Small target detection method, system and device for medical image and storage medium

Abstract

The invention discloses a small target detection method, system and device for a medical image and a storage medium, and the method comprises the steps: constructing a small target detection model of a bronchial medical image, which comprises a backbone network, a neck network and a prediction network, and introducing a residual network and a cavity space convolution pooling pyramid into the backbone network; introducing a global attention module into the neck network; inputting the preprocessed bronchial medical image into the backbone network, performing detail feature supplementation by adopting a residual network and enhancing a receptive field by adopting a void space convolution pooling pyramid to obtain feature maps of various scales output by the backbone network; inputting the multi-scale feature maps into a neck network, and performing global information enhancement by adopting a global attention module to obtain multi-scale tensor data output by the neck network; and carrying out feature fusion on the tensor data of multiple scales by adopting a prediction network, and carrying out small target detection on the fused features. According to the invention, the accuracy of small target detection can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of small target detection, and in particular to a small target detection method, system, device and storage medium for medical images. Background Art

[0002] Bronchial tuberculosis is a rare type of tuberculosis that mainly affects the bronchi and lungs. This bronchial tuberculosis is usually caused by Mycobacterium tuberculosis, and symptoms include cough, sputum, dyspnea, and chest pain. The diagnosis of bronchial tuberculosis is usually performed by line examination, CT scan, bronchoscopy, sputum examination, etc. In recent years, with the advancement and popularization of tuberculosis treatment, the incidence of bronchial tuberculosis has decreased. However, due to its atypical symptoms, there are still challenges in diagnosis and treatment. Therefore, research on bronchial tuberculosis remains of great significance.

[0003] The traditional method of detecting bronchial tuberculosis is mainly through the use of an endoscope, which is then reviewed by a doctor. The waiting time for the endoscope results is too long, and the long working hours will make the doctor extremely tired, which may lead to identification bias and may also reduce the doctor's work efficiency. In the imaging method, the cause of bronchial tuberculosis can be identified because it can clearly show pathological features such as bronchial wall thickening, stenosis, calcification, and lung lesions. This provides a reliable basis for the clinical diagnosis of bronchial tuberculosis. At present, there are relatively few studies on the use of convolutional neural networks combined with CT imaging to diagnose bronchial tuberculosis.

[0004] Traditional CNN usually consists of an input layer, a hidden layer, and an output layer connected in series. The input layer is responsible for receiving the input image, the hidden layer usually consists of multiple convolutional layers and pooling layers, whose function is to extract feature information from the input image, and the output layer is responsible for outputting the classification result. The traditional CNN structure can gradually extract high-level semantic features from the image. However, the traditional CNN method has the following disadvantages: Traditional CNN uses highly semantic information feature maps to extract the final classification prediction. Although it can effectively grasp the overall characteristics of the image, it also ignores some detailed information about the connection between layers, resulting in a decrease in classification accuracy. Summary of the invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art. To this end, the present invention proposes a small target detection method, system, device and storage medium for medical images, which can improve the accuracy of small target detection in bronchial medical images.

[0006] In a first aspect, an embodiment of the present invention provides a small target detection method for medical images, the small target detection method for medical images comprising:

[0007] Acquiring a bronchial medical image, and preprocessing the bronchial medical image to obtain a preprocessed bronchial medical image;

[0008] Construct a small target detection model for bronchial medical images comprising a backbone network, a neck network and a prediction network, introduce a residual network and a dilated spatial convolutional pooling pyramid into the backbone network, and introduce a global attention module into the neck network;

[0009] Inputting the preprocessed bronchial medical image into the backbone network, using the residual network to supplement detail features and using the atrous spatial convolutional pooling pyramid to enhance the receptive field, so as to obtain feature maps of multiple scales output by the backbone network;

[0010] Inputting the feature maps of various scales into the neck network, and using the global attention module to perform global information enhancement to obtain tensor data of various scales output by the neck network;

[0011] The prediction network is used to perform feature fusion on the tensor data of multiple scales to obtain fused features, and small target detection is performed on the fused features to obtain small target detection results of bronchial medical images.

[0012] Compared with the prior art, the first aspect of the present invention has the following beneficial effects:

[0013] This method can enhance the bronchial medical image and improve the visual effect of the image by preprocessing the bronchial medical image; construct a small target detection model for bronchial medical images including a backbone network, a neck network and a prediction network, introduce a residual network and a dilated spatial convolution pooling pyramid into the backbone network, introduce a global attention module into the neck network, input the preprocessed bronchial medical image into the backbone network, use a residual network to supplement the detail features and use a dilated spatial convolution pooling pyramid to enhance the receptive field, input feature maps of multiple scales into the neck network, and use a global attention module to enhance global information, which can provide a good data basis for subsequent small target detection to improve the accuracy and comprehensiveness of small target detection; use a prediction network to perform feature fusion on tensor data of multiple scales to obtain fused features, perform small target detection on the fused features, and obtain small target detection results of bronchial medical images. Using fused features for small target detection can improve the accuracy of small target detection in bronchial medical images.

[0014] According to some embodiments of the present invention, preprocessing the bronchial medical image to obtain the preprocessed bronchial medical image includes:

[0015] The bronchial medical image is preprocessed by using an adaptive contrast pairing method to obtain a preprocessed bronchial medical image.

[0016] According to some embodiments of the present invention, the using the residual network to supplement detail features includes:

[0017] Introducing a residual network between the first convolutional layer and the second CSP network layer in the backbone network;

[0018] Convolving the preprocessed bronchial medical image using the convolution layer to obtain a convolved bronchial medical image;

[0019] Using the CSP network layer to extract detail features from the convolved bronchial medical image to obtain a detail feature map;

[0020] The convolved bronchial medical image and the detail feature map are input to the next convolution layer to supplement the detail features of the image of the next convolution layer.

