Lung adenocarcinoma EGFR-TKIs curative effect intelligent prediction system and method

Abstract

The invention provides a lung adenocarcinoma EGFR-TKIs curative effect intelligent prediction system and method, and the method comprises the steps: excluding and screening non-small cell lung cancer pathological images through early-stage automatic RECIST evaluation, and obtaining a pathological image set after RECIST exclusion; scanning a pathology image set after RECIST sorting by using a same-mode fixed amplification galvanometer scanning module; performing model training parameter learning, model optimization parameter fine tuning and model test identification generalization on the basic convolutional neural network ResNet model to obtain a training generalization convolutional neural network ResNet model; and performing lung adenocarcinoma EGFR-TKIs curative effect intelligent prediction on the non-small cell lung cancer pathological image according to a training generalization convolutional neural network ResNet model by using the image block level ORR objective remission rate thermodynamic diagram and the image block level ORR objective remission rate average value calculation result.

Description

Technical Field

[0001] The present invention relates to the technical field of precise imaging EGFR-TKIs intelligent prediction, and more specifically, to a system and method for intelligent prediction of the efficacy of EGFR-TKIs for lung adenocarcinoma. Background Art

[0002] At present, lung cancer is one of the malignant tumors with the highest morbidity and mortality in the world, among which non-small cell lung cancer is the most common pathological type of lung cancer. Targeted therapy targeting epidermal growth factor receptor (EGFR) is of great significance in the history of lung adenocarcinoma treatment. Research on targeted drugs for EGFR mutations (Epidermal growth factor-receptor tyrosine kinase inhibitors, EGFR-TKIs) continues to emerge. However, even for patients with non-small cell lung cancer with EGFR sensitive mutations, the objective response rate (Objective Response Rate, ORR) can only reach about 40-70% after choosing EGFR-TKIs treatment. Whether it is possible to effectively predict and screen patients with poor efficacy of EGFR-TKIs and adjust the treatment plan in advance is a very critical technical issue.

[0003] For advanced non-small cell lung cancer with EGFR sensitive mutations, EGFR-TKIs targeted therapy is usually performed first. After two cycles of treatment, patients undergo imaging RECIST evaluation to determine complete response (CR), partial response (PR), stable disease (SD), and progressive disease (PD). disease (PD), and determine the next treatment plan based on the evaluation results; the problems existing in the existing technology are: whether it is possible to predict the part of non-small cell lung cancer with poor efficacy of EGFR-TKIs, adjust the treatment plan in advance, optimize the treatment process, and bring overall benefits to patients; avoid losing several months of time waiting for the evaluation of EGFR-TKIs efficacy, resulting in patients losing other treatment opportunities; the efficacy of EGFR-TKIs is related to the mutation status of patients with non-small cell lung cancer, and how to make effective predictions in advance through artificial intelligence deep learning needs to be solved; a large number of subsequent studies after gefitinib was launched confirmed that the efficacy of EGFR-TKIs is mainly related to the EGFR mutation status; the sensitive sites of EGFR mainly include 19Del and L858R, and this type of patients can usually be treated with EGFR-TKIs; although plasma can be used for EGFR dynamic detection or supplementary means, the current guidelines still recommend tissue specimens for EGFR detection first, and the technologies used are mainly polymerase chain reaction (PCR) or next generation sequencing (NGS); the problems encountered in molecular detection of tissue specimens include: if only one wax block is selected for detection, whether there is a problem of tumor heterogeneity;

[0004] In recent years, artificial intelligence deep learning convolutional neural network (CNN) technology has been increasingly used in counting cell numbers, counting mitosis, judging segmental colon glandular atypia, grading breast invasive ductal carcinoma, and lymph node metastasis. In lung cancer, the accuracy of the previous EGFR prediction model was low, and it was difficult to meet the requirements of diagnosis and treatment applications by simply predicting EGFR. The efficacy of EGFR-TKIs is related to multiple factors. The efficacy of targeted therapy may be related to mutation abundance, tumor mutation load, TP53 status, PD-L1 expression level, etc., and the results are often controversial. In addition, studies have explored the morphological and pathological characteristics of lung adenocarcinoma, such as tumor atypia, degree of differentiation, airway dissemination and vascular tumor thrombus, lymphocyte infiltration, etc. Some Research suggests that there is a certain correlation with efficacy; however, most previous efficacy prediction studies used statistical models, and the interpretation process of pathologists is not only time-consuming, but also highly subjective, making it difficult to translate into clinical applications; mainly for the prognosis of non-small cell lung cancer, it is believed that the deep learning method can be used to divide patients into long survival group and short survival group through HE staining sections; in terms of tissue morphology, it is feasible to predict the efficacy or prognosis of patients with non-small cell lung cancer through artificial intelligence methods, but the setting factors are too single and need more optimization, and the results are less accurate and other problems need to be solved; therefore, it is necessary to propose an intelligent prediction system and method for the efficacy of EGFR-TKIs in lung adenocarcinoma to at least partially solve the problems existing in the prior art. Summary of the invention

[0005] A series of simplified concepts are introduced in the content of the invention, which will be further described in detail in the specific implementation method section; the content of the invention of the present invention does not mean to attempt to limit the key features and essential technical features of the technical solution claimed for protection, nor does it mean to attempt to determine the scope of protection of the technical solution claimed for protection.

[0006] In order to at least partially solve the above problems, the present invention provides an intelligent prediction system for the efficacy of EGFR-TKIs for lung adenocarcinoma, comprising:

[0007] The RECIST same-mode magnification scanning subsystem excludes and screens non-small cell lung cancer pathology images through the previous automatic RECIST evaluation, and obtains the RECIST excluded pathology image set; the same-mode fixed magnification galvanometer scanning module scans the RECIST excluded pathology image set, and obtains the non-small cell lung cancer pathology image RECIST excluded scanning data set;

[0008] The internal and external data set acquisition subsystem groups the scanned data set after RECIST selection of pathological images into data set training, verification and test groups according to the set ratio of the data set, and obtains the internal training, verification and test data set after RECIST selection; and includes the case data set of the TCIA database as the external test set of TCIA data;

[0009] The ResNet model training generalization subsystem performs model training parameter learning, model optimization parameter fine-tuning, and model test recognition generalization on the basic convolutional neural network ResNet model based on the internal training, testing, and testing data set selected by RECIST to obtain a training generalized convolutional neural network ResNet model.

[0010] The EGFR-TKIs intelligent prediction subsystem uses the image block-level ORR objective response rate heat map and the image block-level ORR objective response rate average calculation results to train a generalized convolutional neural network ResNet model to intelligently predict the efficacy of EGFR-TKIs in lung adenocarcinoma based on non-small cell lung cancer pathological images.

[0011] Preferably, the RECIST isotype amplification scanning subsystem includes:

[0012] EGFR lung cancer pathology image acquisition subsystem, which acquires pathology images of non-small cell lung cancer with EGFR sensitive mutations and treated with EGFR-TKIs;

[0013] The automatic RECIST evaluation and exclusion subsystem, based on the pathological images of non-small cell lung cancer, excludes and screens the pathological images of non-small cell lung cancer through the previous automatic RECIST evaluation, and obtains the pathological image set after RECIST exclusion;

[0014] The same-mode fixed magnification scanning subsystem scans each pair of RECIST excluded pathology images in the RECIST excluded pathology image set through the same-mode fixed magnification galvanometer scanning module to obtain a RECIST excluded scanning data set of non-small cell lung cancer pathology images.

[0015] Preferably, the internal and external data set acquisition subsystem includes:

[0016] The data set training, verification and testing grouping subsystem performs data set training, verification and testing grouping on the scan data set after RECIST selection of pathological images according to the data set setting ratio, and obtains the internal training, verification and testing data set after RECIST selection; the internal training, verification and testing data set after RECIST selection includes: the internal training data set after RECIST selection, the internal verification data set after RECIST selection and the internal testing data set after RECIST selection; the data set setting ratio includes: 60%: 20%: 20% or 70%: 15%: 15%;

[0017] The TCIA data acquisition subsystem links to the TCIA database to obtain the TCIA database case data set, and incorporates the TCIA database case data set as the TCIA data external test set.

[0018] Preferably, the ResNet model training generalization subsystem includes:

[0019] ResNet architecture setting subsystem, set the convolutional neural network ResNet architecture as the basic architecture of the system deep learning training model, and obtain the basic convolutional neural network ResNet model;

[0020] The ResNet model optimization subsystem performs model training parameter learning, model optimization parameter fine-tuning, and model test recognition generalization on the basic convolutional neural network ResNet model based on the internal training, testing, and test data set after RECIST selection, and obtains the trained generalized convolutional neural network ResNet model.

[0021] Preferably, the EGFR-TKIs intelligent prediction subsystem includes:

[0022] The slice-level ORR labeling subsystem automatically labels the slice-level labels and obtains the slice-level ORR objective response rate results through the slice-level ORR objective response rate empirical model;

[0023] The intelligent learning model interpretation subsystem uses the TCIA data external test set as input to train the generalized convolutional neural network ResNet model. Through weakly supervised intelligent learning, it uses the slice-level ORR objective response rate results as grouping labels and sets pixels to segment non-small cell lung cancer pathology images to obtain non-small cell lung cancer pathology image blocks. It also screens out all non-small cell lung cancer pathology image blocks containing tumors, obtains the image block-level ORR objective response rate heat map, and interprets the training generalized convolutional neural network ResNet model.

