Prediction method and device for pathological complete remission of non-small cell lung cancer neoadjuvant chemotherapy

Abstract

The invention discloses a pathological complete remission prediction method and device for non-small cell lung cancer neoadjuvant chemotherapy, and relates to the field of medical imagines.The method comprises the steps that a preoperative medical image and preoperative clinical data of a patient to be predicted are obtained; inputting the preoperative medical image and the preoperative clinical data into the trained pathology complete remission prediction model, and outputting a pathology complete remission result of the patient to be predicted by the pathology complete remission prediction model; the pathologic complete remission prediction model is constructed by using pre-treatment radiomics features, post-treatment radiomics features, difference-radiomics features, deep learning features and sample clinical data. Whether a non-small cell lung cancer patient receiving neoadjuvant chemotherapy reaches complete pathological remission or not is represented in a non-invasive mode, so that medical staff are assisted in formulating a personalized treatment scheme, and the clinical result of the patient is improved.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging, and particularly to a method and device for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer. Background Art

[0002] Lung cancer is a common malignant tumor globally. Among them, non-small cell lung cancer (NSCLC) is the most common type of lung cancer. The treatment regimens for NSCLC include surgery, radiotherapy, and systemic treatment, etc. Its treatment has entered the era of precision medicine, emphasizing the formulation of individualized treatment plans for patients based on staging, histological type, and molecular characteristics.

[0003] Neoadjuvant immunochemotherapy (NIT) in neoadjuvant treatment is the combined use of immune checkpoint inhibitors (ICIs) and chemotherapy before surgery, and has become a promising treatment regimen to improve the pathological remission rate of NSCLC patients.

[0004] Accurately predicting the treatment effect of patients undergoing NIT before surgery has important clinical value, and can also assist medical staff in formulating individualized treatment plans for patients to ensure the most appropriate and effective treatment for patients. This prediction result can not only improve the cure success rate of patients, but also reduce unnecessary ineffective treatment for patients. In addition, early identification of patients who do not respond well to NIT can also avoid the implementation of unnecessary NIT treatment regimens.

[0005] Although NIT has shown great promise in NSCLC, predicting the treatment effect of NIT for each patient before surgery remains a clinical challenge. Pathological Complete Response (pCR) is usually used as the short-term target endpoint in NIT clinical trials and is defined as the absence of residual tumor cells after evaluation of the resected tumor tissue and regional lymph nodes. Preoperative evaluation often cannot accurately judge and currently still relies on postoperative pathological confirmation because traditional imaging evaluation methods based on tumor size changes may not accurately predict the above treatment effect. For patients who achieve clinical complete remission, not every patient can achieve pCR, and some patients still have a risk of recurrence. Therefore, when determining the surgical indications for NSCLC patients, multiple factors need to be considered comprehensively, including but not limited to the specific condition of the patient, pathological type, stage, and the overall physical condition of the patient, etc. Usually, if a patient is evaluated as clinically completely remitted after NIT and their physical condition permits, surgery is still a treatment option to be considered. However, for certain types of tumors such as rectal cancer, if a patient achieves pCR, further evaluation may be needed to determine whether surgery is required because the prognosis of this part of patients is usually better.

[0006] Therefore, there is an urgent need for a non-invasive and preoperative method for predicting pCR in NSCLC to assist medical staff in determining surgical indications and treatment decisions. Summary of the Invention

[0007] In view of this, the embodiments of the present invention provide a method and device for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer to solve the problem that there is an urgent need for a non-invasive and preoperative method for predicting pCR in NSCLC.

[0008] According to the first aspect, the embodiments of the present invention provide a method for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer, the method comprising: Obtain the preoperative medical images and preoperative clinical data of the patient to be predicted; Input the preoperative medical images and preoperative clinical data into the trained pathological complete remission prediction model, and the pathological complete remission prediction model outputs the pathological complete remission result of the patient to be predicted; The pathological complete remission prediction model is constructed using pre-treatment radiomics features, post-treatment radiomics features, difference - radiomics features, deep learning features, and sample clinical data; The pre-treatment radiomics features and post-treatment radiomics features are radiomics features extracted from the sample medical images based on the region of interest, and the deep learning features are feature information extracted from the post-treatment image features using the trained pre-trained model; The difference-radiomics features are calculated based on the pre-treatment radiomics features and the post-treatment radiomics features, and the difference-radiomics features include: The first difference-radiomics feature for characterizing the absolute change value of the radiomics features before and after neoadjuvant immunochemotherapy, the second difference-radiomics feature for characterizing the percentage change rate of the radiomics features before and after neoadjuvant immunochemotherapy, and the third difference-radiomics feature for the average percentage change rate of each chemotherapy cycle of the radiomics features before and after neoadjuvant immunochemotherapy.

[0009] Combined with the first aspect, in the first embodiment of the first aspect, the pathological complete remission prediction model is trained through the following steps: Obtain the sample medical images of the sample patients; the sample medical images include contrast-enhanced computed tomography images before neoadjuvant chemotherapy and contrast-enhanced computed tomography images after neoadjuvant chemotherapy; Retain the images that meet the preset criteria in the sample medical images to obtain the medical images to be analyzed; Based on the region of interest, extract the lesion regions representing the venous phase and the arterial phase from the images to be analyzed respectively, and use the lesion region representing the venous phase to perform matching calibration of the contours of the venous phase and the arterial phase to obtain the sample target lesion region of the images to be analyzed; Obtain the sample clinical data of the sample patients; Extract radiomics features from the sample target lesion region to obtain pre-treatment radiomics features and post-treatment radiomics features; Based on the pre-treatment and post-treatment radiomics features, construct difference-radiomics features, and extract deep learning features from the post-treatment image features using a trained pre-trained model; Screen out the model training features from the pre-treatment radiomics features, post-treatment radiomics features, difference-radiomics features, and deep learning features, and establish a mapping relationship between the model training features and the treatment effect of the sample patients after neoadjuvant immunochemotherapy to obtain a pathological complete remission prediction model.

[0010] Combined with the first embodiment of the first aspect, in the second embodiment of the first aspect, the step of, based on the region of interest, extracting the lesion regions representing the venous phase and the arterial phase from the images to be analyzed respectively, and using the lesion region representing the venous phase to perform matching calibration of the contours of the venous phase and the arterial phase to obtain the sample target lesion region of the images to be analyzed, specifically includes: Based on the first preset window, determine the corresponding plain scan lung window images from each medical image to be analyzed, and, based on the second preset window, determine the corresponding enhanced mediastinal window images from each medical image to be analyzed; Based on the region of interest, the first target lesion region and the second target lesion region are respectively determined from the plain scan lung window image and the enhanced mediastinal window image; the first target lesion region and the second target lesion region respectively contain the first target lesion and the peripheral region within the first preset range centered on the first target lesion, the second target lesion and the peripheral region within the second preset range centered on the second target lesion; the first target lesion is the largest lesion in the plain scan lung window image, and the second target lesion is the largest lesion in the enhanced mediastinal window image; Eliminate the non-target regions contained in the first target lesion region; Determine the position information of the first target lesion and the second target lesion corresponding to each image to be analyzed respectively. Based on the position information and taking the first target lesion region as the reference, perform matching and calibration on the contour of the lesion region of each image to be analyzed to obtain the sample target lesion region of each image to be analyzed.

