Lung cancer immunotherapy curative effect evaluation and prediction intelligent analysis system

Abstract

The invention discloses a lung cancer immunotherapy curative effect evaluation and prediction intelligent analysis system, which belongs to the technical field of medical image intelligent analysis, and comprises a multi-temporal CT image registration engine, a radiomics feature extraction module, an immunoreaction evaluation module and a curative effect prediction and scheme optimization module, a registration engine carries out precise registration on CT images before and after treatment, a feature extraction module extracts radiomics features from a tumor and a peritumoral region and calculates Delta-radiomics features, and an evaluation module identifies false progress, true progress and super progress based on an immune-related reaction standard algorithm. The prediction module outputs curative effect scores and therapeutic scheme suggestions through an integrated learning model, the system constructs a deep coupling closed-loop cooperation mechanism among registration quality, feature extraction, evaluation accuracy and curative effect prediction, and noninvasive dynamic monitoring and accurate individualized evaluation are achieved.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical image analysis and tumor treatment evaluation technology, specifically to an intelligent analysis system for evaluating and predicting the efficacy of lung cancer immunotherapy based on radiomics and deep learning technologies using computed tomography images. Background Technology

[0002] Lung cancer is one of the leading causes of cancer-related deaths worldwide, with non-small cell lung cancer accounting for approximately 85% of all lung cancer cases. In recent years, immune checkpoint inhibitors, such as programmed death receptor-1 (PD-1) and programmed death ligand-1 (PD-L1) inhibitors, have made groundbreaking progress in lung cancer treatment, significantly improving the survival prognosis of patients with advanced lung cancer. However, the objective response rate of immunotherapy is only 20% to 50%, meaning that most patients do not benefit from immunotherapy, and some patients experience atypical response patterns such as pseudoprogression and hyperprogression, posing significant challenges to clinical decision-making.

[0003] Pseudoprogression refers to a temporary increase in tumor size or the appearance of new lesions at the beginning of immunotherapy, followed by a gradual decrease in tumor burden, with an incidence of approximately 2% to 10%. The mechanism of pseudoprogression may be related to immune cell infiltration, local inflammatory response, and tumor necrosis. Misinterpreting pseudoprogression as disease progression and prematurely discontinuing effective immunotherapy will result in patients missing out on potential treatment benefits. Conversely, hyperprogression refers to a significant acceleration in tumor growth after the start of immunotherapy, with an incidence of approximately 9% to 14%. The median survival of patients with hyperprogression is only 3.4 months, significantly lower than the 6.2 months of patients with conventional progression. For patients with hyperprogression, continuing immunotherapy is not only ineffective but also delays the opportunity for other effective treatments, accelerating disease deterioration. Therefore, accurately identifying pseudoprogression and hyperprogression, and achieving precise assessment and prediction of immunotherapy efficacy, is of great significance for optimizing clinical decision-making.

[0004] The traditional RECIST 1.1 standard for evaluating the efficacy of treatment in solid tumors primarily assesses treatment response based on changes in tumor diameter. However, this standard does not consider information on tumor heterogeneity and struggles to accurately assess atypical response patterns to immunotherapy. To address the evaluation challenges posed by immunotherapy, researchers have proposed novel assessment criteria such as the immune-related response criteria (irRC), the immune-related solid tumor efficacy evaluation criteria (irRECIST), and the immune RECIST (iRECIST). These criteria introduce the concept of "unconfirmed disease progression," requiring a follow-up imaging evaluation after 4 to 12 weeks to confirm progression. However, these criteria still primarily rely on morphological changes in tumor size and lack a deep characterization of the tumor's internal microstructure and functional metabolic features.

[0005] Radiomics technology extracts quantitative features from medical images through high-throughput processing, non-invasively reflecting the biological characteristics of tumors, such as heterogeneity, angiogenesis, metabolic activity, and the immune microenvironment. In recent years, multiple studies have shown a significant correlation between CT radiomics features and lung cancer immunotherapy response. Pre-treatment radiomics features can predict the efficacy of PD-1 / PD-L1 inhibitors, while dynamic changes in radiomics features before and after treatment (Delta-radiomics) can more accurately predict treatment response and survival benefit. Peritumoral radiomics features reflect the tumor microenvironment and immune infiltration status; radiomics models combining intratumoral and peritumoral analysis show superior performance in predicting immunotherapy response compared to intratumoral analysis alone. Furthermore, automated machine learning techniques can automatically select the optimal feature combination from a large number of radiomics features and construct high-performance predictive models with an accuracy of 84% to 89% and an area under the receiver operating characteristic (AUC) of 0.98 to 1.00.

[0006] Existing technology CN119359675A discloses a method for predicting the efficacy of lung cancer immunotherapy based on pathological slide images. This method optimizes the quality of pathological tissue slide images using a neural network algorithm, extracts image feature vectors using a Phikon model, and then predicts the treatment effect using an immunotherapy scoring prediction neural network. The method annotates training samples based on immune infiltration scores and trains a Transformer encoder network using the LA-MIL method. However, this existing technology has the following shortcomings: First, the method is based on the analysis of pathological slide images, which requires obtaining tumor tissue specimens through puncture or surgery, making it an invasive examination and unsuitable for dynamic monitoring during treatment; second, pathological slides can only reflect local information of the sampling site, making it difficult to comprehensively assess the overall heterogeneity and spatial distribution characteristics of the tumor; third, the method only predicts based on image features at a single time point, without utilizing the dynamic changes in imaging features before and after treatment; fourth, the method lacks a mechanism for identifying atypical response patterns to immunotherapy, such as pseudoprogression and hyperprogression; fifth, the method is trained based on immune infiltration scores, which mainly reflect the degree of immune cell infiltration in the tumor microenvironment, and differ to some extent from the actual clinical treatment response, potentially limiting the clinical applicability of the prediction model.

[0007] Therefore, there is an urgent need to develop a non-invasive, dynamic, and comprehensive system for evaluating and predicting the efficacy of lung cancer immunotherapy based on CT images. This system should be able to accurately identify pseudoprogression and hyperprogression, providing reliable technical support for clinical decision-making. Summary of the Invention

[0008] The purpose of this invention is to provide an intelligent analysis system for evaluating and predicting the efficacy of lung cancer immunotherapy, in order to solve the technical problems existing in the prior art, such as the high degree of invasiveness of the evaluation methods, the inability to dynamically monitor, the difficulty in identifying atypical response patterns, and the insufficient accuracy of prediction.

