Bladder cancer muscular layer infiltration and prognosis prediction system and method based on deep learning imageomics

Abstract

The invention discloses a bladder cancer muscular layer infiltration and prognosis prediction system based on deep learning imageomics, relates to the technical field of medical treatment and belongs to auxiliary diagnosis, and adopts an imageomics method and a deep learning model to extract imageomics features and deep learning features of CT images respectively; compared with the prior art, potential meaningful features in the CT image are extracted more comprehensively, the prediction accuracy can be further improved, meanwhile, the features are screened more strictly by using technologies such as intra-class correlation coefficient (ICC) analysis, t inspection, Pearson correlation coefficient, minimum absolute contraction and selection operator (LASSO), PCA dimension reduction and Cox regression, and verification is performed in a plurality of external medical centers. On one hand, the over-fitting risk of the model can be greatly reduced, on the other hand, the excellent generalization of the model is also proved, the use of the model is not excessively influenced by the medical environment and the CT acquisition instrument, and the error generated by the multi-center effect can be effectively solved.

Description

Technical Field

[0001] This invention belongs to the field of auxiliary diagnosis technology, specifically to a system and method for predicting bladder cancer muscle invasion and prognosis based on deep learning radiomics. By combining radiomics features with deep learning features, this system can non-invasively predict the muscle invasion status and prognosis of bladder cancer patients, facilitating personalized precision treatment. Background Art

[0002] Bladder cancer (BLCA) is a common malignancy worldwide and is divided into non-muscle invasive bladder cancer (NMIBC) and muscle invasive bladder cancer (MIBC). Due to significant differences in survival prognosis and treatment options between MIBC and NMIBC, accurately determining the muscle invasive status of bladder cancer is crucial for patient treatment decisions. Furthermore, prognostic prediction for bladder cancer patients is a crucial clinical need. Survival curves and disease progression are often predicted based on patient examination information, clinical data, and pathological indicators.

[0003] Existing methods for identifying muscle invasion and predicting prognosis of bladder cancer have the following problems: 1. Muscle invasion identification: For example, a system for determining the muscle invasion status of bladder cancer and its application, published under the publication number CN113850788A, primarily relies on Pyradiomics to extract radiomics features from CT images and uses a logistic regression model for prediction. However, these methods only extract radiomics features and fail to fully utilize the potential information in CT images. Furthermore, the feature selection method is relatively simple, which can easily lead to model overfitting and poor generalization. At the same time, the above patented technology and other existing technologies are unable to visualize the model's attention to CT image lesions.

[0004] 2. Prognosis prediction: For example, a biomarker for predicting survival prognosis in muscle-invasive bladder cancer and its application, published in publication number CN109884301B, relies primarily on biomarker staining on pathological sections to predict the prognosis of muscle-invasive bladder cancer (MIBC). However, this method requires obtaining pathological tissue sections from patients and performing immunohistochemical staining. Pathological tissue sections require surgery or cystoscopy, which is difficult and costly to obtain, and has low patient compliance. Moreover, this method can only predict the prognosis of muscle-invasive bladder cancer (MIBC), but cannot predict the prognosis of patients with non-muscle-invasive bladder cancer (NMIBC).

[0005] Therefore, there is an urgent need for a non-invasive and accurate method for predicting bladder cancer muscle invasion and prognosis based on deep learning imaging groups to address the shortcomings of existing technologies. Summary of the Invention

[0006] The present invention provides a deep learning imaging genomics model and system for bladder cancer muscle invasion and prognosis prediction based on CT images. The aim is to accurately assess the risk of muscle invasion of bladder cancer and the prognosis of patients through non-invasive methods, and to assist in individualized precision treatment.

[0007] To solve the above technical problems, the present invention provides the following technical solution: a bladder cancer muscle invasion and prognosis prediction system based on deep learning imaging omics, comprising: A data acquisition module is used to collect CT imaging data and clinical data of bladder cancer patients; A preprocessing module, configured to perform denoising, smoothing, data enhancement, super-resolution reconstruction, and ROI delineation on the CT image; A feature extraction module is used to extract radiomics features and deep learning features from the CT images and clinical features from the patient's medical records; A feature screening module is used to screen and merge the radiomics features and deep learning features based on dimensionality reduction methods such as Pearson correlation coefficient and DCA principal component analysis to construct a deep learning radiomics feature set; A model building module, used to build a muscle invasion risk assessment model and a survival prognosis prediction model based on the deep learning radiomics feature set and clinical characteristics; The prediction module is used to input the patient's CT images and clinical characteristics into the prediction model and output the probability of myometrial invasion and the prognostic effect.