[0011] Combined with the second embodiment of the first aspect, in the third embodiment of the first aspect, in the step of respectively extracting the lesion regions representing the venous phase and the arterial phase from the image to be analyzed based on the region of interest, and using the lesion region representing the venous phase to perform matching and calibration on the contours of the venous phase and the arterial phase to obtain the sample target lesion region of the image to be analyzed, the following steps are further included: Determine the intraclass correlation coefficient of the sample target lesion region, and eliminate the sample target lesion regions with a correlation lower than the preset correlation according to the intraclass correlation coefficient.

[0012] Combined with the second embodiment of the first aspect, in the fourth embodiment of the first aspect, the pathological complete remission prediction model is further trained through the following steps: Preprocess the sample medical images.

[0013] Combined with the first embodiment of the first aspect, in the fifth embodiment of the first aspect, the model training features are screened from the pre-treatment imaging features, post-treatment imaging features, difference-imaging features, and deep learning features, and the mapping relationship between the model training features and the treatment effect of the sample patients after receiving neoadjuvant immunochemotherapy is established according to the model training features and the sample clinical data to obtain the pathological complete remission prediction model, which specifically includes: Perform standardization processing on the pre-treatment imaging features, post-treatment imaging features, difference-imaging features, and deep learning features to obtain standardized features; Use the synthetic minority over-sampling technique to perform sample balancing processing on the standardized features to obtain balanced features, and delete the features with a variance of 0 and a Spearman rank correlation coefficient exceeding the preset coefficient in the balanced features; Use the maximum correlation-minimum redundancy algorithm to screen out the preliminary training features from the balanced features; Use the least absolute shrinkage and selection operator algorithm to screen out the model training features from the preliminary training features; According to the model training features and the sample clinical data, and adopting a logistic regression method, establish a mapping relationship with the treatment effect of the sample patients after receiving neoadjuvant immunochemotherapy, and obtain a pathological complete remission prediction model.

[0014] Combined with the first embodiment of the first aspect, in the sixth embodiment of the first aspect, the preset standard is: Only received neoadjuvant immunochemotherapy for non-small cell lung cancer; the time interval between the baseline computed tomography scan and the first treatment does not exceed a first preset interval; the time interval between the preoperative computed tomography scan and the surgery does not exceed a second preset interval; there is no situation in the sample medical image that hinders tumor segmentation and feature extraction.

[0015] According to the second aspect, an embodiment of the present invention further provides a device for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer, and the device includes: A data acquisition module, configured to acquire the preoperative medical image and preoperative clinical data of the patient to be predicted; A model prediction module, configured to input the preoperative medical image and preoperative clinical data into the trained pathological complete remission prediction model, and output the pathological complete remission result of the patient to be predicted by the pathological complete remission prediction model; The pathological complete remission prediction model is constructed by using pre-treatment radiomics features, post-treatment radiomics features, difference-radiomics features, deep learning features, and sample clinical data; The pre-treatment radiomics features and the post-treatment radiomics features are radiomics features extracted from the sample medical image based on the region of interest, and the deep learning features are feature information extracted from the post-treatment image features by using a trained pre-trained model; The difference-radiomics features are calculated according to the pre-treatment radiomics features and the post-treatment radiomics features, and the difference-radiomics features include: A first difference-radiomics feature for characterizing the absolute change value of the radiomics features before and after neoadjuvant immunochemotherapy, a second difference-radiomics feature for characterizing the percentage change rate of the radiomics features before and after neoadjuvant immunochemotherapy, and a third difference-radiomics feature for the average percentage change rate of each chemotherapy cycle of the radiomics features before and after neoadjuvant immunochemotherapy.

[0016] According to a third aspect, an embodiment of the present invention further provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps of the pathological complete remission prediction method for neoadjuvant chemotherapy of non-small cell lung cancer as described in any one of the above are implemented.

[0017] According to a fourth aspect, an embodiment of the present invention further provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the pathological complete remission prediction method for neoadjuvant chemotherapy of non-small cell lung cancer as described in any one of the above are implemented.

[0018] The pathological complete remission prediction method and device for neoadjuvant chemotherapy of non-small cell lung cancer of the present invention perform segmentation on the sample image data before and after NIT based on the data of interest and extract radiomics features to obtain the radiomics features before and after treatment. Then, three difference-radiomics features are constructed based on the radiomics features before and after treatment. Deep learning features are extracted from the post-treatment image features using a trained pre-trained model. The maximum relevance-minimum redundancy algorithm (mRMR) and the least absolute shrinkage and selection operator (LASSO) are used to screen out meaningful features from these features. Then, the screened features are combined with the sample clinical data to construct a final pathological complete remission prediction model. The pathological complete remission prediction model constructed in this way combines three types of features: difference-radiomics, deep learning, and clinical features. The combination of these three types of features provides complementary feature supplementation. Compared with constructing a model using only single-modal information, it can provide better prediction performance. The combination of these three types of features provides complementary feature supplementation, significantly improving the accuracy of prediction. The pathological complete remission prediction model non-invasively characterizes whether NSCLC patients receiving NIT achieve pCR to assist medical staff in formulating personalized treatment plans and ultimately improving the clinical outcomes of non-small cell lung cancer patients. Description of the Drawings

[0019] The features and advantages of the present invention will be more clearly understood by referring to the accompanying drawings. The drawings are schematic and should not be construed as imposing any limitation on the present invention. In the drawings: Figure 1 A flowchart showing the pathological complete remission prediction method for neoadjuvant chemotherapy of non-small cell lung cancer provided by the present invention is shown; Figure 2 A graph showing the mean square error during the construction of the pathological complete remission prediction model in the pathological complete remission prediction method for neoadjuvant chemotherapy of non-small cell lung cancer provided by the present invention is shown; Figure 3 A graph showing the regression coefficient during the construction of the pathological complete remission prediction model in the pathological complete remission prediction method for neoadjuvant chemotherapy of non-small cell lung cancer provided by the present invention is shown; Figure 4 Shows the bar chart of feature coefficients in the process of constructing the pathological complete remission prediction model in the pathological complete remission prediction method for neoadjuvant chemotherapy of non-small cell lung cancer provided by the present invention; Figure 5 Shows the ROC curves of different models in the training set (A) and the test set (B) in the process of constructing the pathological complete remission prediction model in the pathological complete remission prediction method for neoadjuvant chemotherapy of non-small cell lung cancer provided by the present invention; Figure 6 Shows the DCA curves of different models in the training set (C) and the test set (D) in the process of constructing the pathological complete remission prediction model in the pathological complete remission prediction method for neoadjuvant chemotherapy of non-small cell lung cancer provided by the present invention; Figure 7 Shows the SHAP bar chart of the contribution and importance of model training features in the pathological complete remission prediction model in the pathological complete remission prediction method for neoadjuvant chemotherapy of non-small cell lung cancer provided by the present invention; Figure 8 Shows the SHAP dot chart of the influence direction and degree of each model training feature on the prediction result in the pathological complete remission prediction model in the pathological complete remission prediction method for neoadjuvant chemotherapy of non-small cell lung cancer provided by the present invention; Figure 9 Shows the SHAP force chart of the influence of each model training feature in the pathological complete remission prediction model in the pathological complete remission prediction method for neoadjuvant chemotherapy of non-small cell lung cancer provided by the present invention; Figure 10 Shows the structural schematic diagram of the pathological complete remission prediction device for neoadjuvant chemotherapy of non-small cell lung cancer provided by the present invention; Figure 11 Shows the hardware structural schematic diagram of the electronic device provided by the present invention. Detailed implementation manners