[0009] The present invention achieves the above objectives through the following technical solutions:

[0010] This invention provides an intelligent analysis system for evaluating and predicting the efficacy of lung cancer immunotherapy, comprising a multi-temporal CT image registration engine, a radiomics feature extraction module, an immune response assessment module, and an efficacy prediction and treatment plan optimization module. The multi-temporal CT image registration engine acquires baseline CT images of lung cancer patients before treatment and follow-up CT images after treatment. It performs spatial registration on the baseline and follow-up CT images to generate registered image pairs, and determines image resampling parameters and registration accuracy parameters based on the registration quality parameters of the registered image pairs. The radiomics feature extraction module is connected to the multi-temporal CT image registration engine. Based on the image resampling parameters, it performs image standardization processing on the registered image pairs, identifies the tumor region and peritumoral region in the registered image pairs, and extracts radiomics feature sets from the tumor region and peritumoral region. The radiomics feature sets include shape features, first-order statistical features, and texture features. Delta-radiomics features are calculated based on the radiomics feature sets, representing the dynamic changes in radiomics features before and after treatment. The immune response assessment module is connected to the radiomics feature extraction module. Based on Delta-radiomics features and registration accuracy parameters, it evaluates the immunotherapy response type in lung cancer patients using a standard algorithm for immune-related responses. Immunotherapy response types include pseudoprogression, true progression, and hyperprogression. The module generates immune response assessment results, including response type identifiers and response confidence scores. Based on the response confidence scores, it feeds back registration optimization instructions to the multi-temporal CT image registration engine. The efficacy prediction and treatment plan optimization module is connected to the immune response assessment module. It receives the immune response assessment results and a set of radiomics features, inputs the radiomics feature set into a pre-trained efficacy prediction model, and outputs an efficacy score and predicted survival value. Based on the efficacy score and predicted survival value, it generates treatment plan optimization suggestions, including suggestions for continuing immunotherapy, dose adjustment, and combination therapy. The efficacy score is fed back to the immune response assessment module for dynamically adjusting the assessment threshold parameters.

[0011] This invention achieves precise spatial alignment of CT images before and after treatment by setting up a multi-temporal CT image registration engine, laying the foundation for subsequent dynamic feature analysis. The registration engine can adaptively adjust image processing parameters according to registration quality, ensuring that the registration accuracy meets the requirements of radiomics feature extraction. This invention achieves high-throughput quantitative feature extraction of tumor and peritumoral regions through a radiomics feature extraction module. These features can non-invasively reflect tumor heterogeneity, morphological changes, and microenvironment characteristics. In particular, this invention introduces Delta-radiomics features, which, by capturing the dynamic changes in radiomics features before and after treatment, can more accurately reflect changes in tumor internal structure and functional metabolism caused by immunotherapy. This invention achieves intelligent assessment based on immune-related response criteria through an immune response assessment module, accurately identifying different response types such as pseudoprogression, true progression, and hyperprogression. This module combines radiomics features with imaging morphological changes, overcoming the limitations of traditional assessment criteria that rely solely on changes in tumor size. This invention achieves machine learning-based efficacy prediction and personalized treatment plan recommendations through an efficacy prediction and treatment plan optimization module, providing decision support for clinicians.

[0012] The beneficial effects of this invention are as follows: First, this invention is based on CT image analysis, which is a non-invasive examination method that allows for dynamic monitoring throughout the treatment process, avoiding the invasiveness and sampling limitations of traditional pathological biopsies; Second, this invention achieves precise alignment of images before and after treatment through multi-temporal CT image registration, ensuring the accuracy of dynamic feature analysis. The registration quality parameters can provide feedback to optimize the registration strategy, forming an adaptive improvement mechanism for registration accuracy; Third, this invention extracts radiomics features within and around the tumor, comprehensively reflecting the overall heterogeneity and microenvironment characteristics of the tumor. Delta-radiomics features capture dynamic changes before and after treatment, significantly improving the accuracy of efficacy prediction; Fourth, this invention is based on immunology... The relevant response standard algorithm enables intelligent identification of pseudoprogression, true progression, and hyperprogression, solving the technical problem that traditional assessment standards cannot distinguish atypical response patterns and providing a reliable basis for clinical decision-making. Fifth, this invention constructs a closed-loop collaborative mechanism among multiple modules. Registration quality affects the accuracy of feature extraction, assessment confidence feedback optimizes registration parameters, and efficacy scores dynamically adjust assessment thresholds. Each module promotes and optimizes each other, achieving continuous improvement in system performance. Sixth, this invention, through an integrated learning architecture efficacy prediction model, comprehensively utilizes radiomics features, immune response assessment results, and clinical features, achieving a prediction accuracy of over 85%, which can provide precise support for the optimization of individualized treatment plans. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the overall architecture of the intelligent analysis system for evaluating and predicting the efficacy of lung cancer immunotherapy according to the present invention.

[0014] Figure 2 This is a schematic diagram of the multi-temporal CT image registration engine of the present invention.

[0015] Figure 3 This is a schematic diagram of the workflow of the radiomics feature extraction module of the present invention.

[0016] Figure 4 This is a schematic diagram of the evaluation logic of the immune response evaluation module of the present invention.

[0017] Figure 5 This is a schematic diagram of the therapeutic effect prediction and scheme optimization module of the present invention.

[0018] Figure 6 This is a schematic diagram of the Delta-radiomics feature calculation process of the present invention. Detailed Implementation

[0019] Please refer to the attached document. Figures 1-6 To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of this invention.

[0020] Reference Figure 1 , Figure 1 This is a schematic diagram of the overall architecture of the intelligent analysis system for evaluating and predicting the efficacy of lung cancer immunotherapy according to the present invention. The intelligent analysis system for evaluating and predicting the efficacy of lung cancer immunotherapy according to the present invention includes a multi-temporal CT image registration engine 1, a radiomics feature extraction module 2, an immune response assessment module 3, and an efficacy prediction and treatment plan optimization module 4. These four core modules are tightly connected through data flow and control flow, forming a deeply coupled closed-loop collaborative system.

[0021] The multi-temporal CT image registration engine 1 acquires baseline CT images of lung cancer patients before treatment and follow-up CT images at different follow-up time points after treatment. It then performs spatial registration on the baseline and follow-up CT images to generate registered image pairs. Registration engine 1 determines image resampling parameters and registration accuracy parameters based on the registration quality parameters of the registered image pairs and passes these parameters to the radiomics feature extraction module 2. Registration engine 1 also receives registration optimization instructions from the immune response assessment module 3 and dynamically adjusts the registration strategy and parameters according to these instructions to achieve adaptive optimization of registration accuracy. This feedback mechanism ensures that the registration quality meets the accuracy requirements of subsequent feature extraction and evaluation analysis.

[0022] The radiomics feature extraction module 2 is connected to the multi-temporal CT image registration engine 1, receiving registered image pairs and image resampling parameters. Module 2 performs image normalization processing on the registered image pairs based on the image resampling parameters to ensure that CT images acquired at different time points and by different scanning devices have consistent spatial resolution and grayscale range. Module 2 identifies the tumor region and peritumoral region in the registered image pairs, and extracts radiomics feature sets from the tumor region and peritumoral region. The radiomics feature set includes shape features, first-order statistical features, and texture features, which quantitatively describe the morphology, density distribution, and texture pattern of the tumor from multiple dimensions. Module 2 calculates Delta-radiomics features based on the radiomics feature sets before and after treatment. This feature characterizes the dynamic changes in radiomics features before and after treatment and can sensitively reflect changes in the internal structure and functional metabolism of the tumor caused by immunotherapy.

[0023] The immune response assessment module 3 is connected to the radiomics feature extraction module 2, receiving Delta-radiomics features, radiomics feature sets, and registration accuracy parameters. Based on the Delta-radiomics features and registration accuracy parameters, module 3 assesses the immunotherapy response type of lung cancer patients using a standard algorithm for immune-related responses. Immunotherapy response types include three atypical response modes: pseudoprogression, true progression, and hyperprogression. Module 3 generates the immune response assessment result, which includes a response type identifier and response confidence level. The response confidence level reflects the reliability of the assessment result. When the confidence level is low, module 3 feeds back registration optimization instructions to the multi-temporal CT image registration engine 1, instructing the registration engine 1 to adjust the registration strategy or improve the registration accuracy, thereby improving the reliability of the assessment. This feedback mechanism forms a closed-loop optimization between assessment quality and registration quality.

