Crohn disease focus automatic segmentation and activity evaluation system based on deep learning
The Crohn's disease lesion automatic segmentation and activity assessment system, built using deep learning technology, solves the problems of long lesion identification time, inconsistent results, and lack of objective quantification in existing technologies. It achieves efficient and accurate automated assessment and is applicable to CTE imaging equipment and PACS systems.
Patent Information
- Application Number
- CN202511335813.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-18
- Publication Date
- 2026-01-06
AI Technical Summary
In existing technologies, the identification of Crohn's disease lesions relies on manual observation, which is time-consuming and yields inconsistent results. Traditional scoring systems lack objective quantification, and existing methods have small sample sizes, incomplete lesion segmentation, and fail to achieve end-to-end automated assessment.
An automatic segmentation and activity assessment system for Crohn's disease lesions based on deep learning was adopted, including data preprocessing, automatic lesion segmentation, radiomics feature extraction, feature selection and dimensionality reduction, and activity classification. The system utilizes nnU-Net and PyRadiomics toolkits for image processing and feature extraction, and combines a multi-stage feature selection mechanism to construct an end-to-end automated assessment process.
It achieves efficient and accurate automatic segmentation and activity assessment of Crohn's disease lesions, significantly reducing manual operation, improving diagnostic consistency and efficiency, with high segmentation accuracy and superior classification performance. It is suitable for CTE imaging equipment and PACS systems, and supports one-stop assisted diagnosis.
Smart Images

Figure CN121280339A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical artificial intelligence, and in particular to a deep learning-based automatic segmentation and activity assessment system for Crohn's disease lesions. Background Technology
[0002] Crohn's disease is a chronic inflammatory bowel disease, and its diagnosis and activity assessment primarily rely on computed tomography (CTE) images. Currently, the identification of lesions in CTE images in clinical practice still heavily depends on the visual observation of radiologists, a process that takes an average of about 6.8 minutes per case and is significantly influenced by physician experience and subjective judgment, leading to poor consistency of results. Disease activity assessment often employs traditional clinical scoring systems (such as CDAI and SES-CD), lacking objective quantitative standards based on images, thus limiting the accuracy and efficiency of diagnosis. Existing technologies have significant shortcomings: most studies are based on small sample data, resulting in limited model generalization ability; existing methods are mostly limited to the segmentation of partial lesions, failing to comprehensively capture the biological characteristics of the lesions; and an end-to-end automated assessment process from lesion segmentation to feature extraction to activity classification has not yet been achieved, still relying on manual intervention and multi-step independent operations, leading to low efficiency and susceptibility to errors. Summary of the Invention
[0003] The purpose of this invention is to solve the problems of poor generalization ability with small samples, incomplete lesion information capture, and low efficiency and large error caused by reliance on manual operation in the existing technology. Therefore, this invention proposes an automatic segmentation and activity assessment system for Crohn's disease lesions based on deep learning.
[0004] To achieve the above objectives, the present invention adopts the following technical solution:
[0005] The deep learning-based automatic segmentation and activity assessment system for Crohn's disease lesions includes a data preprocessing unit, an automatic lesion segmentation unit, a radiomics feature extraction unit, a feature screening and dimensionality reduction unit, and an activity classification unit.
[0006] The data preprocessing unit includes a resampling module, an intensity normalization module, and a ROI coarse extraction module. The resampling module unifies the input CTE image to the same isotropic resolution. The intensity normalization module normalizes the image's grayscale values to a specific range using Z-Score normalization, allowing the model to focus on relative intensity changes rather than absolute values. The ROI coarse extraction module roughly locates the abdominal region, reducing subsequent computation.
[0007] The automatic lesion segmentation unit includes the nnU-Net core network module and the intelligent optimization module. The nnU-Net core network module processes the preprocessed image to generate a multi-channel feature map. The intelligent optimization module maps the multi-channel feature map to generate a binary segmentation mask. The binary segmentation mask is compared with the manually labeled standard answer through a loss function to generate an evaluation. The evaluation is then conveyed to the nnU-Net core network module for improvement.
[0008] The radiomics feature extraction unit includes the PyRadiomics normalization toolkit, which quantitatively extracts a large number of numerical indicators describing the morphological, intensity and texture features of binary segmentation masks. The PyRadiomics normalization toolkit performs calculations based on predefined mathematical algorithms.
[0009] The feature selection and dimensionality reduction unit includes a selection and dimensionality reduction module. The selection and dimensionality reduction module removes redundant and irrelevant features from high-dimensional features, selects the most discriminative feature subset, prevents model overfitting, and improves generalization ability and computational efficiency.