[0021] According to some embodiments of the present invention, the method of enhancing the receptive field by using the atrous spatial convolutional pooling pyramid includes:

[0022] Using multiple convolutional layers and CSP network layers to extract features of multiple scales from the preprocessed bronchial medical image to obtain first feature maps of multiple scales;

[0023] The atrous spatial convolutional pooling pyramid is used to enhance the receptive field of the first feature maps of various scales to obtain second feature maps of various scales.

[0024] According to some embodiments of the present invention, the adopting the global attention module to perform global information enhancement includes:

[0025] Introducing a global attention module connected to the second CSP network layer in the neck network;

[0026] The global attention module is used to perform global information enhancement on the features of the second CSP network layer.

[0027] According to some embodiments of the present invention, the using the prediction network to perform feature fusion on the tensor data of multiple scales to obtain fusion features, and performing small target detection on the fusion features to obtain small target detection results of bronchial medical images includes:

[0028] Build a prediction network that includes adaptive spatial feature fusion and multiple detection layers;

[0029] Using the adaptive spatial feature fusion to perform feature fusion on the tensor data of the multiple scales to obtain fusion features;

[0030] A plurality of the detection layers are used to perform small target detection on the fusion features to obtain a small target detection result of the bronchial medical image.

[0031] According to some embodiments of the present invention, before using the prediction network to perform small target detection on the fusion features, the small target detection method for medical images further includes training the small target detection model using a mish loss function.

[0032] In a second aspect, an embodiment of the present invention further provides a small target detection system for medical images, the small target detection system for medical images comprising:

[0033] An image processing unit, used for acquiring a bronchial medical image, and preprocessing the bronchial medical image to obtain a preprocessed bronchial medical image;

[0034] A model building unit, used to build a small target detection model for bronchial medical images comprising a backbone network, a neck network and a prediction network, and introduce a residual network and a dilated spatial convolutional pooling pyramid into the backbone network, and introduce a global attention module into the neck network;

[0035] A feature extraction unit, used for inputting the preprocessed bronchial medical image into the backbone network, and using the residual network to supplement the detail features and using the atrous spatial convolutional pooling pyramid to enhance the receptive field, so as to obtain feature maps of multiple scales output by the backbone network;

[0036] An information enhancement unit, used for inputting the feature maps of various scales into the neck network, and performing global information enhancement using the global attention module to obtain tensor data of various scales output by the neck network;

[0037] The target detection unit is used to use the prediction network to perform feature fusion on the tensor data of multiple scales to obtain fused features, and perform small target detection on the fused features to obtain small target detection results of bronchial medical images.

[0038] In the third aspect, an embodiment of the present invention also provides a small target detection device for medical images, comprising at least one control processor and a memory for communicating with the at least one control processor; the memory stores instructions executable by the at least one control processor, and the instructions are executed by the at least one control processor so that the at least one control processor can execute a small target detection method for medical images as described above.

[0039] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the small target detection method for medical images as described above.

[0040] It can be understood that the beneficial effects of the second to fourth aspects compared with the related art are the same as the beneficial effects of the first aspect compared with the related art. Please refer to the relevant description in the first aspect, and no further details will be given here. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0042] Figure 1 is a flow chart of a small target detection method for medical images according to an embodiment of the present invention;

[0043] Figure 2 is a schematic diagram of the architecture of small target detection in bronchial medical images according to an embodiment of the present invention;

[0044] Figure 3 This is a graph showing comparative experimental results of different modified map0.5 of the same model in one embodiment of the present invention;

[0045] Figure 4 is a graph showing comparative experimental results of different modified map0.5 of different models in one embodiment of the present invention;

[0046] Figure 5 This is a comparative experimental result diagram of different modified map0.5:0.95 of the same model in one embodiment of the present invention;

[0047] Figure 6 This is a comparative experimental result diagram of different modified map0.5:0.95 of different models in one embodiment of the present invention;

[0048] Figure 7 This is a graph showing the experimental results of different R modifications of the same model according to an embodiment of the present invention;

[0049] Figure 8 is a graph showing comparative experimental results of different modified Rs of different models according to an embodiment of the present invention;

[0050] Fig. 9 is a graph showing comparative experimental results of different modifications P of different models according to an embodiment of the present invention;

[0051] Fig.10 This is a graph showing the results of a comparative experiment of different modifications P of the same model according to an embodiment of the present invention;

[0052] Fig.11 is a structural diagram of a small object detection system for medical images according to an embodiment of the present invention;

[0053] Fig.12 It is a schematic structural diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0054] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0055] In the description of the present invention, if there is a description of first, second, etc., it is only for the purpose of distinguishing technical features, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.

[0056] In the description of the present invention, it should be understood that descriptions involving orientation, such as orientation or positional relationship indicated as up, down, etc., are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present invention.

[0057] In the description of the present invention, it should be noted that, unless otherwise clearly defined, terms such as setting, installing, connecting, etc. should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above terms in the present invention in combination with the specific content of the technical solution.

[0058] Since the traditional CNN usually consists of an input layer, a hidden layer and an output layer connected in series. The input layer is responsible for receiving the input image, the hidden layer usually consists of multiple convolutional layers and pooling layers, whose function is to extract feature information from the input image, and the output layer is responsible for outputting the classification result. The traditional CNN structure can gradually extract high-level semantic features from the image. However, the traditional CNN method has the following disadvantages: The traditional CNN uses highly semantic information feature maps to extract the final classification prediction. Although it can effectively grasp the overall characteristics of the image, it also ignores some detailed information about the connection between layers, resulting in a decrease in classification accuracy.