[0008] Preferably, the pre-processing module further comprises: A super-resolution reconstruction unit is used to reconstruct low-resolution CT images into high-resolution images using a GAN generative adversarial network; ROI delineation unit is used to extract the region of interest (ROI) based on the self-image segmentation algorithm and adaptive threshold.

[0009] Preferably, the feature extraction module further includes: Radiomics feature extraction unit, used to extract radiomics features from CT images using Pyradiomics tools; The deep learning feature extraction unit is used to extract multi-level deep learning features from CT images using the ResNet101 deep convolutional neural network.

[0010] Preferably, the feature screening module further includes: Feature screening unit, used to screen features using intraclass correlation coefficient (ICC) analysis, t-test, Pearson correlation coefficient, least absolute shrinkage and selection operator (LASSO), and Cox regression; The feature merging unit is used to merge the screened radiomics features and deep learning features to construct a deep learning radiomics feature set.

[0011] Preferably, the model building module further includes: a muscle invasion risk assessment model construction unit, configured to fuse the deep learning radiomics feature set with clinical features to construct a muscle invasion risk assessment model; The survival prognosis prediction model construction unit is used to fuse the deep learning imaging genomics feature set with clinical features to construct a survival prognosis prediction model.

[0012] Preferably, the prediction module further comprises: a muscle invasion probability prediction unit, configured to output the patient's muscle invasion probability according to the muscle invasion risk assessment model; The prognosis effect prediction unit is used to output the patient's prognosis effect according to the survival prognosis prediction model.

[0013] It also includes a visualization module for generating Grad-CAM visualizations to help identify patient tumor regions and enhance model interpretability. The prognosis prediction system was validated by multiple centers to ensure the generalizability and robustness of the model.

[0014] The method for predicting muscle invasion and prognosis of bladder cancer based on deep learning radiomics includes the following steps: S1: Collect and preprocess CT imaging data of bladder cancer patients; S2: extracting radiomics features and deep learning features from the CT images, and extracting clinical features from the patient's medical records; S3: Screen and merge the radiomics features and deep learning features based on dimensionality reduction methods such as Pearson correlation coefficient and DCA principal component analysis to construct a deep learning radiomics feature set; S4: Construct a muscle invasion risk assessment model and a survival prognosis prediction model based on the deep learning radiomics feature set and clinical characteristics; S5: Input the patient's CT images and clinical characteristics into the prediction model, and output the probability of myometrial invasion and the prognostic effect.

[0015] Preferably, the preprocessing step includes denoising, smoothing, data enhancement, super-resolution reconstruction and ROI delineation processing; The feature extraction step includes extracting radiomics features using the Pyradiomics tool and extracting deep learning features using the ResNet101 deep convolutional neural network; The feature screening step includes screening features using intraclass correlation coefficient (ICC) analysis, t-test, Pearson correlation coefficient, least absolute shrinkage and selection operator (LASSO) and Cox regression; The model construction step includes fusing the deep learning radiomics feature set with clinical features to construct a muscle invasion risk assessment model and a survival prognosis prediction model; The prediction step includes outputting the patient's muscle invasion probability according to the muscle invasion risk assessment model, and outputting the patient's prognosis according to the survival prognosis prediction model; It also includes generating Grad-CAM visualizations to help identify patient tumor regions and enhance model interpretability.

[0016] A computer system includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of a method for predicting bladder cancer muscle invasion and prognosis based on deep learning imaging omics are implemented.

[0017] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of a method for predicting bladder cancer muscle invasion and prognosis based on deep learning imaging omics are implemented.

[0018] Compared with the existing technology, the embodiment of the present invention provides a method and application for predicting bladder cancer muscle invasion and prognosis based on deep learning imaging omics, which has the following beneficial effects: 1. The method and system for predicting bladder cancer muscle invasion and prognosis based on deep learning radiomics in the present invention extract radiomics features and deep learning features of CT images through radiomics methods and deep learning models, respectively, and more comprehensively extract potentially meaningful features in CT images, which can further improve prediction accuracy.