[0010] The activity classification unit includes a training module and a judgment module. Based on the selected key radiomics features, the activity classification unit constructs a classification model and automatically classifies patients into "mild activity" or "moderate to severe activity".
[0011] As a preferred option, the nnU-Net core network module in the automatic lesion segmentation unit includes an encoder and a decoder. The encoder acts as a feature extractor, gradually extracting multi-scale contextual features of the image through a series of convolution and pooling operations, while reducing the spatial resolution. The decoder acts as a precise localizer, gradually restoring the spatial resolution through interpolation operations and generating a segmentation map with the same size as the input image.
[0012] As a preferred solution, the intelligent optimization module in the automatic lesion segmentation unit includes an activation function and a loss function. The activation function maps the multi-channel feature map output by the decoder to the final segmentation result, outputs a probability map, and then converts it into a binary segmentation mask through the argmax operation. The loss function receives the binary segmentation mask and the ground truth annotation, and the calculated loss value is backpropagated to the nnU-Net core network module through the optimizer Adam to guide the adjustment of all convolutional layer parameters in the encoder and decoder.
[0013] A deep learning-based automatic segmentation and activity assessment system for Crohn's disease lesions includes an automatic segmentation and activity assessment method, which comprises the following steps:
[0014] S1: Data preprocessing. The CTE image is unified to the same isotropic resolution through the resampling module, and then the gray values of the image are normalized to a specific range through Z-Score normalization of the intensity normalization module. Finally, the required abdominal region is cropped as needed through the ROI coarse extraction module to reduce the amount of computation and generate the processed image.
[0015] S2: Automatic lesion segmentation. After processing, the image is gradually reduced in spatial resolution by the encoder, while the number of feature channels is increased, thereby extracting high-level semantic features from pixel-level information. The high-level semantic features are restored in spatial resolution by the decoder, and a multi-channel feature map with the same size as the input image is reconstructed.
[0016] S3: Evaluation and generation. The multi-channel feature map is mapped to the final binary segmentation mask through the activation function. The loss value is calculated between the binary segmentation mask and the manually labeled real mask through the loss function. Then, the multi-channel feature map is generated again by backpropagating the nnU-Net core network module through the optimizer Adam.
[0017] S4: Feature calculation. The binary segmentation mask is calculated using the PyRadiomics standardized toolkit, and the final output is 1874 features, including shape features, first-order statistical features and texture features. The 1874 features form a high-dimensional feature vector.
[0018] S5: Feature selection and dimensionality reduction. High-dimensional feature vectors are transmitted to the selection and dimensionality reduction module. First, the Pearson correlation coefficient between each pair of features is calculated for preliminary selection. Then, the F test of ANOVA is used for secondary selection. Finally, the feature vectors are input into the LASSO regression model for final selection. In the end, 18 key radiomics features were selected.
[0019] S6: Activity classification. The training module calculates and trains weight coefficients through a large number of historical cases. The judgment module calculates and judges based on the weight coefficients, automatically classifying patients into "mild activity" or "moderate to severe activity".
[0020] As a preferred embodiment, in step S2, the process of converting the processed image into a multi-channel feature map is as follows:
[0021] S21: Inside the encoder, 3D convolution is used to slide and scan on the image to extract local features. The ReLU function is used to introduce non-linearity, enabling the network to learn complex mapping relationships. Downsampling is then performed through 3D max pooling to halve the size, expand the receptive field of subsequent neurons, and generate high-level semantic features.
[0022] S22: Inside the decoder, high-level semantic features are upsampled and enlarged by the size of the 3D transposed convolution to generate semantic information. The encoder obtains downsampled detail information through skip connections. Finally, the convolutional block fuses the detail information from the skip connections and the semantic information from the upsampled convolution, thereby gradually reconstructing a multi-channel feature map with clear boundaries.
[0023] As a preferred approach, the specific steps for feature selection and dimensionality reduction in step S5 are as follows:
[0024] S51: Preliminary screening. For two random feature vectors X and Y, calculate the Pearson correlation coefficient r using the formula and determine whether the absolute value of r is greater than 0.85. If it is, randomly remove one feature vector until the absolute value of r is less than 0.85.
[0025] S52: Secondary screening. Based on the built-in SES-CD scoring system, the research subjects are divided into mild activity group and moderate to severe activity group. The core statistics of analysis of variance are calculated, including between-group sum of squares, within-group sum of squares, total sum of squares, between-group mean square, and within-group mean square. Based on the above statistics, the F-statistic is solved, and the P-value is determined by looking up the F-distribution table according to the corresponding degrees of freedom. Features with P-values greater than 0.05 are removed according to the statistical test criteria. The remaining features are then sorted in descending order of F-value, and the features with the highest ranking are selected.