[0059] To solve the above problems, the present invention can enhance the bronchial tuberculosis bronchial medical images and improve the visual effect of the images by preprocessing the bronchial tuberculosis bronchial medical images; construct a small target detection model for bronchial tuberculosis bronchial medical images comprising a backbone network, a neck network and a prediction network, introduce a residual network and a dilated space convolution pooling pyramid into the backbone network, introduce a global attention module into the neck network, input the preprocessed bronchial medical image into the backbone network, use the residual network to supplement the detail features and use the dilated space convolution pooling pyramid to enhance the receptive field, input feature maps of multiple scales into the neck network, and use the global attention module to enhance the global information, which can provide a good data basis for the subsequent small target detection, so as to improve the accuracy and comprehensiveness of the small target detection; use the prediction network to perform feature fusion on tensor data of multiple scales to obtain fused features, perform small target detection on the fused features, and obtain the small target detection results of the bronchial medical images. The use of fused features for small target detection can improve the accuracy of small target detection of bronchial medical images.

[0060] Reference Figure 1 The embodiment of the present invention provides a small target detection method for medical images. The small target detection method for medical images includes but is not limited to steps S100 to S500, wherein:

[0061] Step S100, acquiring a bronchial medical image, and preprocessing the bronchial medical image to obtain a preprocessed bronchial medical image;

[0062] Step S200, constructing a small target detection model for bronchial medical images including a backbone network, a neck network and a prediction network, and introducing a residual network and a dilated spatial convolutional pooling pyramid into the backbone network, and introducing a global attention module into the neck network;

[0063] Step S300, inputting the preprocessed bronchial medical image into the backbone network, and using the residual network to supplement the detail features and using the dilated spatial convolutional pooling pyramid to enhance the receptive field, so as to obtain feature maps of multiple scales output by the backbone network;

[0064] Step S400: inputting feature maps of various scales into the neck network, and using a global attention module to perform global information enhancement to obtain tensor data of various scales output by the neck network;

[0065] Step S500: using a prediction network to perform feature fusion on tensor data of multiple scales to obtain fused features, and performing small target detection on the fused features to obtain small target detection results of bronchial medical images.

[0066] In this embodiment, in order to improve the visual effect of the image, this embodiment obtains a bronchial medical image and preprocesses the bronchial medical image to obtain a preprocessed bronchial medical image; in order to improve the accuracy and comprehensiveness of small target detection, this embodiment constructs a small target detection model for bronchial medical images including a backbone network, a neck network and a prediction network, introduces a residual network and a dilated spatial convolution pooling pyramid into the backbone network, introduces a global attention module into the neck network, inputs the preprocessed bronchial medical image into the backbone network, and uses a residual network to supplement detail features and a dilated spatial convolution pooling pyramid to enhance the receptive field to obtain feature maps of multiple scales output by the backbone network, inputs the feature maps of multiple scales into the neck network, and uses a global attention module to perform global information enhancement to obtain tensor data of multiple scales output by the neck network, uses a prediction network to perform feature fusion on tensor data of multiple scales to obtain fused features, and performs small target detection on the fused features to obtain small target detection results of bronchial medical images.

[0067] In some embodiments, preprocessing the bronchial medical image to obtain the preprocessed bronchial medical image includes:

[0068] The bronchial medical image is preprocessed by using an adaptive contrast pairing method to obtain a preprocessed bronchial medical image.

[0069] In this embodiment, an adaptive contrast pairing method is used to preprocess the bronchial medical image, and a histogram equalization algorithm, a sharpening algorithm, etc. are used for the input bronchial medical image. By continuously iterating and optimizing parameters, the adaptive contrast pairing method (ACE algorithm) can find the optimal parameter combination, thereby optimizing the enhancement effect of the bronchial medical image and making the features of the bronchial medical image in the bronchial medical image more obvious.

[0070] In some embodiments, a residual network is used to supplement detail features, including:

[0071] Introduce a residual network between the first convolutional layer and the second CSP network layer in the backbone network;

[0072] The preprocessed bronchial medical image is convolved by using a convolution layer to obtain a convolved bronchial medical image;

[0073] The CSP network layer is used to extract detail features of the convolutional bronchial medical image to obtain a detail feature map;

[0074] The convolutional bronchial medical image and detail feature map are input into the next convolutional layer to supplement the detail features of the image of the next convolutional layer.

[0075] In this embodiment, the preprocessed bronchial medical image is input into the improved backbone network. The improved backbone network refers to the connection of the residual network between the output of the first convolution layer of the backbone network (Backbone network) and the second CSP structure. The residual network is mainly set up to give the most original image information to the next convolution layer to avoid the loss of information due to a convolution of the image.

[0076] In some embodiments, a dilated spatial convolutional pooling pyramid is used to enhance the receptive field, including:

[0077] Multiple convolutional layers and CSP network layers are used to extract features of various scales from the preprocessed bronchial medical images to obtain first feature maps of various scales;

[0078] The atrous spatial convolutional pooling pyramid is used to enhance the receptive field of the first feature maps of various scales and obtain the second feature maps of various scales.

[0079] In this embodiment, the spatial pyramid pooling (SPP) is changed to the atrous spatial convolution pooling pyramid (ASPP) in Backbone. ASPP introduces atrous convolution on the basis of SPP. The atrous spatial convolution pooling pyramid performs an enlarged convolution operation on the image processed by multiple convolutions and CSP modules through the atrous spatial convolution pooling pyramid, thereby extracting larger-scale features and then transmitting the image features downward. By introducing holes in the convolution operation, the receptive field of the convolution operation can be expanded, thereby extracting larger-scale features. ASPP adds multi-scale information extraction capabilities on the basis of SPP, making the network more accurate and comprehensive in detecting objects of different scales.

[0080] In some embodiments, a global attention module is used to enhance global information, including:

[0081] Introduce a global attention module connected to the second CSP network layer in the neck network;

[0082] The global attention module is used to enhance the global information of the features of the second CSP network layer.

