Plain scan CT image lesion detection method and related equipment based on back-projection data

Abstract

The disclosed embodiments provide a method and related equipment for detecting lesions in plain scan CT images based on back-projection data, and relate to the technical field of medical image analysis. Specifically, the VVBP data obtained based on the plain scan CT images are sparsely processed to obtain N sparse images used as model inputs, and in the processing of the model, the N sparse images can be feature encoded to obtain N first feature maps, and the image can be reconstructed based on the N first feature maps to obtain a reconstructed image, and the N first feature maps can be integrated by feature fusion to obtain a second feature map, and the first candidate box based on the first feature map and the second candidate box based on the second feature map can be fused to finally output predicted position information indicating the lesion area. The implementation of the disclosed embodiments can effectively utilize the correlation and complementary information of the full-view VVBP data to improve the detection performance.

Description

Technical Field

[0001] The present disclosure relates to the technical field of medical image analysis, and in particular to a method for detecting lesions in plain scan CT images based on back-projection data and related equipment. Background Art

[0002] In the era of digital medicine, a large number of medical images are generated every day, including X-rays, computed tomography (CT), magnetic resonance imaging (MRI), and ultrasound imaging. Clinicians need to spend a lot of time screening and evaluating these images, so the demand for intelligent devices that can assist diagnosis and improve work efficiency is becoming increasingly urgent. In recent years, thanks to the rapid development of computer vision and deep learning algorithms, artificial intelligence (AI) technology has gradually entered the medical field, especially in medical image analysis, helping doctors solve diagnostic problems and improve efficiency. Among them, medical image target detection technology can automatically identify and use bounding boxes to locate lesion areas in images, effectively reducing the risk of misdiagnosis and missed diagnosis, which is of great significance for early detection of diseases and formulation of treatment plans. For example, in lung cancer lesion detection based on CT images and breast cancer lesion detection based on X-ray images, target detection technology plays a vital role.

[0003] Non-invasive medical imaging technology has become an indispensable auxiliary means for cancer diagnosis and treatment. Compared with advanced medical imaging examination methods such as MRI and enhanced CT, plain scan CT plays an important role in basic screening in medical target detection with its relatively low cost, simple and fast examination process and the advantages of no need for contrast agents. Therefore, lesion detection technology based on plain scan CT has important research value. However, plain scan CT has certain limitations in displaying image details, such as relatively low contrast and wide imaging range, which may affect the accuracy and sensitivity of the detection results to a certain extent. In order to overcome these challenges and improve the effectiveness of plain scan CT in medical diagnosis, it is urgent to explore and develop innovative methods. Summary of the invention

[0004] In order to solve at least one of the above technical problems, the present disclosure provides a method and related equipment for detecting lesions in plain scan CT images based on back-projection data. The technical solution is as follows:

[0005] In a first aspect, an embodiment of the present disclosure provides a method for detecting lesions in a plain scan CT image based on back-projection data, comprising:

[0006] Based on the plain scan CT images, the back-projected VVBP data is obtained view by view;

[0007] Performing sparse processing on the VVBP data to obtain N sparse images;

[0008] Processing the N sparse images through a pre-trained lesion detection model to obtain predicted position information;

[0009] The process of processing the N sparse images by using a pre-trained lesion detection model to obtain predicted location information includes:

[0010] Performing feature encoding on the N sparse images to obtain N first feature maps;

[0011] Reconstructing the image based on the N first feature maps to obtain a reconstructed image;

[0012] Fusing the N first feature maps to obtain a second feature map;

[0013] Detecting lesion areas on the first feature map and the second feature map respectively to obtain N groups of first candidate frames and one group of second candidate frames;

[0014] The predicted position information is obtained based on the first candidate box and the second candidate box; the predicted position information is used to indicate the lesion area.

[0015] In a feasible embodiment, obtaining VVBP data based on the plain scan CT image includes:

[0016] Acquire a number of two-dimensional tomographic images of the target tissue structure through plain scan CT;

[0017] Simulating the two-dimensional tomographic image to obtain first VVBP data;

[0018] Performing unified standardization processing on the first VVBP data to obtain a standardized image;

[0019] Based on the set pixel size, the standardized image is resampled using a two-dimensional spline interpolation method to obtain an interpolated image with the same resolution;

[0020] Based on the size of the target tissue structure, the interpolation image is cropped to obtain second VVBP data.

[0021] In a feasible embodiment, sparse processing is performed on the VVBP data to obtain N sparse images, including:

[0022] Based on different initial views, the VVBP data is sampled at the same interval N to obtain N groups of sparse VVBP views;

[0023] For each group of sparse VVBP views, the VVBP views in the group are summed along the view direction to obtain N sparse images.

[0024] In a feasible embodiment, the fusing the N first feature maps to obtain the second feature map includes:

[0025] Performing multi-cross channel attention processing on the N first feature maps to obtain a first fusion result;

[0026] Performing multi-cross spatial attention processing on the first fusion result to obtain a second fusion result;

[0027] Based on the second fusion result, a second feature map is obtained.

[0028] In a feasible embodiment, performing multi-cross channel attention processing on the N first feature maps to obtain a first fusion result includes:

[0029] For each of the first feature maps, generate a query item, a key item, and a value item of channel attention;

[0030] For any two of the first feature graphs, based on a query item of one of the first feature graphs and a key item and a value item of another of the first feature graphs, determining a first interaction result of the two first feature graphs;

[0031] For any two of the first feature maps, a first fusion result of the two first feature maps is determined based on a first interaction result between one of the first feature maps and the two first feature maps.

[0032] In a feasible embodiment, performing multiple cross-spatial attention processing on the first fusion result to obtain the second fusion result includes:

[0033] Based on the first fusion result, generate query items, key items and value items of spatial attention;

[0034] For any two of the first fusion results, based on a query item of one of the first fusion results and a key item and a value item of another of the first fusion results, determine a second interaction result of the two first fusion results;

[0035] For any two of the first fusion results, a second fusion result of the two first fusion results is determined based on one of the first fusion results and a second interaction result of the two first fusion results.

[0036] In a feasible embodiment, the first feature map and the second feature map are respectively subjected to lesion area detection to obtain N groups of first candidate frames and a group of second candidate frames, including performing the following detection operations on the first feature map and the second feature map respectively through a region proposal network RPN:

[0037] For the received feature map, generate a number of predefined anchor boxes at each feature point of the feature map, wherein the anchor boxes include rectangular boxes with different aspect ratios and areas;

[0038] For each anchor box, determine the probability of the contained object and adjust the size and position of the anchor box through the bounding box regression adjuster;

[0039] The overlapping anchor boxes are compared through non-maximum suppression (NMS) to obtain the anchor box with the highest probability of containing the object, and the candidate box is determined based on the anchor box.