[0026] S53: The final screening inputs the feature data processed in the first two steps into the LASSO regression model. Cross-validation is used to find the optimal hyperparameter λ. The larger the λ value, the more features are compressed to zero, and the simpler the model is. Cross-validation will try a series of λ values and select the λ value that makes the model perform best on the validation set. The optimal λ value is used to fit the entire training set, and finally 18 key radiomics features are selected.
[0027] As a preferred embodiment, the activity classification steps in step S6 are as follows:
[0028] S61: Determine the weight coefficients. The training module first randomly assigns weight coefficients to each indicator. The system compares the results with the actual answers. By continuously calculating and changing the weight coefficients, the system gradually converges to the correct answer a little bit each time, and finally stabilizes at a set of optimal values, which are the weight coefficients for subsequent calculations.
[0029] S62: Activity determination. Through linear weighted summation, the model multiplies each patient's 18 feature values by the weight coefficients obtained during training and adds an intercept term to form the Z-Score. The linear score Z is compressed to the (0,1) interval using the Sigmoid function and interpreted as the probability P of belonging to the "moderate to severe" category. The system sets a classification threshold. If P is greater than the classification threshold, the model determines it as "moderate to severe activity". If P is less than the classification threshold, the model determines it as "mild activity".
[0030] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0031] 1. Fully Automated End-to-End Assessment Process: This invention seamlessly integrates the nnU-Net-based automatic lesion segmentation module with the radiomics feature extraction and machine learning classification module to construct a complete CTE image intelligent analysis system. This system achieves full automation from raw image input, preprocessing, lesion segmentation, feature extraction, feature screening to final activity classification, significantly reducing manual operation steps and improving diagnostic consistency and efficiency.
[0032] 2. High-precision segmentation model: Adopting the advanced nnU-Net framework (based on the 3DU-Net architecture), the system achieves pixel-level accurate segmentation while maintaining contextual information through an encoder-decoder structure and skip connection mechanism. On the independent test set, the model's Dice coefficient reaches 0.824, demonstrating excellent segmentation consistency and robustness. At the same time, the system reduces the time for physicians to process each image from 6.8 minutes to 4.6 minutes, significantly improving clinical work efficiency.
[0033] 3. Multi-stage feature selection mechanism: A three-stage feature dimensionality reduction strategy is proposed: First, redundant features are eliminated through Pearson correlation analysis. Then, SelectKBest (based on ANOVAF value) is used to retain features that are significantly related to activity. Finally, LASSO regression is used to further compress the number of features, accurately selecting 18 key features with the most discriminative power from 1874 initial features. This mechanism effectively avoids overfitting and improves the interpretability and generalization ability of the model.
[0034] 4. Superior classification performance: Based on the 18 selected features, the system constructed a variety of machine learning classifiers, including logistic regression, support vector machine, and random forest. Among them, the logistic regression model achieved an AUC of 0.862 on the test set, which is significantly better than the subjective evaluation of radiologists (AUC of 0.729), showing the significant advantage of objective evaluation based on radiomics in the discrimination of Crohn's disease activity.
[0035] 5. System Integration and Clinical Applicability: The system is designed with practicality and compatibility in mind. It can be seamlessly integrated into existing CTE imaging equipment and PACS systems. It supports the DICOM standard interface and provides one-stop auxiliary diagnostic functions. The system output includes lesion segmentation results, feature quantification reports and activity classification suggestions, which significantly improves the standardization level and decision-making efficiency of the diagnostic process. It has strong potential for clinical translation and promotion. Attached Figure Description
[0036] Figure 1 This is a block diagram of the Crohn's disease lesion automatic segmentation and activity assessment system based on deep learning proposed in this invention.
[0037] Figure 2 This is a flowchart of the automatic segmentation and activity assessment method for the Crohn's disease lesion automatic segmentation and activity assessment system based on deep learning proposed in this invention.
[0038] Figure 3 This is a diagram showing the overall data transmission path of the deep learning-based automatic segmentation and activity assessment system for Crohn's disease lesions proposed in this invention.