[0019] 2. This invention uses techniques such as intraclass correlation coefficient (ICC) analysis, t-test, Pearson correlation coefficient, least absolute shrinkage and selection operator (LASSO), PCA dimensionality reduction, and Cox regression to more rigorously screen features. This method has been validated in multiple external medical centers. This significantly reduces the risk of overfitting and demonstrates the model's excellent generalizability. The model's use is not affected by the medical environment or the multitude of CT acquisition equipment, effectively addressing errors caused by the multicenter effect.

[0020] 3. The present invention can simultaneously predict the patient's bladder cancer myometrial invasion status and survival prognosis using only CT images and the patient's clinical information, avoiding invasive examinations and reducing medical expenses and patient trauma.

[0021] 4. The present invention provides a Grad-cam visualization module that can help identify patient tumor areas, enhance model interpretability, and increase doctors' trust in AI-assisted diagnosis results. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0023] Figure 1 It is the overall flow chart of the system of the present invention.

[0024] Figure 2 This is a Grad-CAM visualization diagram. The color-highlighted areas are mainly concentrated in the basal and tumor areas, indicating that the DL model mainly focuses on these areas to extract valuable information.

[0025] Figure 3 SHAP visualization of the model of the present invention and the predictive performance (C, D) and survival analysis (F, G) in the training set and validation set.

[0026] Figure 4 The figure shows the AUC curve and survival analysis diagram of the model of the present invention.

[0027] Figure 5 This figure compares the performance of the DLRN model and the diagnostic ability of doctors alone in diagnosing myometrial invasion.

[0028] Figure 6 This is a data comparison chart showing that the AUC, sensitivity, and specificity of doctors' diagnosis significantly increased with the help of the DLRN model. in: Panel A represents the performance of the DLRN model and the diagnostic ability of six physicians in diagnosing myometrial invasion. Figure B is a statistical analysis of the comparison of diagnostic capabilities, showing that the DLRN model's diagnostic capability is significantly higher than that of manual testing. Figure C shows that with the help of the DLRN model, the six doctors made another diagnosis and their diagnostic abilities improved. Figures EFG show that with the help of the DLRN model, the AUC, sensitivity, and specificity of doctors' diagnosis are significantly improved; The DLRN constructed by the present invention, wherein the Combined score is given by a deep learning imaging omics model.

[0029] Figure 7 Prediction of (non) muscle invasion after clinical data input. DETAILED DESCRIPTION

[0030] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0031] Example 1 Bladder cancer muscle invasion and prognosis prediction method based on deep learning radiomics 1. Data collection and preprocessing Data Collection: CT imaging data of bladder cancer patients were collected from multiple medical institutions, covering non-contrast, arterial, equilibrium, and excretion phases. Clinical data, including age, gender, and tumor size, were also collected.

[0032] Preprocessing: CT images are preprocessed, including denoising, smoothing, data enhancement, super-resolution reconstruction, and ROI (region of interest) delineation. The specific steps are as follows: Super-resolution reconstruction: Using a pre-trained GAN (generative adversarial network) model, low-resolution CT images (e.g., 512×512) are reconstructed into high-resolution images (e.g., 2048×2048) to enhance image details. Normalization: All CT images were normalized and filtered using a −75, 175-75, 175 Hounsfield unit (HU) window level to ensure image data consistency. ROI delineation: Based on the self-image segmentation algorithm and adaptive threshold, the bladder area and tumor area are delineated to generate the ROI image set.

[0033] 2. Feature extraction Radiomics feature extraction: Use the Pyradiomics tool to extract radiomics features from CT images. The specific steps include: 1834 radiomics features were extracted from the ROI images, covering multiple aspects such as shape, texture, and intensity. All features were z-score normalized to reduce variability at different scales.

[0034] Deep learning feature extraction: Use the ResNet101 deep convolutional neural network to extract multi-level deep learning features from CT images. The specific steps include: The preprocessed CT images are input into the ResNet101 model, and the features of the average pooling layer are extracted as deep learning features. Clinical feature extraction: Clinical features, including age, gender, tumor size, etc., are extracted from the patient's medical records.