[0020] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0021] Lung cancer is a common malignant tumor globally. Among them, NSCLC is the most common type of lung cancer. The treatment regimens for NSCLC include surgery, radiotherapy, and systemic treatment, etc. Its treatment has entered the era of precision medicine, emphasizing the formulation of individualized treatment plans for patients based on staging, histological type, and molecular characteristics.

[0022] NIT in neoadjuvant therapy is the combined use of ICIs and chemotherapy before surgery, which has become a promising treatment option to improve the pathological remission rate of NSCLC patients.

[0023] Accurately predicting the treatment effect of patients undergoing NIT before surgery has important clinical value, and can also assist medical staff in formulating personalized treatment plans for patients to ensure the most appropriate and effective treatment for patients. This prediction result can not only improve the cure success rate of patients, but also reduce unnecessary ineffective treatment for patients. In addition, early identification of patients who do not respond well to NIT can also avoid the implementation of unnecessary NIT treatment plans.

[0024] Although NIT has shown great promise in NSCLC, predicting the treatment effect of each patient undergoing NIT before surgery remains a clinical challenge. pCR is usually used as the short-term objective endpoint of NIT clinical trials and is defined as no residual tumor cells after evaluation of resected tumor tissue and regional lymph nodes. Preoperative evaluation often cannot accurately judge and currently still relies on postoperative pathology confirmation because traditional imaging evaluation methods based on tumor size changes may not accurately predict the above treatment effect. For patients who achieve clinical complete remission, not every patient can achieve pCR, and some patients still have a risk of recurrence. Therefore, when determining the surgical indications for NSCLC patients, multiple factors need to be considered comprehensively, including but not limited to the specific condition, pathological type, stage of the patient, and the overall physical condition of the patient, etc. Usually, if the patient is evaluated as clinically completely remitted after NIT and the physical condition permits, surgery will still be a treatment option to consider. However, for certain types of tumors such as rectal cancer, if the patient achieves pCR, further evaluation may be required to determine whether surgery is needed because the prognosis of this part of patients is usually better.

[0025] In summary, how to provide a non-invasive, preoperative method for predicting pCR for NSCLC to assist doctors and patients in determining surgical indications and treatment decisions is an important issue that the industry urgently needs to solve.

[0026] Currently, the main methods for predicting pCR in NSCLC patients after NIT before surgery include the following: Medical imaging: By providing rich quantitative data, it can be used as a potential biomarker; Radiomics: By extracting high-dimensional features from medical images such as computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography (PET) scans, and based on these high-dimensional features, evaluating tumor invasiveness, characterizing tumor microenvironment, and predicting treatment response, traditional radiomics analysis usually relies on imaging at a single time point, which may lead to ignoring the dynamic changes of tumor characteristics after treatment; Delta-radiomics: By analyzing the temporal changes of radiomic features at multiple time points, it provides a non-invasive method to capture tumor changes caused by NIT and predict clinical outcomes; Prediction models built based on deep learning (DL): These prediction models perform well in automatically learning and extracting complex features from medical images, surpassing traditional feature-based methods. They have been proven to perform well in a variety of prediction tasks, especially the prediction models based on the Convolutional Neural Networks (CNNs) architecture.

[0027] However, the use of any of the above methods alone can no longer meet the industry's increasing demand for pCR prediction accuracy.

[0028] In order to solve the above problems, a method for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer is provided in this embodiment, aiming to provide a prediction model that combines difference-radiology, deep learning and clinical characteristics. The prediction results of the prediction model characterize whether NSCLC patients receiving NIT have achieved pCR in a non-invasive manner, so as to assist medical staff in formulating personalized treatment plans and improve the clinical outcomes of patients. The method for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer in the embodiment of the present invention can be used in electronic devices, including but not limited to computers, mobile terminals, etc. Figure 1 is a flow chart of a method for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer according to an embodiment of the present invention, such as Figure 1 As shown, the method may include the following steps: S10, obtaining preoperative medical images and preoperative clinical data of the patient to be predicted. In this embodiment, the preoperative medical images are specifically contrast-enhanced (CE) CT images of the patient before NIT, that is, preoperative contrast-enhanced CT images.

[0029] The preoperative medical images can be stored in advance in an electronic device or obtained by the electronic device from the outside world. For example, the image data obtained by the electronic device from an external acquisition device, or the electronic device obtains it from associated hospitals, medical institutions, etc.

[0030] No specific restrictions are imposed on the specific acquisition form of the preoperative medical images here, as long as it is ensured that the electronic device can obtain the preoperative medical images.

[0031] The preoperative clinical data includes the type of immune checkpoint inhibitor (ICI), patient age, number of treatment cycles, gender, smoking history, pathological classification (squamous cell carcinoma, adenocarcinoma, and other non-small cell lung cancers), clinical stage, serological markers (e.g., cytokeratin 19 fragment, neuron-specific enolase, cancer antigen 125, cancer antigen 199, carcinoembryonic antigen, squamous cell carcinoma antigen, etc.), and the immune response evaluation criteria in solid tumors (iRECIST), etc. Among them, the above-mentioned preoperative clinical data is obtained from the medical records, pathological records, laboratory test results, imaging examination results, etc. of these sample patients.

[0032] It can be understood that the patient to be predicted must be pathologically diagnosed as NSCLC.

[0033] S20: Input the preoperative medical images and the preoperative clinical data into the trained pathological complete response prediction model, and the pathological complete response result of the patient to be predicted is output by the pathological complete response prediction model.