[0024] The ORR calculation efficacy prediction subsystem, based on the interpretation of the trained generalized convolutional neural network ResNet model, intelligently predicts the efficacy of EGFR-TKIs for lung adenocarcinoma on non-small cell lung cancer pathological images; based on the image block level ORR heat map, the average ORR objective response rate at the image block level is calculated to obtain the pixel-level lung adenocarcinoma EGFR-TKIs efficacy prediction ORR objective response rate prediction results.

[0025] A series of simplified concepts are introduced in the content of the invention, which will be further described in detail in the specific implementation method section; the content of the invention of the present invention does not mean to attempt to limit the key features and essential technical features of the technical solution claimed for protection, nor does it mean to attempt to determine the scope of protection of the technical solution claimed for protection.

[0026] In order to at least partially solve the above problems, the present invention provides an intelligent prediction method for the efficacy of EGFR-TKIs in lung adenocarcinoma, comprising:

[0027] S100, excluding and screening non-small cell lung cancer pathological images through the previous automatic RECIST evaluation, obtaining a set of pathological images after RECIST exclusion; scanning the set of pathological images after RECIST exclusion through the same-mode fixed magnifying galvanometer scanning module, obtaining a scanned data set of non-small cell lung cancer pathological images after RECIST exclusion;

[0028] S200, grouping the scanned data set after RECIST selection of pathological images into data set training, verification and test data set according to the set ratio of the data set, and obtaining the internal training, verification and test data set after RECIST selection; and incorporating the case data set of the TCIA database as the external test set of the TCIA data;

[0029] S300, based on the internal training, testing and testing data set selected by RECIST, the basic convolutional neural network ResNet model is trained with model parameters learned, model optimization parameters fine-tuned and model test recognition generalized to obtain a trained generalized convolutional neural network ResNet model;

[0030] S400 uses the block-level ORR objective response rate heat map and the block-level ORR objective response rate average calculation results to train a generalized convolutional neural network ResNet model to intelligently predict the efficacy of lung adenocarcinoma EGFR-TKIs for non-small cell lung cancer pathological images.

[0031] Preferably, S100 includes:

[0032] S101, to obtain pathological images of non-small cell lung cancer with EGFR sensitive mutations and treated with EGFR-TKIs;

[0033] S102, according to the non-small cell lung cancer pathology images, exclude and screen the non-small cell lung cancer pathology images through the previous automatic RECIST evaluation, and obtain a set of pathology images after RECIST exclusion;

[0034] S103, scanning each pair of post-RECIST exclusion pathology images in the post-RECIST exclusion pathology image set by using a fixed magnifying galvanometer scanning module of the same model to obtain a post-RECIST exclusion scanning data set of pathology images of non-small cell lung cancer.

[0035] Preferably, S200 includes:

[0036] S201, grouping the pathological image scan data set after RECIST selection for data set training, verification and testing according to the data set setting ratio, to obtain the RECIST selected internal training, verification and testing data set; the RECIST selected internal training, verification and testing data set includes: a RECIST selected internal training data set, a RECIST selected internal verification data set and a RECIST selected internal testing data set; the data set setting ratio includes: 60%: 20%: 20% or 70%: 15%: 15%;

[0037] S202, linking to the TCIA database to obtain a TCIA database case data set, and incorporating the TCIA database case data set as a TCIA data external test set.

[0038] Preferably, S300 includes:

[0039] S301, setting a convolutional neural network ResNet architecture as a basic architecture of a system deep learning training model, and obtaining a basic convolutional neural network ResNet model;

[0040] S302, based on the internal training, testing and testing data set after RECIST selection, the basic convolutional neural network ResNet model is trained with model parameters, model optimization parameters are fine-tuned, and model testing and recognition generalization is performed to obtain a trained generalized convolutional neural network ResNet model.

[0041] Preferably, S400 includes:

[0042] S401, automatically marking slice-level labels through a slice-level ORR objective response rate empirical model to obtain a slice-level ORR objective response rate result;

[0043] S402: Based on the TCIA data external test set, input the training generalized convolutional neural network ResNet model, use the slice-level ORR objective response rate results as grouping labels through weakly supervised intelligent learning, and use the set pixels to segment the non-small cell lung cancer pathology image to obtain the non-small cell lung cancer pathology image blocks; screen out all non-small cell lung cancer pathology image blocks containing tumors, obtain the image block-level ORR objective response rate heat map, and interpret the training generalized convolutional neural network ResNet model;

[0044] S403, based on the interpretation of the trained generalized convolutional neural network ResNet model, intelligent prediction of the efficacy of EGFR-TKIs for lung adenocarcinoma was performed on the pathological images of non-small cell lung cancer; based on the ORR heat map at the image block level, the average ORR objective response rate at the image block level was calculated to obtain the pixel-level prediction of the ORR objective response rate of lung adenocarcinoma EGFR-TKIs efficacy.

[0045] Compared with the prior art, the present invention has at least the following beneficial effects:

[0046] The present invention discloses an intelligent prediction system and method for the efficacy of EGFR-TKIs for lung adenocarcinoma. The system utilizes a RECIST same-mode magnification scanning subsystem to exclude and screen non-small cell lung cancer pathological images through an early automatic RECIST evaluation to obtain a RECIST excluded pathological image set; the RECIST excluded pathological image set is scanned by a same-mode fixed magnification galvanometer scanning module to obtain a RECIST excluded scanned data set of non-small cell lung cancer pathological images; an internal and external data set acquisition subsystem performs data set training, verification and test grouping on the scanned data set of pathological images after RECIST exclusion according to a data set set ratio to obtain an internal training, verification and test data set after RECIST exclusion; a TCIA database case data set is included as a TCIA data external test set; a ResNet model training generalization subsystem performs a training, verification and test grouping on the base according to the internal training, verification and test data set after RECIST exclusion The convolutional neural network ResNet model performs model training parameter learning, model optimization parameter fine-tuning, and model test identification generalization to obtain a trained generalized convolutional neural network ResNet model; the EGFR-TKIs intelligent prediction subsystem uses the image block-level ORR objective response rate heat map and the image block-level ORR objective response rate average calculation results to perform intelligent prediction of the efficacy of lung adenocarcinoma EGFR-TKIs on non-small cell lung cancer pathological images based on the trained generalized convolutional neural network ResNet model; it can overcome the problem of tumor heterogeneity; the prediction of the efficacy or prognosis setting factors of non-small cell lung cancer patients is enriched and optimized, and the results are more accurate; for advanced non-small cell lung cancer with EGFR sensitive mutations, EGFR-TKIs targeted therapy is performed; after two cycles of treatment, the patient undergoes imaging RECIST evaluation to determine complete remission (complete remission). The system can predict the patients with poor response to EGFR-TKIs, adjust the treatment plan in advance, optimize the treatment process, and bring about the overall prediction of prognosis; avoid losing other treatment opportunities due to the loss of several months of waiting for the evaluation of the efficacy of EGFR-TKIs; automatically perform RECIST evaluation to exclude and screen the pathological images of non-small cell lung cancer for the prediction of non-small cell lung cancer, and scan the pathological image set after RECIST exclusion by the same model fixed magnifying galvanometer, so that the scan data of the pathological image of non-small cell lung cancer after RECIST exclusion has better image features and is more suitable for data training; it can effectively extend the prediction accuracy of sensitive mutations and improve their quality; it can effectively predict and screen out those with poor response to EGFR-TKIs, and predict more effective treatment plans in advance; the grouping of the data set after RECIST exclusion is more accurate, avoiding unreasonable grouping, and the quality of the training, validation and test set data is significantly improved;Including the TCIA database case data set as the TCIA data external test set can greatly improve the overall adaptability of model training; according to the internal training and testing data set after RECIST selection, the basic convolutional neural network ResNet model is trained to learn model training parameters, fine-tune model optimization parameters, and generalize model test identification, significantly improving the generalization ability of the convolutional neural network ResNet model; the image block-level ORR objective response rate heat map and the image block-level ORR objective response rate average calculation results can be effectively used to train the generalized convolutional neural network ResNet model to perform intelligent prediction of lung adenocarcinoma EGFR-TKIs efficacy on non-small cell lung cancer pathological images, significantly improving the prediction accuracy. ;

[0047] The intelligent prediction system and method for the efficacy of EGFR-TKIs for lung adenocarcinoma described in the present invention, and other advantages, objectives and features of the present invention will be partially reflected in the following description, and will also be partially understood by technicians in this field through research and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0049] Figure 1 This is a diagram of an embodiment of the intelligent prediction system architecture for the efficacy of EGFR-TKIs for lung adenocarcinoma described in the present invention.

[0050] Figure 2 This is a diagram showing an embodiment of the framework of an intelligent prediction system for the efficacy of EGFR-TKIs for lung adenocarcinoma described in the present invention.