[0024] The efficacy prediction and treatment plan optimization module 4 is connected to the immune response assessment module 3, receiving the immune response assessment results and a set of radiomics features. Module 4 inputs the radiomics feature set into a pre-trained efficacy prediction model. This model employs an ensemble learning architecture, capable of integrating the prediction results of multiple machine learning algorithms to output an efficacy score and a survival prediction value. The efficacy score quantitatively reflects the patient's response to immunotherapy, while the survival prediction value estimates the patient's progression-free survival and overall survival. Based on the efficacy score, survival prediction value, and immune response assessment results, module 4 generates treatment plan optimization suggestions, including suggestions for continuing immunotherapy, dose adjustment, and combination therapy. Furthermore, module 4 feeds back the efficacy score to the immune response assessment module 3 for dynamically adjusting the assessment threshold parameters. Once sufficient patient data has been accumulated, the system can automatically calibrate the threshold parameters in the immune response assessment module 3 by comparing actual treatment outcomes with predicted results, making the assessment criteria more consistent with real-world clinical scenarios and continuously improving assessment accuracy.

[0025] Reference Figure 2 , Figure 2 This is a schematic diagram of the structure of the multi-temporal CT image registration engine 1 of the present invention. The multi-temporal CT image registration engine 1 includes an image preprocessing unit, a registration strategy selection unit, a registration execution unit, and a registration quality evaluation unit.

[0026] The image preprocessing unit performs noise reduction and contrast enhancement on baseline and follow-up CT images to generate preprocessed images. Specifically, the noise reduction uses a nonlocal mean filtering algorithm, which can effectively preserve edge and detail information while removing image noise. The contrast enhancement employs an adaptive histogram equalization method, which enhances the contrast between the tumor and surrounding normal tissue, facilitating subsequent registration and segmentation operations.

[0027] The registration strategy selection unit selects a suitable registration strategy from rigid registration, affine registration, and non-rigid registration based on the time interval between the baseline CT image and the follow-up CT image, and the magnitude of tumor position change. In a preferred embodiment of the invention, when the time interval is less than 4 weeks and the tumor position change is less than 5 mm, rigid registration or affine registration is sufficient to meet the accuracy requirements; when the time interval is greater than 8 weeks or the tumor position change is greater than 10 mm, non-rigid registration is selected to better handle significant changes in tumor morphology. The magnitude of tumor position change is determined by calculating the Euclidean distance between the tumor centroids in the baseline image and the follow-up image.

[0028] The registration execution unit performs spatial registration on the preprocessed images according to the registration strategy, calculating the transformation matrix to generate registered image pairs. For rigid registration, the unit employs a mutual information-based registration algorithm, maximizing the mutual information between the two images by optimizing rotation and translation parameters. For affine registration, scaling and shearing transformation parameters are added to the rigid transformation. For non-rigid registration, the unit uses a B-spline free deformation model, which describes the local deformation of the image through a control point grid, enabling it to handle nonlinear morphological changes of tumors. The registration process employs a multi-resolution strategy, progressively optimizing from low to high resolution, which improves registration efficiency while avoiding getting trapped in local optima.

[0029] The registration quality evaluation unit calculates the mutual information value and normalized cross-correlation coefficient of the registered image pair, and determines the registration quality parameters based on these values. The mutual information value reflects the information correlation between the two images; a higher mutual information value indicates better registration. The normalized cross-correlation coefficient measures the similarity of the gray-level distributions of the two images, ranging from -1 to 1; a value closer to 1 indicates higher registration accuracy. The registration quality parameters are calculated using a weighted fusion method, comprehensively considering both the mutual information value and the normalized cross-correlation coefficient. Preferably, the weight of the mutual information value is 0.6, and the weight of the normalized cross-correlation coefficient is 0.4. When the registration quality parameter is higher than 0.85, the registration accuracy is considered to meet the requirements; when the registration quality parameter is lower than 0.75, the registration execution unit automatically switches to a higher-level registration strategy or increases the number of iterations to improve registration accuracy.

[0030] Registration quality parameters and registration accuracy parameters are passed to the radiomics feature extraction module 2. The registration accuracy parameters include the root mean square value of the registration error and the number of iterations required for registration convergence. These parameters are used to guide the uncertainty assessment during the feature extraction process. Furthermore, when the response confidence of the immune response assessment module 3 is lower than a preset threshold, the assessment module 3 generates a registration optimization instruction and feeds it back to the registration engine 1. The registration engine 1 adjusts the registration strategy or improves the registration accuracy according to this instruction, for example, switching from affine registration to non-rigid registration, or increasing the number of registration iterations from 100 to 200.

[0031] Reference Figure 3 , Figure 3 This is a schematic diagram of the workflow of the radiomics feature extraction module 2 of the present invention. The radiomics feature extraction module 2 receives registered image pairs and image resampling parameters from the multi-temporal CT image registration engine 1. Module 2 first performs image normalization processing on the registered image pairs based on the image resampling parameters.

[0032] Image normalization processing includes two steps: image resampling and grayscale normalization. Image resampling resamples the registered image pairs to a consistent voxel spacing, preferably 1mm × 1mm × 1mm, using cubic B-spline interpolation to preserve image smoothness and detail. Grayscale normalization maps the grayscale values ​​of the CT image to a uniform window width and level range. For lung cancer CT images, a preferred window width is 1500HU to 2000HU, and a window level is -400HU to -200HU. This window width and level range can clearly display the lung tumor and its surrounding tissue structures.

[0033] Next, module 2 identifies the tumor region and the peritumoral region in the registered image pairs. The tumor region is identified using a three-dimensional convolutional neural network for semantic segmentation. In a preferred embodiment of the invention, a three-dimensional segmentation network with a U-Net architecture is used. This network extracts multi-scale features in the encoder part and recovers spatial details in the decoder part through upsampling and skip connections, enabling accurate tumor segmentation. The segmentation network is pre-trained on a public dataset containing more than 500 lung cancer cases and can be fine-tuned according to specific scenarios in clinical applications. The segmentation results undergo morphological post-processing, including hole filling and boundary smoothing, to determine the tumor boundary. The peritumoral region is defined by extending a certain distance outward from the tumor boundary, preferably between 5mm and 15mm. In a specific embodiment of the invention, the extension distance is set to 10mm, which effectively captures the tumor microenvironment and immune infiltration features while avoiding the inclusion of too much irrelevant normal tissue.

[0034] Module 2 extracts intratumoral radiomics features from the tumor region. The extraction of intratumoral radiomics features utilizes the PyRadiomics open-source software package, which implements the calculation of various radiomics features based on the specifications of the Imaging Biomarkers Standardization Initiative. In a preferred embodiment of the invention, 851 to 1219 intratumoral radiomics features are extracted from the tumor region. These features include 18 first-order statistical features, 14 shape features, and multiple texture features. The first-order statistical features include energy, entropy, kurtosis, skewness, mean, median, maximum, minimum, range, mean absolute deviation, robust mean absolute deviation, root mean square, standard deviation, total energy, homogeneity, variance, 10th percentile, and 90th percentile. These features describe the statistical distribution of voxel gray values ​​within the tumor region. Shape features include volume, surface area, sphericity, compressibility, maximum 2D diameter, elongation, flatness, minimum 2D diameter, principal axis length, secondary axis length, minimum axis length, spherical unconformity, surface area-to-volume ratio, and major-to-minor axis ratio. These features describe the three-dimensional morphological characteristics of the tumor. Texture features include gray-level co-occurrence matrix features, gray-level run-length matrix features, gray-level size region matrix features, adjacent gray-level tone difference matrix features, and gray-level dependency matrix features. There are 24 gray-level co-occurrence matrix features, describing the spatial relationship of gray values ​​of adjacent voxels; 16 gray-level run-length matrix features, describing the pattern of consecutive occurrence of voxels with the same gray value; 16 gray-level size region matrix features, describing the size distribution of connected regions with the same gray value; 5 adjacent gray-level tone difference matrix features, describing the gray-level difference between a voxel and its neighborhood; and 14 gray-level dependency matrix features, describing the degree of dependence of voxel gray values ​​on their neighborhood.