[0034] Among them, the pathological complete remission prediction model is constructed using pre-treatment radiomics features, post-treatment radiomics features, difference-radiomics features, deep learning features, and sample clinical data. Specifically, the pre-treatment radiomics features and post-treatment radiomics features are radiomics features extracted from the sample medical images based on the region of interest. The deep learning features are feature information extracted from the post-treatment image features using a trained pre-trained model. The difference-radiomics features include the first difference-radiomics feature, the second-radiomics feature, and the third-radiomics feature. Among them, the first difference-radiomics feature (Delta 1) is used to characterize the absolute change value of the radiomics features before and after neoadjuvant immunochemotherapy, and its calculation method is the pre-treatment radiomics feature minus the post-treatment radiomics feature; the second difference-radiomics feature (Delta 2) is used to characterize the percentage change rate of the radiomics features before and after neoadjuvant immunochemotherapy, and its calculation method is the difference obtained by subtracting the post-treatment radiomics feature from the pre-treatment radiomics feature and then dividing by the pre-treatment radiomics feature; the third difference-radiomics feature (Delta 3) is used for the average percentage change rate of each chemotherapy cycle of the radiomics features before and after neoadjuvant immunochemotherapy, and its calculation method is the difference obtained by subtracting the post-treatment radiomics feature from the pre-treatment radiomics feature divided by the product of the pre-treatment radiomics feature and, that is: (pre - post) / (pre × number of cycles of neoadjuvant immunochemotherapy) Among them, pre represents the pre-treatment radiomics feature; post represents the post-treatment radiomics feature.

[0035] In this embodiment, the pathological complete remission prediction model is trained through the following steps: S30. Obtain the sample medical images of the sample patients. In this embodiment, the sample medical images include the pre-sample contrast-enhanced (CE) CT images before NIT, that is, the pre-operative sample contrast-enhanced CT images, and the post-sample contrast-enhanced CT images after NIT, that is, the post-operative sample contrast-enhanced CT images.

[0036] It can be understood that these sample patients must be histopathologically diagnosed with NSCLC and the corresponding pathological response is reflected in the postoperative pathology report. Since the output result of the pathological complete remission prediction model is the pathological complete remission result, and the pathological complete remission result can be used to predict whether the patient can achieve pCR after receiving NIT, the sample patients can be divided into a pCR sample group and a non-pCR sample group when selecting.

[0037] To ensure the reliability of the sample medical images, the sample patients need to have received a preset number of cycles (such as 2 - 4 cycles) of NIT.

[0038] Similarly, the sample medical image can be pre-stored in the electronic device or obtained by the electronic device from the outside world. For example, the image data obtained by the electronic device from an external acquisition device, or the image obtained by the electronic device from associated hospitals, medical institutions, etc.

[0039] No specific restrictions are imposed on the specific acquisition form of the sample medical image here, as long as it is ensured that the electronic device can obtain the sample medical image.

[0040] S40. Retain the images in the sample medical image that meet the preset criteria to obtain the medical image to be analyzed.

[0041] In this embodiment, considering that the accuracy of the training data during model training will directly affect the prediction accuracy of the model, when the sample patient has the following situations, the corresponding obtained sample medical image will not be adopted for subsequent model training: Received other anti-tumor treatments; the time interval between the baseline CT scan and the first NIT treatment exceeds the first preset interval (e.g., 4 weeks); the time interval between the preoperative CT scan and the surgery exceeds the second preset interval (e.g., 2 weeks); any situation that hinders tumor segmentation and feature extraction (e.g., there are obvious artifacts, and the lesion has achieved radiological complete remission after NIT).

[0042] Therefore, the preset criteria are as follows: Only received neoadjuvant immunochemotherapy for NSCLC; the time interval between the baseline CT scan and the first treatment does not exceed the first preset interval; the time interval between the preoperative CT scan and the surgery does not exceed the second preset interval; there are no situations in the sample medical image that hinder tumor segmentation and feature extraction.

[0043] S50. Based on the volume of interest (VOI), respectively extract the lesion regions representing the venous phase and the arterial phase from the image to be analyzed, and use the lesion region representing the venous phase to perform matching and calibration of the venous phase and arterial phase contours to obtain the sample target lesion region of the image to be analyzed.

[0044] Specifically, a non-contrast lung window image and a contrast-enhanced mediastinal window image are determined from each medical image to be analyzed. Based on the volume of interest (VOI), a first target lesion area is segmented from the non-contrast lung window image, and a second target lesion area is segmented from the contrast-enhanced mediastinal window image. The first target lesion area segmented from the non-contrast lung window image and the second target lesion area segmented from the contrast-enhanced mediastinal window image can respectively represent the lesion and its surrounding area in the venous phase and the lesion and its surrounding area in the arterial phase.

[0045] In this embodiment, the volume of interest includes areas such as axial, coronal, and sagittal positions. At the same time, the first and second target lesion areas respectively include a first target lesion and a surrounding area (such as blood vessels) within a first preset range centered on the first target lesion, and a second target lesion and a surrounding area (such as enhanced blood vessels) within a second preset range centered on the second target lesion. The first target lesion is the largest lesion in the non-contrast lung window image, and the second target lesion is the largest lesion in the contrast-enhanced mediastinal window image. That is, if there are multiple lesions in the image, the largest one is selected as the target lesion.

[0046] S60. Obtain the sample clinical data of the sample patient. Similarly, the sample clinical data also includes the type of immune checkpoint inhibitor (ICI) of the sample patient, patient age, number of treatment cycles, gender, smoking history, pathological classification (squamous cell carcinoma, adenocarcinoma, and other non-small cell lung cancers), clinical stage, serological markers (such as cytokeratin 19 fragment, neuron-specific enolase, cancer antigen 125, cancer antigen 199, carcinoembryonic antigen, squamous cell carcinoma antigen, etc.), and iRECIST for solid tumor immunotherapy and other data.

[0047] S70. Extract radiomics features from the sample target lesion area to obtain pre-treatment radiomics features and post-treatment radiomics features. Since the image to be analyzed contains pre-treatment and post-treatment image data, the extracted radiomics features include pre-treatment radiomics features and post-treatment radiomics features. All radiomics features are collected to obtain a pre-treatment radiomics feature dataset and a post-treatment radiomics feature dataset.

[0048] In this embodiment, each dataset contains 1037 radiomics features, including: 107 original features, 186 Laplacian of Gaussian (LoG) features, and 744 wavelet features. The original features include: 14 shape features, 18 first-order features, 75 texture features, 24 Gray Level Co-occurrence Matrix (GLCM) features, 14 Gray Level Dependence Matrix (GLDM) features, 16 Gray Level RunLength Matrix (GLRLM) features, 16 Gray Level Size Zone Matrix (GLSZM) features, and 5 Neighboring Gray Tone Difference Matrix (NGTDM) features.

[0049] The LoG features and wavelet features are obtained based on the first-order features and texture features.

[0050] Preferably, the extraction of radiomics features of the sample target lesion area is performed by PyRadiomics in Python version 3.6.

[0051] S80. Based on the radiomics features before and after treatment, construct difference-radiomics features, and extract deep learning features from the post-treatment image features using the trained pre-trained model.

