A thymus disease image intelligent diagnosis system based on a region attention network

Abstract

The application relates to a thymus disease image intelligent diagnosis system based on a region attention network, and belongs to the technical field of image analysis and artificial intelligence, the system comprising a CT image preprocessing module, an abnormal tissue detection and positioning module, an abnormal tissue segmentation module and an abnormal tissue classification and thymoma staging module; the CT image preprocessing module is used for denoising and enhancing the CT image; the abnormal tissue detection and positioning module uses a region attention network and a ViT to detect and position the abnormal tissue in the CT image; the abnormal tissue segmentation module automatically segments the abnormal tissue region by means of a U-Net++ network and extracts a key feature region; the abnormal tissue classification and thymoma staging module is used for extracting features of the abnormal tissue region by the ViT, and then simultaneously performing tissue classification and staging tasks. The system can more accurately capture the lesion region features in the thymus disease image, improve the early thymoma detection capability of the system and reduce errors.

Description

Technical Field

[0001] This invention relates to an intelligent diagnostic system for thymic diseases based on regional attention networks, belonging to the field of image analysis and artificial intelligence technology. Background Technology

[0002] With the development of deep learning technology, especially the application of convolutional neural networks and self-attention mechanisms, artificial intelligence has achieved significant breakthroughs in the field of medical image processing. Intelligent diagnostic systems for CT images of thymic diseases can effectively improve diagnostic accuracy, shorten diagnostic time, and reduce the workload of doctors. However, existing diagnostic systems still have certain limitations:

[0003] Insufficient multi-task learning capability: Existing deep learning models typically focus only on a single task (such as segmentation or classification) and lack an efficient framework that can handle multiple tasks simultaneously, resulting in the overall framework performance not being maximized.

[0004] Small lesions and early lesions are difficult to detect: Early lesions of thymoma often present as tiny lesions, and existing systems may have low sensitivity when detecting these lesions, leading to misdiagnosis or missed diagnosis. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent diagnostic system for thymic diseases based on region-of-interest networks (ROIs). This system utilizes RPIs and deep learning technologies such as U-Net++ to achieve efficient and accurate diagnosis of thymic diseases based on CT images. The system aims to improve the sensitivity and specificity of thymic disease diagnosis, reduce the workload of physicians, shorten diagnosis time, and enhance overall diagnostic efficiency.

[0006] The technical solution of the present invention is as follows:

[0007] A regional attention network-based intelligent imaging diagnostic system for thymic diseases includes a CT image preprocessing module, an abnormal tissue detection and localization module, an abnormal tissue segmentation module, and an abnormal tissue classification and thymoma staging module.

[0008] The CT image preprocessing module is used to denoise and enhance CT images to improve image quality and the effectiveness of subsequent model training. The abnormal tissue detection and localization module uses a region attention network and ViT to detect and localize abnormal tissues in CT images, improving the sensitivity for small lesions and early lesions. The abnormal tissue segmentation module uses the U-Net++ network to automatically segment abnormal tissue regions and extract key feature regions. The abnormal tissue classification and thymoma staging module uses ViT to extract features from abnormal tissue regions, and then performs tissue classification and staging tasks simultaneously.

[0009] Preferably, the CT image preprocessing module removes noise and enhances image contrast by performing the following steps:

[0010] (1) Image denoising: Let the original image be... I raw Mean filtering was used to denoise the CT images, resulting in denoised images. J ;

[0011] (2) Image enhancement: Gamma transform is used to adjust the brightness and contrast of the denoised CT images, making the boundaries of lesions more obvious, thereby improving the visibility of lesions in CT images and facilitating subsequent identification and detection. Enhanced CT images , H in for I height, W in for I width, C in for I The number of channels.