[0083] In this embodiment, a global attention module is added to the neck network part, and the management of global information of features is improved through the global attention module to enhance the information of the first layer of the detection layer (Detect).

[0084] In some embodiments, a prediction network is used to perform feature fusion on tensor data of multiple scales to obtain fused features, and small target detection is performed on the fused features to obtain small target detection results of bronchial medical images, including:

[0085] Build a prediction network that includes adaptive spatial feature fusion and multiple detection layers;

[0086] Adaptive spatial feature fusion is used to fuse the tensor data of various scales to obtain fused features;

[0087] Multiple detection layers are used to perform small target detection on the fused features, and the small target detection results of bronchial medical images are obtained.

[0088] In this embodiment, adaptive spatial feature fusion is used to fuse tensor data of multiple scales, which can integrate and adjust features of other levels to the same resolution, and then train to find the optimal fusion. Multiple detection layers are used to detect small targets on the fused features, which can improve the accuracy of small target detection in bronchial medical images.

[0089] In some embodiments, before using the prediction network to perform small target detection on the fused features, the small target detection method for medical images also includes training the small target detection model using a mish loss function.

[0090] In this embodiment, the mish loss function is used to train the small target detection model, which can increase the speed of obtaining positive samples and improve the convergence speed of the model.

[0091] To facilitate understanding by those skilled in the art, a set of best embodiments is provided below:

[0092] In this embodiment, the basic network architecture and core network components are mainly introduced, and then a series of improvement schemes are formulated for the prediction results of bronchial medical images. Figure 2 This embodiment provides an improved model architecture and elaborates on some implementation details. Specifically:

[0093] 1. Image data acquisition.

[0094] The lung CT images collected in this embodiment are all early hospitalized lung CT images. The experimental data needs to go through a series of processing from acquisition to real input model: first, a series of screening of the patient's CT is performed using dedicated medical image software. The main purpose is to find images containing bronchial tuberculosis lesions (i.e., bronchial medical images). At the same time, dedicated medical image software is also needed to eliminate patient information. Each lung CT image contains detailed information such as the patient's name, age, and gender; then the selected data is marked accordingly. The annotation tool software used in this embodiment is separated by tags.

[0095] 2. Image data processing.

[0096] Adaptive contrast algorithm (ACE) is used to enhance the captured image to enhance the contrast and details of the image. It is an adaptive enhancement method that adjusts the contrast according to the local features of the image to improve the visual effect of the image. Adaptive contrast enhancement technology can enhance the local contrast according to the brightness difference of different areas in the image. ACE mainly includes three steps: 1. Obtain local background information; 2. Calculate standard deviation; 3. Image enhancement. The specific calculation process is as follows:

[0097]

[0098] f(i,j)=m x (i,j)+G(i,j)[x(i,j)-m x (i,j)]

[0099] Among them, x(k,l) represents the pixel value of the coordinate (k,l), m x (i,j) represents the area centered at point (i,j) with a window size of (2n+1)*(2n+1), f(i,j) represents the enhanced pixel, G(i,j) represents the gain, and x(i,j) represents the pixel point at coordinates (i,j).

[0100] 3. The model structure of this embodiment.

[0101] The model is constructed based on the network architecture of YOLO v5. The characteristics of this network structure are that it is suitable for mobile terminal deployment, small model and fast speed. YOLO v5 has four versions: YOLO v5s, YOLO v5m, YOLO v5l, and YOLO v5x. In this embodiment, the structures of these models are basically the same, except that the two parameters of depth_multiple model depth and width_multiple model width are different. The network structure is mainly composed of four parts: input end, Backbone network, Neck network and Prediction network (i.e. prediction network), and the basic components include CBL, CSP1_X, SPP. Specifically include the following contents:

[0102] 1) Input end.

[0103] YOLOv5s uses Mosaic data enhancement at the input end, referring to the CutMix data enhancement method. Mosaic data enhancement increases the number of images from the original two to four for splicing, and randomly scales, randomly crops, and randomly arranges the images. Using data enhancement can improve the problem of imbalanced data among small, medium, and large targets in the data set. The main steps of Mosaic data enhancement are to first randomly select the coordinates of the image splicing reference points, and then randomly select four images. According to the reference points, they are resized and scaled, and placed at the four corners of a large image of a specified size, and then spliced ​​according to the specified horizontal coordinates. There are several advantages to using Mosaic data enhancement:

[0104] (1) Enriching the dataset: We randomly use four images, randomly scale them, and randomly splice them, adding many small objects, which greatly enriches the dataset and improves the robustness of the network.

[0105] (2) Reduce GPU usage: The random splicing method allows one image to calculate the data of four images, reducing the number of each batch. Even with only one GPU, better results can be obtained.

[0106] (3) At the same time, by cropping the identified objects, the model can identify objects based on local features, which helps to detect occluded objects and thus improves the detection ability of the model.

[0107] 2)Backbone network.

[0108] The backbone network mainly includes CBL block, CSP1_X block and SPP module. Among them:

[0109] (1) CBL module.

[0110] The CBL module consists of a convolutional layer, a BN layer, and an activation function. The main function of the BN layer is to prevent overfitting and accelerate convergence, while the activation function mainly increases the nonlinear changes of the data.

[0111] (2)CSP1_X module.

[0112] The structure of the CSP1_X module is mainly composed of a Conv module with a residual structure. This module is the main module for learning residual features. Its structure is divided into two branches. One uses the above-specified multiple residual stacks and then undergoes a convolution. The other only passes through a basic convolution module. Finally, the two branches are concat-operated and then pass through the CBL module. This is to allow the model to learn more features. This method is mainly to divide the feature map into two parts and merge them through the CSP network layer. At the same time, the internal residual network can prevent the problem of gradient disappearance or explosion caused by the poor effect of a certain layer during the training of the network model. At the same time, it solves the problem of gradient information duplication in the network model, reduces the amount of calculation, and enriches the gradient information of this layer.