[0035] 3. Feature screening and merging Feature screening: In order to reduce overfitting and feature redundancy and improve the generalization performance of the model, the following methods are used to screen features: Intraclass correlation coefficient (ICC) analysis: used to evaluate the consistency of features; t-test: used to identify features with significant differences between two groups (p < 0.05); Pearson correlation coefficient: used to evaluate the correlation between features and eliminate highly correlated redundant features; Least Absolute Shrinkage and Selection Operator (LASSO): used to further reduce the feature set and retain the most predictive features; Feature merging: Merge the filtered radiomics features and deep learning features to construct a deep learning radiomics feature set.

[0036] 4. Model construction Myometrial invasion risk assessment model: Deep learning radiomics feature sets are integrated with clinical features to construct a myometrial invasion risk assessment model. Specific steps include: The prediction model was constructed using the XGBoost algorithm, which input the filtered feature set and output the probability of myometrial invasion. Survival prognosis prediction model: The deep learning radiomics feature set is integrated with clinical features to construct a survival prognosis prediction model. The specific steps include: The Cox regression model was used to construct a survival prognosis prediction model, which input the screened feature set and output the patient's prognosis.

[0037] 5. Model Validation and Visualization Model validation: The model is validated in multiple external medical centers to ensure its generalizability and robustness. The specific steps include: The predictive performance of the model was evaluated using indicators such as AUC curve, sensitivity, and specificity; The predictive effect of the survival prognostic model was evaluated by Kaplan-Meier analysis and log-rank test; Visualization: Generate Grad-CAM (gradient-weighted class activation map) visualizations to help identify patient tumor regions and enhance model interpretability. Specific steps include: Grad-CAM technology is used to generate a heat map showing the model's attention to the tumor area in the CT image. The red area indicates the area that the model focuses on.

[0038] 6. Prediction and Output Myometrial invasion probability prediction: The patient's CT images and clinical characteristics are input into the myometrial invasion risk assessment model, which outputs the myometrial invasion probability; Prognosis prediction: The patient's CT images and clinical characteristics are input into the survival prognosis prediction model to output the patient's prognosis.

[0039] Example 2 The present invention provides an artificial intelligence-based deep learning radiomics-assisted diagnosis method for bladder cancer muscle invasion, the method comprising: CT imaging data of bladder cancer patients were collected from multiple medical institutions, including non-contrast, arterial phase, equilibrium phase, and voiding phase images.

[0040] The CT image is subjected to super-resolution reconstruction, normalization and ROI delineation processing to obtain a processed image.

[0041] The processed images are used to construct an auxiliary diagnosis model, which is then fed into a deep convolutional neural network (e.g., ResNet101) to extract multi-layer features. The image features are then integrated with the patient's clinical data to establish a myometrial invasion risk assessment model.

[0042] The auxiliary diagnosis model is used to perform auxiliary diagnosis, including performing CT imaging on bladder cancer patients, outputting muscle invasion risk assessment results, and generating an auxiliary diagnosis report for clinical reference.

[0043] In an implementation, the process of constructing an auxiliary diagnosis model using the processed image includes: By multi-center sampling, the processed images are divided into a plurality of CT image sets according to center differences; Performing ROI extraction based on a self-image segmentation algorithm and an adaptive threshold on each of the CT image sets, cropping a region of interest (ROI), and obtaining multiple ROI image sets; Each of the ROI image sets is subjected to denoising, smoothing, and super-resolution reconstruction processing to obtain multiple image sets to be trained; wherein the super-resolution reconstruction uses a GAN generative adversarial network to reconstruct the CT image from low resolution to high resolution.

[0044] Inputting the multiple image sets to be trained into a deep convolutional neural network (ResNet101) to extract deep learning image features; Inputting the multiple image sets to be trained into pyradiomics to extract radiomics features, and further screening the radiomics features using random forest or support vector machine (SVM); The deep learning imaging features and imaging omics features are integrated with the patient's clinical data (such as age, gender, and tumor size) to construct an auxiliary diagnosis model and a survival prognosis prediction nomogram.

[0045] In an implementation, the auxiliary diagnosis using the auxiliary diagnosis model includes: Obtaining the patient's CT image to be tested; Using the classifier of the auxiliary diagnosis model to classify the CT image to be detected to obtain a classified image; The classified images are input into the Grad-Cam visualization module of the auxiliary diagnosis model for splicing to obtain diagnostic images for the doctor's reference.

[0046] In an implementation, after the step of inputting the classified images into the Grad-Cam visualization module of the auxiliary diagnosis model for splicing to obtain a diagnostic image, the method further includes: Calculating the risk score of each subtype of each classified image with respect to the ROI to obtain multiple probability risk scores; The confidence level of the binary classification of whether the CT image has muscle invasion is determined according to the multiple risk scores.