[0012] Preferably, the abnormal tissue detection and localization module performs the following steps:

[0013] (1) Input: Denoising and enhanced CT images ;

[0014] (2) Process: Multiple candidate regions are generated through the region attention network, and the optimal candidate region, i.e., the abnormal organization region, is selected. R final ;

[0015] (3) The output is: R final Results of anomaly detection and localization D final ;

[0016] In step (2), the region-focused network structure and loss function are as follows:

[0017] (2-1) Candidate Region Generation

[0018] The workflow of using a region focus network to generate potential lesion regions from enhanced feature maps is as follows:

[0019] (2-1-1) Anchor Frame Generation

[0020] In enhanced images I The above uses a sliding window method to generate multiple anchor frames; to adapt to the thymus disease detection task, anchor frames of different sizes and aspect ratios are generated. Further preferred, the size and aspect ratio of the anchor frames are set as follows:

[0021] Set the anchor frame size as follows:

[0022] w 1=32, h 1=32 (small-sized anchor frame)

[0023] w 2=64, h 2=64 (medium-sized anchor frame)

[0024] w 3=128, h 3 = 128 (large anchor frame)

[0025] Set the aspect ratio of the anchor frame as follows:

[0026]

[0027]

[0028]

[0029] The number of anchor frames at each location is The coordinates of each anchor frame are obtained through ( x , y , w , h ) indicates that ( x , y () is the center position of the anchor frame. w and h These are the width and height of the anchor frame;

[0030] (2-1-2) Anchor Frame Feature Extraction

[0031] For each anchor frame ViT is used to extract features at the corresponding locations. The formula is:

[0032] (1)

[0033] in, It is the first Anchor frame In images I Features at the corresponding positions, T For the ViT model;

[0034] (2-1-3) Classification and Regression Tasks

[0035] For each anchor box, the Region Focus Network (RFN) needs to perform two tasks:

[0036] (a) Classification task: Determine whether each anchor box contains abnormal tissue.

[0037] use sigmod The activation function performs binary classification (whether it contains abnormal tissue); assuming the first... i The category output for each anchor box is ,but

[0038] (2)

[0039] in, W cls These are the weights of the classification layer. b cls It is a bias term. yes sigmod Activation function; if This anchor frame is considered one of the candidate regions, thus obtaining the candidate region. R candidate ;

[0040] (b) Regression task: Based on the overlap between the anchor frame and the actual lesion area, perform bounding box regression and adjust the coordinates of the anchor frame;

[0041] The regression output for each anchor box is the offset of that anchor box from the actual border. This is used to adjust the coordinates of the anchor frame so that it matches the target more precisely; This is the offset of the anchor frame center point. The offset of the anchor frame size is given by the regression formula:

[0042] (3)

[0043] in, W reg These are the weights of the regression layer. b reg It is a bias term. This represents the regression output of the anchor box. For the first i Anchor frame anchor i In images I Features at the corresponding positions;

[0044] (2-2) Candidate region screening and positioning

[0045] The non-maximum suppression (NMS) algorithm is used to select the optimal candidate regions, while ensuring that the overlap between candidate regions does not exceed a set threshold. It returns the optimal candidate region, i.e., the abnormal tissue region. R final :

[0046] (4)

[0047] in IoU It measures the degree of overlap between candidate regions and final abnormal tissue regions; it also outputs the location and size of the detected abnormal tissue regions. ;

[0048] (2-3) Loss Function

[0049] (5)

[0050] Let be the smoothing coefficient, where the classification loss is:

[0051] (6)

[0052] in p i It is the first i The real label of the anchor box These are the predicted labels output by the classification task. N This refers to the number of anchor frames;

[0053] The regression loss measures the difference between the regression output of the anchor box and the true bounding box, and is expressed as:

[0054] (7)

[0055] in t i It is the first i The actual border offset of each anchor frame. These are the predicted regression values.