[0113] (3)ASPP module.

[0114] SPP is a spatial pyramid layer, which can convert feature maps of any size into feature vectors of fixed size to increase the receptive field. In this embodiment, spatial pyramid pooling (SPP) is changed to atrous spatial convolution pooling pyramid (ASPP). ASPP introduces atrous convolution on the basis of SPP. Atrous spatial convolution pooling pyramid performs an enlarged convolution operation on the image processed by multiple convolutions and CSP modules through atrous spatial convolution pooling pyramid, thereby extracting larger-scale features and then transmitting the image features downward. By introducing holes in the convolution operation, the receptive field of the convolution operation can be expanded, thereby extracting larger-scale features. ASPP adds multi-scale information extraction capabilities on the basis of SPP, making the network more accurate and comprehensive in detecting objects of different scales.

[0115] 3) Neck feature pyramid.

[0116] Since the size and position of objects in an image are uncertain, a mechanism is needed to handle targets of different scales and sizes. Feature pyramid is a technique for handling multi-scale target detection, which can be achieved by adding feature layers of different scales to the backbone network. In YOLO v5, the FPN (Feature Pyramid Network) feature pyramid structure is used to fuse feature maps of different levels together through upsampling and downsampling operations to generate a multi-scale feature pyramid. The top-down part mainly achieves the fusion of features at different levels by fusing upsampling with coarser-grained feature maps, while the bottom-up part fuses feature maps from different levels by using a convolutional layer.

[0117] The top-down part mainly involves first upsampling the feature map of the last layer to obtain a deeper feature map, then fusing the upsampled feature map with the previous layer to obtain a richer feature expression, and repeating this step until the highest layer. The bottom-up part mainly involves convolving the bottom layer and then fusing it with the previous layer to obtain the final feature map, which is used for detection.

[0118] 4) Prediction network.

[0119] The prediction network (i.e., the target detection head) is used to detect targets on the feature pyramid. It includes some convolutional layers, pooling layers, and fully connected layers. In the YOLO v5 model, the target detection head module is mainly responsible for multi-scale target detection on the feature maps extracted by the backbone network. This module mainly consists of three parts:

[0120] Anchors: used to define target boxes of different sizes and aspect ratios. They are usually obtained by clustering the target boxes of the training set using K-means clustering. They can be calculated before model training and stored in the model to generate detection boxes during prediction.

[0121] Classification: It is used to classify each detection box to determine whether it is a target object. It usually uses a fully connected layer plus a Softmax function to classify features.

[0122] Regression: It is used to regress each detection box to obtain its position and size. It usually regresses the features in the form of a fully connected layer.

[0123] 4. Model training strategy.

[0124] The deep learning-based small target detection method for bronchial medical images proposed in this embodiment adopts the model structure of YOLOv5s-GAM-ASFF-Mish-ACE+ASPP+residual, which includes several improved methods. The residual network and the dilated spatial convolution pooling pyramid are introduced into the backbone network; the GAM attention mechanism is added to the neck network to improve the performance of the deep neural network by reducing information and amplifying the global interaction representation; ASFF is used in the prediction network to supplement multi-level information, and then the final prediction classification is performed, thereby increasing the amount of information between each layer of the network information and increasing the connection between each layer. The complete architecture of several improvements of the YOLOv5s-GAM-ASFF-Mish-ACE+ASPP+residual model proposed in this embodiment is as follows: Figure 2 As shown in the figure, the global attention module (GAM) and adaptive spatial feature fusion (ASFF) specifically include the following contents:

[0125] 1) Global attention module.

[0126] GAM proposes a three-dimensional channel focusing submodule based on multi-layer perceptron based on the convolutional spatial focusing submodule. The specific calculation method is as follows:

[0127]

[0128]

[0129] Among them, F2 represents the intermediate state, M c represents the channel attention map, F1 represents the input feature, Indicates element-wise multiplication operation, F3 indicates output, M s Represents the spatial attention map.

[0130] Considering the importance of interactions between dimensions, the Triple Attention Module (TAM) improves efficiency by weighting the channel, spatial width, and spatial height. The global attention module contains the Channel attention module and the spatial attention module, specifically:

[0131] Channel attention module: Utilizes the inter-channel relationship between features to generate a channel attention map. Since each channel of the feature map is considered as a feature detector, it makes sense to focus on which channels for a given input image. The Channel attention module uses 3D permutation to preserve three-dimensional information, and then uses two layers of MLP (multi-layer perceptron) to amplify cross-dimensional channels and spatial dependencies.

[0132] Spatital attention module: focuses on spatial information and uses two convolutional layers to fuse spatial information. At the same time, the maximum pooling reduces the information and contributes negatively to the information. The spatital attention module of this embodiment removes the pooling to further retain the feature information.

[0133] The main purpose of introducing GAM is to improve the feature expression ability of the model. By using the attention mechanism, the model's perception of the image is enhanced through the processing of Channel attention and spatial attention, so that the model can better understand the important channels and position information in the image, which is conducive to improving the accuracy and stability of the model's feature extraction.

[0134] 2) Adaptive spatial feature fusion technology.

[0135] Pyramid feature representation is a common method to solve the challenge of scale variation in target detection. However, for single-stage detectors based on FPN, the inconsistency between different feature scales is its main limitation. Therefore, this embodiment proposes a new data-driven pyramid feature fusion method, called adaptive spatial feature fusion (ASFF). It learns a method to filter conflicting information spatially to suppress inconsistencies during gradient backpropagation, thereby improving the scale invariance of features and reducing inference overhead.

