Artificial intelligence evaluation method and system for analyzing and predicting colorectal cancer based on lumbar major muscle image
Through deep learning methods and ConvNeXt structure, the accuracy and inefficiency of psoas major muscle image analysis to predict colorectal cancer is solved, and high accuracy and high efficiency automated analysis is achieved, which promotes the early diagnosis and prevention of colorectal cancer.
Patent Information
- Application Number
- CN202411220453.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-02
- Publication Date
- 2024-12-10
AI Technical Summary
The existing methods for predicting colorectal cancer based on psoas major muscle image analysis have problems such as low artificial identification accuracy, poor data acquisition quality and limited computing resources, resulting in low accuracy and inefficiency.
Using the ConvNeXt structure based on deep learning to construct an artificial intelligence evaluation method. Through data collection, preprocessing, model training and model evaluation modules, the CT images of the psoas major muscle are automatically analyzed to reduce human factors interference, improve prediction accuracy and efficiency, and Provides a friendly interface and interaction method through the user interaction module.
It improves the accuracy and efficiency of psoas major muscle image analysis to predict colorectal cancer, reduces the work burden of doctors, enhances the early diagnosis rate, and promotes the development and application of medical technology.
Smart Images
Figure CN119108102A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to an artificial intelligence evaluation method and system, and in particular to an artificial intelligence evaluation method and system for predicting colorectal cancer based on psoas muscle image analysis, and belongs to the technical field of image data processing. Background Art
[0002] In the existing technology, the prediction of colorectal cancer based on image analysis has made significant progress, but the following problems still exist: (1) The method of manual statistical judgment is inaccurate. In traditional medical image analysis, manual statistics and judgment are commonly used methods. Studies have shown that changes in the psoas major muscle are related to the incidence of colorectal cancer. At the same time, changes in the psoas major muscle are related to the prognosis of colorectal cancer. However, it is difficult for manual recognition methods to accurately judge the image changes of the psoas major muscle. It has the defects of strong subjectivity and low accuracy. Due to the influence of factors such as the doctor's experience, knowledge background and fatigue level, the results of manual statistical judgment may have large errors, especially when processing a large amount of image data, the error will be further amplified. In addition, the method of manual statistical judgment is also difficult to cope with complex and changeable image features. The manifestation of colorectal cancer in the psoas major muscle image may vary due to various factors such as individual differences of patients, disease development stage, image acquisition conditions, etc. , which makes manual statistical judgment more difficult, less accurate, and time-consuming; (2) The quality of collected data is poor. Due to the influence of multiple factors such as image acquisition equipment, acquisition conditions, and individual differences among patients, the collected image data may have problems such as noise, artifacts, and inconsistent resolution, which will affect the accuracy and reliability of subsequent image analysis. Some methods choose to directly collect colorectal cancer lesions as the judgment data for disease analysis, but because the position of colorectal cancer lesions in CT images will change, the acquisition process is more complicated and may result in missed diagnosis, and it is impossible to effectively analyze whether the patient is colorectal cancer; (3) The complexity of technical implementation and the limitation of computing resources. Although advanced technologies such as deep learning have provided new ideas and methods for medical image analysis, the implementation of the above technologies requires a lot of computing resources, such as graphics processing units (GPUs) and time costs, which limits the scope of application and popularity of the technology.
[0003] In summary, the current artificial intelligence technology for predicting colorectal cancer based on psoas muscle image analysis still has some problems and defects, and further research and exploration are needed. In order to improve the accuracy and reliability of the technology, an artificial intelligence evaluation method and system for predicting colorectal cancer based on psoas muscle image analysis is needed, which can optimize the algorithm design, improve the quality of collected data, and improve prediction accuracy and efficiency. Summary of the invention
[0004] A brief overview of the present invention is provided below in order to provide a basic understanding of certain aspects of the present invention. It should be understood that this overview is not an exhaustive overview of the present invention. It is not intended to identify key or important parts of the present invention, nor is it intended to limit the scope of the present invention. Its purpose is merely to present certain concepts in a simplified form as a prelude to a more detailed description discussed later.
[0005] In view of this, in order to solve the problems of low manual recognition accuracy and low machine recognition efficiency in the traditional method of predicting colorectal cancer based on psoas muscle image analysis in the prior art, the present invention provides an artificial intelligence evaluation method and system for predicting colorectal cancer based on psoas muscle image analysis.