[0035] To enhance the multi-scale expressive power of the features, Module 2 also extracts radiomics features from the wavelet transform images. The wavelet transform employs a three-dimensional discrete wavelet transform, performing high-pass and low-pass filtering decomposition on the original CT image to generate wavelet sub-band images with eight different frequency combinations. First-order statistical features and texture features are extracted from each wavelet sub-band image, while shape features are extracted only from the original image. Ultimately, the total number of intratumoral radiomics features reaches 851 to 1219, depending on the number of computational directions including wavelet and texture features.

[0036] Module 2 extracts peritumoral radiomics features from the peritumoral region. The extraction method for peritumoral radiomics features is similar to that for intratumoral features, but does not include shape features, as the peritumoral region is an artificially defined extended area whose shape features have no biological significance. Peritumoral radiomics features include first-order statistical features and texture features, used to characterize the heterogeneity of the tumor microenvironment. The tumor microenvironment includes vascular distribution, matrix composition, and immune cell infiltration, factors that significantly influence the response to immunotherapy. The texture features of the peritumoral region can reflect the spatial distribution patterns of these microenvironmental factors.

[0037] Module 2 combines intratumoral and peritumoral radiomics features to form a complete radiomics feature set. During the combination process, to distinguish between intratumoral and peritumoral features, prefixes are added to the feature names; for example, "intra_energy" indicates intratumoral energy features, and "peri_entropy" indicates peritumoral entropy features. The total number of features in the radiomics feature set is typically between 1500 and 2400.

[0038] Next, Module 2 calculates the Delta-radiomics profile based on the pre- and post-treatment radiomics profile sets. (Refer to...) Figure 6 , Figure 6 This is a schematic diagram of the Delta-radiomics feature calculation process of the present invention. Module 2 obtains the baseline feature vector of the baseline CT image and the follow-up feature vector of the follow-up CT image. The baseline feature vector and the follow-up feature vector contain the same type and number of radiomics features, and there is a one-to-one correspondence between the features.

[0039] Module 2 normalizes each eigenvalue in the baseline and follow-up eigenvectors. The normalization process uses Z-score standardization to convert the eigenvalues ​​into a standard normal distribution with a mean of 0 and a standard deviation of 1. The normalization formula is:

[0040] ,

[0041] in, These are the normalized eigenvalues. These are the original eigenvalues. To obtain the mean of this feature in the training set, The standard deviation of this feature in the training set is used. Normalization eliminates the influence of different feature dimensions and numerical ranges, making the features comparable.

[0042] Module 2 calculates the difference between the follow-up feature vector and the baseline feature vector, generating a difference feature vector. The difference feature vector reflects the absolute change in each radiomics feature, and its calculation formula is as follows:

[0043] ,

[0044] in, The difference eigenvector, For follow-up feature vectors, This represents the baseline feature vector. The difference feature directly reflects the magnitude of change in features before and after treatment; positive values ​​indicate an increase in feature values, while negative values ​​indicate a decrease in feature values.

[0045] Module 2 also calculates the ratio of the follow-up feature vector to the baseline feature vector, generating a ratio feature vector. The ratio feature vector reflects the relative rate of change of each radiomics feature, and is calculated using the following formula:

[0046] ,

[0047] in, The ratio eigenvector, The ratio is a very small positive number (e.g., 1e-8) to avoid cases where the denominator is zero. The ratio characteristic can reflect the relative degree of change of the characteristic before and after treatment. For characteristics with small baseline values, the ratio can more sensitively reflect their changes.

[0048] Module 2 combines difference feature vectors and ratio feature vectors to form Delta-radiomics features. The dimension of Delta-radiomics features is twice that of the original radiomics feature set. For example, if the radiomics feature set has 1800 features, the Delta-radiomics feature set will have 3600 features, including 1800 difference features and 1800 ratio features. In practical applications, Delta-radiomics features can sensitively capture the dynamic changes in tumor internal structure, density distribution, and texture patterns induced by immunotherapy. Studies have shown that Delta-radiomics features significantly outperform baseline or follow-up features alone in predicting immunotherapy response.

[0049] To avoid the curse of dimensionality and overfitting caused by excessively high feature dimensionality, module 2 performs feature selection on the radiomics feature set and Delta-radiomics features. Feature selection employs a recursive feature elimination method combined with cross-validation, which can identify the subset of features that contributes most to the prediction task. In a preferred embodiment of the invention, the final number of retained features is 30 to 100, ensuring feature representativeness while controlling model complexity.

[0050] Reference Figure 4 , Figure 4 This is a schematic diagram of the evaluation logic of the immune response evaluation module 3 of the present invention. The immune response evaluation module 3 receives Delta-radiomics features, radiomics feature sets, and registration accuracy parameters from the radiomics feature extraction module 2. Based on the Delta-radiomics features and registration accuracy parameters, module 3 evaluates the type of immunotherapy response in lung cancer patients using a standard algorithm for immune-related responses.

[0051] The immune-related response standard algorithm combines the irRECIST assessment framework with quantitative analysis of radiomic features. Module 3 first measures the maximum diameter of the target tumor lesion based on the registered image pairs. The selection of target lesions follows the irRECIST criteria, prioritizing the largest measurable lesions, with a maximum of 5 lesions selected per organ and a maximum of 10 target lesions selected systemically. Module 3 calculates the sum of the maximum diameters of all target lesions as the baseline tumor burden value. The baseline tumor burden value is denoted as... .

[0052] At the follow-up time point after treatment, Module 3 again measured the sum of the maximum diameters of the target lesions, denoted as... Module 3 calculates the percentage change relative to the baseline tumor burden value using the following formula:

[0053] ,

[0054] in, This represents the percentage change in tumor burden. A positive value indicates an increase in tumor burden, while a negative value indicates a decrease in tumor burden.

[0055] Module 3 combines Delta-radiomics characteristics to comprehensively determine the type of immunotherapy response. The judgment logic is as follows:

[0056] First, if the percentage change in tumor burden shows an increase of more than 25%, that is... If Delta-radiomics characteristics show a decrease in intratumoral heterogeneity indicators, it is identified as pseudoprogression. Intratumoral heterogeneity indicators include entropy, homogeneity, and variance of texture features, which reflect the complexity of the tumor's internal structure. In pseudoprogression, although the tumor volume increases, the internal structure becomes more homogeneous due to immune cell infiltration and inflammatory response, resulting in a decrease in heterogeneity indicators. In a specific embodiment of the present invention, the Delta value of the entropy feature is used as a representative indicator of heterogeneity changes. If the Delta value of entropy is less than -0.3, it indicates a significant decrease in intratumoral heterogeneity, which, combined with an increase in tumor burden, is determined to be pseudoprogression.