[0136] This method enables the network to directly learn how to spatially filter features at other levels so that only useful information is retained for combination. For features at one level, features at other levels are first integrated and adjusted to the same resolution, and then trained to find the optimal fusion.

[0137] At each spatial location, features at different levels are adaptively fused, i.e., some features may be filtered out because they carry contradictory information at this location, and some features may dominate with more discriminative clues. ASFF has several advantages:

[0138] (1) Since the operation of searching for the optimal fusion is differentiable, it can be easily learned using the BP algorithm;

[0139] (2) It is independent of the basic network and can be applied to all single-stage detectors with feature pyramid structures;

[0140] (3) The implementation is simple and the additional computational cost is very small.

[0141] Adaptive Spatial Feature Fusion Mechanism For each layer, features from all other levels are resized to the same shape and spatially fused according to the learned weight map. Different from the previous method that uses element-level sum or concatenation to integrate multi-level features, the key idea of ​​the adaptive spatial feature fusion mechanism is to adaptively learn the spatial weights for the fusion of feature maps at each scale. It includes two steps: identical rescaling and adaptive fusion.

[0142] The feature fusion is performed through the following formula:

[0143]

[0144] in, and Represent the features of level 1, level 2 and level 3 respectively. and Represents the weight parameter.

[0145] Because the addition method is used, the feature size of the output of the level1 to level3 layers needs to be the same when adding, and the number of channels must also be the same. It is necessary to upsample or downsample the features of different layers and adjust the number of channels. For layers that need upsampling, such as ASFF3, level1 needs to be adjusted to the same size as level3. The method used is to first adjust it to the same number of channels as level3 through 1×1 convolution, and then resize it to the same size by interpolation; for layers that need downsampling, such as ASFF1, only a 3×3 convolution with stride=2 is needed for level2 to level1. If it is level3 to level1, a maxpooling with stride=2 needs to be added on top of the 3×3 convolution, so that the size of level3 and level1 can be adjusted to be consistent.

[0146] For the weight parameter and It is obtained by performing a 1×1 convolution on the resized feature maps from level 1 to level 3. And the parameters and After concat, softmax is used to make their range within [0,1] and the sum is 1:

[0147]

[0148] The specific steps of adaptive spatial feature fusion are:

[0149] 1. First, output the l-level feature map, and upsample and downsample the remaining feature maps to obtain feature maps of the same size for subsequent fusion;

[0150] 2. Output the processed three-layer feature map and input 1x1 convolution to obtain three spatial weight vectors;

[0151] 3. Splice the channel direction to obtain the weight fusion map. In order to obtain the weight map with channel 3, perform a 1x1x3 convolution on the above feature map;

[0152] 4. Get the weight vector of the channel softmax operation direction and normalize it;

[0153] 5. Add the 3-vector multiplication to the 3 feature maps to get the fused features Figure 3 The convolution of x3 results in a prediction output layer with 256 output channels.

[0154] For better explanation, the following experiments were performed in this embodiment:

[0155] 1. Evaluation indicators.

[0156] In the object detection of this embodiment, a strategy is used to convert it into a binary classification problem. This means that one class contains positive samples and the other class contains negative samples.

[0157] TP: The number of positive classes correctly predicted as positive.

[0158] FN: The number of positive classes that were incorrectly predicted as negative.

[0159] FP: The number of negative classes that were incorrectly predicted as positive.

[0160] TN: The number of negative classes correctly predicted as negative.

[0161] P (Precision): Precision represents the correct proportion of all targets predicted by the model. It is a measure of accuracy and helps to emphasize the correlation between results. Its calculation formula is as follows:

[0162]

[0163] R (Recall): Recall is the ratio of correctly predicted true (positive) targets to all true targets. Precision and recall are two often conflicting metrics. Generally speaking, when precision is high, recall tends to be low, and vice versa. In other words, when the model has high accuracy, recall may be low, and when recall is high, recall may be low. Its calculation formula is as follows:

[0164]

[0165] The f1 score is defined as the harmonic mean of precision and recall. It is a single metric that combines precision and recall into a single value. It is calculated as follows:

[0166]

[0167] mAP (average precision): It is mapped by calculating the area under the precision and recall curves. Precision refers to the proportion of detected objects that actually belong to this category. The calculation process is to first calculate the area under the precision and recall curves of each category, and then average the areas of all categories. The final result is the average precision. mAP0.5 means the average precision with a relevant threshold of 0.5. mAP 0.5:0.95 means the average precision of the average of the relevant thresholds ld from 0.5 to 0.95. This means that for each query, a relevant threshold from 0.5 to 0.95 is considered, and then the average precision in this range is calculated. The calculation formula for average precision is as follows:

[0168]

[0169] 2. Experimental environment setup.

[0170] The environment of this experiment is based on a windows11 and PyTorch deep learning development framework, and the development language is python. The GPU used in the experiment is NVIDIA GeForce RTX 3060Ti. During the training process, Adam was selected as the optimizer. In order to better detect the lesions in this experiment, the batch size is set to 8, the numerical workers are set to 8, the initial learning rate is set to 0.01, and the epoch is set to 300. The specific experimental steps are as follows:

[0171] (1) This experiment uses two-dimensional slices as training data. In order to ensure the consistency of the input image size of the classifier, these slices are unified.

[0172] (2) Image preprocessing includes format conversion, grayscale normalization, MRI slicing, uniform size, etc.

[0173] (3) An experimental platform for the system was established, and the GAM-ASFF-Yolov5s network model was designed. The data set was randomly input into the training set for training and the training results were obtained.

[0174] (4) Save the optimal model parameters and test the model.

[0175] 3. Test the performance of the model in this embodiment.

