A method for predicting neoadjuvant chemotherapy efficacy of breast cancer based on imagingomics

Abstract

The application discloses a breast cancer neoadjuvant chemotherapy effect prediction method based on imagingomics, comprising the following steps: S1, acquiring double-mode original image data of a breast cancer patient before chemotherapy, wherein the double-mode original image data comprises breast MR image data and breast special PET image data, thereby, the prediction accuracy is significantly improved by fusing multi-parameter breast MR and breast special PET double-mode images, the data set adopted is constructed to cover breast cancer cases of different molecular subtypes and different tumor stages, the generalization ability and clinical adaptability of the model are greatly improved, the optimal double-mode imagingomics features are fused with patient clinical pathological features and immunohistochemical indexes, an image-clinical joint prediction model is constructed, the tumor microscopic image features and patient individual clinical differences are considered, and multiple dimensions of key factors affecting the chemotherapy effect are comprehensively covered, so that objective quantitative basis can be provided for optimizing the breast cancer neoadjuvant chemotherapy scheme.

Description

Technical Field

[0001] This application relates to the technical field of medical imaging, and more particularly to a method for predicting the efficacy of neoadjuvant chemotherapy for breast cancer based on radiomics. Background Technology

[0002] Breast cancer is the most common malignant tumor among women, seriously threatening their lives, health, and quality of life. Neoadjuvant chemotherapy is a treatment for locally advanced breast cancer and early-stage high-risk breast cancer, effectively shrinking tumor lesions and reducing tumor stage, creating conditions for surgical resection and breast-conserving therapy. With the development of radiomics technology, non-invasive efficacy prediction schemes based on medical imaging are gradually being applied to the field of evaluating the efficacy of neoadjuvant chemotherapy for breast cancer. By mining high-throughput quantitative features in imaging data, the biological characteristics of the tumor can be objectively reflected. However, current radiomics prediction methods still have many shortcomings.

[0003] First, existing radiomics feature extraction and screening processes are rudimentary, resulting in insufficient effective feature selection and an excessive number of redundant features. Current technologies extract only a limited range of image features, often simply statistically analyzing basic texture and morphological features, neglecting crucial features that reflect the microscopic, hidden characteristics of tumors, such as higher-order wavelet features and metabolic heterogeneity features. This fails to fully uncover potential imaging biomarkers related to chemotherapy efficacy.

[0004] Secondly, most current research models have limited sample selection in their training datasets, focusing mainly on breast cancer patients with single molecular subtypes and single tumor stages. This narrow sample coverage does not reflect the complex subtypes and diverse stages of breast cancer patients in clinical practice, resulting in overly targeted models with highly fluctuating accuracy in predicting efficacy for patients with different subtypes and stages.

[0005] Furthermore, most existing mainstream prediction models are purely radiomics-based, which suffer from limitations such as limited predictive dimensions and insufficient clinical relevance. The efficacy of neoadjuvant chemotherapy for breast cancer is not only related to the microscopic features of tumor imaging, but also closely related to various clinical factors such as the patient's clinical baseline characteristics and immunohistochemical expression levels. A single imaging feature cannot comprehensively cover all the variables affecting the efficacy of chemotherapy.

[0006] Application content

[0007] This application aims to address, at least to some extent, the technical problems in the related art.

[0008] To achieve the above objectives, this application proposes a radiomics-based method for predicting the efficacy of neoadjuvant chemotherapy in breast cancer, comprising the following steps:

[0009] S1. Obtain bimodal raw image data of the breast cancer patient to be predicted before chemotherapy, wherein the bimodal raw image data includes breast MR image data and breast-specific PET image data;

[0010] S2. The breast MR image data and breast-specific PET image data are preprocessed respectively to obtain standardized dual-modal image data after registration and alignment;

[0011] S3. Based on the preprocessed standardized bimodal image data, the region of interest of breast cancer lesions is delineated, and breast MR radiomics features and breast-specific PET radiomics features are extracted to form a bimodal original radiomics feature set.