[0006] Technical solution 1 is as follows: an artificial intelligence evaluation method for predicting colorectal cancer based on psoas muscle image analysis, comprising the following steps:
[0007] S1. Use the data collection module to obtain psoas muscle CT image data and the corresponding colorectal cancer pathological diagnosis results, use the image annotation tool to mark the psoas muscle CT image data for classification, build a data set, and divide the data set into a training set, a validation set, and a test set;
[0008] S2. Processing the data set using a preprocessing module, that is, cleaning, standardizing and enhancing the psoas major muscle CT image data to obtain a preprocessed data set;
[0009] S3. Use the model training module to select the ConvNeXt structure with the auxiliary head to build a deep learning model, train the deep learning model with the training set, and use the validation set to evaluate the deep learning model in real time to obtain the trained deep learning model and prediction results;
[0010] S4. The model evaluation module is used to perform a final evaluation on the trained deep learning model through the test set. The accuracy and IoU index are calculated based on the prediction results and colorectal cancer pathological diagnosis results.
[0011] S5. The result output module is used to output the accuracy and IoU indicators to the user interaction module in the form of probability.
[0012] Furthermore, in S1, the psoas major muscle CT image data includes the third lumbar vertebra CT images of non-tumor patients and the third lumbar vertebra CT images of colorectal cancer patients. The pathological diagnosis results of colorectal cancer are directly used as label data, the psoas major muscle in the third lumbar vertebra CT image is marked, and it is divided into a non-tumor group and a colorectal cancer group and used as label data. The data not existing in the non-tumor group and the colorectal cancer group is used as the background, and the background is not marked. The label value of the non-tumor group is defined as 1, and the label value of the colorectal cancer group is defined as 2. The pathological diagnosis results of colorectal cancer, the non-tumor group, the colorectal cancer group and the background are integrated to obtain a data set.
[0013] Furthermore, in S2, cleaning is to adjust the window value of the psoas major muscle CT image data to remove interference data, standardization is to adjust the image size of the psoas major muscle CT image data to 512*512, and grayscale and normalize it, and enhancement processing includes random resizing, random cropping, random flipping and random exposure change.
[0014] Furthermore, in S3, the convolution kernel of the deep learning model is 2;
[0015] The auxiliary heads include bilinear interpolation upsampling, deconvolution upsampling, and anti-pooling upsampling;
[0016] The specific steps of the bilinear interpolation upsampling process are as follows: construct a two-dimensional coordinate system and select four known point coordinates for bilinear interpolation;
[0017] The bilinear interpolation formula is expressed as:
[0018] f(a,b)=f(a1,b1)(1-a)(1-b)+f(a2,b1)a(1-b)+f(a1,b2)(1-a)b+f(a2,b2)ab
[0019] Wherein, f is the interpolation function, (a, b) is the point coordinate, a is the horizontal coordinate of the point, b is the vertical coordinate of the point, (a1, b1) is the coordinate of the first known point, (a1, b2) is the coordinate of the second known point, (a2, b1) is the coordinate of the third known point, and (a2, b2) is the coordinate of the fourth known point;
[0020] The specific steps of the deconvolution upsampling process are as follows: set the parameters of the ConvNeXt structure, Channels is the number of channels of the feature map, Channels = 256, Dropout_ratio is a regularization operation, Dropout_ratio = 0.1, In_channels is the number of channels of the input feature map, In_channels = 512, In_index is the channel with the index number selected in the input feature map, In_index = 2;
[0021] Design a feature extraction network based on the ConvNeXt structure with an auxiliary head to extract key features related to colorectal cancer, design a classification layer to map the extracted key features to the corresponding prediction results, use the training set to train the deep learning model, and introduce the loss function L to optimize the parameters of the deep learning model through the back propagation algorithm and gradient descent method;
[0022] The loss function L is expressed as:
[0023]
[0024] Where yi represents the label of sample i in the training set, the positive class is 1, and the negative class is 0. pi represents the probability that sample i in the training set is predicted to be a positive class, i = 1, 2, ... n, N is the data set of n samples in the training set;
[0025] During the optimization process, the initial parameter is 0.4, the feature map output index generated by each state is: [0, 1, 2, 3], the decoder type is UPerHead, the optimizer is AdamW, and the fully connected layer is FCNHead;
[0026] The optimization formula is expressed as:
[0027] moving_avg=alpha*moving_avg+(1-alpha)*w.grad
[0028] w=w-Lr*moving_avg-Lr*wd*w
[0029] Among them, wd is the set hyperparameter, Lr is the learning rate, w.grad is the derivative of the loss function with respect to w, w is the weight parameter, wd*w represents the derivative of the penalty term with respect to w, moving_avg is the moving average, and alpha is a smoothing coefficient between 0 and 1;
[0030] The specific steps of the de-pooling upsampling process are as follows: first, record the maximum value position of the feature map, that is, when performing maximum pooling, it is necessary to record the specific position of the maximum value in each pooling window; second, restore the original size of the pooled feature map to obtain the restored feature map; finally, fill the maximum value, that is, in the restored feature map, put the previously recorded maximum value back to the initial position, and fill the remaining positions with 0.