[0176] This embodiment describes a lung CT image of a patient with bronchial tuberculosis. Although lesions can be quickly found through the doctor's CT image, the demand for doctors has increased. In addition, for patients who have not yet developed the disease, doctors will not consider CT examinations for bronchial tuberculosis, thus missing the best time for treatment. Therefore, it is an expensive challenge for hospitals to have the number and cost of doctors, and it is also a difficult challenge for patients to get treatment as soon as possible. More and more effective data are generated in clinical routine, although the interaction differences between highly trained operators cannot be ignored. Traditional convolutional neural networks CNNS are only trained on routinely collected data and can also be promoted well. However, the accuracy of the traditional model itself cannot guarantee the effectiveness of the model. At the same time, for many current models, the results of bronchial tuberculosis foci examinations are uneven and the accuracy is not high. This embodiment has improved and trained the best models on the market.

[0177] To illustrate the effectiveness of this method, this embodiment selects Yolov5s, Yolov5s-ASFF, Yolov5s-GAM, Yolov5s-GAM-ASFF-Mish, and finally modifies the activation function Yolov5s-GAM-ASFF-Mish. At the same time, data enhancement ACE is used to experiment with the same data set. In 300 rounds of training, the model of this embodiment and other models are compared by precision, recall, and average precision to show the accuracy of each model and verify the effectiveness of the model through the validation set. This embodiment selects three more advanced models for performance comparison on the same data set and uses the same training strategy. The more advanced models include Yolov7, Marsk, and Yolov8. In 300 rounds of training, from Figure 3 , Figure 4 , Figure 5 , Figure 6 , Figure 7 , Figure 8 , Fig. 9 and Fig.10 It can be seen from the experimental comparison results that the improved model of this embodiment (i.e., YOLOv5s-GAM-ASFF-Mish-ACE+ASPP+residual) has better performance than other original models during the training process, and the improved model of this embodiment can be compared with some of the most advanced models at present, and has certain improvements.

[0178] Specifically, compared with the basic model, the improved model of this embodiment has obvious improvements in various classification evaluation indicators. For some difficult-to-distinguish pictures, the improved model has higher classification accuracy. Due to the limitation of the data set size, the performance of the model is best when the ASFF module is introduced and the GAM module is added. Although R is reduced, mAP 0.5 is 7% higher than the original model, mAP 0.5:0.95 is 13.7% higher than the original model, and P is 7.5% higher than the original model.

[0179] As can be seen from Table 1 to Table 2, the performance of all models is very similar, among which the improved model of this embodiment has the best overall performance. Therefore, the improved model of this embodiment can accurately identify the changes in the image of bronchial tuberculosis in the lung CT image. It has a high and accurate ability to identify pathological lesions and can provide objective auxiliary diagnosis suggestions for areas with limited medical resources. This can effectively contribute to the treatment of patients and has important practical value.

[0180] Table 1

[0181]

[0182] Table 2

[0183]

[0184] Reference Fig.11 The embodiment of the present invention further provides a small target detection system for medical images. The small target detection system for medical images includes an image processing unit 100, a model building unit 200, a feature extraction unit 300, an information enhancement unit 400 and a target detection unit 500, wherein:

[0185] The image processing unit 100 is used to acquire a bronchial medical image and preprocess the bronchial medical image to obtain a preprocessed bronchial medical image;

[0186] A model building unit 200 is used to build a small target detection model for bronchial medical images including a backbone network, a neck network and a prediction network, and introduce a residual network and a dilated spatial convolutional pooling pyramid into the backbone network, and introduce a global attention module into the neck network;

[0187] The feature extraction unit 300 is used to input the preprocessed bronchial medical image into the backbone network, and use the residual network to supplement the detail features and use the dilated spatial convolutional pooling pyramid to enhance the receptive field, so as to obtain feature maps of multiple scales output by the backbone network;

[0188] An information enhancement unit 400 is used to input feature maps of various scales into the neck network and perform global information enhancement using a global attention module to obtain tensor data of various scales output by the neck network;

[0189] The target detection unit 500 is used to use a prediction network to perform feature fusion on tensor data of multiple scales to obtain fused features, and perform small target detection on the fused features to obtain small target detection results of bronchial medical images.

[0190] It should be noted that since the small target detection system for medical images in this embodiment and the small target detection method for medical images mentioned above are based on the same inventive concept, the corresponding contents in the method embodiment are also applicable to the system embodiment and will not be described in detail here.

[0191] Reference Fig.12 The embodiment of the present application further provides a small target detection device for medical images, and the small target detection device for medical images includes:

[0192] at least one memory;

[0193] at least one processor;

[0194] at least one program;

[0195] The programs are stored in the memory, and the processor executes at least one program to implement the small target detection method for medical images described above in the present disclosure.

[0196] The electronic device may be any intelligent terminal including a mobile phone, a tablet computer, a personal digital assistant (PDA), a car computer, etc.

[0197] The electronic device according to the embodiment of the present application is described in detail below.

[0198] The processor 1600 may be implemented by a general-purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided by the embodiments of the present disclosure;

[0199] The memory 1700 may be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 1700 may store an operating system and other application programs. When the technical solution provided in the embodiments of this specification is implemented by software or firmware, the relevant program code is stored in the memory 1700, and the processor 1600 calls and executes the small target detection method for medical images in the embodiments of the present disclosure.

[0200] Input / output interface 1800, used to implement information input and output;

[0201] The communication interface 1900 is used to realize the communication interaction between the device and other devices. The communication can be realized through a wired manner (such as USB, network cable, etc.) or a wireless manner (such as mobile network, WIFI, Bluetooth, etc.);

[0202] Bus 2000 , which transmits information between various components of the device (e.g., processor 1600 , memory 1700 , input / output interface 1800 , and communication interface 1900 );

[0203] The processor 1600 , the memory 1700 , the input / output interface 1800 , and the communication interface 1900 are connected to each other in communication within the device via the bus 2000 .