[0040] In a feasible embodiment, obtaining predicted position information based on the first candidate box and the second candidate box includes:

[0041] For N groups of the first candidate boxes, the first candidate boxes whose intersection over union (IoU) values ​​are greater than a preset first threshold are divided into the same cluster, and the bounding boxes in the cluster are merged to obtain first sub-candidate boxes corresponding to the N groups of the first candidate boxes;

[0042] For the first sub-candidate box and the second candidate box, classify the candidate boxes whose IoU values ​​are greater than a preset second threshold into the same cluster, and merge the bounding boxes in the cluster to obtain the second sub-candidate box;

[0043] The second sub-candidate box is processed by a pooling layer and a fully connected layer in the network to obtain predicted position information; the predicted position information includes predicted coordinate values ​​and predicted category probability values ​​of the bounding box.

[0044] In a feasible embodiment, the lesion detection model is trained by the following loss function value:

[0045] Determining a first loss value based on a structural similarity loss SSIM between the plain scan CT image and the reconstructed image;

[0046] Determining a second loss value between the actual position information and the predicted position information of the plain scan CT image;

[0047] Based on the first loss value and the second loss value, a loss function value in the lesion detection model training process is determined.

[0048] In a feasible embodiment, determining the second loss value between the real position information and the predicted position information of the plain scan CT image includes:

[0049] Based on the predicted category probability value, predicted coordinate value and true coordinate value of the candidate box obtained by RPN, the third loss value is determined according to the same method as the loss calculation method of the second stage detector;

[0050] In the second stage detector of the lesion detection model, the second loss value finally used to train the lesion detection model is determined based on the binary cross entropy loss value and the L1 norm loss value calculated when removing the false positive candidate boxes, together with the third loss value.

[0051] In a second aspect, an embodiment of the present disclosure provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method described in the first aspect and any embodiment thereof.

[0052] In a third aspect, an embodiment of the present disclosure provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method described in the first aspect and any embodiment thereof.

[0053] In a fourth aspect, an embodiment of the present disclosure provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the method described in the first aspect and any embodiment thereof.

[0054] The technical solution provided by the embodiments of the present disclosure brings the following beneficial effects:

[0055] The disclosed embodiment provides a method for detecting lesions in plain scan CT images based on back-projection data. Specifically, the disclosed embodiment performs sparse processing on VVBP data obtained based on plain scan CT images to obtain N sparse images used as model inputs, and in the processing of the model, feature encoding can be performed on the N sparse images to obtain N first feature maps, and the image can be reconstructed based on the N first feature maps to obtain a reconstructed image, and the N first feature maps can be integrated by feature fusion to obtain a second feature map, and the first candidate box of the first feature map and the second candidate box of the second feature map can be fused, and finally output predicted position information for indicating the lesion area. The implementation of the disclosed embodiment can effectively utilize the correlation and complementary information of the full-view VVBP data to improve the detection performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure, the drawings required for describing the embodiments of the present disclosure are briefly introduced below.

[0057] Figure 1 A flowchart of a method for detecting lesions in plain scan CT images based on back-projection data provided by an embodiment of the present disclosure;

[0058] Figure 2 A schematic diagram of a plain scan CT lesion detection model based on VVBP data provided in an embodiment of the present disclosure;

[0059] Figure 3Shown is a schematic diagram of the mcCSA module in the model proposed in the embodiment of the present disclosure;

[0060] Figure 4 A visualization image of different ablation experiment results provided by an embodiment of the present disclosure;

[0061] Figure 5 A visualized output feature map of mcCSA with different inputs provided in an embodiment of the present disclosure;

[0062] Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0063] The embodiments of the present disclosure are described below in conjunction with the drawings in the present disclosure. It should be understood that the implementation methods described below in conjunction with the drawings are exemplary descriptions for explaining the technical solutions of the embodiments of the present disclosure and do not constitute a limitation on the technical solutions of the embodiments of the present disclosure.

[0064] It will be understood by those skilled in the art that, unless specifically stated, the singular forms "one", "an", "said" and "the" used herein may also include plural forms. It should be further understood that the terms "including" and "comprising" used in the embodiments of the present disclosure refer to that the corresponding features can be implemented as the displayed features, information, data, steps, operations, elements and / or components, but do not exclude the implementation as other features, information, data, steps, operations, elements, components and / or combinations thereof supported by the technical field. It should be understood that when we say that an element is "connected" or "coupled" to another element, the one element may be directly connected or coupled to the other element, or it may refer to that the one element and the other element establish a connection relationship through an intermediate element. In addition, the "connection" or "coupling" used herein may include wireless connection or wireless coupling. The term "and / or" used herein indicates at least one of the items defined by the term, for example, "A and / or B" may be implemented as "A", or as "B", or as "A and B".

[0065] In order to make the objectives, technical solutions and advantages of the present disclosure more clear, the embodiments of the present disclosure will be further described in detail below with reference to the accompanying drawings.

[0066] The following describes several exemplary embodiments to illustrate the technical solutions of the embodiments of the present disclosure and the technical effects produced by the technical solutions of the present disclosure. It should be noted that the following embodiments can refer to, draw on or combine with each other, and the same terms, similar features and similar implementation steps in different embodiments will not be described repeatedly.

[0067] In CT image reconstruction, the filtered back-projection (FBP) algorithm can be technically divided into three steps: filtering, view-by-view back-projection, and summation. CT images are the sum of the view-by-view back-projections of the original data, which may cause the inherent correlations between all back-projections and the redundancies between views to be compressed and ignored, respectively. Taking these inherent correlations and redundancies into account should be able to solve the challenges of weak imaging expression of CT images, especially low contrast.

[0068] In an embodiment of the present disclosure, a method for detecting lesions in plain scan CT images based on back-projection data is proposed. The method applies a lesion detection model that uses view-by-view back-projection (VVBP) data as input instead of image data, and considers how to process VVBP data and how to construct a detection model to make full use of VVBP data to obtain better detection performance. Specifically, the disclosed embodiment converts VVBP data into N sparse images (N sparse images) through a novel processing method, and adopts the two-stage target detection representative algorithm Faster R-CNN as the basic model, which can be composed of three main modules: ResNet50-Unet, whose encoder ResNet50 is used to extract basic features, and the decoder is used to restore the plain scan CT image to compensate for the information loss of the simulated VVBP data; multi-cross channel-spatial-attention (mcCSA) is used for information interaction of VVBP data in the view direction; Faster R-CNN and weighted candidate Bbox fusion (weighted candidate Bbox fusion, WCBF) are used for lesion detection. The two strategies designed in the disclosed embodiment include feature fusion through mcCSA to interactively integrate the basic features of N sparse images, and candidate Bbox fusion through WCBF in the region proposal network (RPN) of Faster R-CNN, which can make full use of VVBP data to achieve better detection effect.