[0012] S4. Perform feature filtering and optimization on the original bimodal image omics feature set, remove redundant and invalid features, and obtain the optimal bimodal image omics feature subset;

[0013] S5. Input the optimal bimodal radiomics feature subset into the pre-trained neoadjuvant chemotherapy efficacy prediction model for breast cancer, and output the neoadjuvant chemotherapy efficacy prediction results for the patient. The efficacy prediction results include two types of evaluation results: pathological complete remission and non-pathological complete remission.

[0014] In addition, the application may also include the following additional technical features:

[0015] Specifically, in step S1, the breast MR imaging data is multi-parameter breast MR imaging data, including dynamic contrast-enhanced DCE-MRI, diffusion-weighted DWI-MRI, and apparent diffusion coefficient ADC imaging data; the breast-specific PET imaging data is breast-specific positron emission tomography glucose metabolism PET imaging data, including the maximum standardized uptake value of the lesion and the original imaging parameters of metabolic volume.

[0016] Specifically, the preprocessing process in step S2 includes:

[0017] Denoising, grayscale normalization, and slice interpolation resampling were performed on breast MR images and breast-specific PET images respectively to unify the spatial resolution, pixel spacing, and image dimensions of the two types of images.

[0018] Using breast MR images as the reference images, a rigid registration combined with a local elastic registration algorithm is used to perform pixel-level precise registration between breast-specific PET images and breast MR images, eliminating spatial positional deviations between the two modal images.

[0019] By using image thresholding to remove chest wall, breast fat, and blood vessel-irrelevant tissue areas, pure breast lesion image areas are preserved, thus completing the standardized preprocessing of dual-modal images.

[0020] Specifically, in step S3, the region of interest is the solid area of ​​the breast cancer tumor obtained by combining artificial and intelligent segmentation, excluding the normal breast tissue and edema area surrounding the tumor.

[0021] The breast MR radiomics features include first-order statistical features, second-order texture features, morphological features, and higher-order wavelet transform features; the breast-specific PET radiomics features include metabolic intensity features, metabolic heterogeneity texture features, and lesion metabolic distribution morphology features.

[0022] Specifically, the feature selection and optimization in step S4 includes:

[0023] The variance thresholding method is used to remove low-discrimination features with variances below a preset threshold, thus completing the initial screening.

[0024] The Pearson correlation coefficient method was used to remove highly correlated redundant features and retain independent features.

[0025] Based on the LASSO regression algorithm with minimum absolute contraction and selection operator, combined with ten-fold cross-validation, features highly correlated with the efficacy of neoadjuvant chemotherapy for breast cancer were screened out, forming a subset of bimodal radiomics features.

[0026] Specifically, in step S5, the training and construction process of the neoadjuvant chemotherapy efficacy prediction model for breast cancer includes:

[0027] Training, validation, and test datasets were constructed. The dataset samples were all breast cancer patients who received standard neoadjuvant chemotherapy. All samples retained pre-chemotherapy breast MR and breast-specific PET dual-modal imaging data and postoperative pathological efficacy gold standard results.

[0028] All samples in the dataset are processed according to steps S2-S4 to obtain the optimal bimodal image omics feature subset for each sample;

[0029] Using the optimal bimodal radiomics feature subset as input and the chemotherapy efficacy results of the gold standard in pathology as labels, an initial prediction model is constructed based on random forest, support vector machine or deep learning network algorithm;

[0030] The model hyperparameters were iteratively optimized using a validation dataset. The model's accuracy, AUC value, specificity, and sensitivity were validated based on a test dataset. The model with the best performance was selected as the final predictive model for the efficacy of neoadjuvant chemotherapy in breast cancer.

[0031] Specifically, the dataset samples cover breast cancer patients with different molecular subtypes and different tumor stages, and the pathological gold standard results are used to determine pathological complete remission or non-pathological complete remission based on the pathological evaluation criteria for neoadjuvant chemotherapy in breast cancer.

[0032] Specifically, it also includes: fusing the patient's clinicopathological features, immunohistochemical indicators, and the optimal dual-modal radiomics feature subset to construct an image-clinical joint prediction model, thereby achieving accurate prediction of the efficacy of neoadjuvant chemotherapy through multi-dimensional fusion.