[0057] Second, if the percentage change in tumor burden shows an increase in tumor burden exceeding 50%, that is... If the tumor growth rate increases by more than two times compared to pre-treatment levels, it is classified as hyperprogression. The tumor growth rate is calculated using the tumor growth rate ratio method. First, the pre-treatment tumor growth rate is calculated. Calculated based on the changes in tumor diameter and the time interval between the two time points before treatment:

[0058] ,

[0059] in, and These are the maximum tumor diameters at two time points before treatment. and This corresponds to a specific time point. Then, the tumor growth rate after treatment is calculated. :

[0060] ,

[0061] in, The maximum diameter of the tumor at the start of treatment. This refers to the maximum diameter of the tumor at the follow-up after treatment. and For the corresponding time points, calculate the tumor growth rate ratio:

[0062] ,

[0063] like This indicates that the tumor growth rate after treatment is more than twice that before treatment, and combined with an increase in tumor burden of more than 50%, it is judged as hyperprogression.

[0064] Third, if the percentage change in tumor burden shows an increase in tumor burden exceeding 20%, that is... However, if the criteria for pseudoprogression and hyperprogression are not met, it is identified as true progression. True progression refers to the continuous growth of the tumor, the lack of a decrease in heterogeneity in Delta-radiomics features, and the absence of a significant acceleration in the tumor growth rate.

[0065] Fourth, if the percentage change in tumor burden is between -20% and 20%, that is... Fifth, if the percentage change in tumor burden shows a reduction in tumor burden of more than 30%, it is considered stable. The first step is to mark it as partial remission. The sixth step is to mark it as complete remission if all target lesions have completely disappeared.

[0066] Module 3 generates the immune response assessment results, which include a response type identifier and a response confidence level. The response type identifier is one of the following: pseudoprogression, true progression, hyperprogression, stable disease, partial remission, or complete remission. The response confidence level reflects the reliability of the assessment results, taking into account registration quality parameters, feature extraction coverage parameters, and feature change consistency indicators.

[0067] The registration quality parameter, derived from the multi-temporal CT image registration engine, reflects the spatial alignment accuracy of the registered image pairs. Higher registration quality leads to more accurate tumor measurement and feature extraction based on the registered image pairs, resulting in higher reliability of the evaluation results. The feature extraction coverage parameter characterizes the completeness of radiomics feature extraction. If tumor boundaries are unclear or features cannot be extracted from some tumor regions due to artifacts, the coverage parameter will be low, and the reliability of the evaluation results will correspondingly decrease.

[0068] The consistency index of characteristic changes characterizes the degree of consistency in the changing trends of multiple radiomics characteristics. For the same type of treatment response, the changes in different radiomics characteristics should show a consistent trend. For example, for pseudoprogression, multiple heterogeneous related characteristics (such as entropy, homogeneity, and standard deviation) should simultaneously show a decreasing trend. Module 3 calculates the consistency index of characteristic changes as follows: First, based on expert knowledge or statistical analysis, a set of key characteristics related to a specific response type is determined. Then, the Pearson correlation coefficient of the Delta values ​​of these key characteristics is calculated. The absolute value of the correlation coefficient reflects the consistency of characteristic changes. If the absolute value of the correlation coefficient is greater than 0.7, it indicates a high degree of consistency in characteristic changes.

[0069] Module 3 is based on registration quality parameters Feature extraction coverage parameters Consistency index of feature changes The confidence level of the response was determined by weighted fusion. :

[0070] ,

[0071] in, , and For the weighting coefficients, satisfying In a preferred embodiment of the present invention, , , The confidence level ranges from 0 to 1, with a higher value indicating a more reliable assessment result.

[0072] If the response confidence level is lower than a preset threshold (preferably 0.7), module 3 generates a registration optimization instruction, which includes instructions to adjust the image resampling precision or switch the registration strategy. The registration optimization instruction is fed back to the multi-temporal CT image registration engine 1, instructing it to re-execute the registration process or adopt a higher-level registration strategy. For example, if affine registration was initially used, the registration optimization instruction indicates switching to non-rigid registration; if non-rigid registration was already used, the instruction indicates increasing the number of iterations or adjusting the regularization parameters. Registration engine 1 adjusts the registration parameters according to the registration optimization instruction, generating new registered image pairs. The radiomics feature extraction module 2 re-extracts features based on the new registered image pairs, and the immune response assessment module 3 re-evaluates the response. This feedback optimization mechanism forms a closed loop between assessment quality and registration quality, ensuring the reliability of the assessment results.

[0073] Reference Figure 5 , Figure 5 This is a schematic diagram of the structure of the efficacy prediction and treatment plan optimization module 4 of the present invention. The efficacy prediction and treatment plan optimization module 4 receives the immune response assessment results and radiomics feature set from the immune response assessment module 3. Module 4 inputs the radiomics feature set, immune response assessment results, and patient clinical characteristics into the pre-trained efficacy prediction model.

[0074] Patient clinical characteristics include age, sex, smoking history, pathological type, tumor stage, PD-L1 expression level, and gene mutation status. These clinical characteristics are correlated with the efficacy of immunotherapy. For example, PD-L1 expression level is a commonly used biomarker for predicting the efficacy of immunotherapy in clinical practice; patients with high PD-L1 expression levels generally respond better to immunotherapy. Gene mutation status, such as EGFR mutations and ALK fusions, is also associated with the efficacy of immunotherapy. Module 4 fuses these clinical characteristics with a set of radiomics features to form a multimodal feature vector, which is then input into the efficacy prediction model.

[0075] The efficacy prediction model employs an ensemble learning architecture, which includes multiple base learners. These base learners utilize classic machine learning algorithms such as support vector machines, random forests, and gradient boosting decision trees. Each base learner learns and predicts features from a different perspective, complementing its own strengths and weaknesses.

[0076] Support Vector Machines (SVMs) separate different treatment response categories by finding the optimal classification hyperplane, exhibiting good generalization ability and the ability to handle high-dimensional data. In embodiments of this invention, the SVM uses a radial basis function kernel, and the kernel function parameters and penalty parameters are determined through grid search and cross-validation. Random Forests, by constructing multiple decision trees and combining their prediction results, are robust to missing features and outliers. In embodiments of this invention, the random forest contains 100 to 500 decision trees, with a maximum depth of 10 to 20 for each tree. Gradient Boosting Decision Trees (GPRS) train multiple weak learners sequentially, with each new learner trained on the prediction residuals of the previous learner, resulting in high prediction accuracy. In embodiments of this invention, the GPRS is implemented using the XGBoost algorithm, with a learning rate of 0.01 to 0.1 and 100 to 300 iterations.

[0077] Each base learner outputs a predicted probability value, and the efficacy score is obtained by weighted averaging of multiple predicted probabilities. The formula for calculating the efficacy score is:

[0078] ,

[0079] in, To score the treatment efficacy, The number of base learners, For the first The predicted probability values ​​of each base learner For the first The weighting coefficients of each base learner satisfy the following condition: The weighting coefficients are determined based on the performance metrics of each base learner on the validation set; base learners with better performance have larger weights. In a preferred embodiment of the invention, the weight of the support vector machine is 0.35, the weight of the random forest is 0.30, and the weight of the gradient boosting decision tree is 0.35.

[0080] The efficacy score ranges from 0 to 1, with higher values ​​indicating a better response to immunotherapy and more significant treatment benefits. The efficacy prediction model also outputs survival predictions, including progression-free survival and overall survival. Survival prediction uses a Cox proportional hazards regression model, which can handle censoring of survival data and predict the patient's survival function and median survival.