[0051] Figure 3 This is a diagram showing an implementation step of a method for intelligently predicting the efficacy of EGFR-TKIs for lung adenocarcinoma described in the present invention. DETAILED DESCRIPTION

[0052] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments so that those skilled in the art can implement the invention according to the instructions. Figure 1-Figure 3 As shown, the present invention provides an intelligent prediction system for the efficacy of EGFR-TKIs for lung adenocarcinoma, comprising:

[0053] The RECIST same-mode magnification scanning subsystem excludes and screens non-small cell lung cancer pathology images through the previous automatic RECIST evaluation, and obtains the RECIST excluded pathology image set; the same-mode fixed magnification galvanometer scanning module scans the RECIST excluded pathology image set, and obtains the non-small cell lung cancer pathology image RECIST excluded scanning data set;

[0054] The internal and external data set acquisition subsystem groups the scanned data set after RECIST selection of pathological images into data set training, verification and test groups according to the set ratio of the data set, and obtains the internal training, verification and test data set after RECIST selection; and includes the case data set of the TCIA database as the external test set of TCIA data;

[0055] The ResNet model training generalization subsystem performs model training parameter learning, model optimization parameter fine-tuning, and model test recognition generalization on the basic convolutional neural network ResNet model based on the internal training, testing, and testing data set selected by RECIST to obtain a training generalized convolutional neural network ResNet model.

[0056] The EGFR-TKIs intelligent prediction subsystem uses the image block-level ORR objective response rate heat map and the image block-level ORR objective response rate average calculation results to train a generalized convolutional neural network ResNet model to intelligently predict the efficacy of EGFR-TKIs in lung adenocarcinoma based on non-small cell lung cancer pathological images.

[0057] The principle and effect of the above technical solution are as follows: the present invention provides an intelligent prediction system for the efficacy of EGFR-TKIs in lung adenocarcinoma, including: a RECIST same-mode magnification scanning subsystem, which excludes and screens non-small cell lung cancer pathological images through the previous automatic RECIST evaluation, and obtains a RECIST excluded pathological image set; scans the RECIST excluded pathological image set through the same-mode fixed magnification galvanometer scanning module, and obtains a RECIST excluded scanned data set of non-small cell lung cancer pathological images; an internal and external data set acquisition subsystem, which groups the scanned data set of pathological images after RECIST exclusion according to the data set setting ratio, and obtains a RECIST excluded internal training, verification and test data set; incorporates the TCIA database case data set as the TCIA data external test set; and a ResNet model training generalization subsystem, which performs training, verification and test grouping according to the internal ... The test data set is used to learn the model training parameters, fine-tune the model optimization parameters, and generalize the model test identification for the basic convolutional neural network ResNet model to obtain the trained generalized convolutional neural network ResNet model; the EGFR-TKIs intelligent prediction subsystem uses the image block level ORR objective response rate heat map and the image block level ORR objective response rate average calculation results. According to the trained generalized convolutional neural network ResNet model, the lung adenocarcinoma EGFR-TKIs efficacy is intelligently predicted for non-small cell lung cancer pathological images; the problem of tumor heterogeneity can be overcome; the prediction of the efficacy or prognosis setting factors of non-small cell lung cancer patients is enriched and optimized, and the results are more accurate; EGFR-TKIs targeted therapy is performed for advanced non-small cell lung cancer with EGFR sensitive mutations; after two cycles of treatment, the patient is evaluated by imaging RECIST to determine complete remission (complete remission). The system can predict the disease progression (CR), partial response (PR), stable disease (SD) and progressive disease (PD), and determine the next treatment plan based on the evaluation results; it can predict patients with non-small cell lung cancer who have poor response to EGFR-TKIs, adjust the treatment plan in advance, optimize the treatment process, and bring about an overall predicted prognosis effect; it can avoid losing other treatment opportunities due to the loss of several months of waiting for the evaluation of the efficacy of EGFR-TKIs; it can automatically perform RECIST evaluation to exclude and screen non-small cell lung cancer pathology images for non-small cell lung cancer prediction, and scan the RECIST excluded pathology image set through the same model fixed magnifying galvanometer, so that the scan data of non-small cell lung cancer pathology images after RECIST exclusion has better image features and is more suitable for data training; it can effectively extend the prediction accuracy of sensitive mutations and improve their quality; it can effectively predict and screen out patients with poor response to EGFR-TKIs, and predict more effective treatment plans in advance;The grouping of the data set after RECIST exclusion is more accurate, which avoids unreasonable grouping, and the quality of the training, validation and test set data is significantly improved; the inclusion of the TCIA database case data set as the external test set of TCIA data can greatly improve the overall adaptability of model training; according to the internal training, validation and test data set after RECIST exclusion, the basic convolutional neural network ResNet model is trained to learn model training parameters, fine-tune model optimization parameters and generalize model test identification, which significantly improves the generalization ability of the convolutional neural network ResNet model; the image block-level ORR objective response rate heat map and the image block-level ORR objective response rate average calculation results can be effectively used to train the generalized convolutional neural network ResNet model to perform intelligent prediction of lung adenocarcinoma EGFR-TKIs efficacy on non-small cell lung cancer pathological images, significantly improving the prediction accuracy. ;

[0058] In one embodiment, the RECIST same-mode amplification scanning subsystem includes:

[0059] EGFR lung cancer pathology image acquisition subsystem, which acquires pathology images of non-small cell lung cancer with EGFR sensitive mutations and treated with EGFR-TKIs;

[0060] The automatic RECIST evaluation and exclusion subsystem, based on the pathological images of non-small cell lung cancer, excludes and screens the pathological images of non-small cell lung cancer through the previous automatic RECIST evaluation, and obtains the pathological image set after RECIST exclusion;

[0061] The same-mode fixed magnification scanning subsystem scans each pair of RECIST excluded pathology images in the RECIST excluded pathology image set through the same-mode fixed magnification galvanometer scanning module to obtain a RECIST excluded scanning data set of non-small cell lung cancer pathology images;

[0062] Obtaining a set of pathological images after RECIST exclusion through the previous automatic RECIST evaluation and exclusion screening of non-small cell lung cancer pathological images, including: creating an automatic RECIST evaluation and exclusion screening module according to the RECIST evaluation criteria, automatically evaluating and excluding non-small cell lung cancer pathological images, and obtaining a set of pathological images after RECIST exclusion screening;

[0063] The same-mode fixed magnifying galvanometer scanning module includes: a same-mode fixed platform, a flat-field focusing lens unit, a reflection galvanometer scanning galvanometer unit and a scanning focal length adjustment group; the reflection galvanometer scanning galvanometer unit includes: a servo-driven vibration group and a virtual image mirror group; the servo-driven vibration group drives the virtual image mirror group to vibrate back and forth; the same-mode fixed platform selects non-small cell lung cancer pathological images according to the same mode and position fixed image size; the flat-field focusing lens unit is fixedly installed directly above the same-mode fixed platform; the servo-driven vibration group, the scanning focal length adjustment group and the high-power RGB magnifying scanner are arranged in sequence on a horizontal plane; the servo-driven vibration group is fixedly installed above the flat-field focusing lens unit, and its virtual image mirror group is arranged directly above the center line of the flat-field focusing lens unit and has a changeable inclination angle with the horizontal plane; the center of the virtual image mirror group, the center of the scanning focal length adjustment group and the high-power RGB magnifying scanner are fixedly installed above the flat-field focusing lens unit. The center of the large scanner is on a straight line; scanning each pair of RECIST post-selected pathology images in the RECIST post-selected pathology image set through the same-mode fixed magnification galvanometer scanning module includes: selecting non-small cell lung cancer pathology images according to the same mode and same position fixed image size through the same-mode fixed platform; the same mode and same position fixed image size adopts RGB three-channel 24-bit color image mode, 400-fold fixed magnification and set image size, and each channel of the RGB three-channel is 8 bits; during the scanning process, as the scanning mirror of the high-power RGB magnification scanner moves, the servo-driven vibration group drives the virtual image mirror group to reciprocate and vibrate, generating a micro-scale reciprocating change in the scanning focal length; the scanning focal length adjustment group adjusts to compensate for the micro-scale reciprocating change in the scanning focal length, so as to maintain the RECIST post-selected pathology image scanning with precise and consistent same-mode and same position fixed image size.

