A small intestine micro-target detection method based on self-correcting coordinate attention

Abstract

The present application relates to the field of small intestine micro target detection, and discloses a small intestine micro target detection method based on self-correcting coordinate attention, which comprises the following steps: step 1, constructing a small intestine three-lesion dataset; step 2, constructing a small intestine micro target detection network model, and training and verifying; and finally generating a detection result. In step 2, YOLOv8 is used as a basic network model, a self-correcting coordinate attention module is designed in the Backbone part of the model to extract position features of input small intestine lesion images; and a target detection promotion layer is built on the original FPN+PAN structure in the Neck part of the model to perform multi-scale fusion on the extracted features, so that more obvious small target features can be obtained. Based on the above design, the positioning accuracy of the network for smaller targets is improved, and the detection performance for small target lesions is improved.

Description

Technical Field

[0001] This invention relates to the field of small intestinal micro-target detection, and specifically to a method for small intestinal micro-target detection based on self-calibrating coordinate attention. Background Technology

[0002] The incidence of digestive tract diseases in my country is rising year by year. Digestive tract cancers (such as colorectal cancer, gastric cancer, and esophageal cancer) account for approximately 50-60% of all malignant tumors in my country. Studies have shown that early detection and treatment of digestive tract diseases are key to preventing their further deterioration. Currently, conventional handheld endoscopy cannot completely examine the small intestine. In recent years, wireless capsule endoscopy, as a painless and non-invasive imaging tool, has been widely used in the visualization and early screening of gastrointestinal (GI) diseases, especially small intestinal diseases.

[0003] Because each patient's images can number in the tens of thousands, and the complex gastrointestinal environment generates a large number of low-quality WCE images due to interference from factors such as debris, defocus, bubbles, and lighting, these images require manual review. Tiny lesions, such as small polyps, small bleeding points, and small ulcers, are easily overlooked. If these tiny lesions are not detected and treated promptly, they can easily develop into large lesions. Therefore, there is an urgent need for computer-aided diagnosis (CAD) algorithms to assist doctors in making more accurate diagnoses.

[0004] Existing small intestinal lesion detection technologies are not ideal for detecting tiny lesions in the small intestine and are easily affected by the complex intestinal environment, leading to missed detections. The reasons for this are: 1. Lesions appear randomly in the image and vary greatly in shape; current technologies cannot focus on the location information of tiny lesions and have insufficient feature extraction capabilities. 2. High similarity between lesion classes makes network training difficult, resulting in missed and false detections. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by providing a method for detecting small targets in the small intestine based on self-calibrating coordinate attention.

[0006] A method for detecting small targets in the small intestine based on self-calibrating coordinate attention includes the following steps:

[0007] Step 1: Construct a small intestinal lesion dataset including a training set and a validation set; Step 2: Construct a small intestinal micro-target detection network model; Step 3: Use the training set and validation set from Step 1, input them into the small intestinal target detection network model constructed in Step 2, and perform iterative training to obtain the optimal weight model; Step 4: After preprocessing the image to be detected, input it into the optimal weight model described in Step 3 to perform lesion detection and generate detection results, characterized in that:

[0008] The construction of the small intestinal micro-target detection network model described in step 2 is as follows: First, configure a detection network model including three parts: Backbone, Neck and Head. Embed a self-calibrating coordinate attention module in the Backbone part of the detection network model and implant a target detection facilitation layer in the Neck part.

[0009] Step 1 involves constructing a small intestinal lesion dataset, specifically including:

[0010] Image frames of lesions were extracted from original images and videos obtained from clinical settings as data samples to construct a dataset of three types of lesions in the small intestine, including ulcers, lymphangiectasia, and vascular exposure.

[0011] The detection network model configured in step 2 above is the YOLOv8 detection network.

[0012] In step 2, the self-calibrating coordinate attention module includes a residual layer, an X-axis average pooling layer, a Y-axis average pooling layer, a self-calibrating convolutional structure, a first convolutional layer, a normalization layer, two second convolutional layers, two activation functions, and a re-weighting layer.