[0081] Module 4 generates treatment optimization suggestions based on efficacy scores, predicted survival values, and immune response assessment results. These suggestions include recommendations for continuing immunotherapy, dosage adjustments, and combination therapies. The specific decision-making logic is as follows:

[0082] First, if the efficacy score is above a first threshold (preferably 0.7) and the immune response assessment indicates pseudoprogression, a recommendation to continue immunotherapy is generated. Pseudoprogression is a benign response pattern to immunotherapy; although the tumor temporarily increases in size, it subsequently shrinks. For patients with pseudoprogression, premature termination of immunotherapy will result in the loss of treatment benefits. Therefore, when the efficacy score predicts a high probability of treatment response and the immune response assessment identifies pseudoprogression, the system recommends continuing the current immunotherapy regimen and performing a follow-up imaging assessment in 4 to 8 weeks to confirm pseudoprogression.

[0083] Second, if the efficacy score is below the second threshold (preferably 0.3) or the immune response assessment result is hyperprogression, treatment regimen adjustment suggestions are generated. A low efficacy score indicates that the patient is not responding well to the current immunotherapy regimen, and continuing treatment may be ineffective or even harmful. Hyperprogression is an adverse reaction pattern of immunotherapy, characterized by a significantly accelerated tumor growth rate and an extremely poor patient prognosis. For patients with low efficacy scores or hyperprogression, the system recommends adjustments to the treatment regimen, including discontinuing the current immunotherapy, switching to other immunotherapy drugs, or switching to alternative regimens such as chemotherapy.

[0084] Third, if the efficacy score is at a moderate level (between the second and first thresholds), a dose adjustment suggestion or combination therapy suggestion is generated. For patients with moderate efficacy scores, the current immunotherapy regimen has some effect but is not ideal. The system suggests adjusting the dosage of immunotherapy drugs based on the patient's specific situation, or suggests a combination regimen of immunotherapy with chemotherapy, targeted therapy, or anti-angiogenic therapy. Combination therapy can exert a synergistic effect and improve treatment efficacy. In embodiments of the present invention, the system incorporates efficacy data of various combination therapy regimens and can recommend the most suitable combination regimen based on the patient's clinical and radiomic characteristics.

[0085] Module 4 feeds back the efficacy score to Module 3 for immune response assessment, enabling dynamic adjustment of assessment threshold parameters. These threshold parameters include various thresholds used to differentiate between pseudoprogression, true progression, and hyperprogression, such as the percentage change in tumor burden threshold, the tumor growth rate ratio threshold, and the heterogeneity index change threshold. These threshold parameters are set based on literature reports and expert experience during system initialization; however, in actual clinical applications, the optimal thresholds may differ for different patient groups.

[0086] Module 4 collects efficacy scores and actual treatment outcome data from multiple patients. Actual treatment outcome data includes patients' imaging assessments and survival status during subsequent follow-ups. Module 4 calculates correlation indicators between efficacy scores and actual treatment outcomes, including predictive accuracy, sensitivity, specificity, positive predictive value, and negative predictive value. If the correlation indicators show that the predictive accuracy is lower than the target accuracy (preferred value is 85%), it indicates that the current assessment threshold parameters are not accurate enough and need to be recalibrated.

[0087] Module 4 uses receiver operating characteristic (ROC) curve analysis to determine the optimal assessment threshold parameters based on actual treatment outcome data. The ROC curve plots classification performance at different thresholds with the false positive rate on the horizontal axis and the true positive rate on the vertical axis. A larger area under the curve indicates better classification performance. Module 4 selects the threshold that maximizes the Youden index (true positive rate + true negative rate - 1) as the optimal threshold. For example, for the identification of false progression, Module 4 redetermines the threshold for heterogeneity index changes based on accumulated Delta-radiomics data from patients with false and true progression. If the initial threshold is -0.3, the recalibrated optimal threshold might be -0.25 or -0.35, maximizing the accuracy of false progression identification.

[0088] Module 4 feeds back the recalibrated assessment threshold parameters to Module 3, enabling adaptive optimization of the assessment criteria. Module 3 uses the new threshold parameters to assess subsequent patients, improving the accuracy of the assessment. This feedback mechanism forms a closed loop between efficacy prediction and immune response assessment, allowing the system to continuously optimize and improve based on actual clinical data, adapting to the needs of different patient groups and clinical scenarios.

[0089] The core innovation of the intelligent analysis system for evaluating and predicting the efficacy of lung cancer immunotherapy in this invention lies in the construction of a deep coupling closed-loop collaborative mechanism among multiple modules. There is not only positive data flow between the modules, but also reverse feedback control, forming a complete closed loop of "registration-extraction-evaluation-prediction-feedback optimization".

[0090] The first closed loop: a coupling closed loop between registration quality and feature extraction accuracy. The registration quality parameters of the multi-temporal CT image registration engine 1 directly affect the feature extraction accuracy of the radiomics feature extraction module 2. The higher the registration quality, the more accurate the spatial alignment of the registered image pairs, and the more precise the tumor measurement and feature extraction based on the registered image pairs. The registration quality parameters are passed to the feature extraction module 2, which adjusts the uncertainty assessment of feature extraction according to the registration quality parameters, assigning lower confidence weights to features in low-quality registered regions.

[0091] The second closed loop: a feedback loop of evaluation confidence and registration optimization. The response confidence of the immune response evaluation module 3 comprehensively reflects the registration quality, feature extraction coverage, and consistency of feature changes. When the response confidence is lower than a preset threshold, it indicates that the current registration accuracy is insufficient to support reliable evaluation results. The evaluation module 3 feeds back registration optimization instructions to the registration engine 1, instructing it to adjust the registration strategy or improve the registration accuracy. The registration engine 1 responds to this instruction, re-executes the registration process, and generates higher-quality registered image pairs. This feedback mechanism ensures that the registration quality meets the accuracy requirements of the evaluation analysis, forming a closed loop where evaluation quality drives registration optimization.

[0092] The third closed loop: an adaptive closed loop for efficacy prediction and evaluation criteria. The efficacy score of the efficacy prediction and regimen optimization module 4 reflects the actual degree of patient response to immunotherapy. By collecting efficacy scores and actual treatment outcome data from multiple patients, module 4 can assess the evaluation accuracy of the immune response evaluation module 3. When the evaluation accuracy is found to be lower than the target value, module 4 recalibrates the evaluation threshold parameters based on the actual treatment outcome data and feeds the new threshold parameters back to evaluation module 3. Evaluation module 3 uses the new threshold parameters to evaluate subsequent patients, improving the accuracy of the evaluation. This feedback mechanism forms a closed loop where efficacy prediction results drive the optimization of evaluation criteria, enabling the system to adaptively adjust according to actual clinical data and achieve continuous improvement in evaluation accuracy.

[0093] These three closed-loop mechanisms interact to form a system-level synergistic optimization. Improved registration quality directly improves feature extraction accuracy, which in turn increases assessment confidence. Increased assessment confidence reduces the need for feedback on registration optimization, thus lowering the system's computational overhead. Simultaneously, feedback from efficacy prediction results optimizes the assessment criteria, which in turn improves assessment accuracy. Improved assessment accuracy makes efficacy predictions more reliable, creating a virtuous cycle of positive reinforcement.