[0204] The embodiment of the present disclosure also provides a storage medium, which is a computer-readable storage medium. The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the above-mentioned small target detection method for medical images.

[0205] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0206] The embodiments described in the embodiments of the present disclosure are intended to more clearly illustrate the technical solutions of the embodiments of the present disclosure and do not constitute a limitation on the technical solutions provided by the embodiments of the present disclosure. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of the present disclosure are also applicable to similar technical problems.

[0207] Those skilled in the art will appreciate that the technical solutions shown in the figures do not limit the embodiments of the present disclosure and may include more or fewer steps than shown in the figures, or a combination of certain steps, or different steps.

[0208] The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separated, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0209] Those skilled in the art will appreciate that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices may be implemented as software, firmware, hardware, or a suitable combination thereof.

[0210] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0211] It should be understood that in the present application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0212] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0213] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0214] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0215] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including multiple instructions to enable an electronic device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), disk or optical disk and other media that can store programs. The above is a detailed description of the embodiments of the present application in conjunction with the accompanying drawings, but the present application is not limited to the above embodiments. Within the scope of knowledge possessed by ordinary technicians in the relevant technical field, various changes can be made without departing from the purpose of the present application.

Claims

1. A small target detection method for medical images, characterized in that: The small target detection method of the medical image comprises: Acquiring a bronchial medical image, and preprocessing the bronchial medical image to obtain a preprocessed bronchial medical image; Construct a small target detection model for bronchial medical images comprising a backbone network, a neck network and a prediction network, introduce a residual network and a dilated spatial convolutional pooling pyramid into the backbone network, and introduce a global attention module into the neck network; Inputting the preprocessed bronchial medical image into the backbone network, using the residual network to supplement detail features and using the atrous spatial convolutional pooling pyramid to enhance the receptive field, so as to obtain feature maps of multiple scales output by the backbone network; Inputting the feature maps of various scales into the neck network, and using the global attention module to perform global information enhancement to obtain tensor data of various scales output by the neck network; The prediction network is used to perform feature fusion on the tensor data of multiple scales to obtain fused features, and small target detection is performed on the fused features to obtain small target detection results of bronchial medical images.

2. The small target detection method for medical images according to claim 1, characterized in that: The preprocessing of the bronchial medical image to obtain the preprocessed bronchial medical image includes: The bronchial medical image is preprocessed by using an adaptive contrast pairing method to obtain a preprocessed bronchial medical image.

3. The small target detection method for medical images according to claim 1, characterized in that: The using the residual network to supplement detail features includes: Introducing a residual network between the first convolutional layer and the second CSP network layer in the backbone network; Convolving the preprocessed bronchial medical image using the convolution layer to obtain a convolved bronchial medical image; Using the CSP network layer to extract detail features from the convolved bronchial medical image to obtain a detail feature map; The convolved bronchial medical image and the detail feature map are input to the next convolution layer to supplement the detail features of the image of the next convolution layer.

4. The small target detection method for medical images according to claim 3, characterized in that: The method of using the dilated spatial convolutional pooling pyramid to enhance the receptive field includes: Using multiple convolutional layers and CSP network layers to extract features of multiple scales from the preprocessed bronchial medical image to obtain first feature maps of multiple scales; The atrous spatial convolutional pooling pyramid is used to enhance the receptive field of the first feature maps of various scales to obtain second feature maps of various scales.

5. The small target detection method for medical images according to claim 1, characterized in that: The adopting the global attention module to perform global information enhancement includes: Introducing a global attention module connected to the second CSP network layer in the neck network; The global attention module is used to perform global information enhancement on the features of the second CSP network layer.

6. The small target detection method for medical images according to claim 1, characterized in that: The method of using the prediction network to perform feature fusion on the tensor data of multiple scales to obtain fusion features, and performing small target detection on the fusion features to obtain small target detection results of bronchial medical images includes: Build a prediction network that includes adaptive spatial feature fusion and multiple detection layers; Using the adaptive spatial feature fusion to perform feature fusion on the tensor data of the multiple scales to obtain fusion features; A plurality of the detection layers are used to perform small target detection on the fusion features to obtain a small target detection result of the bronchial medical image.

7. The small target detection method for medical images according to claim 1, characterized in that: Before using the prediction network to perform small target detection on the fusion feature, the small target detection method for medical images also includes training the small target detection model using a mish loss function.

8. A small object detection system for medical images, characterized in that: The small target detection system for medical images comprises: An image processing unit, used for acquiring a bronchial medical image, and preprocessing the bronchial medical image to obtain a preprocessed bronchial medical image; A model building unit, used to build a small target detection model for bronchial medical images comprising a backbone network, a neck network and a prediction network, and introduce a residual network and a dilated spatial convolutional pooling pyramid into the backbone network, and introduce a global attention module into the neck network; A feature extraction unit, used for inputting the preprocessed bronchial medical image into the backbone network, and using the residual network to supplement the detail features and using the atrous spatial convolutional pooling pyramid to enhance the receptive field, so as to obtain feature maps of multiple scales output by the backbone network; An information enhancement unit, used for inputting the feature maps of various scales into the neck network, and performing global information enhancement using the global attention module to obtain tensor data of various scales output by the neck network; The target detection unit is used to use the prediction network to perform feature fusion on the tensor data of multiple scales to obtain fused features, and perform small target detection on the fused features to obtain small target detection results of bronchial medical images.

9. A small target detection device for medical images, characterized in that: It includes at least one control processor and a memory for communicating with the at least one control processor; the memory stores instructions that can be executed by the at least one control processor, and the instructions are executed by the at least one control processor so that the at least one control processor can execute the small target detection method for medical images as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the small object detection method for medical images as described in any one of claims 1 to 7.

Images