[0052] In this embodiment, the pre-trained model is a 3D Residual Network 18 (ResNet-18) model. The required deep learning features are the activation values from the final fully connected layer (containing 512 neurons), the high-dimensional feature information learned by the pre-trained model. The binary cross-entropy loss function and Adam optimizer (with an initial learning rate set to 0.001) are used to fine-tune the model parameters of the pre-trained model. The batch size of the pre-trained model is set to 8 to balance memory efficiency and gradient stability.

[0053] Among them, the 3D ResNet-18 model includes a convolutional layer, a first pooling layer, a first residual block layer, a second residual block layer, a third residual block layer, a fourth residual block layer, a second pooling layer, and a fully connected layer connected in sequence.

[0054] S90. Select model training features from the pre-treatment radiomics features, post-treatment radiomics features, difference - radiomics features, and deep learning features, and establish a mapping relationship with the treatment effect of the sample patients after receiving NIT based on the model training features and the sample clinical data to obtain the final pathological complete remission prediction model.

[0055] The pathological complete remission prediction method for neoadjuvant chemotherapy of non - small cell lung cancer in the present invention segmented the sample image data before and after NIT based on the data of interest and extracted radiomics features to obtain the radiomics features before and after treatment. Then, three difference - radiomics features were constructed according to the radiomics features before and after treatment. Deep learning features were extracted from the post - treatment image features using a trained pre - trained model. The maximum relevance - minimum redundancy algorithm (mRMR) and LASSO were used to screen out meaningful features from these features. Then, the selected features were combined with the sample clinical data to construct the final pathological complete remission prediction model. The pathological complete remission prediction model constructed in this way combines three types of features: difference - radiomics, deep learning, and clinical features. The combination of these three features provides complementary feature supplementation. Compared with constructing a model using only single - modality information, it can provide better prediction performance. The combination of these three features provides complementary feature supplementation, significantly improving the prediction accuracy. The pathological complete remission prediction model non - invasively characterizes whether NSCLC patients receiving NIT achieve pCR to assist medical staff in formulating personalized treatment plans and ultimately improving the clinical outcomes of non - small cell lung cancer patients.

[0056] In this embodiment, step S50 specifically includes: S51. Determine the corresponding plain - scan lung window image from each medical image to be analyzed based on the first preset window, and determine the corresponding enhanced mediastinal window image from each medical image to be analyzed based on the second preset window.

[0057] It can be understood that the first preset window is the lung window, and the second preset window is the mediastinal window, and each of these two windows has its corresponding window parameters.

[0058] S52. Based on the region of interest, determine the first target lesion region and the second target lesion region from the plain - scan lung window image and the enhanced mediastinal window image respectively.

[0059] S53. Eliminate the non - target regions included in the first target lesion region. The regions that affect subsequent feature extraction are relatively thick blood vessels and trachea regions, that is, the main blood vessels and trachea regions. In this embodiment, the non - target regions are blood vessels and trachea with a diameter exceeding the preset diameter and the surrounding regions within the third preset range.

[0060] It is understandable that both the first preset range and the second preset range are greater than the third preset range, and the first preset range and the second preset range do not need to be the same.

[0061] Preferably, the non-target regions included in the first target lesion region are extracted by a trained target detection model, and the target detection model can be a prediction model that has been actually applied currently and is used to extract the detection frame and target category information of the target object in the image.

[0062] S54. Determine the position information of the first target lesion and the second target lesion corresponding to each image to be analyzed respectively. Based on the position information and taking the first target lesion region as a reference, perform contour matching and calibration on the lesion region contour of each image to be analyzed, and remove noise information such as blood vessels, so as to obtain the sample target lesion region of each image to be analyzed.

[0063] After removing the non-target regions included in the first target lesion region, the main blood vessels and trachea can be avoided during the contour matching and calibration.

[0064] Affected by the breathing amplitude, the image information of the arterial phase may not be accurate enough. Specifically, transfer the second target lesion region representing the lesion in the arterial phase to the first target lesion region representing the lesion in the venous phase according to the contour, and perform contour matching based on the first target lesion region to match the contour determined in the venous phase, so that the surrounding regions that can be displayed in the second preset window can be removed based on the first target lesion region in the first preset window, and the final sample target lesion region of each image to be analyzed combined with the region of interest can be obtained.

[0065] Preferably, a radiologist with more than ten years of experience in diagnosing chest diseases can also fine-tune the contours of the first and second target lesion regions by referring to the historical images in the calibrated historical enhanced CT sequence, so as to manually correct the VOI region. Similarly, the main blood vessels and trachea will be avoided during this process.

[0066] For example, export or save the image to be analyzed in DICOM format, and then determine the plain scan lung window image and the enhanced mediastinal window image from each medical image to be analyzed based on the first preset window and the second preset window.

[0067] Subsequently, the plain scan lung window image and the enhanced mediastinal window image can be imported into the 3D Slicer software, and a radiologist with more than ten years of experience in diagnosing chest diseases can determine the target lesion and perform segmentation of the region of interest.

[0068] In order to avoid the quality of the extracted radiomics features not meeting the standard due to differences in scanning parameters, reconstruction parameters, etc., before the radiomics feature extraction, the method further includes the following steps: Preprocess the medical images to be analyzed. The medical images to be analyzed after preprocessing are standardized image data. It can be understood that subsequently, the required plain chest window images and enhanced mediastinal window images are determined from each medical image to be analyzed after preprocessing.

[0069] In this embodiment, the preprocessing methods include but are not limited to: filtering, voxelization, gray level processing according to preset parameters, discretization, etc.

[0070] To ensure the consistency and accuracy of the lesion regions segmented based on the regions of interest, in this embodiment, step S50 further includes the following steps: S55. Determine the intraclass correlation coefficient (ICC) of the sample target lesion regions, and remove the sample target lesion regions with a correlation lower than the preset correlation according to the intraclass correlation coefficient. For example, configure the preset correlation as ICC value > 0.75. When the ICC value of the sample target lesion region > 0.75, it indicates that the extracted features have strong consistency.

[0071] In this embodiment, step S90 specifically includes: S91. Standardize the pre-treatment radiomics features, post-treatment radiomics features, difference - radiomics features, and deep learning features to obtain standardized features, so that the features are comparable.

[0072] It can be understood that the standardized features include standardized pre-treatment radiomics features, standardized post-treatment radiomics features, standardized difference - radiomics features, and standardized deep learning features.

[0073] S92. Use the Synthetic Minority Oversampling Technique (SMOTE) to perform sample balancing on the standardized features to obtain balanced features, so that the features of the pCR group and the non - pCR group can achieve a 1:1 ratio balance, and delete the features with a variance of 0 and a spearman correlation coefficient exceeding the preset coefficient in the balanced features, so as to eliminate the collinearity problem existing in the features as much as possible.

[0074] It can be understood that the balanced features after sample balancing and feature deletion include balanced pre-treatment radiomics features, balanced post-treatment radiomics features, balanced difference - radiomics features, and balanced deep learning features.

[0075] S93. Use the max-Relevance and Min-Redundancy (mRMR) algorithm to screen out the preliminary training features with high importance from the balanced features. The prediction ability of these preliminary training features is the strongest among all the balanced features.