[0039] Figure 4 This is a data transmission path diagram of the automatic lesion segmentation unit of the Crohn's disease lesion automatic segmentation and activity assessment system based on deep learning proposed in this invention. Detailed Implementation
[0040] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0041] Example, refer to Figures 1 to 4 The system for automatic segmentation and activity assessment of Crohn's disease lesions based on deep learning includes a data preprocessing unit, an automatic lesion segmentation unit, a radiomics feature extraction unit, a feature screening and dimensionality reduction unit, and an activity classification unit.
[0042] The data preprocessing unit includes a resampling module, an intensity normalization module, and a ROI coarse extraction module. The resampling module unifies the input CTE image to the same isotropic resolution. The intensity normalization module normalizes the image's grayscale values to a specific range using Z-Score normalization, allowing the model to focus on relative intensity changes rather than absolute values. The ROI coarse extraction module roughly locates the abdominal region, reducing subsequent computation.
[0043] The automatic lesion segmentation unit includes the nnU-Net core network module and the intelligent optimization module. The nnU-Net core network module processes the preprocessed image to generate a multi-channel feature map. The intelligent optimization module maps the multi-channel feature map to generate a binary segmentation mask. The binary segmentation mask is compared with the manually labeled standard answer through a loss function to generate an evaluation. The evaluation is then conveyed to the nnU-Net core network module for improvement.
[0044] The radiomics feature extraction unit includes the PyRadiomics normalization toolkit, which quantitatively extracts a large number of numerical indicators describing the morphological, intensity and texture features of binary segmentation masks. The PyRadiomics normalization toolkit performs calculations based on predefined mathematical algorithms.
[0045] The feature selection and dimensionality reduction unit includes a selection and dimensionality reduction module. The selection and dimensionality reduction module removes redundant and irrelevant features from high-dimensional features, selects the most discriminative feature subset, prevents model overfitting, and improves generalization ability and computational efficiency.
[0046] The activity classification unit includes a training module and a judgment module. Based on the selected key radiomics features, the activity classification unit constructs a classification model and automatically classifies patients into "mild activity" or "moderate to severe activity".
[0047] Furthermore, the nnU-Net core network module in the automatic lesion segmentation unit includes an encoder and a decoder. The encoder acts as a feature extractor, gradually extracting multi-scale contextual features of the image through a series of convolution and pooling operations, while reducing the spatial resolution. The decoder acts as a precise localizer, gradually restoring the spatial resolution through interpolation operations and generating a segmentation map with the same size as the input image.
[0048] Furthermore, the intelligent optimization module in the automatic lesion segmentation unit includes an activation function and a loss function. The activation function maps the multi-channel feature map output by the decoder to the final segmentation result, outputs a probability map, and then converts it into a binary segmentation mask through the argmax operation. The loss function receives the binary segmentation mask and the ground truth annotation, and the calculated loss value is backpropagated to the nnU-Net core network module through the optimizer Adam to guide the adjustment of all convolutional layer parameters in the encoder and decoder.
[0049] It should be noted that the purpose of this invention is to provide an automated analysis system and method for Crohn's disease CTE images based on deep learning and radiomics. This system can solve the following problems existing in the prior art: radiologists manually annotating CD lesions in CTE images is time-consuming and greatly affected by subjective experience; traditional clinical scoring lacks objective quantitative standards based on images; existing automated methods have small sample sizes, only segment some lesions, and have not achieved an end-to-end automated process from lesion segmentation to activity assessment.
[0050] To achieve the above objectives, this invention integrates CTE image preprocessing, deep learning segmentation, radiomics feature extraction, and machine learning classification technologies to construct a complete automated analysis workflow. The system first performs standardized preprocessing on CTE images, then uses the nnU-Net deep learning model to automatically and accurately segment intestinal lesion regions. Based on this, it utilizes the PyRadiomics toolkit to extract high-throughput radiomics features, and then eliminates redundant features through a multi-stage feature selection mechanism. Finally, it constructs a machine learning classification model based on the optimal feature subset to achieve an objective and quantitative assessment of Crohn's disease activity. This system boasts advantages such as high segmentation accuracy (Dice coefficient of 0.824 on the test set), excellent classification performance (AUC of 0.862 for the logistic regression model), and high automation (reducing physician processing time from 6.8 minutes to 4.6 minutes). It significantly improves diagnostic efficiency and consistency, and is suitable for clinical auxiliary diagnosis and research analysis in hospital radiology and gastroenterology departments, possessing significant clinical application value and market potential.
[0051] like Figures 2 to 4 As shown, the deep learning-based automatic segmentation and activity assessment system for Crohn's disease lesions includes the following steps:
[0052] S1: Data preprocessing. The CTE image is unified to the same isotropic resolution through the resampling module, and then the gray values of the image are normalized to a specific range through Z-Score normalization of the intensity normalization module. Finally, the required abdominal region is cropped as needed through the ROI coarse extraction module to reduce the amount of computation and generate the processed image.