[0056] Preferably, the abnormal tissue segmentation module uses a U-Net++ network to automatically segment abnormal tissue regions and outputs the abnormal tissue regions. R Accurate segmentation of abnormal tissue areas provides a reliable foundation for subsequent lesion detection, localization, and analysis. The specific steps are as follows:

[0057] Step 1: Encoder for feature extraction and dimensionality reduction

[0058] (1) Convolution operation

[0059] Extracting local features H represents a local feature. F The height is 1, and W is a local feature. F The width is 1, and C is a local feature. F 1 channel number;

[0060] (8)

[0061] inK 1 is the size of convolution kernel, b 1 indicates the bias term. i express I The first in terms of altitude i One location, j express I The first in width j One location, c express I The first one on the passage c One location;

[0062] m and n are the convolution kernels. K 1. Sliding offset in height and width, k The width or height of the convolution kernel. This is the index of the input channel. C in Input the number of channels. C out Number of output channels;

[0063] (2) Pooling operation

[0064] F 1. Through size Step size is s Maximum pooling window, output The expression is as follows:

[0065] (9)

[0066] Step 2: Decoder for Feature Recovery and Fusion

[0067] (1) Upsampling operation

[0068] I After deconvolution kernel To restore the spatial dimension, we get :

[0069] (10)

[0070] in b 2 represents the bias term, and for F 3. Perform bilinear interpolation to obtain ; H in , H out For input height and output height, W in , W out Input width, output width;

[0071] (2) Skip connections and convolution

[0072] See separately F 1. F 2. F 3 and F 4 The decoder fuses encoder features through skip connections, thus... and spliced ;right Convolution is performed Furthermore, multi-scale features are fused to obtain a feature representation that includes both local and global information. F ;

[0073] Step 3, Binary Segmentation

[0074] right F Each pixel in F ( i, j )application sigmod Function and set threshold T =0.5, yielding the binary segmentation result. :

[0075] (11)

[0076] in This indicates the foreground, i.e., the area of ​​abnormal tissue. This indicates the background, i.e., the non-abnormal tissue area;

[0077] Step 4, Loss Function

[0078] The module training uses a binary cross-entropy loss function, which is defined as follows.

[0079] (12)

[0080] in y ij The coordinates in the image are ( i , j The true label of the pixel (abnormal tissue region or non-abnormal tissue region).

[0081] Preferably, the abnormal tissue classification and thymoma staging module performs the following steps:

[0082] (1) Input is an abnormal tissue region R CT images after denoising and enhancement ;

[0083] (2) The process is as follows: ViT is used to extract global features of abnormal tissue regions, followed by simultaneous classification of abnormal tissues and... Masaoka Phased tasks;

[0084] (3) The output is the classification result of abnormal tissue regions. C type and Masaoka Phased Results P ;

[0085] Step (2) specifically includes:

[0086] (2-1) Feature extraction of abnormal tissue regions

[0087] (13)

[0088] (2-2) Classification task

[0089] Classification is performed using a fully connected layer, outputting the abnormal organization type. :

[0090] (14)

[0091] in These are the weights of the classification layer. It is a bias term. Softmax It is an activation function. It is the first i ViT features of abnormal tissue regions If the value is 0, then it is a thymic cyst. =1, then it is a thymoma;

[0092] (2-3) Phased tasks

[0093] If the abnormal tissue area is a thymoma, then regression analysis is performed to stage the thymoma using a regression network:

[0094] (15)

[0095] in These are the weights of the regression layer. It is the bias term of the regression layer. It is the predicted number i A thymoma Masaoka The phased results take values ​​of ;

[0096] (2-4) Loss Function

[0097] (16)

[0098] The classification loss is:

[0099] (17)

[0100] Ci The abnormal organization is actually of the organizational type;

[0101] The regression loss is:

[0102] (18)

[0103] P i True thymoma Masaoka Installment payments.

[0104] This invention uses a sliding window approach to generate multiple anchor frames; the use of anchor frames with different sizes and aspect ratios is more suitable for early thymoma detection tasks. This system is comprehensive in its tasks covering various pathological types of thymomas and thymic cysts, and supports auxiliary diagnosis.