[0064] The principle and effect of the above technical solution are as follows: using RECIST same-mode magnification scanning, the EGFR lung cancer pathology image acquisition subsystem obtains pathology images of non-small cell lung cancer with EGFR sensitive mutations and treated with EGFR-TKIs; the automatic RECIST evaluation and exclusion subsystem excludes and screens non-small cell lung cancer pathology images through the previous automatic RECIST evaluation according to the non-small cell lung cancer pathology images, and obtains the RECIST excluded pathology image set; the same-mode fixed magnification scanning subsystem scans each pair of RECIST excluded pathology images in the RECIST excluded pathology image set through the same-mode fixed magnification galvanometer scanning module to obtain the non-small cell lung cancer pathology image RECIST. IST post-selection scanning data set; excluding and screening non-small cell lung cancer pathological images through the previous automatic RECIST evaluation, obtaining the RECIST post-selection pathological image set includes: creating an automatic RECIST evaluation exclusion screening module according to the RECIST evaluation standard, automatically evaluating and excluding non-small cell lung cancer pathological images, and obtaining the RECIST post-selection pathological image set; the same-mode fixed magnifying galvanometer scanning module includes: a same-mode fixed platform, a flat-field focusing lens unit, a reflection galvanometer scanning galvanometer unit and a scanning focus adjustment group; the reflection galvanometer scanning galvanometer unit includes: a servo-driven vibration group and a virtual image mirror group; the servo-driven vibration group drives the virtual image mirror group to reciprocate; the same-mode fixed magnifying galvanometer scanning module includes: a same-mode fixed platform, a flat-field focusing lens unit, a reflection galvanometer scanning galvanometer unit and a scanning focus adjustment group; the reflection galvanometer scanning galvanometer unit includes: a servo-driven vibration group and a virtual image mirror group; the servo-driven vibration group drives the virtual image mirror group to vibrate reciprocatingly; The fixed platform selects non-small cell lung cancer pathological images according to the same mode and position fixed image size; the flat-field focusing lens unit is fixedly installed just above the same-mode fixed platform; the servo-driven vibration group, the scanning focus adjustment group and the high-power RGB magnification scanner are arranged in sequence on a horizontal plane; the servo-driven vibration group is fixedly installed above the flat-field focusing lens unit, and its virtual image mirror group is arranged just above the center line of the flat-field focusing lens unit and has a variable angle of inclination with the horizontal plane; the center of the virtual image mirror group, the center of the scanning focus adjustment group and the center of the high-power RGB magnification scanner are on a straight line; each RECIST pathological image set after RECIST selection is scanned by the same-mode fixed magnification galvanometer scanning module The pathological images after CIST selection include: selecting non-small cell lung cancer pathological images through the same model fixed platform according to the same mode and position fixed image size; the same mode and position fixed image size adopts RGB three-channel 24-bit color image mode, 400 times fixed magnification and set image size, and each channel of the RGB three-channel is 8 bits; during the scanning process, as the scanning mirror of the high-power RGB magnification scanner moves, the servo-driven vibration group drives the virtual image mirror group to vibrate back and forth, resulting in micro-scale reciprocating changes in the scanning focus; the scanning focus adjustment group adjusts to compensate for the micro-scale reciprocating changes in the scanning focus, so as to maintain the precise and consistent same mode and position fixed image size for the pathological image scan after RECIST selection;It can automatically evaluate and exclude non-small cell lung cancer pathology images for non-small cell lung cancer prediction. By scanning the RECIST excluded pathology image set with the same model fixed magnifying galvanometer, the scanned data of non-small cell lung cancer pathology images after RECIST exclusion has better image features and is more suitable for data training. ;

[0065] In one embodiment, the internal and external data set acquisition subsystem includes:

[0066] The data set training, verification and testing grouping subsystem performs data set training, verification and testing grouping on the scan data set after RECIST selection of pathological images according to the data set setting ratio, and obtains the internal training, verification and testing data set after RECIST selection; the internal training, verification and testing data set after RECIST selection includes: the internal training data set after RECIST selection, the internal verification data set after RECIST selection and the internal testing data set after RECIST selection; the data set setting ratio includes: 60%: 20%: 20% or 70%: 15%: 15%;

[0067] The TCIA data acquisition subsystem links to the TCIA database to obtain the TCIA database case data set, and incorporates the TCIA database case data set as the TCIA data external test set.

[0068] The principle and effect of the above technical solution are as follows: by utilizing the internal and external data set acquisition, the data set training, verification and testing grouping subsystem groups the pathological image RECIST-selected scan data set according to the data set setting ratio to perform data set training, verification and testing grouping, and obtain the RECIST-selected internal training, verification and testing data set; the RECIST-selected internal training, verification and testing data set includes: the RECIST-selected internal training data set, the RECIST-selected internal verification data set and the RECIST-selected internal test data set; the data set setting ratio includes: 60%: 20%: 20% or 70%: 15%: 15%; the TCIA data acquisition subsystem links the TCIA database to obtain the TCIA database case data set, and includes the TCIA database case data set as the TCIA data external test set; the RECIST-selected data set is grouped more accurately to avoid unreasonable grouping, and the quality of the training, verification and test set data is significantly improved; including the TCIA database case data set as the TCIA data external test set can greatly improve the comprehensive adaptability of model training.

[0069] In one embodiment, the ResNet model training generalization subsystem includes:

[0070] ResNet architecture setting subsystem, set the convolutional neural network ResNet architecture as the basic architecture of the system deep learning training model, and obtain the basic convolutional neural network ResNet model;

[0071] The ResNet model optimization subsystem performs model training parameter learning, model optimization parameter fine-tuning, and model test recognition generalization on the basic convolutional neural network ResNet model based on the internal training, testing, and test data set after RECIST selection, and obtains the trained generalized convolutional neural network ResNet model.

[0072] The principle and effect of the above technical solution are as follows: using the ResNet model training generalization, the ResNet architecture setting subsystem sets the convolutional neural network ResNet architecture as the basic architecture of the system deep learning training model to obtain the basic convolutional neural network ResNet model; the ResNet model optimization subsystem performs model training parameter learning, model optimization parameter fine-tuning, and model test recognition generalization on the basic convolutional neural network ResNet model according to the internal training, verification, and test data set after RECIST selection to obtain the training generalized convolutional neural network ResNet model; according to the internal training, verification, and test data set after RECIST selection, the basic convolutional neural network ResNet model performs model training parameter learning, model optimization parameter fine-tuning, and model test recognition generalization, significantly improving the generalization ability of the convolutional neural network ResNet model.

[0073] In one embodiment, the EGFR-TKIs intelligent prediction subsystem includes:

[0074] The slice-level ORR labeling subsystem automatically labels the slice-level labels and obtains the slice-level ORR objective response rate results through the slice-level ORR objective response rate empirical model;

[0075] The intelligent learning model interpretation subsystem uses the TCIA data external test set as input to train the generalized convolutional neural network ResNet model. Through weakly supervised intelligent learning, it uses the slice-level ORR objective response rate results as grouping labels and sets pixels to segment non-small cell lung cancer pathology images to obtain non-small cell lung cancer pathology image blocks. It also screens out all non-small cell lung cancer pathology image blocks containing tumors, obtains the image block-level ORR objective response rate heat map, and interprets the training generalized convolutional neural network ResNet model.

[0076] The ORR calculation efficacy prediction subsystem performs intelligent prediction of the efficacy of EGFR-TKIs for lung adenocarcinoma based on the interpretation of the trained generalized convolutional neural network ResNet model for non-small cell lung cancer pathological images; calculates the average ORR objective response rate at the image block level based on the ORR heat map at the image block level, and obtains the ORR objective response rate prediction result of EGFR-TKIs for lung adenocarcinoma at the pixel level;

[0077] The training of the generalized convolutional neural network ResNet model interpretation includes: obtaining the previous EGFR prediction information, differentiation degree identification information, microenvironment parameter learning information, setting the EGFR-TKIs resistance PFS value and mutation targeted drug sensitivity level value of multiple mutation points corresponding to the targeted drug; the previous EGFR prediction information includes: previous EGFR prediction information, differentiation degree identification information, microenvironment parameter learning information; the previous EGFR prediction information includes: extracellular ligand binding region prediction curve, transmembrane region prediction curve, intracellular kinase region prediction curve and EGFR-TKIs resistance progression curve; the horizontal axis of the EGFR-TKIs resistance progression curve is the EGFR-TKIs resistance PFS value, and the vertical axis is the mutation type of the mutation point corresponding to the lesion progression ratio. The mutation types of the mutation point include: 19Del mutation, L858R mutation and other types of mutations; lesion progression includes: dual primary lesion metastasis progression, primary lesion progression and metastatic lesion progression; the targeted drug sensitivity level values ​​of multiple mutation points are divided into multiple distribution intervals, and the intersection of any two distribution intervals is an empty set;

[0078] The sensitivity level values ​​of the targeted drugs of the multiple mutation points and the sensitivity level values ​​of the corresponding multiple mutation points on the EGFR-TKIs resistance progression curve are respectively fitted in the multiple distribution intervals to establish the EGFR-TKIs resistance ResNet interpretation model, including:

[0079] If the previous EGFR prediction information is an extracellular ligand binding region prediction curve, determining the 19Del mutation targeted drug sensitivity level values ​​corresponding to the multiple mutation points according to the extracellular ligand binding region prediction curve;

[0080] In the multiple distribution intervals, the sensitivity level value of the targeted drug of the 19Del mutation is fitted with the sensitivity level value of the targeted drug of multiple mutation points to establish a ResNet interpretation model of EGFR-TKIs resistance;

[0081] If the previous EGFR prediction information is a transmembrane region prediction curve, determining the L858R mutation targeted drug sensitivity level values ​​corresponding to the multiple mutation points according to the transmembrane region prediction curve;

[0082] In the multiple distribution intervals, the sensitivity level value of the L858R mutation targeted drug is fitted with the sensitivity level values ​​of the targeted drug of multiple mutation points to establish a ResNet explanation model of EGFR-TKIs resistance;

[0083] If the previous EGFR prediction information is an intracellular kinase region prediction curve, determining the sensitivity level values ​​of other mutation targeted drugs corresponding to the multiple mutation points according to the intracellular kinase region prediction curve;

[0084] In the multiple distribution intervals, the sensitivity level values ​​of other mutation targeted drugs are fitted with the sensitivity level values ​​of the targeted drugs of multiple mutation points to establish a ResNet explanation model of EGFR-TKIs resistance;

[0085] The targeted drug sensitivity level values ​​of multiple mutation points are calculated according to the EGFR-TKIs resistance ResNet interpretation model, including: if the previous EGFR prediction information includes at least two curves in the EGFR-TKIs resistance progression curve, the sensitivity level is calculated separately according to each established EGFR-TKIs resistance ResNet interpretation model; the average sensitivity level is calculated, and the average sensitivity level is used as the targeted drug sensitivity level of multiple mutation points.

