. Five-fold cross-validation shows that the detection performance mAP50 of the super-resolution image generated by HLIP is 0.5887±0.0158, the mAP50 value of the original image detection is 0.3012±0.0035, and the mAP50 value of the image detection processed by bicubic interpolation is 0.3413±0.0062. The effect of HP detection after the image is processed by the method disclosed in the present invention to obtain a super-resolution image is significantly higher than that of no processing or the existing processing method (see Figure 3 ). In addition, the visualization results show that the detection results of the super-resolution images generated by HLIP are highly consistent with the expert annotations, with a consistency rate of 100%. (See Figure 4 ), which shows that HLIP improves the recall rate of HP detection and effectively reduces the number of false and missed detections. From the perspective of target detection, pathologists' annotations may still have the possibility of missed diagnoses. Therefore, the present invention pays more attention to the recall rate and notices that some suspected HP objects are ignored by pathologists during the annotation process (see Figure 5 , indicated by the black arrow). Notably, these examples were independently reviewed by two associate chief physician pathologists and confirmed as correct detections. These results demonstrate that super-resolution images generated by HLIP enable superior HP detection on pathology slides.
[0127] Example 2
[0128] The present invention relates to a method for detecting HP in pathological sections, some steps of which can be run on a computer device, such as a desktop host, a laptop computer, etc. The computer device executes the detection method of this embodiment to detect HP in digital sections. The object of the detection is a digital section, and the detection target is HP. However, this is not limited to this, so any object that needs to detect the target object can use the method of this embodiment to complete the detection. The present invention can also be used in combination with an electronic device, which may include: one or more processors and one or more memories, wherein the one or more memories store a computer program, and the computer program is loaded and executed by the one or more processors to implement the above-mentioned super-resolution image generation method and HP detection target detection method, thereby ensuring better implementation and use of the present invention.
[0129] The specific embodiments described herein are merely illustrative of the principles and effects of the present invention and are not intended to limit the present invention. Anyone skilled in the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by one of ordinary skill in the art without departing from the spirit and technical concepts disclosed herein are intended to be covered by the claims of the present invention.
Claims
1. A method for detecting Helicobacter pylori in digital slices, characterized in that: The steps include: 1) Scanning the pathological sections with a digital slide scanner and converting the obtained digital slides into image blocks as original images; 2) Processing the original image from step 1) using a neural network-based super-resolution image processing method and a hierarchical deep learning framework to obtain a super-resolution image. The super-resolution image has feature information of the original image and is used to provide detailed features for the trained object detection model, thereby facilitating computer recognition and detection of HP and reducing the workload of manual microscopic step-by-step screening. 3) Using the target detection model trained in step 2) to detect Helicobacter pylori to obtain HP detection results.
2. The method for detecting Helicobacter pylori in digital slices according to claim 1, wherein: The digital slice in step 1) is a pathological slice of gastric mucosal biopsy tissue. The pathological slice is complete and a full scan image. The pathological slice is stained with HE staining to facilitate image processing in step 2).
3. The method for detecting Helicobacter pylori in digital slices according to claim 1, wherein: The algorithm of the neural network-based super-resolution image processing method in step 2) specifically includes the following steps: 1) Fourier transform is used to extract the high-frequency enhanced image of the image block, and the semantic features of the image block are extracted in parallel; 2) extracting a pixel feature map of the high-frequency enhanced image using a convolutional neural network, and concurrently extracting a 2D morphological feature map using stacked staggered blocks; 3) merging the semantic features in step 1) and the pixel feature map in step 2) to obtain gated pixel features, and concurrently merging the 2D morphological feature map and the semantic features to obtain gated morphological features; 4) The gated pixel features in step 3) are combined with the gated features to obtain an embedded feature map, which is further calculated to obtain a super-resolution feature map, and finally a super-resolution image is generated.
4. The method for detecting Helicobacter pylori in digital slices according to claim 1, wherein: The target detection model trained by the hierarchical deep learning framework in step 2) is defined as HLIP, which extracts hierarchical features from the high-frequency enhanced image based on the intrinsic features of the pathological image, and then establishes a super-resolution image based on the local pathological image pattern.