[0105] The beneficial effects of this invention are as follows:

[0106] This invention, by introducing a Region Focus Network (RFN) and combining it with existing U-Net++ and Vision Transformer (ViT) networks, can more accurately capture the lesion region features in CT images of thymic diseases, improving the system's ability to detect early thymomas and reducing errors. This invention can accomplish the tasks of detecting, segmenting, and classifying thymic diseases based on CT images, improving the overall efficiency and accuracy of diagnosis, significantly reducing the workload of physicians, and has broad prospects for practical clinical applications.

[0107] This invention targets the detection, segmentation, classification, and staging of small lesions in the early stages of thymoma. The tasks it can perform are more comprehensive, and the detection of these small lesions has a wider range of applications in physical examinations. Attached Figure Description

[0108] Figure 1 This is the system roadmap of the present invention. Detailed Implementation

[0109] The present invention will be further described below with reference to the embodiments and accompanying drawings, but is not limited thereto.

[0110] Example 1:

[0111] A regional attention network-based intelligent imaging diagnostic system for thymic diseases, such as Figure 1 As shown, the system includes a CT image preprocessing module, an abnormal tissue detection and localization module, an abnormal tissue segmentation module, and an abnormal tissue classification and thymoma staging module.

[0112] The CT image preprocessing module is used to denoise and enhance CT images to improve image quality and the effectiveness of subsequent model training. The abnormal tissue detection and localization module uses a region attention network and ViT to detect and localize abnormal tissues in CT images, improving the sensitivity for small lesions and early lesions. The abnormal tissue segmentation module uses the U-Net++ network to automatically segment abnormal tissue regions and extract key feature regions. The abnormal tissue classification and thymoma staging module uses ViT to extract features from abnormal tissue regions, and then performs tissue classification and staging tasks simultaneously.

[0113] The CT image preprocessing module removes noise and enhances image contrast by performing the following steps:

[0114] (1) Image denoising: Let the original image be... I raw Mean filtering was used to denoise the CT images, resulting in denoised images. J .

[0115] (2) Image enhancement: Gamma transform is used to adjust the brightness and contrast of the denoised CT images, making the boundaries of lesions more obvious, thereby improving the visibility of lesions in CT images and facilitating subsequent identification and detection. Enhanced CT images , H in for I height, W in for I width, C in for I The number of channels.

[0116] The abnormal tissue detection and localization module performs the following steps:

[0117] (1) Input: Denoising and enhanced CT images .

[0118] (2) Process: Multiple candidate regions are generated through the region attention network, and the optimal candidate region, i.e., the abnormal organization region, is selected. R final .

[0119] In step (2), the region-focused network structure and loss function are as follows:

[0120] (2-1) Candidate Region Generation

[0121] The workflow of using a region focus network to generate potential lesion regions from enhanced feature maps is as follows:

[0122] (2-1-1) Anchor Frame Generation

[0123] In enhanced images I The above uses a sliding window method to generate multiple anchor frames; to adapt to the detection task of thymus disease, anchor frames of different sizes and aspect ratios are generated. Further preferred, the size and aspect ratio of the anchor frames are set as follows:

[0124] Set the anchor frame size as follows:

[0125] w 1=32, h 1=32 (small-sized anchor frame)

[0126] w 2=64, h 2=64 (medium-sized anchor frame)

[0127] w 3=128, h 3 = 128 (large anchor frame)

[0128] Set the aspect ratio of the anchor frame as follows:

[0129]

[0130]

[0131]

[0132] The number of anchor frames at each location is The coordinates of each anchor frame are obtained through ( x , y , w , h ) indicates that ( x , y () is the center position of the anchor frame. w and h These are the width and height of the anchor frame.

[0133] (2-1-2) Anchor Frame Feature Extraction

[0134] For each anchor frame ViT is used to extract features at the corresponding locations. The formula is:

[0135] (1)

[0136] in, It is the first Anchor frame In images I Features at the corresponding positions, T This is the ViT model.