[0047] In an implementation, the classifying and extracting processing of the CT images includes: Selecting a bladder region and related tumor regions from the CT image based on a trained U-Net deep learning model; Accurate segmentation of bladder and tumor tissue is achieved through multi-scale feature extraction of convolutional and deconvolutional layers; generating a corresponding segmentation mask image, wherein the mask image marks the bladder region and the tumor region; Based on the morphological characteristics and segmentation volume of the tumor area, the grading categories are determined, including: non-muscle invasive bladder cancer (NMIBC), muscle invasive bladder cancer (MIBC); Image labels are added to the segmentation results, including: bladder region label, muscle-invasive tumor region label, and non-muscle-invasive tumor region label.

[0048] Input the multiple images into a preset convolutional neural network (ResNet101) and pyradiomics to extract deep learning features and radiomics features, respectively; The deep learning imaging features and radiomics features are integrated to generate an auxiliary diagnosis model, which can output a myometrial invasion risk score.

[0049] The output muscle invasion risk score is fused with the patient's clinical data (such as age, gender, and tumor size) to construct a survival prognosis prediction nomogram.

[0050] In an implementation, the step of acquiring CT images of a bladder cancer patient includes: Collecting case information of a number of patients and extracting CT images of the patients from each case information; The PACS system is called to export the CT images into DICOM format.

[0051] In an implementation, after the step of constructing an auxiliary diagnosis model using the processed image, the method further includes: The auxiliary diagnosis accuracy of the auxiliary diagnosis model is obtained, and the auxiliary diagnosis accuracy is used to perform model verification.

[0052] Example 3: like Figure 1 As shown, this embodiment discloses a method for predicting bladder cancer muscle invasion and its prognosis based on deep learning imaging omics, including: Pyradiomics and Resnet101 were used to extract radiomics features and deep learning features from CT images, and clinical features were extracted from patient medical records. With the aim of reducing overfitting and feature redundancy and improving generalization performance, the imaging omics features and deep learning features were screened and merged based on dimensionality reduction methods such as Pearson correlation coefficient and DCA principal component analysis, and the deep learning imaging omics feature set was combined with the clinical features to construct a survival prognosis feature set.

[0053] A prediction model for predicting bladder cancer muscle invasion and survival prognosis is constructed based on the deep learning imaging genomics feature set and the survival prognosis feature set, and the patient's CT images and clinical characteristics are obtained. The images and clinical characteristics are input into the prediction model to obtain the probability of muscle invasion and the prognostic effect.

[0054] In this embodiment, the prognosis prediction model can be any one of a machine learning model, a deep learning model, and a regression model. The machine learning model can be a deep neural network model, and the regression model can be a linear fitting model or a nomogram.

[0055] In addition, in this embodiment, a computer system is also disclosed, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above method are implemented.

[0056] The present invention's method and system for predicting bladder cancer muscle invasion and its prognosis based on deep learning imaging genomics replaces pathological sections with imaging genomics features and deep learning features, combines clinical prognosis information with redundant features, establishes a muscle invasion and prognosis prediction model from both clinical indicators and CT images, and then uses this prognosis prediction model for prediction. Compared with existing technologies, using CT images instead of pathological sections can predict patients' muscle invasion and survival prognosis under non-invasive conditions, greatly advancing the decision-making time point for diagnosis and treatment; on the other hand, this method combines deep learning and imaging genomics-related technologies, implements a strict feature screening strategy, and has been verified in multiple medical centers. Therefore, the use of the prognosis prediction model in the present invention can effectively solve the lack of generalization caused by the overfitting effect.

[0057] 2. Image acquisition and processing All patients underwent enhanced CT scans covering the bladder area before surgery, including cystourethrography, combined chest and abdomen CT, and abdominal and pelvic CT. Delayed bladder filling images were retrieved from the picture archiving and communication system (PACS). In order to improve model performance and reduce overfitting, a super-resolution reconstruction method was applied to the training set. The pre-trained GAN deep learning model was trained on a large dataset of paired high-resolution and low-resolution images to extract key features from low-resolution images and reconstruct missing details in high-resolution images. The super-resolution sampling technology enhanced the original 512×512 resolution CT images to 2048×2048 resolution; For each patient, the largest two-dimensional (2D) tumor slice along the z-axis was selected as input data, and all images were filtered using a [-75, 175] Hounsfield unit (HU) window level.