[0053] S2: Automatic lesion segmentation. After processing, the image is gradually reduced in spatial resolution by the encoder, while the number of feature channels is increased, thereby extracting high-level semantic features from pixel-level information. The high-level semantic features are restored in spatial resolution by the decoder, and a multi-channel feature map with the same size as the input image is reconstructed.
[0054] Furthermore, in step S2, the process of converting the processed image into a multi-channel feature map is as follows:
[0055] S21: Inside the encoder, 3D convolution is used to slide and scan on the image to extract local features. The ReLU function is used to introduce non-linearity, enabling the network to learn complex mapping relationships. Downsampling is then performed through 3D max pooling to halve the size, expand the receptive field of subsequent neurons, and generate high-level semantic features.
[0056] S22: Inside the decoder, high-level semantic features are upsampled and enlarged by the size of the 3D transposed convolution to generate semantic information. The encoder obtains downsampled detail information through skip connections. Finally, the convolutional block fuses the detail information from the skip connections and the semantic information from the upsampled convolution to gradually reconstruct a multi-channel feature map with clear boundaries.
[0057] Furthermore, automatic lesion segmentation is a crucial step in CTE image processing, achieving accurate lesion region division through a deep learning model. It employs an encoder-decoder architecture and skip connection mechanism to maintain feature representation capabilities while ensuring spatial localization accuracy, providing reliable binary mask support for subsequent radiomics feature extraction. This segmentation process effectively handles complex situations such as differences in image quality and varied lesion morphologies, guaranteeing the accuracy and stability of the segmentation results.
[0058] It should be noted that the segmentation system of this invention adopts a deep learning architecture based on nnU-Net. The encoder path is responsible for feature extraction and abstraction, gradually expanding the receptive field and extracting multi-level features through convolution and downsampling operations; the decoder path gradually restores spatial details through upsampling and feature fusion, ultimately outputting a refined segmentation result. During processing, a skip connection mechanism fuses the feature maps of each stage of the encoder with the corresponding stage of the decoder, effectively solving the problems of gradient vanishing and detail loss in deep network training. In addition, network parameters and hyperparameters can be adaptively adjusted according to different data characteristics and clinical needs, optimizing computational efficiency while ensuring segmentation accuracy, providing a reliable foundation for subsequent radiomics analysis and activity assessment.
[0059] S3: Evaluation and generation. The multi-channel feature map is mapped to the final binary segmentation mask through the activation function. The loss value is calculated between the binary segmentation mask and the manually labeled real mask through the loss function. Then, the multi-channel feature map is generated again by backpropagating the nnU-Net core network module through the optimizer Adam.
[0060] S4: Feature calculation. The binary segmentation mask is calculated using the PyRadiomics standardized toolkit, and the final output is 1874 features, including shape features, first-order statistical features and texture features. The 1874 features form a high-dimensional feature vector.
[0061] S5: Feature selection and dimensionality reduction. High-dimensional feature vectors are transmitted to the selection and dimensionality reduction module. First, the Pearson correlation coefficient between each pair of features is calculated for preliminary selection. Then, the F test of ANOVA is used for secondary selection. Finally, the feature vectors are input into the LASSO regression model for final selection. In the end, 18 key radiomics features were selected.
[0062] Furthermore, in step S5, the specific steps of feature selection and dimensionality reduction are as follows:
[0063] S51: Preliminary screening. For two random feature vectors X and Y, calculate the Pearson correlation coefficient r using the formula and determine whether the absolute value of r is greater than 0.85. If it is, randomly remove one feature vector until the absolute value of r is less than 0.85.
[0064] S52: Secondary screening. Based on the built-in SES-CD scoring system, the research subjects are divided into mild activity group and moderate to severe activity group. The core statistics of analysis of variance are calculated, including between-group sum of squares, within-group sum of squares, total sum of squares, between-group mean square, and within-group mean square. Based on the above statistics, the F-statistic is solved, and the P-value is determined by looking up the F-distribution table according to the corresponding degrees of freedom. Features with P-values greater than 0.05 are removed according to the statistical test criteria. The remaining features are then sorted in descending order of F-value, and the features with the highest ranking are selected.