[0031] Furthermore, in S4, the accuracy Acc is expressed as:
[0032]
[0033] Among them, TP is a true positive example, that is, the prediction result of the deep learning model is a positive sample, and the actual result is also a positive sample; TN is a true negative example, that is, the prediction result of the deep learning model is a negative sample, and the actual result is also a negative sample; FP is a false positive example, that is, the prediction result of the deep learning model is a positive sample, and the actual result is a negative sample; FN is a false negative example, that is, the prediction result of the deep learning model is a negative sample, and the actual result is a positive sample;
[0034] The IoU metric (Intersection over Union) is expressed as:
[0035]
[0036] Among them, A1 is the predicted result and A2 is the actual result.
[0037] Technical solution 2 is as follows: an artificial intelligence evaluation system for predicting colorectal cancer based on psoas muscle image analysis, which is used to implement the artificial intelligence evaluation method for predicting colorectal cancer based on psoas muscle image analysis described in technical solution 1, including a data collection module, a preprocessing module, a model training module, a model evaluation module, a result output module and a user interaction module;
[0038] The data collection module, preprocessing module, model training module, model evaluation module and result output module are connected in sequence;
[0039] The data collection module and the result output module are respectively connected to the user interaction module;
[0040] The data collection module is used to collect psoas muscle image data and related diagnostic data, and obtain uploaded data;
[0041] The preprocessing module is used to clean, standardize and enhance the data collected and acquired by the data collection module;
[0042] The model training module is used to build a deep learning model, train it with the data processed by the preprocessing module, and evaluate and optimize the deep learning model in real time through the validation set, and output the prediction results;
[0043] The model evaluation module is used to evaluate the trained model using independent test set data and generate a detailed evaluation report;
[0044] The result output module is used to display the prediction results to the user and provide the prediction basis and credibility evaluation of the prediction results;
[0045] The user interaction module is used to select functions, set parameters, view prediction results and upload data.
[0046] The beneficial effects of the present invention are as follows: the present invention overcomes the shortcomings of traditional statistical methods. When judging the relationship between psoas major muscle CT images and colorectal cancer, traditional manual analysis methods are easily affected by factors such as doctor experience and fatigue, resulting in inaccurate prediction results. The present invention uses artificial intelligence technology to automatically analyze and extract features of psoas major muscle CT images through a deep learning model, thereby reducing the interference of human factors and improving prediction accuracy and efficiency. In the face of massive medical image data, traditional manual analysis methods consume a lot of manpower and time and are inefficient. The present invention realizes rapid analysis and prediction of psoas major muscle CT images through an automated processing flow, greatly improving analysis efficiency. Traditional methods usually lack a friendly user interface and interactive mode, which makes doctors feel inconvenient during use. The present invention designs a user-friendly interface. The user interface and interactive method are convenient for doctors to upload images and obtain prediction results, and a detailed explanation of the prediction results is provided, thereby optimizing the user experience. Traditional methods may have instability and reliability problems when processing complex and changeable image data. The present invention strictly screens and optimizes the input data and models through a preprocessing module, a model training module, and a model evaluation module to ensure the stability and reliability of the system. Through advanced technologies such as deep learning, the automatic analysis and feature extraction of psoas major muscle images are realized, thereby providing strong support for the early diagnosis and prevention of colorectal cancer. In summary, the present invention uses artificial intelligence technology to perform in-depth analysis and prediction of psoas major muscle images, which not only improves the prediction accuracy and efficiency, reduces the workload of doctors, and improves the early diagnosis rate, but also promotes the development and application of medical technology. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:
[0048] Figure 1 A flowchart of an artificial intelligence evaluation method for predicting colorectal cancer based on psoas muscle image analysis;
[0049] Figure 2 A schematic diagram of the structure of an artificial intelligence evaluation system for predicting colorectal cancer based on psoas muscle image analysis;
[0050] Figure 3 Schematic diagram of the deep learning model. DETAILED DESCRIPTION
[0051] In order to make the technical solutions and advantages of the embodiments of the present invention more clearly understood, the exemplary embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than an exhaustive list of all the embodiments. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.