[0033] The beneficial effects of the radiomics-based method for predicting the efficacy of neoadjuvant chemotherapy for breast cancer proposed in this application are as follows:

[0034] 1. It integrates multi-parameter breast MR and breast-specific PET dual-modal images, and simultaneously extracts multiple quantitative parameters such as morphology, texture, statistics, wavelet higher-order features, and tumor metabolic intensity and metabolic heterogeneity. This fully explores the hidden heterogeneity information of the tumor microscopic layer, ensures the high effectiveness of the features input to the model, and significantly improves the prediction accuracy.

[0035] 2. The dataset used in the construction covers breast cancer cases with different molecular subtypes and tumor stages, which closely matches the distribution characteristics of real clinical cases. By combining multi-algorithm modeling, hyperparameter iterative optimization, and multi-dimensional index verification and selection, stable and reliable efficacy prediction can be achieved for various types of breast cancer patients, which greatly improves the generalization ability and clinical adaptability of the model.

[0036] 3. By integrating the optimal dual-modal radiomics features with the patient's clinicopathological features and immunohistochemical indicators, an imaging-clinical joint prediction model is constructed. This model takes into account both the tumor's microscopic imaging features and the individual clinical differences of patients, comprehensively covering the multi-dimensional key factors affecting the efficacy of chemotherapy. It can provide objective quantitative evidence for optimizing neoadjuvant chemotherapy regimens for breast cancer in clinical practice. Attached Figure Description

[0037] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0038] Figure 1 This is a schematic diagram of a radiomics-based method for predicting the efficacy of neoadjuvant chemotherapy for breast cancer, as proposed in this application.

[0039] Figure 2 This is a flowchart illustrating the feature screening and optimization process for a radiomics-based method for predicting the efficacy of neoadjuvant chemotherapy for breast cancer, as described in this application.

[0040] Figure 3 This is a flowchart illustrating the model construction process for a radiomics-based method for predicting the efficacy of neoadjuvant chemotherapy for breast cancer, as described in this application. Detailed Implementation

[0041] To make the technical means, inventive features, objectives, and effects of this application easier to understand, the application is further described below with reference to specific illustrations. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0042] The present application will now be described in further detail with reference to the accompanying drawings.

[0043] like Figure 1 As shown in the figure, an embodiment of this application provides a method for predicting the efficacy of neoadjuvant chemotherapy for breast cancer based on radiomics, which includes the following steps:

[0044] S1. Obtain bimodal raw imaging data of patients with breast cancer to be predicted before chemotherapy. The bimodal raw imaging data includes breast MR imaging data and breast-specific PET imaging data.

[0045] It should be noted that the specific subjects of the data collection were patients diagnosed with breast cancer who were scheduled to receive neoadjuvant chemotherapy. All imaging data were collected before the patients started neoadjuvant chemotherapy to avoid changes in lesion characteristics caused by chemotherapy drugs and to ensure the authenticity of the data and the accuracy of the prediction benchmark.

[0046] A dual-modal image fusion analysis scheme was employed, with the acquired raw image data specifically including breast MR imaging data and breast-specific PET imaging data. Breast MR imaging offers the advantage of high soft tissue resolution, clearly presenting the morphology, boundary extent, degree of infiltration, size, and involvement of surrounding glandular tissue in breast lesions, accurately capturing the anatomical features of the lesions. Breast-specific PET imaging, based on metabolic imaging principles, effectively reflects the functional characteristics of breast cancer lesions, such as cell proliferation activity and glucose metabolism levels, and can identify microlesions, occult metastases, and the malignant activity of lesions. The two modalities complement each other, covering both lesion anatomical and functional metabolic information.

[0047] S2. Preprocess the breast MR image data and breast-specific PET image data separately to obtain standardized dual-modal image data after registration and alignment.

[0048] It should be noted that separate preprocessing was performed for each of the two types of image data before fusion and registration. First, the raw breast MR images were preprocessed to eliminate image interference caused by equipment noise and motion artifacts, and to unify the spatial resolution, grayscale, and pixel size of the images. Then, the raw breast-specific PET images underwent corresponding preprocessing, including metabolic value correction, background noise removal, image smoothing, and resolution standardization to correct metabolic value deviations in the PET images and optimize image clarity. After the single-modal image standardization was completed, a high-precision image registration algorithm was used, with the breast anatomy as a reference, to precisely register and align the preprocessed MR and PET images at the pixel level and spatial dimension, eliminating spatial positional deviations between the two types of images and obtaining standardized dual-modal image data.