[0069] The following is a detailed description of the method for detecting lesions in plain scan CT images based on back-projection data in an embodiment of the present disclosure.

[0070] Specifically, Figure 1 As shown, the method may include S101 to S103:

[0071] S101 . Obtaining view-by-view back-projection VVBP data based on the plain scan CT image.

[0072] S102: Perform sparse processing on the VVBP data to obtain N sparse images.

[0073] S103: Process the N sparse images using a pre-trained lesion detection model to obtain predicted position information.

[0074] Among them, in S103, the N sparse images are processed by a pre-trained lesion detection model to obtain predicted position information, including steps A1 to A5:

[0075] Step A1: perform feature encoding on the N sparse images to obtain N first feature maps.

[0076] Step A2: reconstruct the image based on the N first feature maps to obtain a reconstructed image.

[0077] Step A3: fuse the N first feature maps to obtain a second feature map.

[0078] Step A4: Detect the lesion area on the first feature map and the second feature map respectively to obtain N groups of first candidate frames and one group of second candidate frames.

[0079] Step A5: obtaining the predicted position information based on the first candidate box and the second candidate box; the predicted position information is used to indicate the lesion area based on a reconstructed image, and the reconstructed image is obtained based on the N first feature maps.

[0080] Optionally, when obtaining VVBP data based on plain scan CT images, CT image reconstruction (such as filtered back projection FBP algorithm) is involved, which is a process of reconstructing the internal structure of an object from a series of projection data. The projection data is usually collected by ray projection of the object at different angles, reflecting the density distribution information of the object in different directions. When acquiring VVBP data, after filtering the original projection data, the filtered projection data can be back-projected according to each projection angle (in this process, the back-projection data of each view will be processed separately and stacked into a third-order tensor, that is, VVBP data, such as Figure 2 as shown).

[0081] Optionally, sparse processing of VVBP data can utilize information such as the correlation and complementarity of full-view VVBP data, while reducing redundancy and reducing the computational burden. In the disclosed embodiment, considering that the size of the VVBP data of the full view is too large, such as the VVBP tensor size of 1152 full-view CT scan images is 512×512×1152, it is difficult to directly mine rich information from such a large three-dimensional tensor using a two-dimensional modeling method, especially correlation and complementarity, and it is easy to ignore redundancy along the view direction. In addition, the use of three-dimensional models such as convolutional neural networks (CNNs) also requires a lot of computing power. Therefore, by sparse processing of VVBP data, the detection performance can be improved while utilizing information such as the correlation and complementarity of full-view VVBP data for doctors to make accurate diagnoses based on the processing results, while reducing information redundancy, speeding up processing and reducing storage costs.

[0082] In the disclosed embodiment, a lesion detection model of plain scan CT images based on VVBP data is proposed, such as Figure 2 As shown in the figure, the model uses Faster R-CNN (such as a two-stage target detection representative algorithm) as a baseline, and consists of three main modules: ResNet50-Unet (such as a deep learning model that combines the residual network ResNet50 and the fully convolutional network Unet), mcCSA (Multi-Cross-channel-spatial-attention) and Faster R-CNN with WCBF (weighted candidate Bbox fusion).

[0083] Among them, in the ResNet50-Unet architecture, the N sparse images generated by S102 are encoded by ResNet50 (such as encoder Encoder) to generate N first feature maps (also called basic feature maps) These feature maps are reconstructed through upsampling and deconvolution decoding (such as decoder) to compensate for the information loss of simulated VVBP data, and can also be interactively fused into a second feature map (also called a high-level feature map) through mcCSA. To optimize processing.

[0084] Among them, the second feature map obtained in step A3 and the N first feature maps obtained in step A1 Input RPN (Region Poposal Network) respectively to get a set of second candidate boxes (also called advanced candidate boxes) And N groups of first candidate boxes (also called basic candidate boxes) And through WCBF, these N groups of candidate frames are interactively fused into one group of candidate frames .after, and Further merging can obtain the final predicted location information, such as the proposed candidate box .at last, All proposed candidate boxes in In the algorithm, the alignment is performed by ROIPooling (Region of Interest Pooling). The aligned features of each proposed candidate box are expanded into a fixed-size feature vector and input into the Fast R-CNN detector based on the fully connected layer to obtain the detection results.

[0085] Optionally, the reconstructed image obtained by feature encoding N sparse images can make up for the information loss of the simulated VVBP data, effectively improving the accuracy of the processing results.

[0086] Optionally, the predicted position information obtained by processing N sparse images through the model can be displayed based on the reconstructed image, such as indicating the lesion area in the reconstructed image, and can also be indicated in the output image obtained based on the reconstructed image. Exemplarily, when plain scan CT is used to examine the patient's abdominal organs, a reconstructed image marked with predicted position information can be output through model processing, and the organ with lesions, such as the pancreas, can be indicated on the reconstructed image by marking boxes of different shapes such as rectangular boxes. In one example, the reconstructed image output by the model can be an image corresponding to the original image (such as a non-contrast computed tomography image, Non-contrast computed tomography, NCCT). Figure 2 In Figure 1, the output of the model Fast R-CNN detector shows a local portion of an NCCT image. In an example application, the reconstructed image can include the same image content as the original image.

[0087] In a feasible embodiment, S101 obtains VVBP data based on the plain scan CT image, including S101a to S101e:

[0088] S101a, obtaining a plurality of two-dimensional tomographic images of the target tissue structure through plain scan CT.

[0089] Optionally, in the acquisition of plain scan CT images in S101a, the embodiment of the present disclosure collects and analyzes plain scan CT images of 861 patients from four medical centers in China (Nanfang Hospital of Southern Medical University, Zhujiang Hospital Affiliated to Southern Medical University, Changzhou First People's Hospital, and Zhejiang Provincial People's Hospital) from June 2010 to July 2020. The specific distribution is: 431 cases in Center 1 (Nanfang Hospital of Southern Medical University), 275 cases in Center 2 (Zhujiang Hospital Affiliated to Southern Medical University), 126 cases in Center 3 (Changzhou First People's Hospital), and 29 cases in Center 4 (Zhejiang Provincial People's Hospital). Among them, a total of 511 patients, including 210 patients from Center 1, 159 patients from Center 2, 113 patients from Center 3, and 29 patients from Center 4, underwent radical surgery and were pathologically diagnosed with a single pancreatic lesion; a total of 256 patients, including 221 patients from Center 1, 22 patients from Center 2, and 13 patients from Center 3, underwent puncture exploration and were diagnosed with a single pancreatic lesion. In addition, 94 normal patients from Center 2 were radiologically diagnosed with normal pancreatic tissue by enhanced CT. The study covered various pathological types of pancreatic lesions, but did not include cystic lesions outside the pancreatic area, as these lesions are easily distinguished on plain CT. All participants were of Asian descent, with an average age of 56.9 years and an age range of 14 to 88 years. Male patients accounted for approximately 54.5% (470 people), and the rest were female.