[0052] Example 1: Reference Figure 1 and Figure 2 The present embodiment is described in detail, and the artificial intelligence evaluation method for predicting colorectal cancer based on psoas muscle image analysis specifically includes the following steps:
[0053] S1. Use the data collection module to obtain psoas muscle CT image data and the corresponding colorectal cancer pathological diagnosis results, use the image annotation tool to mark the psoas muscle CT image data for classification, build a data set, and divide the data set into a training set, a validation set, and a test set;
[0054] S2. Processing the data set using a preprocessing module, that is, cleaning, standardizing and enhancing the psoas major muscle CT image data to obtain a preprocessed data set;
[0055] S3. Use the model training module to select the ConvNeXt structure with the auxiliary head to build a deep learning model, train the deep learning model with the training set, and use the validation set to evaluate the deep learning model in real time to obtain the trained deep learning model and prediction results;
[0056] S4. The model evaluation module is used to perform a final evaluation on the trained deep learning model through the test set. The accuracy and IoU index are calculated based on the prediction results and colorectal cancer pathological diagnosis results.
[0057] S5. The result output module is used to output the accuracy and IoU indicators to the user interaction module in the form of probability.
[0058] Further, in S1, the psoas major muscle CT image data includes the third lumbar vertebra CT images of non-tumor patients and the third lumbar vertebra CT images of colorectal cancer patients, the pathological diagnosis results of colorectal cancer are directly used as label data, the psoas major muscle in the third lumbar vertebra CT image is marked, and it is divided into a non-tumor group and a colorectal cancer group and used as label data, the data not existing in the non-tumor group and the colorectal cancer group is used as background, the background is not marked, the label value of the non-tumor group is defined as 1, the label value of the colorectal cancer group is defined as 2, and the pathological diagnosis results of colorectal cancer, the non-tumor group, the colorectal cancer group and the background are integrated to obtain a data set;
[0059] Specifically, we collected 1500 third lumbar vertebrae CT images (L3) of non-tumor patients and 1457 third lumbar vertebrae CT images of colorectal cancer patients to ensure that the data covered patients of different ages, genders, disease severity and corresponding control healthy groups. At the same time, we used the image annotation tool Labelme to mark the psoas major muscle of L3 in the two groups of data, group 1 was the non-tumor group and group 2 was the colorectal cancer group. The data sets were divided into training set, validation set and test set in a ratio of 7:2:1 to ensure that the test set was unknown in the subsequent model training process.
[0060] Furthermore, in S2, cleaning is to adjust the window value of the psoas major muscle CT image data, remove interference data, standardize the image data to 512*512, and grayscale and normalize, and enhance the image data including random size change, random cropping, random flipping and random exposure change;
[0061] Specifically, preprocessing the image data can improve the robustness of the model data, cleaning can ensure image quality, standardization processing can facilitate subsequent model processing, and enhancement processing performs a series of operations on the image, such as random resizing, random cropping, random flipping, and random exposure changes. Random resizing means adjusting the image size. Adjusting the image size to 512*512 helps the model learn targets of different scales and improves the model's generalization ability for targets of different sizes. Randomly cropping the image means adjusting the window value, which increases the model's attention to different areas in the image and helps the model better learn the spatial context and local features of the target. Randomly flipping the image is generally horizontal flipping, which helps the model learn features that are insensitive to direction, because targets in the real world may appear in any direction. Randomly changing the exposure, such as changes in brightness, contrast, saturation, and hue, simulates changes in lighting conditions in the real world and improves the robustness of the model under different lighting conditions.