[0049] S3. Based on the preprocessed standardized bimodal image data, the region of interest for breast cancer lesions is delineated, and breast MR radiomics features and breast-specific PET radiomics features are extracted to form a bimodal original radiomics feature set.

[0050] It should be noted that the delineation work can be completed by senior radiologists using a combination of manual delineation and intelligent assisted segmentation algorithms, based on pathological diagnosis results and imaging anatomical features. This ensures that the delineation closely matches the actual boundaries of the breast cancer lesion, covers the main tumor area, and avoids irrelevant tissue areas such as normal breast glands, fat, blood vessels, and chest wall, thus guaranteeing the accuracy and uniqueness of the region of interest and preventing irrelevant areas from interfering with the feature extraction results.

[0051] After accurately delineating the ROI region, multi-dimensional radiomics features were extracted for both modal images. For breast MR images, the focus was on extracting morphological features, texture features, gray-level statistical features, and wavelet transform features of the lesions, specifically including anatomical morphology and gray-level texture quantification features such as lesion perimeter, area, volume, roundness, concavity / convexity, gray-level mean, gray-level variance, texture roughness, contrast, and correlation.

[0052] For breast-specific PET images, the focus is on extracting metabolic-related radiomics features of lesions, including maximum standardized uptake, mean standardized uptake, lesion metabolic volume, total lesion glycolysis, metabolic texture features, and metabolic heterogeneity features. The extracted MR radiomics features are then integrated and summarized with the PET radiomics features, completing feature classification and preliminary organization to construct a dual-modal raw radiomics feature set encompassing multidimensional features of breast cancer lesions, including macroscopic morphology, microscopic texture, and metabolic activity.

[0053] S4. Perform feature screening and optimization on the original bimodal image omics feature set, remove redundant and invalid features, and obtain the optimal bimodal image omics feature subset.

[0054] It should be noted that a multi-level optimization strategy of preliminary filtering, correlation analysis, and precise screening was adopted to purify the original feature set layer by layer. First, preliminary filtering was carried out to remove invalid features with too many missing values, numerical variance close to zero, and no discriminative power. These types of features cannot reflect the differences in lesions among different patients and do not contribute to the prediction of treatment efficacy.

[0055] Subsequently, feature correlation analysis was performed. Using methods such as Pearson correlation coefficient and variance inflation factor, highly correlated redundant features were eliminated to avoid the repeated accumulation of similar features that could interfere with model training. Finally, machine learning feature selection algorithms were combined, relying on feature importance scoring, recursive feature elimination, and LASSO regularization to select features that were highly correlated with the efficacy of neoadjuvant chemotherapy for breast cancer, possessed strong discriminative ability, and exhibited high stability. Key features with significant predictive value for both pathological and non-pathological complete remission outcomes were retained.

[0056] After multiple rounds of screening, redundancy removal, noise reduction, optimization and verification, all invalid, redundant and low-contribution features were eliminated, and the optimal dual-modal image omics feature subset was finally obtained. While retaining the core prediction information, the feature dimension was significantly reduced to meet the input requirements of the prediction model.

[0057] S5. Input the optimal bimodal radiomics feature subset into the pre-trained neoadjuvant chemotherapy efficacy prediction model for breast cancer, and output the patient's neoadjuvant chemotherapy efficacy prediction results. The efficacy prediction results include two types of evaluation results: pathological complete remission and non-pathological complete remission.

[0058] It should be noted that during the model training phase, a large sample size of preoperative bimodal imaging feature data, clinical pathology data, and neoadjuvant chemotherapy follow-up efficacy data of breast cancer patients were collected to complete the iterative optimization of model parameters. At the same time, model validation was completed through cross-validation, external dataset testing, and other methods to ensure that the model has good fitting ability, generalization ability, and prediction accuracy, and can stably adapt to the lesion feature data of different patients.

[0059] After the optimal bimodal radiomics feature subset obtained in step S4 is adapted and normalized according to the model input specifications, it is input into the pre-trained efficacy prediction model. The model, through its built-in algorithm logic, performs deep computation, feature correlation analysis, and efficacy probability discrimination on the input radiomics features, automatically uncovering the potential correlation between lesion imaging features and neoadjuvant chemotherapy efficacy, and outputting the patient's neoadjuvant chemotherapy efficacy prediction results.