[0086] The principle and effect of the above technical solution are as follows: using EGFR-TKIs intelligent prediction, the slice-level ORR labeling subsystem automatically labels the slice-level labels through the slice-level ORR objective response rate empirical model to obtain the slice-level ORR objective response rate results; the intelligent learning model interpretation subsystem inputs the training generalized convolutional neural network ResNet model based on the TCIA data external test set, uses weakly supervised intelligent learning, uses the slice-level ORR objective response rate results as grouping labels, and sets pixels to perform non-small cell lung cancer pathology image segmentation to obtain non-small cell lung cancer pathology images. Block; filter out all NSCLC pathology image blocks containing tumors, obtain the ORR objective response rate heat map at the image block level, and interpret the training generalized convolutional neural network ResNet model; ORR calculation efficacy prediction subsystem, based on the interpretation of the training generalized convolutional neural network ResNet model, intelligently predict the efficacy of lung adenocarcinoma EGFR-TKIs for NSCLC pathology images; calculate the average ORR objective response rate at the image block level based on the image block level ORR heat map, and obtain the pixel-level lung adenocarcinoma EGFR-TKIs efficacy prediction ORR objective response rate prediction result;

[0087] The training of the generalized convolutional neural network ResNet model interpretation includes: obtaining the previous EGFR prediction information, differentiation degree identification information, microenvironment parameter learning information, setting the EGFR-TKIs resistance PFS value and mutation targeted drug sensitivity level value of multiple mutation points corresponding to the targeted drug; the previous EGFR prediction information includes: previous EGFR prediction information, differentiation degree identification information, microenvironment parameter learning information; the previous EGFR prediction information includes: extracellular ligand binding region prediction curve, transmembrane region prediction curve, intracellular kinase region prediction curve and EGFR-TKIs resistance progression curve; the horizontal axis of the EGFR-TKIs resistance progression curve is the EGFR-TKIs resistance PFS value, and the vertical axis is the mutation type of the mutation point corresponding to the lesion progression ratio. The mutation types of the mutation point include: 19Del mutation, L858R mutation and other types of mutations; lesion progression includes: dual primary lesion metastasis progression, primary lesion progression and metastatic lesion progression; the targeted drug sensitivity level values ​​of multiple mutation points are divided into multiple distribution intervals, and the intersection of any two distribution intervals is an empty set;

[0088] The sensitivity level values ​​of targeted drugs at multiple mutation points are respectively fitted with the sensitivity level values ​​corresponding to the multiple mutation points on the EGFR-TKIs resistance progression curve in multiple distribution intervals to establish an EGFR-TKIs resistance ResNet interpretation model, including: if the previous EGFR prediction information is an extracellular ligand binding region prediction curve, the sensitivity level values ​​of 19Del mutation targeted drugs corresponding to the multiple mutation points are determined according to the extracellular ligand binding region prediction curve; within the multiple distribution intervals, the sensitivity level values ​​of 19Del mutation targeted drugs are fitted with multiple The targeted drug sensitivity level values ​​of the mutation points are fitted to establish the EGFR-TKIs resistance ResNet interpretation model; if the previous EGFR prediction information is a transmembrane region prediction curve, the L858R mutation targeted drug sensitivity level values ​​corresponding to the multiple mutation points are determined according to the transmembrane region prediction curve; within the multiple distribution intervals, the L858R mutation targeted drug sensitivity level value is fitted with the targeted drug sensitivity level values ​​of the multiple mutation points to establish the EGFR-TKIs resistance ResNet interpretation model; if the previous EGFR prediction information is a cell The method comprises the following steps: an intracellular kinase region prediction curve is used to determine the sensitivity level values ​​of other mutation targeted drugs corresponding to the multiple mutation points according to the intracellular kinase region prediction curve; within the multiple distribution intervals, the sensitivity level values ​​of other mutation targeted drugs are fitted with the sensitivity level values ​​of the targeted drugs at the multiple mutation points to establish an EGFR-TKIs resistance ResNet interpretation model; the sensitivity level values ​​of the targeted drugs at the multiple mutation points are calculated according to the EGFR-TKIs resistance ResNet interpretation model, including: if the previous EGFR prediction information includes at least two curves in the EGFR-TKIs resistance progression curve, the sensitivity level is calculated respectively according to each established EGFR-TKIs resistance ResNet interpretation model; the sensitivity level average value is calculated, and the sensitivity level average value is used as the sensitivity level of the targeted drugs for the multiple mutation points; the image block level ORR objective response rate heat map and the image block level ORR objective response rate average value calculation results can be effectively utilized, and the lung adenocarcinoma EGFR-TKIs efficacy intelligent prediction can be performed on the non-small cell lung cancer pathological images by training the generalized convolutional neural network ResNet model, thereby significantly improving the prediction accuracy.

[0089] The present invention provides an intelligent prediction method for the efficacy of EGFR-TKIs for lung adenocarcinoma, comprising:

[0090] S100, excluding and screening non-small cell lung cancer pathological images through the previous automatic RECIST evaluation, obtaining a set of pathological images after RECIST exclusion; scanning the set of pathological images after RECIST exclusion through the same-mode fixed magnifying galvanometer scanning module, obtaining a scanned data set of non-small cell lung cancer pathological images after RECIST exclusion;

[0091] S200, grouping the scanned data set after RECIST selection of pathological images into data set training, verification and test data set according to the set ratio of the data set, and obtaining the internal training, verification and test data set after RECIST selection; and incorporating the case data set of the TCIA database as the external test set of the TCIA data;

[0092] S300, based on the internal training, testing and testing data set selected by RECIST, the basic convolutional neural network ResNet model is trained with model parameters learned, model optimization parameters fine-tuned and model test recognition generalized to obtain a trained generalized convolutional neural network ResNet model;

[0093] S400 uses the block-level ORR objective response rate heat map and the block-level ORR objective response rate average calculation results to train a generalized convolutional neural network ResNet model to intelligently predict the efficacy of lung adenocarcinoma EGFR-TKIs for non-small cell lung cancer pathological images.

[0094] The principle and effect of the above technical solution are as follows: the present invention provides an intelligent prediction method for the efficacy of EGFR-TKIs in lung adenocarcinoma, comprising: excluding and screening non-small cell lung cancer pathological images through the preliminary automatic RECIST evaluation to obtain a set of pathological images after RECIST exclusion; scanning the set of pathological images after RECIST exclusion through the same-mode fixed magnifying galvanometer scanning module to obtain a scanned data set of pathological images after RECIST exclusion of non-small cell lung cancer; performing data set training, verification and test grouping on the scanned data set of pathological images after RECIST exclusion according to the set ratio of the data set to obtain an internal training, verification and test data set after RECIST exclusion; incorporating the case data set of the TCIA database as an external test set of TCIA data; and performing basic convolutional neural network training, verification and test on the internal training, verification and test data set after RECIST exclusion according to the internal training, verification and test data set after RECIST exclusion. The ResNet model is used to learn model training parameters, fine-tune model optimization parameters, and generalize model testing identification to obtain a trained generalized convolutional neural network ResNet model; the ORR objective response rate heat map at the image block level and the average ORR objective response rate at the image block level are used to calculate the results, and the trained generalized convolutional neural network ResNet model is used to intelligently predict the efficacy of lung adenocarcinoma EGFR-TKIs for non-small cell lung cancer pathological images; the problem of tumor heterogeneity can be overcome; the factors for predicting the efficacy or prognosis of patients with non-small cell lung cancer are enriched and optimized, and the results are more accurate; EGFR-TKIs targeted therapy is performed for advanced non-small cell lung cancer with EGFR sensitive mutations; after two cycles of treatment, the patient undergoes imaging RECIST evaluation to determine complete remission (complete remission). The system can predict the patients with poor response to EGFR-TKIs, adjust the treatment plan in advance, optimize the treatment process, and bring about the overall prediction of prognosis; avoid losing other treatment opportunities due to the loss of several months of waiting for the evaluation of EGFR-TKIs efficacy; automatically perform RECIST evaluation to exclude and screen the pathological images of non-small cell lung cancer for the prediction of non-small cell lung cancer, and scan the pathological image set after RECIST exclusion by the same model fixed magnifying galvanometer, so that the scan data of the pathological image of non-small cell lung cancer after RECIST exclusion has better image features and is more suitable for data training; it can effectively extend the prediction accuracy of sensitive mutations and improve its quality; it can effectively predict and screen out those with poor response to EGFR-TKIs, and predict more effective treatment plans in advance; the grouping of the data set after RECIST exclusion is more accurate, avoiding unreasonable grouping, and the quality of the training, validation and test set data is significantly improved; incorporating the case data set of the TCIA database as the external test set of TCIA data can greatly improve the overall adaptability of model training;According to the internal training and testing data set after RECIST selection, the basic convolutional neural network ResNet model is trained with model training parameters, model optimization parameters are fine-tuned, and model test recognition is generalized, which significantly improves the generalization ability of the convolutional neural network ResNet model; the image block-level ORR objective response rate heat map and the image block-level ORR objective response rate average calculation results can be effectively used to train the generalized convolutional neural network ResNet model to perform intelligent prediction of lung adenocarcinoma EGFR-TKIs efficacy on non-small cell lung cancer pathological images, significantly improving the prediction accuracy. ;