[0090] Exemplarily, since enhanced CT is the preferred imaging modality for pancreatic diseases and plain scan CT alone is relatively seldom used, all plain scan CT images in the disclosed embodiments are obtained from the plain scan phase of the enhanced CT scan. The imaging equipment used by Centers 1-4 include a Brilliance iCT256 scanner, a LightSpeed ​​RT 16 CT scanner, a SOMATOM Definition AS CT scanner, and a SOMATOM Definition Flash CT scanner. The scanning scheme has specific parameters, such as a tube voltage of 120 kV, a rotation time of 0.5 seconds or 0.75 seconds, an X-ray tube current of 150-500 mA, a matrix of 512×512, a slice thickness of 1-7 mm, and a detector single collimation width of 0.625 mm or 0.6 mm.

[0091] For example, patient A has chest pain. To clearly analyze the cause of the pain, a CT scanner can be used to perform a cross-sectional scan of patient A's chest to obtain lung window and mediastinal window images of each cross section of the chest.

[0092] S101b, simulating the two-dimensional tomographic image to obtain first VVBP data.

[0093] Optionally, the plain scan CT image acquired by S101a is simulated to obtain VVBP data. The VVBP data is simulated by the Michigan Image Reconstruction Toolbox (MIRT) based on Radon transform and Hanning filter. The specific parameters are set as follows: the projection view is a full view with a total of 1152 views, the detector channels are 222, the distance from the source to the origin is 541 mm, and the distance from the source to the detector is 949.075 mm.

[0094] S101c: performing unified standardization processing on the first VVBP data to obtain a standardized image.

[0095] S101d. Based on the set pixel size, the standardized image is resampled using a two-dimensional spline interpolation method to obtain an interpolated image with the same resolution.

[0096] S101e. Based on the size of the target tissue structure, crop the interpolation image to obtain second VVBP data.

[0097] Optionally, S101c to S101e are preprocessing of plain scan CT and VVBP data for the experiment. Exemplarily, first, all original plain scan CT images and simulated VVBP data are uniformly standardized, with the window level set to 10 HU (Hounsfield Units, HU) and the window width set to 400 HU to reduce the heterogeneity generated by different CT scanning schemes. Subsequently, all standardized images of plain scan CT images and simulated VVBP data are resampled to the same resolution using a two-dimensional spline interpolation method with a pixel size of 0.5 mm×0.5 mm. All interpolated images are then cropped into patches of size 340×270 covering the entire target tissue structure (such as a pancreatic organ). By replacing the entire image, patches of N sparse images generated by plain scan CT images and simulated VVBP data are used as model inputs to construct corresponding detection models, including the model proposed in the disclosed embodiment and all comparative models, to reduce the impact of complex anatomical backgrounds and computational burdens.

[0098] Among them, the window level determines the display range of data in the visualization chart or interface. By setting a specific window level, it can ensure that the data is displayed within an appropriate range to avoid data being too dense or sparse, which affects the readability and analysis effect of the data. In the unified standardization process, setting a consistent window level can ensure that different data sets are analyzed and compared under the same conditions, thereby improving the consistency and comparability of the data. The window width directly affects the clarity of the data in the visualization chart. Setting an appropriate window width can ensure that the intervals between data points are moderate, so that the data can be clearly displayed in the chart, which is convenient for users to observe and compare. The setting of the window width also affects the efficiency of data analysis. If the window width is too narrow, the data points will be too dense, making the analysis process cumbersome and time-consuming. If the window width is too wide, some important data details may be missed. Therefore, setting an appropriate window width can balance the relationship between analysis accuracy and efficiency.

[0099] In a feasible embodiment, in S102, sparse processing is performed on the VVBP data to obtain N sparse images, including S102a to S102b:

[0100] S102a, based on different initial views, the VVBP data is sampled at the same interval N to obtain N groups of sparse VVBP views.

[0101] S102b: For each group of sparse VVBP views, sum the VVBP views in the group along the view direction to obtain N sparse images.

[0102] Optionally, the VVBP data preprocessed in S102 is processed into N sparse images. First, the preprocessed full-view VVBP is divided into N groups of sparse sub-VVBPs (such as sub-VVBPs), and each group of sub-VVBPs is sampled at the same interval N starting from a different initial view (i.e., the first view, the second view, ... the Nth view). Subsequently, each group of sub-VVBPs is summed along the view direction to generate N sparse images. The generated N sparse images will replace the full-view VVBP as the model input.

[0103] Optionally, the basic data processed by S102a to S102b may be the first VVBP data or the second VVBP data.

[0104] In a feasible embodiment, the N first feature maps are fused in step A3 to obtain a second feature map, including steps A31 to A33:

[0105] Step A31: perform multi-cross-channel attention processing on the N first feature maps to obtain a first fusion result.

[0106] Step A32: perform multi-cross spatial attention processing on the first fusion result to obtain a second fusion result.

[0107] Step A33: Obtain a second feature map based on the second fusion result.

[0108] Optionally, based on the ResNet50-Unet architecture, the N sparse images generated by S102 are feature encoded by ResNet50 to generate N basic feature maps , these feature maps are up-sampled and deconvolutionally decoded to restore the information lost by simulating VVBP data. Interactive fusion into a high-level feature map through the novel mcCSA module .like Figure 2 and Figure 3 As shown in the figure, the mcCSA module adopts a two-stage series structure to connect N basic feature maps As input, these feature maps are interactively processed by combining multi-cross channel attention and multi-cross spatial attention.

[0109] Optionally, in step A31, multi-cross channel attention processing is performed on the N first feature maps to obtain a first fusion result, including steps A311 to A313:

[0110] Step A311: For each of the first feature graphs, generate a query item, a key item, and a value item of channel attention.

[0111] Step A312: For any two of the first feature graphs, based on the query item of one of the first feature graphs and the key item and value item of the other first feature graph, determine a first interaction result of the two first feature graphs.

[0112] Step A313: For any two of the first feature maps, based on a first interaction result between one of the first feature maps and the two first feature maps, determine a first fusion result of the two first feature maps.

[0113] Optional, based on Generate query in channel attention mechanism ( ),key( ) and value ( ) matrices, which are obtained by ( ) by applying a 3x3 convolution with three channels and layer normalization (layernorm) on it. and The interaction result between them is defined as follows:

[0114] ......Formula (1)

[0115] Among them, d is the smoothing parameter of the softmax function, This definition shows that By using Follow the query To obtain the key and value information of and The relationship between is The result of self-attention (SA). Next, for and ( The multi-cross channel attention result is defined as follows:

[0116] ...Formula (2)

[0117] The results of the above multi-cross channel attention processing can be used as the input of the second layer multi-cross spatial attention of mcCSA.