[0060] Complete pathological remission means that after neoadjuvant chemotherapy, there are no residual invasive cancer cells in the primary breast tumor and regional lymph nodes, and the chemotherapy effect is excellent; non-complete pathological remission means that there are still residual tumor lesions after chemotherapy, and the chemotherapy effect has not reached the standard of complete remission.

[0061] In one embodiment of this application, in step S1, the breast MR imaging data is multi-parameter breast MR imaging data, including dynamic contrast-enhanced DCE-MRI, diffusion-weighted DWI-MRI, and apparent diffusion coefficient ADC imaging data; the breast-specific PET imaging data is breast-specific positron emission tomography glucose metabolism PET imaging data, including the maximum standardized uptake value of the lesion and the raw imaging parameters of metabolic volume.

[0062] It should be noted that DCE-MRI can dynamically capture the blood perfusion characteristics, angiogenesis, and contrast agent metabolism patterns of breast lesions, accurately reflecting the proliferative activity and malignancy of tumor tissue; DWI-MRI can detect the restriction of free diffusion of water molecules within the lesion, effectively distinguishing between solid tumor lesions, edematous tissue, and normal glandular tissue; ADC imaging data are quantitative derivatives of DWI-MRI, enabling quantitative characterization of lesion diffusion features and providing quantitative numerical basis for subsequent feature extraction.

[0063] This PET (Positron Emission Tomography) imaging data of glucose metabolism is used for precise scanning of breast lesions. The images are focused on the breast lesion area, offering higher scanning resolution and superior data accuracy, enabling precise capture of abnormal glucose metabolism characteristics in breast cancer lesions. The core image parameters acquired include the maximum normalized uptake value (MNU) of the lesion and raw metabolic volume parameters. The MNU quantifies the glucose metabolic activity of tumor tissue, directly reflecting the proliferation and invasiveness of tumor cells; the metabolic volume parameters precisely delineate the actual spatial extent of highly metabolic tumor lesions, fully covering the active tumor area, providing core raw data support for subsequent analysis of lesion metabolic heterogeneity and evaluation of chemotherapy efficacy.

[0064] In one embodiment of this application, the preprocessing step S2 specifically includes:

[0065] Denoising, grayscale normalization, and slice interpolation resampling were performed on breast MR images and breast-specific PET images respectively to unify the spatial resolution, pixel spacing, and image dimensions of the two types of images.

[0066] Using breast MR images as the reference images, a rigid registration combined with a local elastic registration algorithm is used to perform pixel-level precise registration between breast-specific PET images and breast MR images, eliminating spatial positional deviations between the two modal images.

[0067] By using image thresholding to remove chest wall, breast fat, and blood vessel-irrelevant tissue areas, pure breast lesion image areas are preserved, thus completing the standardized preprocessing of dual-modal images.

[0068] It should be noted that the denoising process uses an adaptive Gaussian filtering algorithm, which can effectively remove the loss of lesion information caused by excessive filtering during image acquisition; the grayscale normalization process maps the grayscale values ​​of the two types of images to a fixed range, eliminating the grayscale scale differences caused by different scanning devices and different scanning parameters, and achieving uniform image grayscale standards; the slice interpolation resampling uses a bicubic interpolation algorithm to uniformly correct the slice thickness and pixel size of the two types of images, and standardize the spatial resolution, pixel spacing and image dimensions of the two types of images.

[0069] Rigid registration is responsible for coarse matching of overall image translation, rotation, and scaling, eliminating overall spatial positional offset between the two types of images; local elastic registration is designed for the deformation characteristics of breast soft tissue, performing fine deformation correction on local lesion areas, accurately adapting to the physiological deformation differences of breast soft tissue, achieving pixel-level precise registration between breast PET images and MR images, ensuring that the same lesion area corresponds completely in the two types of images.

[0070] An adaptive image thresholding segmentation algorithm is used to remove background regions from the registered fused images. Combined with breast anatomical features, irrelevant tissue areas such as chest wall muscles, breast fat, and normal blood vessels are precisely segmented and removed, avoiding interference from non-lesion tissues in subsequent feature extraction. The entire preprocessing process retains only highly pure breast lesion image regions, achieving standardized and refined preprocessing of dual-modal images, maximizing the purification of effective data, and improving the effectiveness and accuracy of subsequent radiomics features.