[0013] The output of the residual layer is directly connected to the re-weight layer. Simultaneously, the output of the residual layer is connected to a parallel X-axis average pooling layer and a parallel Y-axis average pooling layer. The outputs of the parallel X-axis average pooling layer and the parallel Y-axis average pooling layer are connected to the self-calibrating convolutional structure. The output of the self-calibrating convolutional structure is connected to the first convolutional layer. The output of the first convolutional layer is connected to the input of the normalization layer. The output of the normalization layer is connected in parallel to two combined layers. Each combined layer includes a second convolutional layer and an activation function in series. The outputs of the activation functions of both combined layers are connected to the re-weight layer.

[0014] The self-calibrating coordinate attention module described above can better integrate feature maps under different receptive fields, effectively expand the receptive field of each spatial location, effectively enhance feature extraction of targets at different scales, and help the network to more accurately locate the target area.

[0015] In step 2, a target layer is implanted in the Neck region of YOLOv8, and a neighborhood erasure module is designed between the feature layer of the aforementioned Backbone region and the target layer. The target layer and the neighborhood erasure module form a target detection facilitation layer, specifically including:

[0016] In the Neck section of YOLOv8, the original PAN-FPN structure's three feature maps are expanded to four, and a neighborhood erasure module is introduced to convert the output p of the first-level C2f and second-level C2f in the Backbone section. s and p s+1 As input to the neighborhood erasure module, after erasure processing, the output features of the neighborhood erasure module are... The output feature O is further upsampled with the third set of feature maps. u The layers are spliced ​​together and finally output through the C2f module, forming the target detection facilitation layer.

[0017] The aforementioned target detection enhancement layer erases the features of relatively large targets from the first-level C2f output of the Backbone part of the YOLOv8 detection network model, highlighting the features of smaller targets. The first-level C2f of the Backbone part is a low-feature layer with strong positional information. This target detection enhancement layer improves the network's accuracy in locating smaller targets and reduces missed detections.

[0018] In step 2, the size of the feature map output by the target detection facilitation layer is 160*160.

[0019] Further preferred, the technical solution adopted in this invention includes the following steps:

[0020] Step 1: Construct a small intestine three-lesion dataset. Professional physicians manually labeled the lesion areas of the original images, using three labels: ulcer, lymphangiectasis, and vascular exposure. Image augmentation algorithms were used to flip, rotate, and then flip and rotate again to create the small intestine three-lesion dataset. Each category of cases was randomly divided into training, validation, and test sets in a 6:2:2 ratio.

[0021] Step 2: A coordinate attention module is embedded in the backbone of the network, and a self-calibration convolutional layer is designed within the coordinate attention module. This proposes a novel attention mechanism, Self-Calibration Coordinate Attention (SCCA), which can help detect lesions at random locations. A small target layer is added to the neck part, and a Neighbor Erasing Module (NEM) is designed between the shallow feature layer and the small target layer, resulting in a novel object detection facilitation layer.

[0022] An object detection enhancement layer is added to the existing PAN-FPN structure of Neck, expanding the original 3 sets of feature maps to 4 sets, and embedding an NEM neighborhood erasure module. The outputs p of the first-level C2f and second-level C2f in the Backbone part are then processed. s and p s+1 As input to the NEM, after erasing, the output features of the NEM are... The output feature O is upsampled from the previous layer feature map. u The layers are spliced ​​together, and the final result of this layer is obtained through a C2f module.

[0023] A NEM neighborhood erasure module is designed in the small target layer. The NEM neighborhood erasure module upsamples the adjacent deep features to the same size as the shallow features, and then subtracts the result of multiplying the upsampled deep features and the shallow feature matrix from the shallow features to erase the features of the large target and obtain more obvious small target features.

[0024] Step 3: Train the network. Use the pre-divided dataset and input it into the small intestine micro-target detection network based on self-calibrating coordinate attention of this invention for iterative training to obtain the optimal weight model.

[0025] Step 4: After preprocessing the image to be detected, input it into the optimal weight model for lesion detection and generate detection results.