[0062] Furthermore, in S3, the convolution kernel of the deep learning model is 2;
[0063] The auxiliary heads include bilinear interpolation upsampling, deconvolution upsampling, and anti-pooling upsampling;
[0064] The specific steps of the bilinear interpolation upsampling process are as follows: construct a two-dimensional coordinate system and select four known point coordinates for bilinear interpolation;
[0065] The bilinear interpolation formula is expressed as:
[0066] f(a,b)=f(a1,b1)(1-a)(1-b)+f(a2,b1)a(1-b)+f(a1,b2)(1-a)b+f(a2,b2)ab
[0067] Wherein, f is the interpolation function, (a, b) is the point coordinate, a is the horizontal coordinate of the point, b is the vertical coordinate of the point, (a1, b1) is the coordinate of the first known point, (a1, b2) is the coordinate of the second known point, (a2, b1) is the coordinate of the third known point, and (a2, b2) is the coordinate of the fourth known point;
[0068] The specific steps of the deconvolution upsampling process are as follows: Set the parameters of the ConvNeXt structure, Channels is the number of channels of the feature map, Channels = 256, Dropout_ratio is a regularization operation used to reduce overfitting in the neural network. When Dropout_ratio = 0.1 is set, it means that each neuron has a 10% probability of being randomly discarded during training. In_channels is the number of channels of the input feature map, In_channels = 512, and In_index = 2 is to select the channel with index 2 in the input feature map;
[0069] Design a feature extraction network based on the ConvNeXt structure with an auxiliary head to extract key features related to colorectal cancer, design a classification layer (UPerHead, Upper Pyramid Head), map the extracted key features to the corresponding prediction results, use the training set to train the deep learning model, and introduce the loss function L to optimize the parameters of the deep learning model through the back propagation algorithm and gradient descent method;
[0070] The loss function L (CrossEntropyLoss) is expressed as:
[0071]
[0072] Where yi represents the label of sample i in the training set, the positive class is 1, and the negative class is 0. pi represents the probability that sample i in the training set is predicted to be a positive class, i = 1, 2, ... n, N is the data set of n samples in the training set;
[0073] During the optimization process, the initial parameter is 0.4, the feature map output index generated by each state is: [0, 1, 2, 3], the decoder type is UPerHead, the optimizer is AdamW, and the fully connected layer is FCNHead;
[0074] The optimization formula is expressed as:
[0075] moving_avg=alpha*moving_avg+(1-alpha)*w.grad
[0076] w=w-Lr*moving_avg-Lr*wd*w
[0077] Among them, wd is the set hyperparameter, Lr is the learning rate, w.grad is the derivative of the loss function with respect to w, w is the weight parameter, wd*w represents the derivative of the penalty term with respect to w, moving_avg is the moving average, and alpha is a smoothing coefficient between 0 and 1;
[0078] De-pooling upsampling is used to restore the spatial dimension information lost in the pooling process when processing image data. The specific steps of the de-pooling upsampling process are as follows: First, record the maximum position of the feature map, that is, when performing maximum pooling, it is necessary to record the specific position of the maximum value in each pooling window. Secondly, restore the original size of the pooled feature map, which usually means multiplying each dimension of the feature map by the size of the pooling window to obtain the restored feature map. Finally, fill the maximum value, that is, in the restored feature map, put the previously recorded maximum value back to the initial position, and fill the remaining positions with 0;
[0079] Specifically, in this embodiment, Lr=0.0001, wd=0.05;
[0080] refer to Figure 3 , the psoas muscle CT images with a uniform resolution of 512×512 are input into the deep learning model, 512×512×3 represents the RGB three-layer channels of the image, D7×7 represents the 7×7 deep convolution, LN represents the technique of normalizing each layer of each sample, Gelu represents the nonlinear activation function, 128 represents the number of channels of the output feature map, 1×1 represents convolution, 512 represents the number of channels of the output feature map, and + represents the judgment mechanism, whether to apply this layer or the next layer;
[0081] The ConvNeXt structure is an architecture based on a convolutional neural network (CNN). It combines the traditional convolutional neural network design with certain features of a neural network (Transformer) based on a self-attention mechanism, aiming to improve the performance of the model in image recognition and other computer vision tasks. The following are the main features of the ConvNeXt structure: (1) Deeply stacked convolutional layers: The ConvNeXt structure builds a network by deeply stacking multiple convolutional layers. Each convolutional layer is responsible for learning specific features in the input image. The convolutional layer can capture various features from low-level to high-level, including edges, textures, and shapes. (2) Pyramid structure: The ConvNeXt structure adopts a pyramid structure, which means that as the network goes deeper, the resolution of the feature map gradually decreases, but the number of channels, that is, the depth of the feature map, gradually increases, which helps to capture multi-scale information in the image, thereby accurately detecting and segmenting objects of different sizes. (3) Residual connection Connections): Residual connections are introduced in the ConvNeXt structure, which is an effective technique to prevent gradient vanishing and model degradation in deep networks. Residual connections make it easier to learn identity mappings, i.e., input equals output, making it easier to optimize deep networks. (4) Normalization and activation functions: The ConvNeXt structure also uses normalization layers (such as BatchNormalization) and activation functions (such as ReLU functions or Swish functions) to speed up training and improve model performance. Normalization layers help reduce internal covariate shifts, while activation functions increase the nonlinearity of the network. (5) Group convolution and dense connections: In order to further improve the performance of the model, the ConvNeXt structure also introduces group convolution and dense connections. Connection) and other technologies. Grouped convolution reduces computational complexity and memory usage by dividing feature maps into multiple groups and performing convolution operations separately, while dense connection enhances feature propagation and reuse by connecting the output of each convolution layer with the input of all subsequent layers. (6) Efficient training strategy: The ConvNeXt structure adopts a variety of efficient strategies during training, such as data enhancement, learning rate scheduling, and weight decay, which helps to improve the generalization ability and robustness of the model.