[0095] In one embodiment, S100 includes:

[0096] S101, to obtain pathological images of non-small cell lung cancer with EGFR sensitive mutations and treated with EGFR-TKIs;

[0097] S102, according to the non-small cell lung cancer pathology images, exclude and screen the non-small cell lung cancer pathology images through the previous automatic RECIST evaluation, and obtain a set of pathology images after RECIST exclusion;

[0098] S103, scanning each pair of RECIST excluded pathology images in the RECIST excluded pathology image set by using a fixed magnifying galvanometer scanning module of the same model to obtain a RECIST excluded scan data set of pathology images of non-small cell lung cancer;

[0099] Obtaining a set of pathological images after RECIST exclusion through the previous automatic RECIST evaluation and exclusion screening of non-small cell lung cancer pathological images, including: creating an automatic RECIST evaluation and exclusion screening module according to the RECIST evaluation criteria, automatically evaluating and excluding non-small cell lung cancer pathological images, and obtaining a set of pathological images after RECIST exclusion screening;

[0100] The same-mode fixed magnifying galvanometer scanning module includes: a same-mode fixed platform, a flat-field focusing lens unit, a reflection galvanometer scanning galvanometer unit and a scanning focal length adjustment group; the reflection galvanometer scanning galvanometer unit includes: a servo-driven vibration group and a virtual image mirror group; the servo-driven vibration group drives the virtual image mirror group to vibrate back and forth; the same-mode fixed platform selects non-small cell lung cancer pathological images according to the same mode and position fixed image size; the flat-field focusing lens unit is fixedly installed directly above the same-mode fixed platform; the servo-driven vibration group, the scanning focal length adjustment group and the high-power RGB magnifying scanner are arranged in sequence on a horizontal plane; the servo-driven vibration group is fixedly installed above the flat-field focusing lens unit, and its virtual image mirror group is arranged directly above the center line of the flat-field focusing lens unit and has a changeable inclination angle with the horizontal plane; the center of the virtual image mirror group, the center of the scanning focal length adjustment group and the high-power RGB magnifying scanner are fixedly installed above the flat-field focusing lens unit. The center of the large scanner is on a straight line; scanning each pair of RECIST post-selected pathology images in the RECIST post-selected pathology image set through the same-mode fixed magnification galvanometer scanning module includes: selecting non-small cell lung cancer pathology images according to the same mode and same position fixed image size through the same-mode fixed platform; the same mode and same position fixed image size adopts RGB three-channel 24-bit color image mode, 400-fold fixed magnification and set image size, and each channel of the RGB three-channel is 8 bits; during the scanning process, as the scanning mirror of the high-power RGB magnification scanner moves, the servo-driven vibration group drives the virtual image mirror group to reciprocate and vibrate, generating a micro-scale reciprocating change in the scanning focal length; the scanning focal length adjustment group adjusts to compensate for the micro-scale reciprocating change in the scanning focal length, so as to maintain the RECIST post-selected pathology image scanning with precise and consistent same-mode and same position fixed image size.

[0101] The principle and effect of the above technical solution are: obtaining pathological images of non-small cell lung cancer with EGFR sensitive mutation and treated with EGFR-TKIs; according to the pathological images of non-small cell lung cancer, excluding and screening the pathological images of non-small cell lung cancer through the previous automatic RECIST evaluation, and obtaining the pathological image set after RECIST exclusion; scanning each pair of pathological images after RECIST exclusion in the pathological image set after RECIST exclusion by the same model fixed magnifying galvanometer scanning module, and obtaining the non-small cell lung cancer pathological image RECIST exclusion scanning data set; excluding and screening the pathological images of non-small cell lung cancer through the previous automatic RECIST evaluation, and obtaining the pathological image set after RECIST exclusion includes: according to the RECIST evaluation standard The automatic RECIST evaluation exclusion screening module is created to automatically evaluate and exclude non-small cell lung cancer pathological images and obtain the pathological image set after RECIST exclusion. The same-mode fixed magnifying galvanometer scanning module includes: a same-mode fixed platform, a flat-field focusing lens unit, a reflection galvanometer scanning galvanometer unit and a scanning focus adjustment group. The reflection galvanometer scanning galvanometer unit includes: a servo-driven vibration group and a virtual image mirror group. The servo-driven vibration group drives the virtual image mirror group to vibrate back and forth. The same-mode fixed platform selects non-small cell lung cancer pathological images according to the same mode and position fixed image size. The flat-field focusing lens unit is fixedly installed just above the same-mode fixed platform. The servo-driven vibration group, the scanning focus adjustment group and the high-power RGB magnifying scanner are arranged in sequence. The servo-driven vibration group is fixedly installed above the flat-field focusing lens unit, and the virtual image mirror group is arranged just above the center line of the flat-field focusing lens unit and has a variable angle of inclination with the horizontal plane; the center of the virtual image mirror group, the center of the scanning focus adjustment group and the center of the high-power RGB magnification scanner are in a straight line; each pair of RECIST post-selection pathology images in the RECIST post-selection pathology image set is scanned by the same-mode fixed magnification galvanometer scanning module, including: selecting non-small cell lung cancer pathology images according to the same mode and position fixed image size through the same-mode fixed platform; the same mode and position fixed image size adopts RGB three-channel 24-bit color image mode, 400 times fixed magnification and set image size inch, each of the three RGB channels is 8 bits; during the scanning process, as the scanning mirror of the high-power RGB magnification scanner moves, the servo-driven vibration group drives the virtual image mirror group to vibrate back and forth, resulting in micro-scale reciprocating changes in the scanning focal length; the scanning focal length adjustment group adjusts to compensate for the micro-scale reciprocating changes in the scanning focal length, and maintains the precise and consistent image size of the same mode and position fixed after RECIST screening of pathological images after RECIST screening; automatic RECIST evaluation can be performed for non-small cell lung cancer prediction to exclude and screen non-small cell lung cancer pathological images, and the RECIST screened pathological image set after RECIST screening is scanned by the same-mode fixed magnification galvanometer, so that the non-small cell lung cancer pathological image RECIST screening scan data has better image features and is more suitable for data training.

[0102] In one embodiment, S200 includes:

[0103] S201, grouping the pathological image scan data set after RECIST selection for data set training, verification and testing according to the data set setting ratio, to obtain the RECIST selected internal training, verification and testing data set; the RECIST selected internal training, verification and testing data set includes: a RECIST selected internal training data set, a RECIST selected internal verification data set and a RECIST selected internal testing data set; the data set setting ratio includes: 60%: 20%: 20% or 70%: 15%: 15%;

[0104] S202, linking to the TCIA database to obtain a TCIA database case data set, and incorporating the TCIA database case data set into a TCIA data external test set.

[0105] The principle and effect of the above technical solution are: grouping the pathological image RECIST selection scan data set according to the data set setting ratio for data set training, verification and test to obtain the RECIST selection internal training, verification and test data set; the RECIST selection internal training, verification and test data set includes: RECIST selection internal training data set, RECIST selection internal verification data set and RECIST selection internal test data set; the data set setting ratio includes: 60%: 20%: 20% or 70%: 15%: 15%; linking the TCIA database to obtain the TCIA database case data set, and incorporating the TCIA database case data set as the TCIA data external test set; the RECIST selection data set is more accurately grouped to avoid unreasonable grouping, and the quality of the training, verification and test set data is significantly improved; incorporating the TCIA database case data set as the TCIA data external test set can greatly improve the comprehensive adaptability of model training.

[0106] In one embodiment, S300 includes:

[0107] S301, setting a convolutional neural network ResNet architecture as a basic architecture of a system deep learning training model, and obtaining a basic convolutional neural network ResNet model;

[0108] S302, based on the internal training, testing and testing data set after RECIST selection, the basic convolutional neural network ResNet model is trained with model parameters, model optimization parameters are fine-tuned, and model testing and recognition generalization is performed to obtain a trained generalized convolutional neural network ResNet model.

[0109] The principle and effect of the above technical solution are as follows: setting the convolutional neural network ResNet architecture as the basic architecture of the system deep learning training model to obtain a basic convolutional neural network ResNet model; performing model training parameter learning, model optimization parameter fine-tuning, and model test recognition generalization on the basic convolutional neural network ResNet model according to the internal training, verification, and test data set after RECIST selection to obtain a trained generalized convolutional neural network ResNet model; performing model training parameter learning, model optimization parameter fine-tuning, and model test recognition generalization on the basic convolutional neural network ResNet model according to the internal training, verification, and test data set after RECIST selection to significantly improve the generalization ability of the convolutional neural network ResNet model.