[0058] 3. Model procedure and selection Handcrafted radiomic features were extracted using an in-house feature analysis program implemented in Pyradiomics. Feature selection was performed on the training set using intraclass correlation coefficient (ICC) analysis, t-test, Pearson correlation coefficient, least absolute shrinkage and selection operator (LASSO) Cox regression. The selected features were combined with their respective weights to calculate the radiomic score (Rad-score). The XGBoost algorithm was selected to build a predictive model using the identified features. Radiomic features within the tumor region of interest (ROI) were extracted using the Pyradiomics tool in Python (version 2.1.2). A total of 1,834 features were extracted (Figure S3A, B). All features were normalized using z-scores to minimize variability at different scales. During feature selection, t-tests were used to identify features that were significantly different between the two risk groups (p < 0.05); Pearson correlation coefficients were calculated to assess the correlation between features, and the least absolute shrinkage and selection operator (LASSO) logistic regression model was applied to further reduce the feature set, retaining the most predictive features for model building ( Figure S3C,D ).

[0059] Features extracted from the average pooling layer of ResNet101 were used as deep learning features. The deep learning model demonstrated good ability to distinguish muscle-invasive bladder cancer (MIBC) from non-muscle-invasive bladder cancer (NMIBC) in an internal validation cohort (Figure S5A). The risk score distribution generated by the deep learning (DL) model and the gradient-weighted class activation map (Grad-CAM) of the original image is shown in Figure 5. Figure 2 The red highlighted areas are mainly concentrated in the basal and tumor regions, indicating that the DL model mainly focuses on these areas to extract valuable information.

[0060] A hybrid model deep learning radiomics nomogram model for predicting bladder cancer muscle invasion was established by combining radiomics score (RL) with deep learning features (DL). Figure 3 ), providing personalized assessment of the overall survival of patients after bladder cancer surgery. Analyze the SHAP visualization, AUC curve, and survival analysis of the Nomogram model of the present invention. Among them, the SHAP visualization analyzes the weight influence of each feature in the deep learning imaging omics feature set on the model results ( Figure 4 ). AUC curve of the nomogram model and survival analysis Figure 5 As shown, the prognostic prediction effect was verified by external multicenter cohorts, of which two cohorts verified the prediction effect of muscle invasion as shown in Figure 5 C. Figure 5 D, and the prognostic prediction effect was verified by an external multicenter cohort, among which the prognostic prediction effect was as follows Figure 5 As shown in F\G, Figure 5 It can be seen that the AUC curve and survival analysis of the DLRN model of the present invention both prove that the DLRN model of the present invention has good myometrial invasion prediction and prognosis prediction capabilities.

[0061] Conclusion: This paper demonstrates various implementations of a deep learning radiomics-based system for predicting muscle invasion and prognosis in bladder cancer. These implementations incorporate advanced technologies such as multimodal data fusion, time series analysis, multitask learning, and federated learning. These technologies effectively improve prediction accuracy, reduce the risk of overfitting, and ensure the generalizability and robustness of the model through multicenter validation. Furthermore, the present invention provides a Grad-CAM visualization module, which enhances the model's interpretability and helps physicians better understand AI-assisted diagnostic results.

[0062] Each embodiment in this specification is described in a related manner. Similar parts between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences between the other embodiments. In particular, the system embodiment is generally similar to the method embodiment, so the description is relatively simple. For related parts, refer to the description of the method embodiment.

[0063] The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention are included in the scope of protection of the present invention.

Claims

1. A bladder cancer muscle invasion and prognosis prediction system based on deep learning imaging omics, characterized by: include: A data acquisition module is used to collect CT imaging data and clinical data of bladder cancer patients; A preprocessing module, configured to perform denoising, smoothing, data enhancement, super-resolution reconstruction, and ROI delineation on the CT image; A feature extraction module is used to extract radiomics features and deep learning features from the CT images and clinical features from the patient's medical records; A feature screening module is used to screen and merge the radiomics features and deep learning features based on dimensionality reduction methods such as Pearson correlation coefficient and DCA principal component analysis to construct a deep learning radiomics feature set; A model building module, used to build a muscle invasion risk assessment model and a survival prognosis prediction model based on the deep learning radiomics feature set and clinical characteristics; The prediction module is used to input the patient's CT images and clinical characteristics into the prediction model and output the probability of myometrial invasion and the prognostic effect.