[0082] The main improvements of the traditional ConvNeXt structure include Macro design, ResNeXt, Invertedbottleneck, Invertedbottleneck, Large kernel size and Various layer-wise Micro designs. Macro design adjusts the number of stacking times of each stage of ResNet to make it similar to the stacking ratio of the Transformer network, and draws on the idea of ResNeXt to improve the performance of the network. Invertedbottleneck is similar to the design of MobileNet v2 to improve network efficiency. Large kernel size uses larger convolution kernels to improve the network's perception. Various layer-wise Micro designs include improvements in normalization methods, activation functions, etc.
[0083] The ConvNeXt structure itself does not have an auxiliary head. The present invention adds an auxiliary head to the ConvNexXt structure. The auxiliary head is usually used to introduce additional supervision signals in the deep neural network to help train deeper networks or solve problems such as gradient disappearance. In target detection models such as YOLOv7, the design of the auxiliary head is used to improve the performance of the model. The task type of the auxiliary head in the present invention is to segment the psoas major muscle CT image data and judge the category at the same time. The auxiliary head improves the robustness and accuracy of the model. Real-time evaluation of the model using the verification set can timely discover problems and make adjustments.
[0084] Furthermore, in S4, the accuracy Acc (Accuracy) is expressed as:
[0085]
[0086] Among them, TP (True Positives) is a true positive example, that is, the prediction result of the deep learning model is a positive sample, and the actual result is also a positive sample; TN (True Negatives) is a true negative example, that is, the prediction result of the deep learning model is a negative sample, and the actual result is also a negative sample; FP (False Positives) is a false positive example, that is, the prediction result of the deep learning model is a positive sample, and the actual result is a negative sample (false positive); FN (False Negatives) is a false negative example, that is, the prediction result of the deep learning model is a negative sample, and the actual result is a positive sample (false negative);
[0087] The IoU metric (Intersection over Union) is expressed as:
[0088]
[0089] Among them, A1 is the predicted result, and A2 is the actual result;
[0090] Sensitivity is expressed as:
[0091]
[0092] Specificity Sensitivity is expressed as:
[0093]
[0094] Specifically, accuracy is one of the most commonly used evaluation indicators in classification problems, which indicates the proportion of samples predicted correctly by the model to the total number of samples;
[0095] The IoU indicator is a commonly used evaluation indicator in target detection problems, which is used to measure the degree of overlap between the predicted box (Bounding Box) and the actual box (Ground Truth);
[0096] Sensitivity indicates the proportion of true positive samples in the actual result positive samples, and specificity indicates the proportion of false negative samples in the actual result negative samples. The purpose of calculating sensitivity and specificity is to measure the key parameters of the evaluation method and system performance of the present invention, which respectively reflect the accuracy and reliability of the model in identifying colorectal cancer patients and non-colorectal cancer patients. Sensitivity can evaluate the ability of the system to identify colorectal cancer patients. High sensitivity means that the system can more effectively screen out true colorectal cancer patients from the suspected population and reduce the missed diagnosis rate. Specificity can evaluate the ability of the system to distinguish non-colorectal cancer patients. High specificity means that the system can more accurately identify people who do not have colorectal cancer and reduce the misdiagnosis rate.
[0097] Table 1
[0098] type IoU Acc background 99.88 99.94 Colorectal cancer group 89.11 94.59 Non-tumor group 86.83 92.99
[0099] In this embodiment, Table 1 shows the accuracy and IoU indicators.