[0026] This invention presents a method for detecting small intestinal micro-targets based on self-calibrating coordinate attention. Using YOLOv8 as the base network model, a self-calibrating coordinate attention module is designed in the backbone to extract location features from the input small intestinal lesion image. A target detection facilitation layer is built on the existing FPN+PAN structure in the neck region to perform multi-scale fusion of the extracted features. A prediction head is added after the newly added detection layer, combining with three other prediction heads to detect targets at multiple scales. A small intestinal three-lesion dataset is created and input into the network for training until the network stabilizes. When a small intestinal lesion image to be detected is input, the stable network performs lesion detection.

[0027] Compared with the prior art, the present invention has the following advantages:

[0028] (1) A method for detecting small targets in the small intestine based on self-calibrating coordinate attention is proposed to improve the accuracy of detecting small intestinal lesions.

[0029] (2) To address the problem of uneven and irregular distribution of lesion locations, a self-calibrating coordinate attention module is designed in the Backbone part of the network in this invention. This fully utilizes the captured location information to accurately locate the lesion region of interest.

[0030] (3) To address the issue of high similarity among small target classes, this invention constructs a target detection enhancement layer in the Neck section. First, a small target layer is added. Then, an NEM neighborhood erasure module is embedded between the first and second level C2f modules and the small target layer in the Backbone section to erase the salient features of large targets and emphasize the features of small targets in the shallow layer. This, combined with the other three detection layers, achieves effective fusion of multi-scale information, improving the detection performance of small target lesions. Attached Figure Description

[0031] Figure 1 This is a flowchart of the method of the present invention;

[0032] Figure 2 This is the complete network structure diagram of the present invention;

[0033] Figure 3 This is a structural diagram of the self-calibrating coordinate attention module in this invention;

[0034] Figure 4 This is a diagram of the self-correcting convolution structure in this invention;

[0035] Figure 5 This is a structural diagram of the NEM neighborhood erasure module in this invention;

[0036] Figure 6 This is a partial image detection result diagram from the present invention. Detailed Implementation

[0037] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the scope of protection of the present invention is not limited thereto.

[0038] See Figure 1 and Figure 2 This invention relates to a method for detecting small targets in the small intestine based on self-calibrating coordinate attention, comprising the following steps:

[0039] Step 1: Dataset Creation. Original clinical images and videos were obtained from collaborating hospitals, and image frames of lesions were extracted from the videos as data samples. A small intestine three-lesion dataset was constructed, with lesion regions in the original images manually labeled by professional physicians. The labels were divided into three types: ulcer, lymphangiectasis, and vascular exposure. Image enhancement algorithms were used to flip, rotate, and then flip and rotate the images again to create a small intestine three-lesion dataset. Each type of case was randomly divided into training, validation, and test sets in a 6:2:2 ratio. The number of labeled images is shown in Table 1.

[0040] Table 1

[0041]

[0042] For example, there were 57 ulcer cases, which were divided into a training set of 35 cases, a validation set of 11 cases, and a test set of 11 cases in a 6:2:2 ratio. The 35 cases provided 896 images.

[0043] Step 2: Based on the YOLOv8 detection network model, construct a small intestinal micro-target detection network model.

[0044] The YOLOv8 network model consists of three parts: Backbone, Neck, and Head.

[0045] A coordinate attention module is embedded in the backbone of the network, and a self-calibration convolutional layer is designed within the coordinate attention module to propose a novel attention mechanism, Self-Calibration Coordinate Attention (SCCA), which can help discover lesions at random locations. A small target layer is added to the neck part, and a Neighbor Erasing Module (NEM) is designed between the shallow feature layer and the small target layer to obtain a novel object detection facilitation layer.

[0046] Step 3: Train the network. Use the pre-divided dataset and input it into the small intestine micro-target detection network based on position attention enhancement of the present invention for iterative training to obtain the optimal weight model.

[0047] Step 4: After preprocessing the image to be detected, input it into the optimal weight model for lesion detection and generate detection results.

[0048] Preferably, in step 2, a self-calibrating coordinate attention (SCCA) module is added after the fourth C2f module in the Backbone part of the YOLOv8 algorithm model.