2. The bladder cancer muscle invasion and prognosis prediction system based on deep learning imaging omics according to claim 1, characterized in that: The pre-processing module further comprises: A super-resolution reconstruction unit is used to reconstruct low-resolution CT images into high-resolution images using a GAN generative adversarial network; ROI delineation unit is used to extract the region of interest (ROI) based on the self-image segmentation algorithm and adaptive threshold.

3. The bladder cancer muscle invasion and prognosis prediction system based on deep learning imaging omics according to claim 1, characterized in that: The feature extraction module further comprises: Radiomics feature extraction unit, used to extract radiomics features from CT images using Pyradiomics tools; The deep learning feature extraction unit is used to extract multi-level deep learning features from CT images using the ResNet101 deep convolutional neural network.

4. The bladder cancer muscle invasion and prognosis prediction system based on deep learning imaging omics according to claim 1, characterized in that: The feature screening module further includes: Feature screening unit, used to screen features using intraclass correlation coefficient (ICC) analysis, t-test, Pearson correlation coefficient, least absolute shrinkage and selection operator (LASSO), and Cox regression; The feature merging unit is used to merge the screened radiomics features and deep learning features to construct a deep learning radiomics feature set.

5. The bladder cancer muscle invasion and prognosis prediction system based on deep learning imaging omics according to claim 1, characterized in that: The model building module further includes: a muscle invasion risk assessment model construction unit, configured to fuse the deep learning radiomics feature set with clinical features to construct a muscle invasion risk assessment model; The survival prognosis prediction model construction unit is used to fuse the deep learning imaging genomics feature set with clinical features to construct a survival prognosis prediction model.

6. The bladder cancer muscle invasion and prognosis prediction system based on deep learning imaging omics according to claim 1, characterized in that: The prediction module further comprises: a muscle invasion probability prediction unit, configured to output the patient's muscle invasion probability according to the muscle invasion risk assessment model; A prognosis effect prediction unit, configured to output the patient's prognosis effect according to the survival prognosis prediction model; It also includes a visualization module for generating Grad-CAM visualizations to help identify patient tumor regions and enhance model interpretability. The prognosis prediction system was validated by multiple centers to ensure the generalizability and robustness of the model.

7. A method for predicting bladder cancer muscle invasion and prognosis based on deep learning imaging omics, characterized in that: The following steps are involved: S1: Collect and preprocess CT imaging data of bladder cancer patients; S2: extracting radiomics features and deep learning features from the CT images, and extracting clinical features from the patient's medical records; S3: Screen and merge the radiomics features and deep learning features based on dimensionality reduction methods such as Pearson correlation coefficient and DCA principal component analysis to construct a deep learning radiomics feature set; S4: Construct a muscle invasion risk assessment model and a survival prognosis prediction model based on the deep learning radiomics feature set and clinical characteristics; S5: Input the patient's CT images and clinical characteristics into the prediction model, and output the probability of myometrial invasion and the prognostic effect.

8. The method for predicting bladder cancer muscle invasion and prognosis based on deep learning imaging omics according to claim 7, characterized in that: The preprocessing steps include denoising, smoothing, data enhancement, super-resolution reconstruction and ROI delineation; The feature extraction step includes extracting radiomics features using the Pyradiomics tool and extracting deep learning features using the ResNet101 deep convolutional neural network; The feature screening step includes screening features using intraclass correlation coefficient (ICC) analysis, t-test, Pearson correlation coefficient, least absolute shrinkage and selection operator (LASSO) and Cox regression; The model construction step includes fusing the deep learning radiomics feature set with clinical features to construct a muscle invasion risk assessment model and a survival prognosis prediction model; The prediction step includes outputting the patient's muscle invasion probability according to the muscle invasion risk assessment model, and outputting the patient's prognosis according to the survival prognosis prediction model; It also includes generating Grad-CAM visualizations to help identify patient tumor regions and enhance model interpretability.

9. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method for predicting bladder cancer muscle invasion and prognosis based on deep learning imaging omics according to any one of claims 7 and 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for predicting bladder cancer muscle invasion and prognosis based on deep learning imaging omics according to any one of claims 7 and 8 are implemented.

Images