[0110] In one embodiment, S400 includes:

[0111] S401, automatically marking slice-level labels through a slice-level ORR objective response rate empirical model to obtain a slice-level ORR objective response rate result;

[0112] S402: Based on the TCIA data external test set, input the training generalized convolutional neural network ResNet model, use the slice-level ORR objective response rate results as grouping labels through weakly supervised intelligent learning, and use the set pixels to segment the non-small cell lung cancer pathology image to obtain the non-small cell lung cancer pathology image blocks; screen out all non-small cell lung cancer pathology image blocks containing tumors, obtain the image block-level ORR objective response rate heat map, and interpret the training generalized convolutional neural network ResNet model;

[0113] S403, based on the interpretation of the trained generalized convolutional neural network ResNet model, intelligent prediction of the efficacy of EGFR-TKIs for lung adenocarcinoma was performed on the pathological images of non-small cell lung cancer; based on the ORR heat map at the image block level, the average ORR objective response rate at the image block level was calculated to obtain the prediction result of the ORR objective response rate of EGFR-TKIs for lung adenocarcinoma at the pixel level;

[0114] The training of the generalized convolutional neural network ResNet model interpretation includes: obtaining the previous EGFR prediction information, differentiation degree identification information, microenvironment parameter learning information, setting the EGFR-TKIs resistance PFS value and mutation targeted drug sensitivity level value of multiple mutation points corresponding to the targeted drug; the previous EGFR prediction information includes: previous EGFR prediction information, differentiation degree identification information, microenvironment parameter learning information; the previous EGFR prediction information includes: extracellular ligand binding region prediction curve, transmembrane region prediction curve, intracellular kinase region prediction curve and EGFR-TKIs resistance progression curve; the horizontal axis of the EGFR-TKIs resistance progression curve is the EGFR-TKIs resistance PFS value, and the vertical axis is the mutation type of the mutation point corresponding to the lesion progression ratio. The mutation types of the mutation point include: 19Del mutation, L858R mutation and other types of mutations; lesion progression includes: dual primary lesion metastasis progression, primary lesion progression and metastatic lesion progression; the targeted drug sensitivity level values ​​of multiple mutation points are divided into multiple distribution intervals, and the intersection of any two distribution intervals is an empty set;

[0115] The sensitivity level values ​​of the targeted drugs of the multiple mutation points and the sensitivity level values ​​of the corresponding multiple mutation points on the EGFR-TKIs resistance progression curve are respectively fitted in the multiple distribution intervals to establish the EGFR-TKIs resistance ResNet interpretation model, including:

[0116] If the previous EGFR prediction information is an extracellular ligand binding region prediction curve, determining the 19Del mutation targeted drug sensitivity level values ​​corresponding to the multiple mutation points according to the extracellular ligand binding region prediction curve;

[0117] In the multiple distribution intervals, the sensitivity level value of the targeted drug of the 19Del mutation is fitted with the sensitivity level value of the targeted drug of multiple mutation points to establish a ResNet interpretation model of EGFR-TKIs resistance;

[0118] If the previous EGFR prediction information is a transmembrane region prediction curve, determining the L858R mutation targeted drug sensitivity level values ​​corresponding to the multiple mutation points according to the transmembrane region prediction curve;

[0119] In the multiple distribution intervals, the sensitivity level value of the L858R mutation targeted drug is fitted with the sensitivity level values ​​of the targeted drug of multiple mutation points to establish a ResNet explanation model of EGFR-TKIs resistance;

[0120] If the previous EGFR prediction information is an intracellular kinase region prediction curve, determining the sensitivity level values ​​of other mutation targeted drugs corresponding to the multiple mutation points according to the intracellular kinase region prediction curve;

[0121] In the multiple distribution intervals, the sensitivity level values ​​of other mutation targeted drugs are fitted with the sensitivity level values ​​of the targeted drugs of multiple mutation points to establish a ResNet explanation model of EGFR-TKIs resistance;

[0122] The targeted drug sensitivity level values ​​of multiple mutation points are calculated according to the EGFR-TKIs resistance ResNet interpretation model, including: if the previous EGFR prediction information includes at least two curves in the EGFR-TKIs resistance progression curve, the sensitivity level is calculated separately according to each established EGFR-TKIs resistance ResNet interpretation model; the average sensitivity level is calculated, and the average sensitivity level is used as the targeted drug sensitivity level of multiple mutation points.

[0123] The principle and effect of the above technical solution are as follows: through the slice-level ORR objective response rate empirical model, the slice-level labels are automatically marked to obtain the slice-level ORR objective response rate results; according to the TCIA data external test set, the training generalized convolutional neural network ResNet model is input, and through weakly supervised intelligent learning, the slice-level ORR objective response rate results are used as grouping labels, and the non-small cell lung cancer pathology image is segmented with set pixels to obtain the non-small cell lung cancer pathology image blocks; all non-small cell lung cancer pathology image blocks containing tumors are screened out, and the image block-level ORR objective response rate heat map is obtained, and the training generalized convolutional neural network ResNet model is interpreted; according to the training generalized convolutional neural network ResNet model interpretation, the non-small cell lung cancer pathology images are intelligently predicted for the efficacy of lung adenocarcinoma EGFR-TKIs; according to the image block-level ORR heat map, the image block-level ORR objective response rate is calculated. The average value of the response rate is obtained to obtain the pixel-level lung adenocarcinoma EGFR-TKIs efficacy prediction ORR objective response rate prediction result; the training generalized convolutional neural network ResNet model interpretation includes: obtaining the previous EGFR prediction information, differentiation degree identification information, microenvironment parameter learning information, and setting the EGFR-TKIs resistance PFS value and mutation targeted drug sensitivity level value of multiple mutation points corresponding to the targeted drug; the previous EGFR prediction information includes: previous EGFR prediction information, differentiation degree identification information, microenvironment parameter learning information; the previous EGFR prediction information includes: extracellular ligand binding region prediction curve, transmembrane region prediction curve, intracellular kinase region prediction curve and EGFR-TKIs resistance progression curve; the abscissa of the EGFR-TKIs resistance progression curve is the EGFR-TKIs resistance PFS value, and the ordinate is the lesion progression ratio corresponding to the mutation type of the mutation point. The mutation types of mutation points include: 19Del mutation, L858R mutation and other types of mutations; lesion progression includes: dual primary lesion metastasis progression, primary lesion progression and metastasis progression; the sensitivity level values ​​of targeted drugs of multiple mutation points are divided into multiple distribution intervals, and the intersection of any two distribution intervals is an empty set;

[0124] The sensitivity level values ​​of targeted drugs at multiple mutation points are respectively fitted with the sensitivity level values ​​corresponding to the multiple mutation points on the EGFR-TKIs resistance progression curve in multiple distribution intervals to establish an EGFR-TKIs resistance ResNet interpretation model, including: if the previous EGFR prediction information is an extracellular ligand binding region prediction curve, the sensitivity level values ​​of 19Del mutation targeted drugs corresponding to the multiple mutation points are determined according to the extracellular ligand binding region prediction curve; within the multiple distribution intervals, the sensitivity level values ​​of 19Del mutation targeted drugs are fitted with multiple The targeted drug sensitivity level values ​​of the mutation points are fitted to establish the EGFR-TKIs resistance ResNet interpretation model; if the previous EGFR prediction information is a transmembrane region prediction curve, the L858R mutation targeted drug sensitivity level values ​​corresponding to the multiple mutation points are determined according to the transmembrane region prediction curve; within the multiple distribution intervals, the L858R mutation targeted drug sensitivity level value is fitted with the targeted drug sensitivity level values ​​of the multiple mutation points to establish the EGFR-TKIs resistance ResNet interpretation model; if the previous EGFR prediction information is a cell The method comprises the following steps: an intracellular kinase region prediction curve is used to determine the sensitivity level values ​​of other mutation targeted drugs corresponding to the multiple mutation points according to the intracellular kinase region prediction curve; within the multiple distribution intervals, the sensitivity level values ​​of other mutation targeted drugs are fitted with the sensitivity level values ​​of the targeted drugs at the multiple mutation points to establish an EGFR-TKIs resistance ResNet interpretation model; the sensitivity level values ​​of the targeted drugs at the multiple mutation points are calculated according to the EGFR-TKIs resistance ResNet interpretation model, including: if the previous EGFR prediction information includes at least two curves in the EGFR-TKIs resistance progression curve, the sensitivity level is calculated respectively according to each established EGFR-TKIs resistance ResNet interpretation model; the sensitivity level average value is calculated, and the sensitivity level average value is used as the sensitivity level of the targeted drugs for the multiple mutation points; the image block level ORR objective response rate heat map and the image block level ORR objective response rate average value calculation results can be effectively utilized, and the lung adenocarcinoma EGFR-TKIs efficacy intelligent prediction can be performed on the non-small cell lung cancer pathological images by training the generalized convolutional neural network ResNet model, thereby significantly improving the prediction accuracy.