[0118] Optional, such as Figure 3 As shown in the figure, in the multi-cross channel attention, for the first feature map, a 3×3 convolution is first performed through the normalization layer (Norm). The layout of this layer can improve the generalization ability and robustness of the network. Then, the query item, key item and value item of the first feature map are generated. In the processing of multi-cross channel attention, the relationship between the query matrix, the key matrix and the value matrix is ​​calculated by matrix multiplication (matmul), and an attention score matrix can be obtained to represent the correlation between each element in the query matrix and each element in the key matrix. Among them, the attention score matrix can be converted into a probability distribution through the softmax function, such as mapping each element in the input vector to the interval 0-1 so that the sum of all elements is 1; in the multi-cross channel attention mechanism, the softmax function converts each element in the attention score matrix into a probability value, and the probability value can represent the attention weight of each element in the query matrix for each element in the key matrix. Then, after a 1×1 convolution is performed through the convolution layer, the first fusion result can be output.

[0119] Optionally, in step A32, multiple cross-spatial attention processing is performed on the first fusion result to obtain a second fusion result, including steps A321 to A323:

[0120] Step A321: Based on the first fusion result, generate query items, key items and value items of spatial attention.

[0121] Step A322: For any two of the first fusion results, based on the query item of one of the first fusion results and the key item and value item of the other first fusion result, determine the second interaction result of the two first fusion results.

[0122] Step A323: For any two of the first fusion results, based on one of the first fusion results and the second interaction result of the two first fusion results, determine a second fusion result of the two first fusion results.

[0123] Optional, assuming Based on The query, key, and value matrices in the spatial self-attention of These operations include position embedding, linear projection, and layer normalization. , which is consistent with all feature maps ( The multi-cross spatial attention result between is defined as follows:

[0124] ...Formula (3)

[0125] Optionally, you can sum The output of mcCSA is obtained, which is used as a high-level feature map for high-level candidate box generation in RPN and lesion detection in Fast R-CNN.

[0126] Optional, such as Figure 3 As shown, in the multi-cross spatial attention, linear projection and position embedding (Pos_embed) can be performed first through a normalization layer (such as Norm), and then the query item, key item and value item of the first fusion result are generated. Then, as in the processing of multi-cross channel attention, the matmul (matrix multiplication) layer, softmax layer, matmul layer and convolution (conv) layer connected in series can be processed to output the processing result of the multi-cross spatial attention (such as the second fusion result), and through the summation process, the second feature map can be finally output.

[0127] In a feasible embodiment, in step A4, the first feature map and the second feature map are respectively subjected to lesion area detection to obtain N groups of first candidate frames and a group of second candidate frames, including performing the following detection operations of steps A41 to A43 on the first feature map and the second feature map respectively through a region proposal network RPN:

[0128] Step A41: for the received feature map, generate several predefined anchor boxes at each feature point of the feature map, wherein the anchor boxes include rectangular boxes with different aspect ratios and areas.

[0129] Step A42: For each anchor box, determine the probability of the contained object and adjust the size and position of the anchor box through a bounding box regression adjuster.

[0130] Step A43: Compare overlapping anchor frames through non-maximum suppression (NMS) to obtain the anchor frame with the highest probability of containing the object, and determine the candidate frame based on the anchor frame.

[0131] Optionally, the basic feature map generated in step A1 and the one obtained in step A3 It is used as the feature map of Faster R-CNN detection and is input into the RPN module to generate N+1 sets of candidate bounding boxes (Bboxes), which are used express The corresponding bounding box (also called the first candidate box), and express The corresponding bounding box (also called the second candidate box).

[0132] Optionally, when the feature map is detected by RPN, RPN first receives the feature map extracted by the convolutional neural network (CNN) as input, and the convolutional neural network can be any valid architecture, such as VGG, ResNet, etc. At each position of the feature map, RPN generates multiple predefined anchor boxes (anchors), that is, rectangular boxes with different aspect ratios and areas (the shape of the box can be set according to actual conditions). The anchor box can be used to cover objects of different shapes and sizes (such as tissue structures) as the initial detection box. For each anchor box, RPN can predict two outputs, one of which is the probability that the anchor box contains the object (such as the object score), which is used to distinguish the foreground (containing the object) from the background, which can be achieved by a softmax classifier. The other is a bounding box regression adjuster, which is used to fine-tune the position of the anchor box so that it matches the real object more accurately. The bounding box regression can be a linear regression process. Then, considering that the same object may be covered by multiple anchor boxes, the embodiment of the present disclosure also selects the best anchor box through NMS. NMS compares the overlapping anchor boxes and suppresses the anchor boxes with high overlap and low scores, and finally selects the best anchor box corresponding to each object.

[0133] Optionally, RPN can process the feature map to output a series of candidate boxes with a high probability of containing the object, and these candidate boxes can be used to subsequently determine the predicted position information. Exemplarily, RPN processes N first feature maps to output N groups of first candidate boxes, and processes the second feature map to output the second candidate box. Figure 2 As shown, the input of RPN includes N first feature maps output by the encoder, and the second feature map output by Fc (such as the fully connected layer, Fully Connected Layer) based on the second fusion result.

[0134] In a feasible embodiment, obtaining predicted position information based on the first candidate box and the second candidate box in step A5 includes steps A51 to A53:

[0135] Step A51: for N groups of first candidate boxes, the first candidate boxes whose intersection over union (IoU) values ​​are greater than a preset first threshold are divided into the same cluster, and the bounding boxes in the cluster are merged to obtain first sub-candidate boxes corresponding to the N groups of first candidate boxes.

[0136] Step A52: for the first sub-candidate box and the second candidate box, classify the candidate boxes whose IoU values ​​are greater than a preset second threshold into the same cluster, and merge the bounding boxes in the cluster to obtain the second sub-candidate box.

[0137] Step A53: For the second sub-candidate box, the pooling layer and the fully connected layer in the network are processed to obtain predicted position information; the predicted position information includes the predicted coordinate value and the predicted category probability value of the bounding box.

[0138] Alternatively, you can use the WCBF method, Fusion into a set of high-level candidate bounding boxes, denoted as . With the high-level feature map Advanced Further merge to obtain the final proposed candidate box . The WCBF method is designed based on the WBF (Weighted Box Fusion) method. First, all candidate bounding boxes are merged into a sorted list in descending order according to their confidence scores. Next, based on IoU (Intersection over Union), a threshold of 0.7 (just an example, it can be adjusted according to actual conditions) is used to group these bounding boxes into different clusters. Then, the bounding boxes within each cluster are merged into a single bounding box by weighted summing up the coordinate values ​​of each bounding box within the cluster. The clustering method and weight calculation in WCBF are the same as those in WBF.

[0139] Optionally, the merging of the first candidate box and the second candidate box may also be to use the first sub-candidate box and the second candidate box output in step A51 as elements of the second sub-candidate box set.