[0076] It can be understood that the preliminary training features include the screened pre-treatment radiomics features, the screened post-treatment radiomics features, the screened difference - radiomics features, and the screened deep learning features.

[0077] S94. Use the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm to screen out the model training features with high importance from the preliminary training features again, that is, a total of two feature selections are performed through the mRMR algorithm and the LASSO algorithm.

[0078] It can be understood that the preliminary training features include the screened pre-treatment radiomics features, the screened post-treatment radiomics features, the screened difference - radiomics features, and the screened deep learning features.

[0079] More specifically, the LASSO algorithm compresses the regression coefficients through L1 regularization, eliminates redundant features, and retains stable features.

[0080] Among them, after the pre-treatment radiomics features are processed through steps S91 to S94, a total of 1 model training feature is obtained; after the post-treatment radiomics features are processed through steps S91 to S94, a total of 1 model training feature is obtained. After the difference - radiomics features are processed through steps S91 to S94, the first difference - radiomics features obtain a total of 17 model training features, the second difference - radiomics features obtain a total of 10 model training features, and the third difference - radiomics features obtain a total of 9 model training features; after the deep learning features are processed through steps S91 to S94, a total of 7 model training features are obtained.

[0081] S95. According to the model training features and the sample clinical data, and using the logistic regression method, establish the mapping relationship with the treatment effect of the sample patients after receiving neoadjuvant chemotherapy, construct a pathological complete remission prediction model, and the pathological complete remission result output by the pathological complete remission prediction model is the pathological remission status (pCR or non - pCR) of the patient to be predicted.

[0082] For the features obtained from radiomics in the model training features (pre-treatment radiomics features, post-treatment radiomics features, and difference - radiomics features), the radiomics score, i.e., Rad-score, is used to determine the prediction probability of each feature. For the features obtained from deep learning in the model training features (deep learning features), the deep learning score, i.e., Deep-score, is used to determine the prediction probability of each feature. Then, by combining the model training features with the sample clinical data, 5-fold cross-validation is used to fine-tune the model parameters and determine the optimal regularization strength value to ensure the optimal model performance and good generalization ability of the model.

[0083] Please refer to Figures 2 to 9 , and the receiver operating characteristic curve (ROC), the area under the ROC curve (AUC), and the decision curve analysis curve (DCA) are used to evaluate the model performance and clinical utility. Among them, the optimal cut-off value can be determined by the Youden index of the ROC curve, and then key indicators such as accuracy, sensitivity, specificity, positive predictive value, and F1 value are calculated.

[0084] More specifically, a total of 8 models were trained in the training stage, namely Model 1: trained based on the model training features in the pre-treatment radiomics features, Model 2: trained based on the model training features in the post-treatment radiomics features, Model 3: trained based on the model training features in the first difference - radiomics features, Model 4: trained based on the model training features in the second difference - radiomics features, Model 5: trained based on the model training features in the third difference - radiomics features, Model 6: trained based on the model training features in the deep learning features, Model 7: trained based on the model training features in the post-treatment radiomics features and the first difference - radiomics features, Model 8: trained based on all model training features and sample clinical data.

[0085] In this embodiment, the SHAP (Shapley Additive Explanations) method is used to quantify the contribution of each model training feature to the model prediction during logistic regression, realizing a dual mechanism of global interpretation and local interpretation. Among them, global interpretation reveals the behavior pattern of the model in the entire dataset, highlighting the key features of data-driven prediction; local interpretation focuses on the individual cases of patients, showing the impact of specific features on single-case prediction. This dual interpretation mechanism not only ensures the transparency of the decision-making of the pathological complete remission prediction model but also provides a reliable basis for clinical work integration.

[0086] Among them, Figure 7 the SHAP bar chart in Figure 7 shows the contribution degrees of these key features in the model training features of the pathological complete remission prediction model; Figure 8 the SHAP dot plot in Figure 8 presents the direction and magnitude of the influence of each model training feature on the prediction, Figure 8 wherein red indicates an increase in the prediction probability and blue indicates a decrease in the prediction probability; Figure 9 the SHAP force plot in Figure 9 comprehensively shows how the Shapley values of each model training feature affect the prediction result, reflecting the increasing or decreasing effect of each model training feature on the prediction probability relative to the expected output of the model.

[0087] Finally, it is found that the radiomics score (Rad-score) is the core driving factor of the pathological complete remission prediction model, and the deep learning score (Deep-score) provides complementary prediction information. The pathological complete remission prediction model finally selects model 8, which is trained based on all model training features and sample clinical data.

[0088] Next, a device for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer provided by an embodiment of the present invention will be described. The device for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer described below can be mutually corresponded and referred to with the method for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer described above.

[0089] To solve the above problems, in this embodiment, a device for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer is provided, aiming to provide a prediction model that combines difference-radiomics, deep learning, and clinical features. The prediction result of this prediction model non-invasively characterizes whether NSCLC patients receiving NIT achieve pCR, so as to assist medical staff in formulating personalized treatment plans and improving the clinical outcomes of patients. Figure 10 is a schematic structural diagram of a method for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer according to an embodiment of the present invention. As Figure 10 shown, the device may include: A data acquisition module 10, configured to acquire preoperative medical images and preoperative clinical data of a patient to be predicted. In this embodiment, the preoperative medical images are specifically contrast-enhanced CT images before the patient receives NIT, that is, preoperative contrast-enhanced CT images.

[0090] The preoperative medical images may be stored in an electronic device in advance, or may be obtained by the electronic device from the outside. For example, the image data obtained by the electronic device from an external acquisition device, or the data obtained by the electronic device from associated hospitals, medical institutions, etc.

[0091] There is no specific limitation on the specific acquisition form of preoperative medical images here, as long as it is ensured that the electronic device can obtain preoperative medical images.

[0092] Preoperative clinical data includes the type of immune checkpoint inhibitor (ICI), patient age, number of treatment cycles, gender, smoking history, pathological classification (squamous cell carcinoma, adenocarcinoma, and other non-small cell lung cancers), clinical stage, serological markers (such as cytokeratin 19 fragment, neuron-specific enolase, cancer antigen 125, cancer antigen 199, carcinoembryonic antigen, squamous cell carcinoma antigen, etc.), and iRECIST, etc. data. Among them, the above-mentioned preoperative clinical data is obtained from the medical records, medical records, laboratory test results, imaging examination results, etc. of these sample patients.

[0093] It can be understood that the patient to be predicted must be pathologically diagnosed as NSCLC.

[0094] The model prediction module 20 is used to input the preoperative medical images and preoperative clinical data into the trained pathological complete remission prediction model, and the pathological complete remission result of the patient to be predicted is output by the pathological complete remission prediction model.