[0137] (2-1-3) Classification and Regression Tasks

[0138] For each anchor box, the Region Focus Network (RFN) needs to perform two tasks:

[0139] (a) Classification task: Determine whether each anchor box contains abnormal tissue.

[0140] use sigmod The activation function performs binary classification (whether it contains abnormal tissue); assuming the first... i The category output for each anchor box is ,but

[0141] (2)

[0142] in, W cls These are the weights of the classification layer. b cls It is a bias term. yes sigmod Activation function; if This anchor frame is considered one of the candidate regions, thus obtaining the candidate region. R candidate .

[0143] (b) Regression task: Based on the overlap between the anchor frame and the actual lesion area, perform bounding box regression and adjust the coordinates of the anchor frame.

[0144] The regression output for each anchor box is the offset of that anchor box from the actual border. This is used to adjust the coordinates of the anchor frame so that it matches the target more precisely; This is the offset of the anchor frame center point. The offset of the anchor frame size is given by the regression formula:

[0145] (3)

[0146] in, W reg These are the weights of the regression layer. b reg It is a bias term. This represents the regression output of the anchor box. For the first i Anchor frame anchor i In images I Features at the corresponding positions.

[0147] (2-2) Candidate region screening and positioning

[0148] The non-maximum suppression (NMS) algorithm is used to select the optimal candidate regions, while ensuring that the overlap between candidate regions does not exceed a set threshold. It returns the optimal candidate region, i.e., the abnormal tissue region. R final :

[0149] (4)

[0150] in IoU It measures the degree of overlap between candidate regions and final abnormal tissue regions; it also outputs the location and size of the detected abnormal tissue regions. .

[0151] (2-3) Loss Function

[0152] (5)

[0153] Let be the smoothing coefficient, where the classification loss is:

[0154] (6)

[0155] in p i It is the first i The real label of the anchor box These are the predicted labels output by the classification task. N This represents the number of anchor frames.

[0156] The regression loss measures the difference between the regression output of the anchor box and the true bounding box, and is expressed as:

[0157] (7)

[0158] in t i It is the first i The actual border offset of each anchor frame. These are the predicted regression values.

[0159] (3) The output is: R final Results of anomaly detection and localization D final .

[0160] The abnormal tissue segmentation module uses the U-Net++ network to automatically segment abnormal tissue regions and outputs the abnormal tissue regions. R Accurate segmentation of abnormal tissue areas provides a reliable foundation for subsequent lesion detection, localization, and analysis. The specific steps are as follows:

[0161] Step 1: Encoder for feature extraction and dimensionality reduction

[0162] (1) Convolution operation

[0163] Extracting local features H represents a local feature. F The height is 1, and W is a local feature. F The width is 1, and C is a local feature. F 1 channel number;

[0164] (8)

[0165] in K 1 is the size of convolution kernel, b 1 indicates the bias term. i express I The first in terms of altitude i One location, j express I The first in width j One location, c express I The first one on the passage c One location;

[0166] m and n are the convolution kernels. K 1. Sliding offset in height and width, k The width or height of the convolution kernel. This is the index of the input channel. C in Input the number of channels. C out This represents the number of output channels.

[0167] (2) Pooling operation

[0168] F 1. Through size Step size is s Maximum pooling window, output The expression is as follows:

[0169] (9)

[0170] Step 2: Decoder for Feature Recovery and Fusion

[0171] (1) Upsampling operation

[0172] I After deconvolution kernel To restore the spatial dimension, we get :

[0173] (10)

[0174] in b 2 represents the bias term, and for F 3. Perform bilinear interpolation to obtain ; H in , H out For input height and output height, W in , W out For input width and output width.

[0175] (2) Skip connections and convolution

[0176] See separately F 1. F 2. F 3 and F 4 The decoder fuses encoder features through skip connections, thus... and spliced ;right Convolution is performed Furthermore, multi-scale features are fused to obtain a feature representation that includes both local and global information. F .