[0140] Optional, such as Figure 2 As shown, All proposed candidate boxes in In the algorithm, the alignment is performed by ROI pooling. The alignment features corresponding to each proposed candidate box are expanded into a fixed-size feature vector and input into the FastR-CNN detector based on the fully connected layer. The false positives are removed by the classifier (classification prediction, Cls_pred) based on the fully connected layer, and the position of the detected bounding box is optimized by the regressor (bounding box regression, Bbox_reg) to obtain the detection result. Among them, the classifier predicts the category of each candidate region (RoI) according to the feature vector output by the fully connected layer. For example, cls_pred calculates the probability that each candidate region belongs to each category through the softmax function or other classifiers, thereby determining the most likely category label for the region. Among them, the regressor shares the output of the fully connected layer with the classifier. The regressor is used to adjust the position and size of the candidate region so that it more accurately surrounds the target tissue structure. For example, Bbox_reg fine-tunes the candidate region by calculating the offset between the candidate region and the true target bounding box, thereby obtaining a more accurate target detection frame.

[0141] Optionally, the predicted position information may include predicted coordinate values ​​and predicted category probability values ​​of the bounding box. Exemplarily, the bounding box may be identified in the reconstructed image by the predicted coordinate values, and the predicted category probability may be displayed near the bounding box of the reconstructed image to indicate the confidence level.

[0142] In a feasible embodiment, the lesion detection model provided by the embodiment of the present disclosure is obtained by training the loss function values ​​determined by the following steps B1 to B3:

[0143] Step B1: Determine a first loss value based on the structural similarity loss SSIM between the plain scan CT image and the reconstructed image.

[0144] Step B2: determining a second loss value between the actual position information and the predicted position information of the plain scan CT image.

[0145] Step B3: Based on the first loss value and the second loss value, determine the loss function value in the lesion detection model training process.

[0146] Optionally, the model uses the RMSprop optimizer to optimize network parameters. The loss function consists of two parts: the loss of the original Faster R-CNN and the loss of rebuilding branches Among them, the reconstruction loss is specifically designed using the structural similarity loss (SSIM) between the original plain scan CT image and the reconstructed image.

[0147] Optionally, the total loss function for model training is as follows:

[0148] ...Formula (4)

[0149] Optionally, determining the second loss value between the real position information and the predicted position information of the plain scan CT image in step B2 includes step B21 and / or step B22:

[0150] Step B21, based on the predicted category probability value, predicted coordinate value and true coordinate value of the candidate box obtained by RPN, determine the third loss value according to the same method as the loss calculated by the second stage detector.

[0151] Step B22: In the second stage detector of the lesion detection model, based on the binary cross entropy loss value and the L1 norm loss value calculated when removing the false positive candidate boxes, together with the third loss value, determine the second loss value ultimately used to train the lesion detection model.

[0152] Optionally, the Faster R-CNN loss function can be composed of four parts: a binary cross entropy loss for two classifiers to remove false positive Bboxes in the RPN and the second-stage detector, and a smoothed L1 norm loss for two regressors to optimize the Bbox coordinates in the RPN and the second-stage detector.

[0153] In RPN, it is assumed that is the predicted probability of the i-th anchor box Bbox, whose category label is y∈{0,1}, which is the training sample of the classifier; Represents the predicted coordinates of the jth anchor box Bbox, is the true value box coordinate, forming the training sample of the regressor. Then, the loss function is defined as follows:

[0154] ...Formula (5)

[0155] in, and are the number of training samples for the classifier and the regressor, respectively.

[0156] ,

[0157]

[0158] Optional, The third loss value calculated in step B21 above can be obtained. Then, in the detector used in the second stage, the loss function can be obtained by replacing the training samples in the RPN with the corresponding training samples in the second stage detector. The total loss function of the Faster R-CNN training process (such as the second loss value calculated) is .

[0159] The following describes the data set and related experimental data used for model training in the embodiments of the present disclosure.

[0160] In the data set division, the disclosed embodiment focuses on the data of four medical centers (i.e., centers 1-4). First, a total of 519 lesion patients and 4255 images from center 1 (421 cases) and center 2 (98 cases), which have the largest number of cases, are used as the core data basis. In order to further optimize the model, a small number of patients (center 1-30 patients, 138 images; center 2-29 patients, 421 images) are randomly selected from these two centers for model verification, and the remaining 460 patients and their corresponding 3696 images are used for model training. Afterwards, an independent test set is set up, containing 304 patients and their corresponding 1743 images to ensure data independence. The internal test set specifically selects 169 patients from center 2 (75 lesion patients + 94 normal cases), and the external test set covers 109 patients from center 3 and 26 lesion patients from center 4, of which 94 normal cases are dedicated to testing to avoid their images affecting the training process, while the lesion patients provide the necessary positive and negative bounding box training data for the classifier. Table 1 below records the data distribution of this process in detail.

[0161] Table 1 Dataset partition table

[0162]

[0163] After the data set is divided, experiments can be conducted on this basis. First, by adjusting different values ​​of the parameter N, different sparse image sets are generated. Using different numbers of sparse images as the input of the model, the experimental results on the validation set show that when N=3, the model achieves the best detection results. Among them, the accuracy at the image level is 41.4%, the recall rate is 75.7%, the F1 score is 53.5%, the F2 score is 64.9%, and the Ap50 is 56.2%. At the patient level, the accuracy is 66.9%, the recall rate is 90.5%, the F1 score is 76.9%, the F2 score is 84.5%, and the Ap50 is 76.9%. Therefore, the subsequent comparative experiments and ablation experiments all use N=3 for the choice of N.

[0164] The image-level detection results and patient-level detection results of the model provided by the embodiment of the present disclosure and other comparative models in the independent test set are shown in Table 2.

[0165] Table 2 Comparison results with existing methods at the image and patient level

[0166]

[0167] In the experiment, N=3 VVBP data (VVBP_ours) was used as the input of the model provided by the embodiment of the present disclosure. The results showed that the results of copying the plain scan CT image data into the same size, i.e., N=3 copies (Image_ours), as the input of the model were basically better than the results of copying the plain scan CT image data into the same size, i.e., N=3 copies (Image_ours) as the input of the model in various evaluation indicators. These findings indicate that the VVBP data contains relevant information that is crucial for pancreatic lesion detection, which may be compressed or ignored in the image data. In addition, compared with existing methods for image-based target detection, including Faster R-CNN, SSD, RetinaNet, YOLOv7, YOLOv10, and DeformableDETR, and the three-dimensional volume-based Bbox detection method (nnDetection), the model provided by the embodiment of the present disclosure outperforms all other models in terms of F1 score, F2 score, and Ap50, and obtains the second best precision and recall rate at the image level. At the patient level, it also obtains the best recall rate, F1 score, F2 score, and Ap50, as well as the second highest precision. These results clearly demonstrate the effectiveness and advantages of the model provided by the embodiment of the present disclosure. Compared with the image-based model (Image) with the same input size of N=3, the model provided by the embodiment of the present disclosure shows better performance when inputting N (N=3) sparse images. Specifically, at the image level and patient level, the model provided by the embodiment of the present disclosure achieved 0.7% to 4.8% improvement in various indicators compared with the model. These results highlight the importance of the information contained in the VVBP data for the accurate detection of pancreatic lesions, which may not be fully utilized when using image data.