[0094] Furthermore, deep coupling exists between modules at the parameter, state, and logic levels. Parameter-level coupling is reflected in the transfer and use of registration quality parameters, image resampling parameters, and registration accuracy parameters between modules. The output parameters of one module directly serve as key input parameters for the next, influencing the processing strategy and results of the latter. State-level coupling is reflected in the registration states (rigid / affine / non-rigid) of registration engine 1, the feature selection states of feature extraction module 2, and the evaluation threshold states of evaluation module 3. These states are dynamically adjusted during feedback optimization, and the state changes of each module are interconnected. Logic-level coupling is reflected in the fact that the evaluation logic of evaluation module 3 depends on the feature types of feature extraction module 2 and the registration strategy of registration engine 1, while the decision logic of prediction module 4 comprehensively considers the evaluation results of evaluation module 3 and the feature quality of feature extraction module 2.

[0095] Through this deeply coupled closed-loop collaborative mechanism, the system of this invention achieves the following synergistic effects: First, a mutually reinforcing effect: improved registration quality promotes improved feature extraction accuracy, improved feature extraction accuracy promotes improved assessment accuracy, and improved assessment accuracy promotes enhanced reliability of efficacy prediction; Second, a synergistic effect: the overall performance generated by the collaborative work of multiple modules exceeds the simple sum of the performance of each module working independently, and the system's efficacy prediction accuracy reaches over 85%, significantly higher than the performance of a single module or simple combination; Third, a contradiction-resolving effect: the contradiction between registration accuracy and computational efficiency is balanced through a feedback mechanism, triggering high-precision registration only when the assessment confidence is low, ensuring necessary registration quality while avoiding unnecessary computational overhead; Fourth, an adaptive adjustment effect: the system can adaptively adjust the assessment threshold parameters and registration strategy according to actual clinical data, adapting to the needs of different patient groups and clinical scenarios, and achieving continuous performance optimization.

[0096] The efficacy prediction model of this invention is trained as follows: First, a training sample set is collected, including multi-temporal CT images, radiomics features, immune response assessment results, and treatment outcome data. The training sample set comes from lung cancer immunotherapy patients from multiple medical centers, with a sample size preferably exceeding 500 cases to ensure the model's generalization ability. Treatment outcome data includes patients' imaging assessment results, progression-free survival, and overall survival at different time points after treatment.

[0097] Secondly, data augmentation is performed on the training sample set. Due to the low incidence of pseudoprogression and hyperprogression, the number of these two classes in the training sample set is significantly less than that of samples with stable disease and partial remission, resulting in class imbalance. Data augmentation includes synthetic minority oversampling and adaptive synthetic sampling. Synthetic minority oversampling generates synthetic samples for minority class samples by interpolating between minority class samples to increase the number of minority class samples. Adaptive synthetic sampling, building upon synthetic minority oversampling, performs more synthesis for samples that are difficult to classify, improving the model's ability to identify boundary samples. Through data augmentation, the number of samples in each class in the training sample set tends to be balanced, preventing the model from being biased towards the majority class.

[0098] Next, cross-validation is used to divide the training sample set into training and validation subsets. Five-fold cross-validation is employed, randomly dividing the training sample set into five subsets. Each time, four subsets are used as the training subset, and the remaining subset is used as the validation subset. This process is repeated five times, ensuring that each subset serves as a validation subset. Cross-validation effectively utilizes limited training data and improves the reliability of model performance evaluation.

[0099] Next, hyperparameter optimization is performed for each base learner. The hyperparameter optimization employs the Bayesian optimization algorithm, which intelligently searches for the optimal combination of hyperparameters in the hyperparameter space by constructing a Gaussian process surrogate model within the hyperparameter space. Compared to grid search and random search, Bayesian optimization can find near-optimal hyperparameters in fewer iterations, significantly improving optimization efficiency. For Support Vector Machines, the optimized hyperparameters include the kernel function parameters and penalty parameters; for Random Forests, the optimized hyperparameters include the number of decision trees, the maximum depth, and the minimum number of split samples; for Gradient Boosting Decision Trees, the optimized hyperparameters include the learning rate, the number of iterations, the maximum depth, and the regularization parameter.

[0100] Then, each base learner is trained using a training subset. The training process uses standard training algorithms for each base learner: Support Vector Machines use sequential minimum optimization, Random Forests use bootstrap aggregation and random feature selection, and Gradient Boosting Decision Trees use gradient boosting and second-order Taylor expansion. After training, the performance metrics of each base learner are evaluated on a validation subset. Performance metrics include the area under the receiver operating characteristic (AUC), F1 score, and calibration curve. The AUC reflects the overall classification performance of the model, ranging from 0.5 to 1.0, with higher values ​​indicating better performance. The F1 score is the harmonic mean of precision and recall, comprehensively reflecting the model's ability to identify positive samples. The calibration curve reflects the consistency between predicted probabilities and actual occurrence rates; the diagonal line represents perfect calibration.

[0101] Finally, the weighting coefficients for each base learner are determined based on performance metrics, and a therapeutic efficacy prediction model is constructed. The weighting coefficients are determined as follows: first, the area under the receiver operating characteristic (ROC) curve for each base learner on the validation set is calculated, denoted as _{area}_. Then calculate the normalized weighting coefficients:

[0102] ,

[0103] in, For the first The weighting coefficients of each base learner The total number of base learners. Base learners with better performance receive higher weights and play a greater role in ensemble prediction. The efficacy prediction model is finally evaluated on an independent test set containing patient data from different medical centers to ensure the model's generalization ability in real-world clinical scenarios. In embodiments of the present invention, the efficacy prediction model achieves an area under the receiver operating characteristic (ROC) curve of 0.88 to 0.92 on the test set and a prediction accuracy of 85% to 89%, significantly outperforming single models and traditional clinical assessment methods.

[0104] The system of this invention also includes a data storage module and a visualization module. The data storage module stores intermediate data and final results such as registered image pairs, radiomics feature sets, immune response assessment results, and efficacy scores, facilitating data management, retrieval, and retrospective analysis. The data storage module employs a combination of relational database and file system storage. Structured data such as patient information, assessment results, and efficacy scores are stored in the relational database, while image data and radiomics feature files are stored in the file system. The database records the storage paths of the files, enabling data association and indexing.

[0105] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can be modified and varied in various ways. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent analysis system for evaluating and predicting the efficacy of lung cancer immunotherapy, characterized in that, include: A multi-temporal CT image registration engine is used to acquire baseline CT images and follow-up CT images of lung cancer patients before treatment and after treatment. The engine performs spatial registration on the baseline CT images and the follow-up CT images to generate registered image pairs. Based on the registration quality parameters of the registered image pairs, image resampling parameters and registration accuracy parameters are determined. A radiomics feature extraction module, connected to the multi-temporal CT image registration engine, is used to perform image normalization processing on the registered image pairs based on the image resampling parameters, identify the tumor region and the peritumoral region in the registered image pairs respectively, extract a set of radiomics features from the tumor region and the peritumoral region, the set of radiomics features includes shape features, first-order statistical features, and texture features, and calculate Delta-radiomics features based on the set of radiomics features, the Delta-radiomics features representing the dynamic changes in radiomics features before and after treatment; An immune response assessment module, connected to the radiomics feature extraction module, is used to assess the type of immunotherapy response in lung cancer patients based on the Delta-radiomics features and the registration accuracy parameters using an immune-related response standard algorithm. The immunotherapy response types include pseudoprogression, true progression, and hyperprogression. The module generates an immune response assessment result, which includes a response type identifier and a response confidence score. Based on the response confidence score, the module feeds back registration optimization instructions to the multi-temporal CT image registration engine. The efficacy prediction and treatment plan optimization module, connected to the immune response assessment module, receives the immune response assessment results and the radiomics feature set, inputs the radiomics feature set into a pre-trained efficacy prediction model, the efficacy prediction model outputs an efficacy score and a survival prediction value, and generates treatment plan optimization suggestions based on the efficacy score and the survival prediction value. The treatment plan optimization suggestions include suggestions for continuing immunotherapy, dose adjustment, and combination therapy. The efficacy score is fed back to the immune response assessment module for dynamically adjusting the assessment threshold parameters.