[0177] Step 3, Binary Segmentation

[0178] right F Each pixel in F ( i, j )application sigmod Function and set threshold T =0.5, yielding the binary segmentation result. :

[0179] (11)

[0180] in This indicates the foreground, i.e., the area of ​​abnormal tissue. This indicates the background, i.e., the non-abnormal tissue area.

[0181] Step 4, Loss Function

[0182] The module training uses a binary cross-entropy loss function, which is defined as follows.

[0183] (12)

[0184] in y ij The coordinates in the image are ( i ,j The true label of the pixel (abnormal tissue region or non-abnormal tissue region).

[0185] The abnormal tissue classification and thymoma staging module performs the following steps:

[0186] (1) Input is an abnormal tissue region R CT images after denoising and enhancement .

[0187] (2) The process is as follows: ViT is used to extract global features of abnormal tissue regions, followed by simultaneous classification of abnormal tissues and... Masaoka Phased tasks.

[0188] Step (2) specifically includes:

[0189] (2-1) Feature extraction of abnormal tissue regions

[0190] (13)

[0191] (2-2) Classification task

[0192] Classification is performed using a fully connected layer, outputting the abnormal organization type. :

[0193] (14)

[0194] in These are the weights of the classification layer. It is a bias term. Softmax It is an activation function. It is the first i ViT features of abnormal tissue regions If the value is 0, then it is a thymic cyst. =1 indicates a thymoma.

[0195] (2-3) Phased tasks

[0196] If the abnormal tissue area is a thymoma, then regression analysis is performed to stage the thymoma using a regression network:

[0197] (15)

[0198] in These are the weights of the regression layer. It is the bias term of the regression layer. It is the predicted number i A thymoma Masaoka The phased results take values ​​of ;

[0199] (2-4) Loss Function

[0200] (16)

[0201] The classification loss is:

[0202] (17)

[0203] C i This represents the true organizational type of the abnormal organization.

[0204] The regression loss is:

[0205] (18)

[0206] P i True thymoma Masaoka Installment payments.

[0207] (3) The output is the classification result of abnormal tissue regions. C type and Masaoka Phased Results P .

[0208] Using this system, abnormal tissue detection and localization, abnormal tissue segmentation, abnormal tissue classification, and thymoma staging were performed, and the following evaluation and verification were conducted.

[0209] Abnormal tissue detection and localization involves automated detection and annotation of abnormal areas to ensure accurate tumor identification. The evaluation criteria are as follows:

[0210] Sensitivity: Evaluates the system's ability to detect lesions in abnormal areas. Higher sensitivity means fewer missed detections. The calculation formula is:

[0211] (19)

[0212] in, TP For the detected abnormal areas, FN These are the abnormal areas that were missed during detection.

[0213] Specificity: Assess the system's ability to identify abnormal regions. Higher specificity means a lower false positive rate for healthy tissue. The calculation formula is:

[0214] (20)

[0215] in, TN For pixels correctly classified as non-abnormal regions, FP Pixels that were incorrectly classified as abnormal regions.

[0216] The task of aberrant tissue segmentation is to accurately segment abnormal regions from CT images for subsequent tumor detection and analysis. The evaluation criteria are as follows:

[0217] set up A For automatically segmented abnormal regions, B These are manually labeled outlier regions. Dice similarity coefficient: This metric measures the degree of overlap between automatically segmented results and manually labeled data. The value ranges from 0 to 1; a higher value indicates more accurate segmentation. It can be expressed as:

[0218] (twenty one)

[0219] Intersection over Union (IoU): Evaluates the degree of overlap between segmented regions. IoU A larger value indicates a better segmentation result. The formula is:

[0220] (twenty two)

[0221] Accuracy: Measures whether each pixel is accurately classified; the calculation formula is:

[0222] (twenty three)

[0223] in, TP Pixels that are correctly classified as abnormal regions TN Pixels that are correctly classified as non-abnormal regions FP and FN All of these represent misclassified pixels.