[0049] according to Figure 3 As shown, self-calibrating coordinate attention is accomplished through three steps: information embedding, self-calibration, and attention generation. Information embedding: To capture remote spatial interactions with precise location information, this invention decomposes global average pooling into a one-to-one one-dimensional feature encoding operation according to the following formula:

[0050]

[0051] Among them, z c This is the output associated with the c-th channel, where H represents the height of the input image, W represents the width of the input image, and x... c Let represent the original input image, and 'c' represent the number of channels. Specifically, given an input X, each channel is first encoded along the horizontal and vertical coordinates using a pooling kernel of size (H, 1) or (1, W). This involves performing global average pooling on the input feature map in both the width and height directions, obtaining feature maps in both directions. Assuming the input feature layer has a shape of [C, H, W], after average pooling in the width direction, the resulting feature layer has a shape of [C, H, 1], mapping the features to the higher dimension; after average pooling in the height direction, the resulting feature layer has a shape of [C, 1, W], mapping the features to the wider dimension. Therefore, the output of the c-th channel with height h is represented as:

[0052]

[0053] in This represents the output of the c-th channel at height h.

[0054] Similarly, the output of the c-th channel with width w can be written as:

[0055]

[0056] in This represents the output of the c-th channel at height w.

[0057] After one-dimensional global pooling in both vertical and horizontal directions, the feature maps are combined and then passed through a self-calibrating convolutional layer. Instead of performing convolution on all outputs in the space, the self-calibrating convolution first downsamples the input to a low-dimensional embedding. Then, the low-dimensional embedding transformed from the filter portion is used to correct the convolution transformation of the other filter portion. This convolution can expand the receptive field of the convolution through the intrinsic communication of features, thereby enhancing the diversity of output features.

[0058] according to Figure 4As shown, the input X is first uniformly divided into two parts {X1, X2}, and then each part undergoes different processing to collect different types of contextual information. The original filter is divided into four parts {K1, K2, K3, K4}, each responsible for a different function. X1 is fed into the self-calibration branch, and self-calibration is performed on X1 using {K1, K2, K3}: given input X1, a filter of size r×r and stride r is used, followed by downsampling, K2 convolutional feature transformation, upsampling, and then addition with the input X1. Finally, it is passed through a sigmoid function to obtain the spatial process as follows:

[0059] T1 = AvgPool r (X1) (4)

[0060] Where T1 represents the result of downsampling the input X1.

[0061] Perform feature transformation on X1 using K2:

[0062] X'1=Up(F2(T1))=Up(T1*K2) (5)

[0063] Where Up(·) is bilinear interpolation, mapping the intermediate reference point from the small-scale space to the original feature space. The calibration operation can be represented as follows:

[0064] Y1'=F3(X1)·σ(X1+X'1) (6)

[0065] Where F3(X1) = X1 * K1, σ is the Sigmoid function, and X'1 is used as the residual to form the weights for correction. The final output after correction is:

[0066] Y1=F4(Y1')=Y1'*K4 (7)

[0067] Feed X2 into the regular convolution branch and perform a simple convolution operation:

[0068] Y2=F1(X2)=X2*K1 (8)

[0069] The goal is to preserve the spatial environment. Finally, {Y1, Y2} are concatenated together to form the output Y.

[0070] The method of this invention can effectively obtain the global receptive field and encode precise location information through the above transformation. In order to utilize the resulting representation, the third step, attention generation, is then performed.

[0071] After information embedding and self-correction transformations, this part merges the two parallel stages mentioned above, transposes the width and height to the same dimension, and then stacks them to merge the width and height features. The resulting feature layer is [C, 1, H+W]. Features are obtained using 1×1 convolution + normalization + activation function. The transformation operations are as follows:

[0072]

[0073] Where δ is h_swish, F1 is a 1×1 convolution transform function, and f is an intermediate feature map that encodes spatial information in the horizontal and vertical directions. Then, the two parallel stages are separated again along the spatial dimension, and f is decomposed into f... h and f w Using two additional 1×1 convolution transformations F h and F w For f respectively h and f w By increasing the dimensionality and then combining this with the sigmoid activation function to obtain the attention cases in the width and height dimensions, we obtain the attention vector as follows: and

[0074]

[0075]

[0076] Here, σ is the sigmoid activation function.