[0168] Table 3 shows the ablation experiment results at the image and patient levels.

[0169] Table 3

[0170]

[0171] Among all the models that underwent ablation experiments, the model provided by the embodiment of the present disclosure integrates three modules (reconstruction branch (Recon), mcSCA and WCBF), and achieves the best precision, recall rate, F1 score, F2 score and AP50 results at the image level. At the same time, at the patient level, the recall rate, F1 score and F2 score results of the model are the best, and the precision and AP50 are ranked second. These results fully demonstrate the necessity and effectiveness of each module in the model provided by the embodiment of the present disclosure. Figure 4The visualization results of the ablation experiment of six images randomly selected from the independent test set are shown. The columns represent the performance of the different models shown in Table 3 above in each ablation experiment, where the column (h) represents the performance of the model provided by the embodiment of the present disclosure in each ablation experiment, and the rows correspond to the results of different patients. The green box represents the true value box, and the red box represents the detected bounding box, and the confidence score is displayed above it. The visualization results show that the model provided by the embodiment of the present disclosure effectively reduces missed detections and significantly improves the detection accuracy.

[0172] Table 4 shows the experimental results based on different input methods at the image level and patient level.

[0173] Table 4

[0174]

[0175] Specifically, VVBP_3_sparse_n represents the result of extracting the nth image from three sparse images, copying it three times and inputting it into the model; while VVBP_3_sparse represents the result of directly inputting all three sparse images into the model at the same time. At the image level, VVBP_3_sparse achieved the highest precision, F1 score, F2 score, and AP50 score, as well as the second highest recall rate. At the patient level, VVBP_3_sparse achieved the highest precision, recall rate, F1 score, F2 score, and AP50 score. These results show that directly inputting all sparse images into the model (VVBP_3_sparse) performs better than using a single repeated image (VVBP_3_sparse_n). These results highlight the significant advantages of using all sparse images over just copying a single image, and verify the effectiveness of the method provided by the embodiment of the present disclosure in utilizing the complementarity and correlation of data from different angles in VVBP data.

[0176] The disclosed embodiment verifies the effectiveness of the model in processing seven images containing lesions selected from an independent test subset (e.g. Figure 5 Specifically, by analyzing the feature map output by mcCSA Visualization can explain the effectiveness of the model. The visualization results show that compared with Image_ours ( Figure 5 (b) compared to VVBP_ours ( Figure 5 The feature map shown in the middle row (c) locates the lesion area more accurately in space. In addition, VVBP_ours also reduces the interference from non-lesion areas, making lesion identification clearer and more accurate; Figure 5 The middle row (a) shows the baseline results. Comparative analysis further shows that using VVBP data instead of images as input has significant advantages in lesion focusing and interference suppression.

[0177] The disclosed embodiment proposes a method for detecting lesions in plain scan CT images based on back-projection data, which uses VVBP data instead of traditional CT image data as input. Specifically, the disclosed embodiment proposes a novel method for processing VVBP data, which converts it into N sparse images as model input. Based on this input, the disclosed embodiment designs two strategies to make full use of its information: first, feature fusion is performed through mcCSA to interactively integrate the basic features of N sparse images; second, candidate Bbox fusion is performed through WCBF in the RPN of Faster R-CNN. This method utilizes the correlation and complementary information of full-view VVBP data, while reducing redundancy and alleviating the computational burden, thereby improving detection performance.

[0178] The purpose of the disclosed embodiment is to address the problem of poor contrast and unsatisfactory image performance of plain scan CT in medical image target detection, and to propose a lesion detection method using VVBP data as input. This method will provide key technical support for the research of lesion screening algorithms based on plain scan CT, thereby facilitating early detection and early diagnosis of diseases.

[0179] It should be noted that in the optional embodiments of the present disclosure, the information involved (such as plain scan CT images, etc.), when the above embodiments of the present disclosure are applied to specific products or technologies, needs to obtain permission or consent from the user, and the collection, use and processing of relevant data need to comply with the relevant laws, regulations and standards of the relevant countries and regions. In other words, if the embodiments of the present disclosure involve data related to the object, these data need to be obtained with the authorization and consent of the object and in compliance with the relevant laws, regulations and standards of the country and region.

[0180] In the embodiments of the present disclosure, an electronic device is provided, which may be a medical device. The host of the device includes a memory, a processor and a computer program stored in the memory. The processor executes the above computer program to implement the steps of the method for detecting lesions in plain scan CT images based on back-projection data. Compared with the related art, the following can be achieved: the embodiment of the present disclosure performs sparse processing on the VVBP data obtained based on the plain scan CT image to obtain N sparse images used as model input, and in the processing of the model, the N sparse images can be feature encoded to obtain N first feature maps, and the image can be reconstructed based on the N first feature maps to obtain a reconstructed image, and the N first feature maps can be integrated by feature fusion to obtain a second feature map, and the first candidate box of the first feature map and the second candidate box of the second feature map can be fused, and finally output the predicted position information for indicating the lesion area. The implementation of the embodiments of the present disclosure can effectively utilize the correlation and complementary information of the full-view VVBP data to improve the detection performance.

[0181] In an alternative embodiment, an electronic device is provided, such as Figure 6 As shown, Figure 6 The electronic device 4000 shown includes: a processor 4001 and a memory 4003. The processor 4001 and the memory 4003 are connected, such as through a bus 4002. Optionally, the electronic device 4000 may also include a transceiver 4004, which may be used for data interaction between the electronic device and other electronic devices, such as data transmission and / or data reception. It should be noted that in actual applications, the transceiver 4004 is not limited to one, and the structure of the electronic device 4000 does not constitute a limitation on the embodiments of the present disclosure.

[0182] Processor 4001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. It may implement or execute various exemplary logic blocks, modules and circuits described in conjunction with the disclosure of the present invention. Processor 4001 may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0183] The bus 4002 may include a path to transmit information between the above components. The bus 4002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus. The bus 4002 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 6 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0184] The memory 4003 may be a ROM (Read Only Memory) or other types of static storage devices that can store static information and instructions, a RAM (Random Access Memory) or other types of dynamic storage devices that can store information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, optical disk storage (including compressed optical disk, laser disk, optical disk, digital versatile disk, Blu-ray disk, etc.), magnetic disk storage media, other magnetic storage devices, or any other medium that can be used to carry or store computer programs and can be read by a computer, without limitation herein.