[0224] The abnormal tissue classification and thymoma staging function aims to identify cystic and solid tissue types and thymomas in detected abnormal tissues. Masaoka Staged. System performance can be evaluated using the accuracy, sensitivity, and specificity metrics described above. This system has good detection capabilities; after five-fold cross-validation, its sensitivity is no less than 95%, and its diagnostic accuracy can reach over 92%.

Claims

1. A smart imaging diagnostic system for thymic diseases based on regional attention networks, characterized in that, The system includes a CT image preprocessing module, an abnormal tissue detection and localization module, an abnormal tissue segmentation module, and an abnormal tissue classification and thymoma staging module; The CT image preprocessing module is used for denoising and enhancing CT images; the abnormal tissue detection and localization module uses a region attention network and ViT to detect and localize abnormal tissues in CT images; the abnormal tissue segmentation module uses the U-Net++ network to automatically segment abnormal tissue regions, extract key feature regions, and output the abnormal tissue regions. The abnormal tissue classification and thymoma staging module is used to extract features from abnormal tissue regions using ViT, and then simultaneously perform tissue classification and staging tasks. The abnormal tissue classification and thymoma staging module performs the following steps: (1) The input is the abnormal tissue region. CT images after denoising and enhancement (2) The process is as follows: use ViT to extract global features of abnormal tissue regions, and then simultaneously perform abnormal tissue classification and... Phased tasks; (3) Output is the classification result of abnormal tissue regions. and Phased Results ; The abnormal tissue detection and localization module performs the following steps: (1) Input: Denoising and enhanced CT images ; (2) Process: Multiple candidate regions are generated through the region attention network, and the optimal candidate region, i.e., the abnormal organization region, is selected. ; (3) The output is: Results of anomaly detection and localization ; In step (2) of the abnormal tissue detection and localization module, the region-focused network structure and loss function are as follows: (2-1) Candidate Region Generation The workflow of using a region focus network to generate potential lesion regions from enhanced feature maps is as follows: (2-1-1) Anchor Frame Generation In enhanced images Above, multiple anchor boxes are generated using a sliding window method; Generate anchor frames of different sizes and aspect ratios; The coordinates of each anchor frame are obtained through... It means that, among them It is the center position of the anchor frame. and These are the width and height of the anchor frame; (2-1-2) Anchor Frame Feature Extraction For each anchor frame ViT is used to extract features at the corresponding locations. The formula is: （1） in, It is the first Anchor frame In images Features at the corresponding positions, For the ViT model; (2-1-3) Classification and Regression Tasks For each anchor box, the region focus network needs to perform two tasks: (a) Classification task: Determine whether each anchor box contains abnormal tissue. use The activation function performs binary classification; assuming the first... The category output for each anchor box is ,but （2） in, These are the weights of the classification layer. It is a bias term. yes Activation function; if This anchor frame is considered one of the candidate regions, thus obtaining the candidate region. ; (b) Regression task: Based on the overlap between the anchor frame and the actual lesion area, perform bounding box regression and adjust the coordinates of the anchor frame; The regression output for each anchor box is the offset of that anchor box from the actual border. , used to adjust the coordinates of the anchor frame; This is the offset of the anchor frame center point. The offset of the anchor frame size is given by the regression formula: （3） in, These are the weights of the regression layer. It is a bias term. This represents the regression output of the anchor box. For the first Anchor frame In images Features at the corresponding positions; (2-2) Candidate region screening and positioning The non-maximum suppression algorithm is used to select the optimal candidate regions, while ensuring that the overlap between candidate regions does not exceed a set threshold. It returns the optimal candidate region, i.e., the abnormal tissue region. : （4） in It measures the degree of overlap between candidate regions and final abnormal tissue regions; it also outputs the location and size of the detected abnormal tissue regions. ; (2-3) Loss Function （5） Let be the smoothing coefficient, where the classification loss is: = （6） in It is the first The real label of the anchor box These are the predicted labels output by the classification task. This refers to the number of anchor frames; The regression loss measures the difference between the regression output of the anchor box and the true bounding box, and is expressed as: （7） in It is the first The actual border offset of each anchor frame. These are the predicted regression values.