[0065] S53: The final screening inputs the feature data processed in the first two steps into the LASSO regression model. Cross-validation is used to find the optimal hyperparameter λ. The larger the λ value, the more features are compressed to zero, and the simpler the model is. Cross-validation will try a series of λ values and select the λ value that makes the model perform best on the validation set. The optimal λ value is used to fit the entire training set, and finally 18 key radiomics features are selected.
[0066] It should be noted that the formula in step S51 is...
[0067]
[0068] The meaning of each part of the formula is X i ,Y i : The values of the i-th patient on features X and Y;
[0069] The average of all values for features X and Y;
[0070] n: Total number of patients (sample size);
[0071] ∑: Summation symbol, indicating that calculations are performed on all patients and then summed.
[0072] In step S52, the study participants are divided into a mild activity group and a moderate-to-severe activity group based on the built-in SES-CD scoring system. Mild-to-severe activity typically corresponds to a lower SES-CD score (e.g., an SES-CD score less than or equal to 3 or less than or equal to 5, depending on the study's established threshold). Moderate-to-severe activity typically corresponds to a higher SES-CD score (e.g., an SES-CD score greater than 3 or greater than 5). The thresholds used will be clearly defined in the research paper.
[0073] The formulas used in step S52 are:
[0074] Within-group sum of squares formula
[0075]
[0076] The meaning of each part of the formula is SS within Square within a group;
[0077] X i,group : The specific numerical value of the i-th patient in the i-th group for this current feature;
[0078] The mean of this characteristic for all patients in group 1, where group 2 represents group 1 (mild activity) and group 2 (moderate to severe activity).
[0079] Between-groups sum of squares formula
[0080]
[0081] The meaning of each part of the formula is SS between Between-group squares;
[0082] n1, n2: The number of patients in group 1 and group 2;
[0083] The group averages of Group 1 and Group 2;
[0084] Overall mean, the average of all patients (mild + moderate to severe) on this characteristic;
[0085] Total Sum of Squares Formula
[0086] SS total =SS between +SS within
[0087] The meaning of each part of the formula is SS total Total sum of squares;
[0088] SS between Between-group squares;
[0089] SS within Square within a group;
[0090] Formula for calculating mean square intervals
[0091]
[0092] The meaning of each part of the formula is MS between : Mean square value;
[0093] SS betweeb Between-group squares;
[0094] df between : represents the degrees of freedom, where the formula for calculating the degrees of freedom is df between =k-1, k: number of groups;
[0095] In-mean-square calculation formula
[0096]
[0097] The meaning of each part of the formula is MS within : Mean squared internal value;
[0098] SS within Within-group squares, where the degrees of freedom df within =n total -k,n total Total number of patients;
[0099] F-statistic calculation formula
[0100]
[0101] The meanings of each part of the formula are as follows: F: statistic;
[0102] MS between Mean square value;
[0103] MS winthin : Mean squared internal value;
[0104] Loss function in step S53:
[0105] Loss=MSE+λ*Σ|βj|
[0106] The meaning of each part of the formula is as follows: MSE: measures the difference between the model's predicted value and the actual value, ensuring the accuracy of the model;
[0107] λ*Σ|βj| (L1 regularization term): penalizes the absolute value of the coefficients in the model (βj is the coefficient of the j-th feature);
[0108] λ is a hyperparameter that controls the intensity of the penalty. This term forces the model to tend toward a sparse solution, i.e., many βj become 0, thereby achieving feature selection.
[0109] After step S5, a small, optimal subset of features was obtained. Ultimately, 18 key radiomics features were selected from the system. These features proved to have high discriminative value in differentiating Crohn's disease activity.
[0110] Based on the above, the feature selection and dimensionality reduction method of this invention achieves accurate mapping from a high-dimensional feature space to a low-dimensional feature subset through a triple screening mechanism. This method effectively solves the "curse of dimensionality" problem in high-dimensional data by progressively eliminating redundant features, retaining salient features, and finally selecting the most discriminative features, significantly improving the performance and interpretability of subsequent classification models.
[0111] It is worth noting that the triple screening strategy employed in this invention has a clear hierarchical relationship and a sound theoretical foundation: correlation analysis ensures feature independence, analysis of variance ensures feature significance, and LASSO regression ensures feature sparsity and model generalization ability. This multi-stage, multi-angle screening approach has significant advantages over single screening methods, avoiding the omission of important features and effectively eliminating redundant features, thus laying a solid foundation for building a high-performance activity classification model.
[0112] S6: Activity classification. The training module calculates and trains weight coefficients through a large number of historical cases. The judgment module calculates and judges based on the weight coefficients, automatically classifying patients into "mild activity" or "moderate to severe activity".