[0095] Among them, the pathological complete remission prediction model is constructed using pre-treatment radiomics features, post-treatment radiomics features, difference-radiomics features, deep learning features, and sample clinical data. Specifically, the pre-treatment radiomics features and post-treatment radiomics features are radiomics features extracted from the sample medical images based on the region of interest, and the deep learning features are feature information extracted from the post-treatment image features using the trained pre-trained model. The difference-radiomics features include the first difference-radiomics feature, the second-radiomics feature, and the third-radiomics feature. Among them, the first difference-radiomics feature (Delta 1) is used to characterize the absolute change value of the radiomics features before and after neoadjuvant immunochemotherapy, and its calculation method is the pre-treatment radiomics feature minus the post-treatment radiomics feature; the second difference-radiomics feature (Delta 2) is used to characterize the percentage change rate of the radiomics features before and after neoadjuvant immunochemotherapy, and its calculation method is the difference obtained by subtracting the post-treatment radiomics feature from the pre-treatment radiomics feature and then dividing by the pre-treatment radiomics feature; the third difference-radiomics feature (Delta 3) is used for the average percentage change rate of each chemotherapy cycle of the radiomics features before and after neoadjuvant immunochemotherapy, and its calculation method is the difference obtained by subtracting the post-treatment radiomics feature from the pre-treatment radiomics feature divided by the product of the pre-treatment radiomics feature.

[0096] The pathological complete remission prediction device for neoadjuvant chemotherapy of non-small cell lung cancer according to the present invention segmented the sample image data before and after NIT based on the data of interest and extracted radiomics features therefrom to obtain the radiomics features before and after treatment. Then, three difference-radiomics features were constructed based on the radiomics features before and after treatment. The trained pre-trained model was used to extract deep learning features from the post-treatment image features. The maximum relevance-minimum redundancy algorithm (mRMR) and the least absolute shrinkage and selection operator (LASSO) were used to screen out meaningful features from these features. Then, the screened features were combined with the sample clinical data to construct the final pathological complete remission prediction model. The pathological complete remission prediction model constructed in this way combines three types of features: difference-radiomics, deep learning, and clinical features. The combination of these three features provides complementary feature supplementation. Compared with constructing a model using only single-modal information, it can provide better prediction performance. The combination of these three features provides complementary feature supplementation, significantly improving the prediction accuracy. The pathological complete remission prediction model non-invasively characterizes whether NSCLC patients receiving NIT achieve pCR to assist medical staff in formulating personalized treatment plans and ultimately improving the clinical outcomes of non-small cell lung cancer patients.

[0097] Figure 11 Illustrates a schematic physical structure diagram of an electronic device, such as Figure 11 shown, the electronic device may include: a processor 1110, a communication interface 1120, a memory 1130, and a communication bus 1140. Among them, the processor 1110, the communication interface 1120, and the memory 1130 complete mutual communication through the communication bus 1140. The processor 1110 can call the logical commands in the memory 1130 to execute the method for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer, and the method includes: Obtain the preoperative medical image and preoperative clinical data of the patient to be predicted; Input the preoperative medical image and preoperative clinical data into the trained pathological complete remission prediction model, and the pathological complete remission prediction model outputs the pathological complete remission result of the patient to be predicted; The pathological complete remission prediction model is constructed by using pre-treatment radiomics features, post-treatment radiomics features, difference-radiomics features, deep learning features, and sample clinical data; The pre-treatment radiomics features and the post-treatment radiomics features are radiomics features extracted from the sample medical images based on the region of interest, and the deep learning features are feature information extracted from the post-treatment image features by using the trained pre-trained model; The difference-radiomics features are obtained based on the pre-treatment radiomics features and the post-treatment radiomics features. The difference-radiomics features include: The first difference-radiomics feature for characterizing the absolute change value of the radiomics features before and after neoadjuvant immunochemotherapy, the second difference-radiomics feature for characterizing the percentage change rate of the radiomics features before and after neoadjuvant immunochemotherapy, and the third difference-radiomics feature for the average percentage change rate of each chemotherapy cycle of the radiomics features before and after neoadjuvant immunochemotherapy.

[0098] In addition, when the logical instructions in the above-mentioned memory 1130 are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or a part of this technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memor), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0099] On the other hand, the present invention also provides a computer program product. The computer program product includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the method for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer provided by the above-mentioned various methods. The method includes: Obtain the preoperative medical images and preoperative clinical data of the patient to be predicted; Input the preoperative medical images and preoperative clinical data into the trained pathological complete remission prediction model, and the pathological complete remission prediction model outputs the pathological complete remission result of the patient to be predicted; The pathological complete remission prediction model is constructed by using pre-treatment radiomics features, post-treatment radiomics features, difference-radiomics features, deep learning features, and sample clinical data; The pre-treatment radiomics features and the post-treatment radiomics features are radiomics features extracted from the sample medical images based on the region of interest. The deep learning features are feature information extracted from the post-treatment image features by using a trained pre-trained model; The difference-radiomics features are obtained based on the pre-treatment radiomics features and the post-treatment radiomics features. The difference-radiomics features include: The first difference-radiomics feature for characterizing the absolute change value of the radiomics features before and after neoadjuvant immunochemotherapy, the second difference-radiomics feature for characterizing the percentage change rate of the radiomics features before and after neoadjuvant immunochemotherapy, and the third difference-radiomics feature for the average percentage change rate of the radiomics features in each chemotherapy cycle before and after neoadjuvant immunochemotherapy.

[0100] In another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it is implemented to execute the above-provided method for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer. The method includes: Obtain the preoperative medical images and preoperative clinical data of the patient to be predicted; Input the preoperative medical images and preoperative clinical data into the trained pathological complete remission prediction model, and the pathological complete remission prediction model outputs the pathological complete remission result of the patient to be predicted; The pathological complete remission prediction model is constructed by using pre-treatment radiomics features, post-treatment radiomics features, difference-radiomics features, deep learning features, and sample clinical data; The pre-treatment radiomics features and the post-treatment radiomics features are radiomics features extracted from the sample medical images based on the region of interest. The deep learning features are feature information extracted from the post-treatment image features by using the trained pre-trained model; The difference-radiomics features are obtained based on the pre-treatment radiomics features and the post-treatment radiomics features. The difference-radiomics features include: The first difference-radiomics feature for characterizing the absolute change value of the radiomics features before and after neoadjuvant immunochemotherapy, the second difference-radiomics feature for characterizing the percentage change rate of the radiomics features before and after neoadjuvant immunochemotherapy, and the third difference-radiomics feature for the average percentage change rate of the radiomics features in each chemotherapy cycle before and after neoadjuvant immunochemotherapy.