2. The intelligent imaging diagnostic system for thymic diseases based on regional attention networks according to claim 1, characterized in that, The CT image preprocessing module removes noise and enhances image contrast by performing the following steps: (1) Image denoising: Let the original image be Mean filtering was used to denoise the CT images, resulting in denoised images. ; (2) Image enhancement: Gamma transform is used to adjust the brightness and contrast of the denoised CT images to enhance the CT images. , for height, for width, for The number of channels.

3. The intelligent imaging diagnostic system for thymic diseases based on regional attention networks according to claim 1, characterized in that, The abnormal tissue segmentation module uses the U-Net++ network to automatically segment abnormal tissue regions and outputs the abnormal tissue regions. The specific steps are as follows: Step 1: Encoder for feature extraction and dimensionality reduction (1) Convolution operation Extracting local features H represents a local feature. The height, W is a local feature. The width of C, where C is a local feature. The number of channels; （8） in It is the size of convolution kernel, Indicates the bias term. express The first in terms of altitude One location, The first in width One location, The first one on the passage One location; m and n are the convolution kernels K 1. Sliding offset in height and width, k The width or height of the convolution kernel. This is the index of the input channel. Input the number of channels. Number of output channels; (2) Pooling operation Through size Step size is Maximum pooling window, output The expression is as follows: （9） Step 2: Decoder for Feature Recovery and Fusion (1) Upsampling operation After deconvolution kernel To restore the spatial dimension, we get : （10） in Represents the bias term, and for Perform bilinear interpolation to obtain ; , For input height and output height, , Input width, output width; (2) Skip connections and convolution See separately for ; The decoder fuses encoder features through skip connections. and spliced ;right Convolution is performed Furthermore, multi-scale features are fused to obtain a feature representation that includes both local and global information. ; Step 3, Binary Segmentation right Each pixel in application Function and set threshold The binary segmentation result is obtained. : （11） in This indicates the foreground, i.e., the area of ​​abnormal tissue. This indicates the background, i.e., the non-abnormal tissue area; Step 4, Loss Function The module training uses a binary cross-entropy loss function, which is defined as follows. （12） in The coordinates in the image are The true label of the pixels.

4. The intelligent imaging diagnostic system for thymic diseases based on regional attention networks according to claim 1, characterized in that, The steps (2) of the abnormal tissue classification and thymoma staging module specifically include: (2-1) Feature extraction of abnormal tissue regions （13） (2-2) Classification task Classification is performed using a fully connected layer, outputting the abnormal organization type. : （14） in These are the weights of the classification layer. It is a bias term. It is an activation function. It is the first ViT features of abnormal tissue regions If the value is 0, then it is a thymic cyst. =1, then it is a thymoma; (2-3) Phased tasks If the abnormal tissue area is a thymoma, then regression analysis is performed to stage the thymoma using a regression network: （15） in These are the weights of the regression layer. It is the bias term of the regression layer. It is the predicted number i The Masaoka staging results for a thymoma are as follows, with values ​​ranging from 1 to 10. ; (2-4) Loss Function （16） The classification loss is: （17） The abnormal organization is the actual organization type; The regression loss is: （18） For thymoma Installment payments.

5. The intelligent imaging diagnostic system for thymic diseases based on regional attention networks according to claim 3, characterized in that, In step (2-1-1), the dimensions and aspect ratio of the anchor frame are set as follows: Set the anchor frame size as follows: (Small-sized anchor frame) · (Medium-sized anchor frame) (Large-sized anchor frame) Set the aspect ratio of the anchor frame as follows: The number of anchor frames at each location is .

Images