[0100] Example 2: Reference Figure 2 This embodiment is described in detail, the artificial intelligence evaluation system for predicting colorectal cancer based on psoas muscle image analysis is used to implement the artificial intelligence evaluation method for predicting colorectal cancer based on psoas muscle image analysis described in Example 1, including a data collection module, a preprocessing module, a model training module, a model evaluation module, a result output module and a user interaction module;
[0101] The data collection module, preprocessing module, model training module, model evaluation module and result output module are connected in sequence;
[0102] The data collection module and the result output module are respectively connected to the user interaction module;
[0103] The data collection module is used to collect psoas muscle image data and related diagnostic data, and obtain uploaded data to ensure that the system has sufficient data for subsequent analysis;
[0104] The preprocessing module is used to clean, standardize and enhance the data collected and acquired by the data collection module to improve the quality of the data and the training effect of the model;
[0105] The model training module is used to build a deep learning model, train it with the data processed by the preprocessing module, and evaluate and optimize the deep learning model in real time through the validation set, and output the prediction results;
[0106] The model evaluation module is used to evaluate the trained model using independent test set data and generate a detailed evaluation report to help users understand the performance of the model;
[0107] The result output module is used to display the prediction results to the user in an intuitive and easy-to-understand manner, and provide the prediction basis and credibility evaluation of the prediction results to help the user better understand the prediction results;
[0108] The user interaction module is used to select functions, set parameters, view prediction results and upload data. It provides a friendly user interface to enhance user experience and enables users to interact with other system components, i.e., it is convenient for users to upload new L3 psoas muscle image data and obtain prediction results, and provides user feedback and opinion collection functions to continuously improve and optimize system parameters;
[0109] Specifically, the entire system process starts with data collection, goes through preprocessing, model training and evaluation, and finally presents the prediction results to the user, forming a complete closed loop. Each module plays an indispensable role, ensuring that the system can accurately and efficiently predict the risk of colorectal cancer and provide strong support for medical diagnosis and patient treatment.
[0110] Although the present invention has been described according to a limited number of embodiments, it will be apparent to those skilled in the art, with the benefit of the above description, that other embodiments may be envisioned within the scope of the invention thus described. In addition, it should be noted that the language used in this specification is selected primarily for readability and teaching purposes, rather than for explaining or defining the subject matter of the present invention. Therefore, many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the appended claims. The disclosure of the present invention is illustrative, not restrictive, with respect to the scope of the present invention, which is defined by the appended claims.
Claims
1. An artificial intelligence evaluation method for predicting colorectal cancer based on psoas muscle image analysis, characterized in that: The following steps are involved: S1. Use the data collection module to obtain psoas muscle CT image data and the corresponding colorectal cancer pathological diagnosis results; use the image annotation tool to mark the psoas muscle CT image data for classification, build a data set, and divide the data set into a training set, a validation set, and a test set; S2. Processing the data set using a preprocessing module, that is, cleaning, standardizing and enhancing the psoas major muscle CT image data to obtain a preprocessed data set; S3. Use the model training module to select the ConvNeXt structure with the auxiliary head to build a deep learning model, train the deep learning model through the training set; use the validation set to evaluate the deep learning model in real time to obtain the trained deep learning model and prediction results; S4. The model evaluation module is used to perform a final evaluation on the trained deep learning model through the test set. The accuracy and IoU index are calculated based on the prediction results and colorectal cancer pathological diagnosis results. S5. The result output module is used to output the accuracy and IoU indicators to the user interaction module in the form of probability.
2. The artificial intelligence evaluation method for predicting colorectal cancer based on psoas muscle image analysis according to claim 1, characterized in that: In S1, the psoas major muscle CT image data includes the third lumbar vertebra CT images of non-tumor patients and the third lumbar vertebra CT images of colorectal cancer patients. The pathological diagnosis results of colorectal cancer are directly used as label data, the psoas major muscle in the third lumbar vertebra CT image is marked, and it is divided into a non-tumor group and a colorectal cancer group and used as label data. The data not existing in the non-tumor group and the colorectal cancer group is used as the background, and the background is not marked. The label value of the non-tumor group is defined as 1, and the label value of the colorectal cancer group is defined as 2. The pathological diagnosis results of colorectal cancer, the non-tumor group, the colorectal cancer group and the background are integrated to obtain a data set.
3. The artificial intelligence evaluation method for predicting colorectal cancer based on psoas muscle image analysis according to claim 2, characterized in that: In S2, cleaning is to adjust the window value of the psoas major muscle CT image data to remove interference data, standardization is to adjust the image size of the psoas major muscle CT image data to 512*512, and grayscale and normalize it, and enhancement processing includes random resizing, random cropping, random flipping and random exposure change.