[0185] The memory 4003 is used to store the computer program for executing the embodiment of the present disclosure, and the execution is controlled by the processor 4001. The processor 4001 is used to execute the computer program stored in the memory 4003 to implement the steps shown in the above method embodiment.

[0186] An embodiment of the present disclosure provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps and corresponding contents of the aforementioned method embodiment can be implemented.

[0187] The embodiments of the present disclosure also provide a computer program product, including a computer program, which can implement the steps and corresponding contents of the aforementioned method embodiments when executed by a processor.

[0188] The terms "first", "second", "third", "fourth", "1", "2", etc. (if any) in the specification and claims of the present disclosure 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 numbers used in this way can be interchanged where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than that shown or described in the text.

[0189] It should be understood that, although the flowchart of the embodiment of the present disclosure indicates each operation step by arrows, the implementation order of these steps is not limited to the order indicated by the arrows. Unless clearly stated herein, in some implementation scenarios of the embodiment of the present disclosure, the implementation steps in each flowchart can be executed in other orders according to demand. In addition, some or all of the steps in each flowchart may include multiple sub-steps or multiple stages based on the actual implementation scenario. Some or all of these sub-steps or stages can be executed at the same time, and each sub-step or stage in these sub-steps or stages can also be executed at different times. In scenarios with different execution times, the execution order of these sub-steps or stages can be flexibly configured according to demand, and the embodiment of the present disclosure does not limit this.

[0190] The above is only an optional implementation method for some implementation scenarios of the present disclosure. It should be pointed out that for ordinary technicians in this technical field, without departing from the technical concept of the scheme of the present disclosure, other similar implementation methods based on the technical ideas of the present disclosure are also within the protection scope of the embodiments of the present disclosure.

Claims

1. A method for detecting lesions in plain scan CT images based on back-projection data, characterized in that: include: Based on the plain scan CT images, the back-projected VVBP data is obtained view by view; Performing sparse processing on the VVBP data to obtain N sparse images; Processing the N sparse images through a pre-trained lesion detection model to obtain predicted position information; The process of processing the N sparse images by using a pre-trained lesion detection model to obtain predicted location information includes: Performing feature encoding on the N sparse images to obtain N first feature maps; Reconstructing the image based on the N first feature maps to obtain a reconstructed image; Fusing the N first feature maps to obtain a second feature map; Detecting the lesion area on the first feature map and the second feature map respectively to obtain N groups of first candidate frames and a group of second candidate frames, including performing the following detection operations on the first feature map and the second feature map respectively through a region proposal network RPN: generating a number of predefined anchor frames on each feature point of the received feature map, the anchor frames including rectangular frames with different aspect ratios and areas; determining the probability of the included object for each anchor frame and adjusting the size and position of the anchor frame through a bounding box regression adjuster; comparing overlapping anchor frames through non-maximum suppression (NMS) to obtain an anchor frame with the highest probability of the included object, and determining a candidate frame based on the anchor frame; Based on the first candidate box and the second candidate box, the predicted position information is obtained, including: for N groups of the first candidate boxes, the first candidate boxes whose intersection over union (IoU) values ​​are greater than a preset first threshold are divided into the same cluster, and the bounding boxes in the cluster are merged to obtain the first sub-candidate boxes corresponding to the N groups of the first candidate boxes; for the first sub-candidate box and the second candidate box, the candidate boxes whose IoU values ​​are greater than a preset second threshold are divided into the same cluster, and the bounding boxes in the cluster are merged to obtain the second sub-candidate box; for the second sub-candidate box, the predicted position information is obtained by processing through the pooling layer and the fully connected layer in the network; the predicted position information includes the predicted coordinate value and the predicted category probability value of the bounding box; the predicted position information is used to indicate the lesion area based on the reconstructed image.

2. The method according to claim 1, characterized in that The method of obtaining VVBP data based on the plain scan CT image includes: Acquire a number of two-dimensional tomographic images of the target tissue structure through plain scan CT; Simulating the two-dimensional tomographic image to obtain first VVBP data; Performing unified standardization processing on the first VVBP data to obtain a standardized image; Based on the set pixel size, the standardized image is resampled using a two-dimensional spline interpolation method to obtain an interpolated image with the same resolution; Based on the size of the target tissue structure, the interpolation image is cropped to obtain second VVBP data.

3. The method according to claim 1, characterized in that The VVBP data is subjected to sparse processing to obtain N sparse images, including: Based on different initial views, the VVBP data is sampled at the same interval N to obtain N groups of sparse VVBP views; For each group of sparse VVBP views, the VVBP views in the group are summed along the view direction to obtain N sparse images.

4. The method according to claim 1, characterized in that The fusing the N first feature maps to obtain a second feature map includes: Performing multi-cross channel attention processing on the N first feature maps to obtain a first fusion result; Performing multi-cross spatial attention processing on the first fusion result to obtain a second fusion result; Based on the second fusion result, a second feature map is obtained.

5. The method according to claim 4, characterized in that The performing multi-cross channel attention processing on the N first feature maps to obtain a first fusion result includes: For each of the first feature maps, generate a query item, a key item, and a value item of channel attention; For any two of the first feature graphs, based on a query item of one of the first feature graphs and a key item and a value item of another first feature graph, determining a first interaction result of the two first feature graphs; For any two of the first feature maps, a first fusion result of the two first feature maps is determined based on a first interaction result between one of the first feature maps and the two first feature maps.

6. The method according to claim 4, characterized in that The performing multiple cross-spatial attention processing on the first fusion result to obtain a second fusion result includes: Based on the first fusion result, generate query items, key items and value items of spatial attention; For any two of the first fusion results, based on a query item of one of the first fusion results and a key item and a value item of another of the first fusion results, determine a second interaction result of the two first fusion results; For any two of the first fusion results, a second fusion result of the two first fusion results is determined based on one of the first fusion results and a second interaction result of the two first fusion results.

7. The method according to any one of claims 1 to 6, characterized in that The lesion detection model is trained by the following loss function value: Determining a first loss value based on a structural similarity loss SSIM between the plain scan CT image and the reconstructed image; Determining a second loss value between the actual position information and the predicted position information of the plain scan CT image; Based on the first loss value and the second loss value, a loss function value in the lesion detection model training process is determined.

8. The method according to claim 7, characterized in that The determining of the second loss value between the real position information and the predicted position information of the plain scan CT image comprises: Based on the predicted category probability value, predicted coordinate value and true coordinate value of the candidate box obtained by RPN, the third loss value is determined according to the same method as the loss calculation method of the second stage detector; In the second stage detector of the lesion detection model, the second loss value finally used to train the lesion detection model is determined based on the binary cross entropy loss value and the L1 norm loss value calculated when removing the false positive candidate boxes, together with the third loss value.

9. An electronic device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

11. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images