[0077] Finally, the output Y of the self-calibrating coordinate attention module can be written as:

[0078]

[0079] Preferably, a small target layer is added to the original PAN-FPN structure of the YOLOv8 network model Neck, expanding the original 3 sets of feature maps to 4 sets, and introducing a NEM neighborhood erasure module to obtain a new target detection enhancement layer. The outputs p of the first-level C2f and second-level C2f of the Backbone part are then combined. s and p s+1 As input to the NEM, after erasing, the output features of the NEM are... The output feature O is upsampled from the previous layer feature map. u The layers are concatenated, and the final result for this layer is obtained through a C2f module. The structure of the domain erasure module is as follows: Figure 5 As shown.

[0080] In step 2, the feature map output by the target detection facilitation layer is 160*160 in size.

[0081] Shallow features contain both small and large target features. In the NEM neighborhood erasure module, adjacent deep features are upsampled to the same size as the shallow features. Then, the shallow features are subtracted from the result of multiplying the upsampled deep and shallow feature matrices, erasing the features of large targets and obtaining more obvious small target features. The output p of the first-level C2f in the Backbone part. s The output p of the adjacent second stage C2f s+1 Together, they are input into the NEM module, where p s+1 Upsampling yields p', ensuring consistent spatial resolution among features:

[0082] p'=Up(p s+1 (13)

[0083] Simultaneously use p s+1 To guide p s Feature erasure is performed by multiplying the deep feature matrix after Up sampling with the shallow feature matrix to obtain p. es :

[0084] p es =p′⊙p s =Up(p s+1 )⊙p s (14)

[0085] Here, ⊙ represents pixel-by-pixel multiplication.

[0086] Finally, the above results are erased using shallow features to obtain the small target features:

[0087]

[0088] The object detection facilitation layer integrates features from other target layers at multiple scales, thereby improving the detection performance of the network model.

[0089] Preferably, the loss function of the YOLOv8 algorithm model consists of classification loss and regression loss. The classification loss uses Binary Cross Entropy (BCE), and the regression loss consists of CIouLoss + DFLLoss.

[0090] Cross-entropy loss L BCE The calculation formula is as follows:

[0091]

[0092] In the formula, y i ∈{0,1}、p i ∈[0,1] represent the true label value and the pixel prediction probability value, respectively.

[0093] The regression loss is divided into two parts. The first part calculates the IoU between the predicted box and the target box using CIoU loss.

[0094] The formula for CIoULoss is as follows:

[0095]

[0096] in, d1 represents the Euclidean distance between the center points of the predicted bounding box and the target bounding box, and d2 represents the diagonal distance of the minimum bounding rectangle. and These represent the aspect ratios of the target bounding box and the predicted bounding box, respectively.

[0097]

[0098] The second part uses the following DFLLoss calculation formula:

[0099] L DFL (S i S i+1 )=-((y i+1 -y)log(S i )+(yy i )log(S i+1 ))(18)

[0100] Where y is the target, y i and y i+1 y represents the left and right integer values ​​of the floating-point value y, and S represents the output distribution.

[0101]

[0102] Finally, the joint loss function L total It can be represented as L total =0.75L BCE +0.05L CIoU +0.15L DFL .

[0103] This invention provides an experimental example:

[0104] (1) Experimental conditions

[0105] The experiment used a workstation configured with an Intel(R) Xeon(R) Silver4210R CPU @ 2.40GHz, 64GB of RAM, a Windows operating system, and two Nvidia GTX 3080 Ti graphics cards. The model was implemented using the PyTorch deep learning framework, version 1.11.0, and Python version 3.7. The batch size of training samples fed into the network was 16, and the training iterations were 300.

[0106] (2) Experimental Results

[0107] The model evaluation metrics used in this invention mainly include mean precision (mAP), recall, and precision. It should be noted that mean precision (mAP) is the sum of the average precision of all classes divided by the average precision of all classes in the dataset. mAP is also an important metric for evaluating model performance; a higher value indicates better model performance. As shown in the following formula, AP is the area under the PR curve.