[0101] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer, characterized in that, The method includes: Obtaining the preoperative medical images and preoperative clinical data of the patient to be predicted; Inputting the preoperative medical images and preoperative clinical data into the trained pathological complete remission prediction model, and outputting the pathological complete remission result of the patient to be predicted by the pathological complete remission prediction model; The pathological complete remission prediction model is constructed by using pre-treatment radiomics features, post-treatment radiomics features, difference-radiomics features, deep learning features, and sample clinical data; The pre-treatment radiomics features and post-treatment radiomics features are radiomics features extracted from the sample medical images based on the region of interest, and the deep learning features are feature information extracted from the post-treatment image features by using the trained pre-trained model; The difference-radiomics features are calculated according to the pre-treatment radiomics features and post-treatment radiomics features, and the difference-radiomics features include: The first difference-radiomics feature for characterizing the absolute change value of the radiomics features before and after neoadjuvant immunochemotherapy, the second difference-radiomics feature for characterizing the percentage change rate of the radiomics features before and after neoadjuvant immunochemotherapy, and the third difference-radiomics feature for the average percentage change rate of the radiomics features in each chemotherapy cycle before and after neoadjuvant immunochemotherapy.

2. The pathological complete remission prediction method for neoadjuvant chemotherapy of non-small cell lung cancer according to claim 1, wherein, The pathological complete remission prediction model is trained through the following steps: Obtaining the sample medical images of the sample patients; the sample medical images include contrast-enhanced computed tomography images before receiving neoadjuvant chemotherapy and contrast-enhanced computed tomography images after receiving neoadjuvant chemotherapy; Retaining the images that meet the preset criteria in the sample medical images to obtain the medical images to be analyzed; Based on the region of interest, respectively extracting the lesion regions representing the venous phase and arterial phase from the images to be analyzed, and using the lesion region representing the venous phase to perform matching calibration of the venous phase and arterial phase contours to obtain the sample target lesion region of the images to be analyzed; Obtaining the sample clinical data of the sample patients; Extracting radiomics features from the sample target lesion region to obtain pre-treatment radiomics features and post-treatment radiomics features; Based on the pre-treatment and post-treatment radiomics features, constructing difference-radiomics features, and using the trained pre-trained model to extract deep learning features from the post-treatment image features; Screening out the model training features from the pre-treatment radiomics features, post-treatment radiomics features, difference-radiomics features, and deep learning features, and establishing a mapping relationship with the treatment effect after the sample patients receive neoadjuvant immunochemotherapy according to the model training features and sample clinical data to obtain the pathological complete remission prediction model.

3. The method for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer according to claim 2, wherein The step of respectively extracting the lesion regions representing the venous phase and arterial phase from the images to be analyzed based on the region of interest, and using the lesion region representing the venous phase to perform matching calibration of the venous phase and arterial phase contours to obtain the sample target lesion region of the images to be analyzed specifically includes: Determining the corresponding plain lung window images from each medical image to be analyzed based on the first preset window, and determining the corresponding enhanced mediastinal window images from each medical image to be analyzed based on the second preset window; Based on the region of interest, the first target lesion region and the second target lesion region are respectively determined from the plain lung window image and the enhanced mediastinal window image; the first target lesion region and the second target lesion region respectively include the first target lesion and the peripheral region within the first preset range centered on the first target lesion, the second target lesion and the peripheral region within the second preset range centered on the second target lesion; the first target lesion is the largest lesion in the plain lung window image, and the second target lesion is the largest lesion in the enhanced mediastinal window image; Eliminate the non-target regions included in the first target lesion region; Determine the position information of the first target lesion and the second target lesion corresponding to each image to be analyzed respectively. Based on the position information and taking the first target lesion region as the reference, perform matching and calibration of the lesion region contours of each image to be analyzed to obtain the sample target lesion region of each image to be analyzed.

4. The method for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer according to claim 3, wherein The step of respectively extracting the lesion regions representing the venous phase and the arterial phase from the images to be analyzed based on the region of interest, and using the lesion region representing the venous phase to perform matching and calibration of the venous phase and arterial phase contours to obtain the sample target lesion region of the image to be analyzed further includes the following steps: Determine the intraclass correlation coefficient of the sample target lesion region, and eliminate the sample target lesion regions with a correlation lower than the preset correlation according to the intraclass correlation coefficient.

5. The method for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer according to claim 3, wherein The pathological complete remission prediction model is also trained through the following steps: Preprocess the sample medical images.

6. The method for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer according to claim 2, wherein Screen out the model training features from the pre-treatment imaging features, post-treatment imaging features, difference-imaging features, and deep learning features, and establish a mapping relationship with the treatment effect of the sample patients after receiving neoadjuvant immunochemotherapy according to the model training features and the sample clinical data to obtain the pathological complete remission prediction model, specifically including: Perform standardization processing on the pre-treatment imaging features, post-treatment imaging features, difference-imaging features, and deep learning features to obtain standardized features; Use the synthetic minority over-sampling technique to perform sample balancing processing on the standardized features to obtain balanced features, and delete the features with a variance of 0 and a Spearman rank correlation coefficient exceeding the preset coefficient in the balanced features; Use the maximum correlation-minimum redundancy algorithm to screen out the preliminary training features from the balanced features; Use the least absolute shrinkage and selection operator to screen out the model training features from the preliminary training features; According to the model training features and the sample clinical data and using the logistic regression method, establish a mapping relationship with the treatment effect of the sample patients after receiving neoadjuvant immunochemotherapy to obtain the pathological complete remission prediction model.

7. The method for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer according to claim 2, wherein The preset criteria are: Only received neoadjuvant immunochemotherapy for non-small cell lung cancer; the time interval between the baseline computed tomography scan and the first treatment does not exceed the first preset interval; the time interval between the preoperative computed tomography scan and the surgery does not exceed the second preset interval; there are no situations in the sample medical images that prevent tumor segmentation and feature extraction.

8. A pathological complete remission prediction device for neoadjuvant chemotherapy of non-small cell lung cancer, characterized in that, The device includes: A data acquisition module for acquiring the preoperative medical images and preoperative clinical data of the patient to be predicted; A model prediction module, configured to input pre-operative medical images and pre-operative clinical data into a trained pathological complete remission prediction model, and output the pathological complete remission result of the patient to be predicted by the pathological complete remission prediction model; The pathological complete remission prediction model is constructed by using pre-treatment radiomics features, post-treatment radiomics features, difference-radiomics features, deep learning features, and sample clinical data; The pre-treatment radiomics features and the post-treatment radiomics features are radiomics features extracted from the sample medical images based on the region of interest, and the deep learning features are feature information extracted from the post-treatment image features by using a trained pre-trained model; The difference-radiomics features are obtained based on the pre-treatment radiomics features and the post-treatment radiomics features, and the difference-radiomics features include: A first difference-radiomics feature for characterizing the absolute change value of the radiomics features before and after neoadjuvant immunochemotherapy, a second difference-radiomics feature for characterizing the percentage change rate of the radiomics features before and after neoadjuvant immunochemotherapy, and a third difference-radiomics feature for the average percentage change rate of each chemotherapy cycle of the radiomics features before and after neoadjuvant immunochemotherapy.

9. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, the steps of the method for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer according to any one of claims 1 to 7 are implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the steps of the method for predicting pathological complete remission of neoadjuvant chemotherapy for non-small cell lung cancer according to any one of claims 1 to 7 are implemented.

Images