[0113] Furthermore, in step S6, the activity classification steps are as follows:
[0114] S61: Determine the weight coefficients. The training module first randomly assigns weight coefficients to each indicator. The system compares the results with the actual answers. By continuously calculating and changing the weight coefficients, the system gradually converges to the correct answer a little bit each time, and finally stabilizes at a set of optimal values, which are the weight coefficients for subsequent calculations.
[0115] S62: Activity determination. Through linear weighted summation, the model multiplies each patient's 18 feature values by the weight coefficients obtained during training and adds an intercept term to obtain the Z-Score. The linear score Z is compressed to the (0,1) interval using the Sigmoid function and interpreted as the probability P of belonging to the "moderate to severe" category. The system sets a classification threshold. If P is greater than the classification threshold, the model determines it as "moderate to severe activity". If P is less than the classification threshold, the model determines it as "mild activity".
[0116] Furthermore, the linear weighted summation formula in step S62 is:
[0117] Z=β0+(β1×x1)+(β2×x2)+...+(β 18 ×x 18 )
[0118] The meanings of each part of the formula are as follows: Z: weighted sum of evidence, β0 (intercept term): baseline score of the model;
[0119] When all eigenvalues x1 to x 18 When both are 0, the model has an inherent tendency to favor a certain class;
[0120] x1,x2,...,x 18 (Eigenvalues): Specific numerical values of 18 selected radiomics features for a new patient;
[0121] β1,β2,...,β 18 (Weight coefficients): These are the core parameters learned by the model during training. Each weight corresponds to a feature, representing the magnitude and direction of that feature's contribution to the final decision;
[0122] The Sigmoid function transformation maps the Z-value to a probability (P-value) using the following formula:
[0123]
[0124] The meaning of each part of the formula is as follows: P(Y=1): the probability predicted by the model that the patient belongs to "moderate to severe activity" ("moderate to severe" is defined as category 1);
[0125] Value range: (0,1). For example, P = 0.85 means that the model has an 85% confidence level that the patient is of moderate to severe condition;
[0126] e: natural constant;
[0127] -Z: The negative of the linear score Z;
[0128] It should be noted that this classification threshold is not static. In clinical practice, if it is believed that "misclassifying moderate to severe as mild" has more serious consequences than "misclassifying mild as moderate to severe," the classification threshold can be lowered (e.g., set to 0.3). This makes the model more "sensitive," classifying as moderate to severe with only a 30% certainty, thus reducing the risk of missed diagnoses. However, this comes at the cost of increasing the risk of misdiagnosis. The system compares the probability P with the threshold (default 0.5) and outputs the final classification result as "mild" or "moderate to severe." The system will report the classification result and its confidence probability (i.e., P-value), for example: Diagnosis: Moderate to severe activity (confidence: 92%).
[0129] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A deep learning-based automatic segmentation of Crohn's disease lesions and assessment of disease activity system, characterized in that, The system comprises a data preprocessing unit, a lesion automatic segmentation unit, an imageomics feature extraction unit, a feature screening and dimension reduction unit, and an activity classification unit. The data preprocessing unit comprises a resampling module, an intensity normalization module, and an ROI rough extraction module. The lesion automatic segmentation unit comprises an nnU-Net core network module and an intelligent optimization module. The imageomics feature extraction unit comprises a PyRadiomics standardized toolkit. The feature screening and dimension reduction unit comprises a screening and dimension reduction module. The activity classification unit comprises a training module and a judgment module. 2.The deep learning-based Crohn's disease lesion automatic segmentation and activity assessment system of claim 1, wherein, The nnU-Net core network module in the lesion automatic segmentation unit comprises an encoder and a decoder. 3.The deep learning-based Crohn's disease lesion automatic segmentation and activity assessment system of claim 1, wherein, The intelligent optimization module in the lesion automatic segmentation unit comprises an activation function and a loss function. 4.The deep learning-based Crohn's disease lesion automatic segmentation and activity assessment system according to claim 1, comprising an automatic segmentation and activity assessment method, characterized in that, The automatic segmentation and activity evaluation method comprises the following steps: S1: data preprocessing, the CTE image is unified to the same isotropic resolution through the resampling module, the gray value of the image is standardized to a specific range through the Z-Score normalization of the intensity normalization module, and finally the ROI rough extraction module is used to extract the required abdominal region according to the needs, reduce the calculation amount, and generate the processed picture; S2: automatic lesion segmentation, the processed picture is gradually reduced in spatial resolution by the encoder while increasing the number of feature channels, thereby extracting high-level semantic features from pixel-level information, and the high-level semantic features recover their spatial resolution through the decoder to reconstruct a multi-channel feature map with the same size as the input