4. The artificial intelligence evaluation method for predicting colorectal cancer based on psoas muscle image analysis according to claim 3 is characterized in that: In S3, the convolution kernel of the deep learning model is 2; The auxiliary heads include bilinear interpolation upsampling, deconvolution upsampling, and anti-pooling upsampling; The specific steps of the bilinear interpolation upsampling process are as follows: construct a two-dimensional coordinate system and select four known point coordinates for bilinear interpolation; The bilinear interpolation formula is expressed as: f(a;b)=f(a1;b1)(1-a)(1-b)+f(a2;b1)a(1-b)+f(a1;b2)(1-a)b+f(a2;b2)ab Wherein, f is the interpolation function, (a; b) is the point coordinates, a is the horizontal coordinate of the point, b is the vertical coordinate of the point, (a1; b1) is the coordinates of the first known point, (a1; b2) is the coordinates of the second known point, (a2; b1) is the coordinates of the third known point, and (a2; b2) is the coordinates of the fourth known point; The specific steps of the deconvolution upsampling process are as follows: set the parameters of the ConvNeXt structure, Channels is the number of channels of the feature map, Channels = 256, Dropout_ratio is a regularization operation, Dropout_ratio = 0.1, In_channels is the number of channels of the input feature map, In_channels = 512, In_index is the channel with the index number selected in the input feature map, In_index = 2; Design a feature extraction network based on the ConvNeXt structure with an auxiliary head to extract key features related to colorectal cancer, design a classification layer to map the extracted key features to the corresponding prediction results, use the training set to train the deep learning model, and introduce the loss function L to optimize the parameters of the deep learning model through the back propagation algorithm and gradient descent method; The loss function L is expressed as: Where yi represents the label of sample i in the training set, the positive class is 1, and the negative class is 0. pi represents the probability that sample i in the training set is predicted to be a positive class, i = 1, 2, ... n, N is the data set of n samples in the training set; During the optimization process, the initial parameter is 0.4, the feature map output index generated by each state is: [0; 1; 2; 3], the decoder type is UPerHead, the optimizer is AdamW, and the fully connected layer is FCNHead; The optimization formula is expressed as: moving_avg=alpha*moving_avg+(1-alpha)*w.grad w=w-Lr*moving_avg-Lr*wd*w Among them, wd is the set hyperparameter, Lr is the learning rate, w.grad is the derivative of the loss function with respect to w, w is the weight parameter, wd*w represents the derivative of the penalty term with respect to w, moving_avg is the moving average, and alpha is a smoothing coefficient between 0 and 1; The specific steps of the de-pooling upsampling process are as follows: first, record the maximum value position of the feature map, that is, when performing maximum pooling, it is necessary to record the specific position of the maximum value in each pooling window; second, restore the original size of the pooled feature map to obtain the restored feature map; finally, fill the maximum value, that is, in the restored feature map, put the previously recorded maximum value back to the initial position, and fill the remaining positions with 0.
5. The artificial intelligence evaluation method for predicting colorectal cancer based on psoas muscle image analysis according to claim 4, characterized in that: In S4, the accuracy Acc is expressed as: Among them, TP is a true positive example, that is, the prediction result of the deep learning model is a positive sample, and the actual result is also a positive sample; TN is a true negative example, that is, the prediction result of the deep learning model is a negative sample, and the actual result is also a negative sample; FP is a false positive example, that is, the prediction result of the deep learning model is a positive sample, and the actual result is a negative sample; FN is a false negative example, that is, the prediction result of the deep learning model is a negative sample, and the actual result is a positive sample; The IoU metric (Intersection over Union) is expressed as: Among them, A1 is the predicted result and A2 is the actual result.
6. An artificial intelligence evaluation system for predicting colorectal cancer based on psoas muscle image analysis, characterized in that: An artificial intelligence evaluation method for predicting colorectal cancer based on psoas muscle image analysis according to any one of claims 1 to 5, comprising a data collection module, a preprocessing module, a model training module, a model evaluation module, a result output module and a user interaction module; The data collection module, preprocessing module, model training module, model evaluation module and result output module are connected in sequence; The data collection module and the result output module are respectively connected to the user interaction module; The data collection module is used to collect psoas muscle image data and related diagnostic data, and obtain uploaded data; The preprocessing module is used to clean, standardize and enhance the data collected and acquired by the data collection module; The model training module is used to build a deep learning model, train it with the data processed by the preprocessing module, and evaluate and optimize the deep learning model in real time through the validation set, and output the prediction results; The model evaluation module is used to evaluate the trained model using independent test set data and generate a detailed evaluation report; The result output module is used to display the prediction results to the user and provide the prediction basis and credibility evaluation of the prediction results; The user interaction module is used to select functions, set parameters, view prediction results and upload data.