2. The intelligent analysis system for evaluating and predicting the efficacy of lung cancer immunotherapy according to claim 1, characterized in that, The multi-temporal CT image registration engine includes: The image preprocessing unit is used to perform noise reduction and contrast enhancement processing on the baseline CT images and the follow-up CT images to generate preprocessed images; The registration strategy selection unit is used to select a registration strategy from rigid registration, affine registration, and non-rigid registration based on the time interval between the baseline CT image and the follow-up CT image and the magnitude of the tumor location change; A registration execution unit is configured to perform spatial registration on the preprocessed image according to the registration strategy, and calculate a transformation matrix to generate the registered image pair; The registration quality assessment unit is used to calculate the mutual information value and normalized cross-correlation coefficient of the registered image pair, and determine the registration quality parameters based on the mutual information value and the normalized cross-correlation coefficient.

3. The intelligent analysis system for evaluating and predicting the efficacy of lung cancer immunotherapy according to claim 1 or 2, characterized in that, The radiomics feature extraction module extracts the radiomics feature set as follows: A three-dimensional convolutional neural network is used to perform semantic segmentation on the tumor region to determine the tumor boundary. The peritumoral region is then defined by extending 5mm to 15mm outward from the tumor boundary. From the tumor region, 851 to 1219 intratumoral radiomic features were extracted, including 18 first-order statistical features, 14 shape features, and multiple texture features; Peritumoral radiomics features were extracted from the peritumoral region, and these features were used to characterize the heterogeneity of the tumor microenvironment. The radiomics features are combined with the intratumoral radiomics features to form the radiomics feature set.

4. The intelligent analysis system for evaluating and predicting the efficacy of lung cancer immunotherapy according to claim 1, characterized in that, The radiomics feature extraction module calculates Delta-radiomics features as follows: The baseline feature vector of the baseline CT image and the follow-up feature vector of the follow-up CT image are obtained respectively; Normalize each feature value in the baseline feature vector and the follow-up feature vector; Calculate the difference between the follow-up feature vector and the baseline feature vector to generate a difference feature vector; Calculate the ratio of the follow-up feature vector to the baseline feature vector to generate a ratio feature vector; The difference feature vector and the ratio feature vector are combined to form the Delta-radiomics feature.

5. The intelligent analysis system for evaluating and predicting the efficacy of lung cancer immunotherapy according to claim 1, characterized in that, When the immune response assessment module evaluates using the standard algorithm for immune-related responses: Based on the registered images, the maximum diameter of the target tumor lesion is measured, and the sum of the diameters of the target lesions is calculated as the baseline value of the tumor burden; At the follow-up time point after treatment, the sum of the diameters of the target lesions was measured again, and the percentage change relative to the baseline value of the tumor burden was calculated; If the percentage change indicates an increase in tumor burden exceeding 25%, and the Delta-radiomics profile shows a decrease in intratumoral heterogeneity, then it is identified as pseudoprogression; If the percentage change indicates an increase in tumor burden exceeding 50% and a tumor growth rate that is more than twice the rate before treatment, it is identified as hyperprogression; If the percentage change indicates an increase in tumor burden of more than 20% but does not meet the criteria for pseudoprogression or hyperprogression, it is identified as true progression.

6. The intelligent analysis system for evaluating and predicting the efficacy of lung cancer immunotherapy according to claim 5, characterized in that, When calculating the confidence level of the immune response, the immune response assessment module: Obtain the registration quality parameters and feature extraction coverage parameters; A consistency index of feature changes is calculated based on the Delta-radiomics characteristics, and the consistency index of feature changes characterizes the degree of consistency in the changing trends of multiple radiomics characteristics; The response confidence level is determined by weighted fusion based on the registration quality parameters, the feature extraction coverage parameters, and the feature change consistency index. If the confidence level of the response is lower than a preset threshold, the registration optimization instruction is generated. The registration optimization instruction includes instruction information for adjusting the image resampling accuracy or switching the registration strategy.

7. The intelligent analysis system for evaluating and predicting the efficacy of lung cancer immunotherapy according to claim 1, characterized in that, When generating treatment plan optimization suggestions, the efficacy prediction and treatment plan optimization module: The radiomics feature set, the immune response assessment results, and the patient's clinical characteristics are input into the efficacy prediction model; The efficacy prediction model employs an ensemble learning architecture, which includes multiple base learners, such as support vector machines, random forests, and gradient boosting decision trees. Each base learner outputs a predicted probability value, and the efficacy score is obtained by weighted averaging of multiple predicted probability values. Based on the efficacy score and the immune response assessment results, if the efficacy score is higher than a first threshold and the immune response assessment results indicate false progression, a recommendation to continue the immunotherapy is generated; if the efficacy score is lower than a second threshold or the immune response assessment results indicate hyperprogression, a recommendation to adjust the treatment plan is generated.

8. The intelligent analysis system for evaluating and predicting the efficacy of lung cancer immunotherapy according to claim 7, characterized in that, The efficacy prediction model was trained in the following way: A training sample set containing multi-temporal CT images, radiomics features, immune response assessment results, and treatment outcome data was collected. The training sample set is subjected to data augmentation processing, which includes synthetic minority class oversampling and adaptive synthetic sampling; The training sample set is divided into a training subset and a validation subset using cross-validation. Hyperparameter optimization is performed for each of the base learners using a Bayesian optimization algorithm. The base learner is trained using the training subset and its performance metrics are evaluated on the validation subset, including the area under the receiver operating characteristic (ROC) curve, F1 score, and calibration curve. The weighting coefficients of each base learner are determined based on the performance metrics, and the efficacy prediction model is constructed.

9. The intelligent analysis system for evaluating and predicting the efficacy of lung cancer immunotherapy according to claim 1, characterized in that, The system also includes: The data storage module is used to store the registered image pairs, the radiomics feature set, the immune response assessment results, and the efficacy score; The visualization module generates a visual interface that includes 3D reconstructed tumor images, radiomics feature heatmaps, efficacy prediction curves, and reports on treatment optimization recommendations.

10. The intelligent analysis system for evaluating and predicting the efficacy of lung cancer immunotherapy according to claim 1, characterized in that, When the efficacy prediction and treatment optimization module dynamically adjusts the evaluation threshold parameters: Collect efficacy scores and actual treatment outcomes data from multiple patients; Calculate the correlation index between the efficacy score and the actual treatment outcome; If the correlation index shows that the prediction accuracy is lower than the target accuracy, the threshold parameter used to distinguish between pseudoprogression, true progression, and hyperprogression in the immune response assessment module is recalibrated based on the actual treatment outcome data; The recalibrated threshold parameters are fed back to the immune response assessment module to achieve adaptive optimization of the assessment criteria.

Images