image; S3: evaluation and generation, the multi-channel feature map is mapped to the final binary segmentation mask through the activation function, the binary segmentation mask and the real mask annotated by artificial labeling are calculated by the loss function to obtain the loss value, and the optimizer Adam is used to modify and optimize the nnU-Net core network module to generate the multi-channel feature map again; S4 : feature calculation, the binary segmentation mask is calculated by the PyRadiomics standardization toolkit to finally output 1874 features, including shape features, first-order statistical features and texture features, and the 1874 features form a high-dimensional feature vector; S5: feature screening and dimensionality reduction, the high-dimensional feature vector is transmitted into the screening and dimensionality reduction module, the Pearson correlation coefficient between each two features is first calculated for preliminary screening, the F test of analysis of variance ANOVA is then used for secondary screening, and finally the LASSO regression model is input for the final screening, and 18 key image features are finally selected; S6: activity classification, the training module calculates the weight coefficient by training a large number of historical cases, and the judgment module calculates and judges based on the weight coefficient to automatically classify the patient as "mild activity" or "moderate to severe activity". 5.The deep learning-based Crohn's disease lesion automatic segmentation and activity assessment system according to claim 4, characterized in that, In step S2, the steps of converting the processed picture into a multi-channel feature map are as follows: S21: inside the encoder, 3D convolution is used to slide and scan the image to extract local features, ReLU function is used to introduce nonlinearity to enable the network to learn complex mapping relationships, and 3D maximum pooling is used for downsampling to reduce the size by half and expand the receptive field of subsequent neurons to generate high-level semantic features; S22: inside the decoder, the high-level semantic features enter the upsampling through the size expansion of the 3D transposed convolution to generate semantic information, the encoder obtains the down-sampled detail information through the jump connection, and finally the convolution block fuses the detail information from the jump connection and the semantic information from the upsampling to gradually reconstruct a multi-channel feature map with clear boundaries. 6.The deep learning-based Crohn's disease lesion automatic segmentation and activity assessment system according to claim 4, characterized in that, In step S5, the specific steps of feature screening and dimensionality reduction are as follows: S51: preliminary screening, for two random feature vectors X and Y, the absolute value of r is calculated by the calculation formula of Pearson correlation coefficient r, and if the absolute value of r is greater than 0.85, one of the feature vectors is randomly removed until the absolute value of r is less than 0.85; S52: Secondary screening, according to the built-in SES-CD scoring system, the subjects are divided into mild activity group and moderate to severe activity group, calculate the analysis of variance core statistics, including between-group sum of squares, within-group sum of squares, total sum of squares, between-group mean square and within-group mean square, based on the above statistical quantities to solve the F statistic, and according to the corresponding degree of freedom to query the F distribution table to determine the P value, according to the statistical test standard to eliminate the characteristics with P value greater than 0.05, and then arrange the remaining characteristics in descending order of F value, select the top-ranked characteristics; S53: The last screening inputs the feature data processed by the previous two steps into the LASSO regression model, and finds the optimal hyperparameter λ through cross-validation. The larger the λ value, the more features are compressed to zero, and the simpler the model. Cross-validation will try a series of λ values, and select the one that makes the model perform best on the validation set. Use the optimal λ value to fit the entire training set, and finally select 18 key imaging features. 7.The deep learning-based Crohn's disease lesion automatic segmentation and activity assessment system of claim 4, wherein, In step S6, the steps of activity classification are as follows: S61: Determine the weight coefficient. The training module first randomly assigns a weight coefficient to each indicator. The system compares the results with the true answers. By constantly calculating and changing the weight coefficient, the correct answer is gradually converged each time. Finally, it stabilizes at a set of optimal numerical values, which are the weight coefficients for subsequent calculations. S62: Activity judgment. Through linear weighted summation, the model multiplies each patient's 18 feature values by the weight coefficients obtained by training and adds an intercept term to become Z-Score. Linear score Z is compressed to the (0, 1) interval through the Sigmoid function, which is interpreted as the probability P of belonging to the "moderate to severe" category. The system sets a classification threshold. If P is greater than the classification threshold, the model judges it as "moderate to severe activity". If P is less than the classification threshold, the model judges it as "mild activity".
Citation Information
Cited By
Hysteroscope focus segmentation quantification method and system based on deep learning
CN121883515A
A Deep Learning-Based Method and System for Hysteroscopic Lesion Segmentation and Quantification
CN121883515B