5. The method for detecting Helicobacter pylori in digital slices according to claim 4, characterized in that: The specific steps of establishing the super-resolution image are as follows: 1) First, the original image is converted into a high-frequency enhanced image to recover the high-frequency texture features that are highly relevant to pathological diagnosis, and used as the input of the HLIP model; 2) For the input pathology image p with width W and height H, apply a two-dimensional Fourier transform to convert the pixel value p_i(x,y) of channel i in RGB space to frequency space: Where j is the imaginary unit and (x,y) is the coordinate of a single pixel; 3) After frequency calibration, place the zero-frequency component at the edge of the feature map; 4) By designing high-pass filters (HPFs) with different cutoff frequencies, the m k The mask is used to perform spectrum filtering to obtain the multi-level high-frequency characteristic spectrum G of each channel. i,k * (u,v): G i,k * (u,v)=G i (u,v)⊙m k Where ⊙ represents the Hadamard product of the matrix; 5) Obtain the corresponding high-frequency spatial feature map h through inverse Fourier transform i,k (x,y): 6) Using the original image g i (x,y) and multiple high-frequency spatial feature maps h i,k (x,y) is superimposed to obtain the high-frequency enhanced image q: q(x,y)=concat(g1(x,y),g2(x,y),…g c (x,y),h 1,1 (x,y),h 1,2 (x,y),…,h c,d (x,y)) Where c is the number of original channels, d is the number of high-frequency spatial feature maps; 7) Using attention-based methods to integrate hierarchical features, including semantic features and extracting features using stacked convolutional neural networks; 8) Using the general base model for computational pathology INI Extract the semantic features V of a single image block in the original image p ∈R W ′×H′×F The model is pre-trained on the Mass-100K dataset and, in parallel, a single-layer CNN is used to extract V based on the enhanced image q. m ∈R W′×H′×F 2D morphological feature map, and further based on the pixel features, the stacked residual blocks with CNN backbone are used to extract V m ∈R W′×H′×F 2D morphological feature map; v s =f UNI (p) V p =CNN(p) V m =StackedCNN(V p ); 9) Take the semantic feature as the query object and the 2D morphological feature map as the key value to obtain the gated morphological feature V m,gated , for the gated pixel feature V m,gated The same construction procedure is applied to the pixel feature map of : V p,gated =CrossAttention(query=v s ,key=V p ,value=V p ) V m,gated =CrossAttention(query=v s ,key=V m ,value=V m ) 10) Aggregate the gated morphological features and gated pixel features to obtain an embedded feature map M∈R with super resolution W×H×F : M=V p,gated +V m,gated ; Attention-based hierarchical feature integration is used to upsample local implicit image blocks to generate a super-resolution feature map Z∈R with a given magnification ratio δ. δW×δH×F , and obtain the final super-resolution image based on the feature map.
6. The method for detecting Helicobacter pylori in digital slices according to claim 5, characterized in that: The specific steps to obtain a super-resolution image based on the super-resolution feature map are as follows: 1) Map M(a,b) to Z(δa,δb), where a∈{1,2,…,W} and b∈{1,2,…,H}, and resize the embedded feature map M to the size of Z, expressed as: Z(δa,δb)=M(a,b) 2) By taking the weighted sum of the features of the four nearest coordinates that are divisible by δ, the features corresponding to the other coordinates are generated. For the feature z with coordinates (x, y) in the super-resolution feature map, x,y ∈R F , where x∈{1,2,…,δW} and y∈{1,2,…,δH}, get the coordinates of the nearest upper left corner (x * ,y * )as follows: 3) The four closest features: upper left M(x * ,y * )、Upper right M(x * +1,y * )、lower left M(x * ,y * +1) and the lower right M(x * +1,y * +1) as the weighted sum of spatial neighbors, the weight S is the target coordinate (x, y) and the opposite spatial neighbor coordinate (ie M(x * ,y * ) is weighted by (x,y) and (x * +1,y * +1); the super-resolution feature z x,y The calculation process is as follows: d t =‖(x t ,y t )-(x,y)‖2 Where (x t ,y t )∈{(x * ,y * ),(x * +1,y * ),(x * ,y * +1),(x * +1,y * +1)}, f is a linear feature embedding layer; 4) Use another linear layer with input and output dimensions of F and 3 respectively to convert the feature map Z into RGB pixels and obtain the super-resolution image p s ; 5) During the training process, an adversarial strategy is used to carry out HLIP, by introducing a discriminator D θ (·), the discriminator consists of 9 convolutional neural networks, which can obtain the probability that each image belongs to the generated image. For the generated super-resolution image p s , discriminator loss L d It can be calculated as follows: L d =-log(D θ (p s )) Among them L d is the loss function used for discriminator optimization.
7. The method for detecting Helicobacter pylori in digital slices according to claim 4, wherein: The target detection model is constructed by an overall loss L consisting of pixel loss and content loss. g Training is performed with pixel loss L pixel Expressed as the l_1-distance between each pixel of the super-resolution image and the high-resolution image, in order to obtain a semantically indistinguishable super-resolution image, the latent features of the super-resolution and high-resolution images derived by HLIP are minimized. Feature extraction is performed using ResNet-50 (denoted as E(·)) pre-trained on ImageNet, and a content loss L is adopted. content , specifically derived from the Jensen-Shannon (JS) divergence between potential features, the overall loss function for calculating HLIP is expressed as: L pixel =‖p s -p h ‖1 THE g =L pixel +γL content Where p s and p h are super-resolution and high-resolution images respectively, E(p s )∈R 2048 and E(p h )∈R 2048 are the latent features extracted by ResNet-50.
8. The method for detecting Helicobacter pylori in digital slices according to claim 7, wherein: The target detection model is the YOLOv5 model.