[0125] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and the implementation methods. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and the illustrations shown and described herein.

Claims

1. An intelligent prediction system for the efficacy of EGFR-TKIs in lung adenocarcinoma, characterized in that: include: The RECIST same-mode magnification scanning subsystem excludes and screens non-small cell lung cancer pathology images through the previous automatic RECIST evaluation, and obtains the RECIST excluded pathology image set; the same-mode fixed magnification galvanometer scanning module scans the RECIST excluded pathology image set, and obtains the non-small cell lung cancer pathology image RECIST excluded scanning data set; The internal and external data set acquisition subsystem groups the scanned data set after RECIST selection of pathological images into data set training, verification and test groups according to the set ratio of the data set, and obtains the internal training, verification and test data set after RECIST selection; and includes the case data set of the TCIA database as the external test set of TCIA data; The ResNet model training generalization subsystem performs model training parameter learning, model optimization parameter fine-tuning, and model test recognition generalization on the basic convolutional neural network ResNet model based on the internal training, testing, and testing data set selected by RECIST to obtain a training generalized convolutional neural network ResNet model. The EGFR-TKIs intelligent prediction subsystem uses the image block-level ORR objective response rate heat map and the image block-level ORR objective response rate average calculation results to train a generalized convolutional neural network ResNet model to intelligently predict the efficacy of EGFR-TKIs in lung adenocarcinoma based on non-small cell lung cancer pathological images.

2. The intelligent prediction system for the efficacy of EGFR-TKIs for lung adenocarcinoma according to claim 1, characterized in that: RECIST same-mode amplification scanning subsystem includes: EGFR lung cancer pathology image acquisition subsystem, which acquires pathology images of non-small cell lung cancer with EGFR sensitive mutations and treated with EGFR-TKIs; The automatic RECIST evaluation and exclusion subsystem, based on the pathological images of non-small cell lung cancer, excludes and screens the pathological images of non-small cell lung cancer through the previous automatic RECIST evaluation, and obtains the pathological image set after RECIST exclusion; The same-mode fixed magnification scanning subsystem scans each pair of RECIST excluded pathology images in the RECIST excluded pathology image set through the same-mode fixed magnification galvanometer scanning module to obtain a RECIST excluded scanning data set of non-small cell lung cancer pathology images.

3. The intelligent prediction system for the efficacy of EGFR-TKIs for lung adenocarcinoma according to claim 1, characterized in that: The internal and external data set acquisition subsystem includes: The data set training, verification and testing grouping subsystem performs data set training, verification and testing grouping on the scan data set after RECIST selection of pathological images according to the data set setting ratio, and obtains the internal training, verification and testing data set after RECIST selection; the internal training, verification and testing data set after RECIST selection includes: the internal training data set after RECIST selection, the internal verification data set after RECIST selection and the internal testing data set after RECIST selection; the data set setting ratio includes: 60%: 20%: 20% or 70%: 15%: 15%; The TCIA data acquisition subsystem links to the TCIA database to obtain the TCIA database case data set, and incorporates the TCIA database case data set as the TCIA data external test set.

4. The intelligent prediction system for the efficacy of EGFR-TKIs for lung adenocarcinoma according to claim 1, characterized in that: The ResNet model training generalization subsystem includes: ResNet architecture setting subsystem, set the convolutional neural network ResNet architecture as the basic architecture of the system deep learning training model, and obtain the basic convolutional neural network ResNet model; The ResNet model optimization subsystem performs model training parameter learning, model optimization parameter fine-tuning, and model test recognition generalization on the basic convolutional neural network ResNet model based on the internal training, testing, and test data set after RECIST selection, and obtains the trained generalized convolutional neural network ResNet model.

5. The intelligent prediction system for the efficacy of EGFR-TKIs for lung adenocarcinoma according to claim 1, characterized in that: EGFR-TKIs intelligent prediction subsystem includes: The slice-level ORR labeling subsystem automatically labels the slice-level labels and obtains the slice-level ORR objective response rate results through the slice-level ORR objective response rate empirical model; The intelligent learning model interpretation subsystem uses the TCIA data external test set as input to train the generalized convolutional neural network ResNet model. Through weakly supervised intelligent learning, it uses the slice-level ORR objective response rate results as grouping labels and sets pixels to segment non-small cell lung cancer pathology images to obtain non-small cell lung cancer pathology image blocks. It also screens out all non-small cell lung cancer pathology image blocks containing tumors, obtains the image block-level ORR objective response rate heat map, and interprets the training generalized convolutional neural network ResNet model. The ORR calculation efficacy prediction subsystem, based on the interpretation of the trained generalized convolutional neural network ResNet model, intelligently predicts the efficacy of EGFR-TKIs for lung adenocarcinoma on non-small cell lung cancer pathological images; based on the image block level ORR heat map, the average ORR objective response rate at the image block level is calculated to obtain the pixel-level lung adenocarcinoma EGFR-TKIs efficacy prediction ORR objective response rate prediction results.

6. An intelligent prediction method for the efficacy of EGFR-TKIs in lung adenocarcinoma, characterized in that: include: S100, excluding and screening non-small cell lung cancer pathological images through the previous automatic RECIST evaluation, obtaining a set of pathological images after RECIST exclusion; scanning the set of pathological images after RECIST exclusion through the same-mode fixed magnifying galvanometer scanning module, obtaining a scanned data set of non-small cell lung cancer pathological images after RECIST exclusion; S200, grouping the scanned data set after RECIST selection of pathological images into data set training, verification and test data set according to the set ratio of the data set, and obtaining the internal training, verification and test data set after RECIST selection; and incorporating the case data set of the TCIA database as the external test set of the TCIA data; S300, based on the internal training, testing and testing data set selected by RECIST, the basic convolutional neural network ResNet model is trained with model parameters learned, model optimization parameters fine-tuned and model test recognition generalized to obtain a trained generalized convolutional neural network ResNet model; S400 uses the block-level ORR objective response rate heat map and the block-level ORR objective response rate average calculation results to train a generalized convolutional neural network ResNet model to intelligently predict the efficacy of lung adenocarcinoma EGFR-TKIs for non-small cell lung cancer pathological images.

7. The method for intelligent prediction of the efficacy of EGFR-TKIs for lung adenocarcinoma according to claim 6, characterized in that: S100 includes: S101, to obtain pathological images of non-small cell lung cancer with EGFR sensitive mutations and treated with EGFR-TKIs; S102, according to the non-small cell lung cancer pathology images, exclude and screen the non-small cell lung cancer pathology images through the previous automatic RECIST evaluation, and obtain a set of pathology images after RECIST exclusion; S103, scanning each pair of post-RECIST exclusion pathology images in the post-RECIST exclusion pathology image set by using a fixed magnifying galvanometer scanning module of the same model to obtain a post-RECIST exclusion scanning data set of pathology images of non-small cell lung cancer.

8. The method for intelligent prediction of the efficacy of EGFR-TKIs for lung adenocarcinoma according to claim 6, characterized in that: S200 includes: S201, grouping the pathological image scan data set after RECIST selection for data set training, verification and testing according to the data set setting ratio, to obtain the RECIST selected internal training, verification and testing data set; the RECIST selected internal training, verification and testing data set includes: a RECIST selected internal training data set, a RECIST selected internal verification data set and a RECIST selected internal testing data set; the data set setting ratio includes: 60%: 20%: 20% or 70%: 15%: 15%; S202, linking to the TCIA database to obtain a TCIA database case data set, and incorporating the TCIA database case data set as a TCIA data external test set.

9. The method for intelligent prediction of the efficacy of EGFR-TKIs for lung adenocarcinoma according to claim 6, characterized in that: S300 includes: S301, setting a convolutional neural network ResNet architecture as a basic architecture of a system deep learning training model, and obtaining a basic convolutional neural network ResNet model; S302, based on the internal training, testing and testing data set after RECIST selection, the basic convolutional neural network ResNet model is trained with model parameters, model optimization parameters are fine-tuned, and model testing and recognition generalization is performed to obtain a trained generalized convolutional neural network ResNet model.

10. The method for intelligent prediction of the efficacy of EGFR-TKIs for lung adenocarcinoma according to claim 6, characterized in that: S400 includes: S401, automatically marking slice-level labels through a slice-level ORR objective response rate empirical model to obtain a slice-level ORR objective response rate result; S402: Based on the TCIA data external test set, input the training generalized convolutional neural network ResNet model, use the slice-level ORR objective response rate results as grouping labels through weakly supervised intelligent learning, and use the set pixels to segment the non-small cell lung cancer pathology image to obtain the non-small cell lung cancer pathology image blocks; screen out all non-small cell lung cancer pathology image blocks containing tumors, obtain the image block-level ORR objective response rate heat map, and interpret the training generalized convolutional neural network ResNet model; S403, based on the interpretation of the trained generalized convolutional neural network ResNet model, intelligent prediction of the efficacy of EGFR-TKIs for lung adenocarcinoma was performed on the pathological images of non-small cell lung cancer; based on the ORR heat map at the image block level, the average ORR objective response rate at the image block level was calculated to obtain the pixel-level prediction of the ORR objective response rate of lung adenocarcinoma EGFR-TKIs efficacy.

Images