[0108]

[0109] Recall, which is the probability that the correct class in a sample is correctly predicted by the model, is expressed by the following formula:

[0110]

[0111] Accuracy is calculated by dividing the number of correctly predicted positive samples in the predicted dataset by the actual number of positive samples. The formula is as follows:

[0112]

[0113] Wherein, TP represents the number of positive samples predicted as positive, FP represents the number of negative samples predicted as positive, TN represents the number of negative samples predicted as negative, and FN represents the number of positive samples predicted as negative. Table 2 shows a comparison of the metrics of the network constructed for this invention with other neural networks.

[0114] Table 2

[0115]

[0116] As shown in Table 2, compared with existing models, the model proposed in this invention achieved the highest scores in the mAP index for all three types of lesions: ulceration (85.2%), lymphatic dilation (78.2%), and vascular exposure (77.7%). Overall, this data demonstrates that the method proposed in this invention outperforms other methods in detecting small intestinal lesions. It can not only detect large target lesions normally, but also has good detection effects on small target lesions with irregular and uneven locations. The method of this invention can be better applied to the auxiliary diagnosis of small intestinal lesions and has greater practical engineering application value.

[0117] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0118] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0119] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0120] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a realizable process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the function specified in one or more boxes.

Claims

1. A method for detecting small targets in the small intestine based on self-calibrating coordinate attention, comprising the following steps: Step 1: Construct a small intestinal lesion dataset including a training set and a validation set; Step 2: Construct a small intestinal micro-target detection network model; Step 3: Use the training set and validation set from Step 1, input them into the small intestinal target detection network model constructed in Step 2, and perform iterative training to obtain the optimal weight model; Step 4: After preprocessing the image to be detected, input it into the optimal weight model described in Step 3 to perform lesion detection and generate detection results, characterized in that: The construction of the small intestinal micro-target detection network model in step 2 is as follows: First, configure a detection network model including three parts: Backbone, Neck and Head. Embed a self-calibrating coordinate attention module in the Backbone part of the detection network model and implant a target detection facilitating layer in the Neck part. The detection network model mentioned is the YOLOv8 detection network model; The self-calibrating coordinate attention module includes a residual layer, an X-axis average pooling layer, a Y-axis average pooling layer, a self-calibrating convolutional structure, a first convolutional layer, a normalization layer, two second convolutional layers, two activation functions, and a re-weighting layer. The output of the residual layer is directly connected to the re-weight layer. Simultaneously, the output of the residual layer is connected to a parallel X-axis average pooling layer and a parallel Y-axis average pooling layer. The outputs of the parallel X-axis average pooling layer and the parallel Y-axis average pooling layer are connected to the self-calibrating convolutional structure. The output of the self-calibrating convolutional structure is connected to the first convolutional layer. The output of the first convolutional layer is connected to the input of the normalization layer. The output of the normalization layer is connected in parallel to two combined layers. Each combined layer includes a second convolutional layer and an activation function in series. The outputs of the activation functions of both combined layers are connected to the re-weight layer. An object detection enhancement layer is implanted in the Neck region, specifically including: In the Neck part of the YOLOv8 detection network model, the original PAN-FPN structure's three sets of feature maps are expanded to four sets of feature maps, and a neighborhood erasure module is introduced to modify the outputs of the first-level C2f and second-level C2f of the Backbone part of the YOLOv8 detection network model. and As input to the neighborhood erasure module, after erasure processing, the output features of the neighborhood erasure module are... The output features are further upsampled with the feature mapping of the third group. The layers are spliced ​​together and finally output through the C2f module, forming the target detection facilitation layer.

2. The method for detecting small targets in the small intestine based on self-calibrating coordinate attention according to claim 1, characterized in that, Step 1 involves constructing a small bowel lesion dataset, specifically including: Image frames of lesions were extracted from original images and videos obtained from clinical settings as data samples to construct a dataset of three types of lesions in the small intestine, including ulcers, lymphangiectasia, and vascular exposure.

3. The method for detecting small targets in the small intestine based on self-calibrating coordinate attention according to claim 1, characterized in that, In step 2, the size of the output feature map of the target detection facilitation layer is 160*160.

Images