[0071] In one embodiment of this application, in step S3, the region of interest is the solid area of ​​the breast cancer tumor obtained by combining artificial and intelligent segmentation, excluding the normal breast tissue and edema area surrounding the tumor.

[0072] Among them, breast MR radiomics features include first-order statistical features, second-order texture features, morphological features, and higher-order wavelet transform features; breast-specific PET radiomics features include metabolic intensity features, metabolic heterogeneity texture features, and lesion metabolic distribution morphology features.

[0073] It should be noted that intelligent segmentation relies on a deep learning segmentation model to achieve rapid initial segmentation of the lesion area and accurately locate the core area of ​​the tumor; manual delineation is carried out by senior radiologists who combine pathological results and imaging features to finely correct the segmentation boundaries, remove edema and normal tissue interference areas, and achieve high-precision calibration of the lesion's solid area.

[0074] Breast MR imaging features encompass four categories: first-order statistical features primarily reflect the overall distribution pattern of grayscale in lesion images, including basic quantitative parameters such as grayscale mean, variance, skewness, and kurtosis, characterizing the overall grayscale differences of lesions; second-order texture features focus on the spatial distribution and correlation patterns of pixels within lesions, including parameters such as contrast, correlation, entropy, and roughness, reflecting the texture heterogeneity within tumors; morphological features quantify the geometric morphological characteristics of tumor lesions, covering indicators such as lesion area, perimeter, volume, roundness, and concavity / convexity, reflecting the growth morphology and invasive characteristics of tumors; and higher-order wavelet transform features, through multi-scale wavelet decomposition, uncover hidden high-frequency and low-frequency features in images, capturing subtle pathological differences in lesions that cannot be identified by conventional features.

[0075] Metabolic intensity characteristics, with standardized uptake values ​​as the core, include a series of parameters such as mean, peak, and extreme values ​​to quantify the overall metabolic activity of the tumor; metabolic heterogeneity texture characteristics analyze the spatial distribution differences of glucose metabolism in different regions within the lesion to uncover the heterogeneous characteristics of metabolism within the tumor, accurately reflecting the heterogeneous characteristics of tumor cell proliferation and differentiation; lesion metabolic distribution morphology characteristics combine the geometric shape and distribution range of metabolic regions to characterize the spatial growth pattern of hypermetabolic tumor lesions, comprehensively covering the metabolic pathological characteristics of breast cancer lesions.

[0076] In one embodiment of this application, such as Figure 2 As shown, the feature selection and optimization in step S4 specifically includes:

[0077] The variance thresholding method is used to remove low-discrimination features with variances below a preset threshold, thus completing the initial screening.

[0078] The Pearson correlation coefficient method was used to remove highly correlated redundant features and retain independent features.

[0079] Based on the LASSO regression algorithm with minimum absolute contraction and selection operator, combined with ten-fold cross-validation, features highly correlated with the efficacy of neoadjuvant chemotherapy for breast cancer were screened out, forming a subset of bimodal radiomics features.

[0080] It should be noted that the first stage is the initial feature screening, which uses a variance thresholding method to remove low-quality features. By batch calculating the variance of all extracted bimodal radiomics features across the entire sample set, a reasonable preset variance threshold is set, and features with variance values ​​below the threshold are directly removed. These low-variance features exhibit extremely small numerical fluctuations across different patients and different treatment efficacy samples, resulting in very low discriminative power. They cannot effectively reflect differences in lesions and chemotherapy efficacy, and are therefore considered ineffective features. Removing them can initially reduce the feature dimensionality and simplify subsequent calculations.

[0081] The second stage involves redundant feature removal, using the Pearson correlation coefficient method to screen for feature independence. Pairwise correlation analysis is performed on the features retained after the initial screening. The Pearson correlation coefficient between any two features is calculated. When the correlation coefficient is higher than a preset high correlation threshold, the two features are considered to have high information overlap and extremely high redundancy. At this point, the core features with higher relevance to the efficacy label are retained, while overlapping redundant features are removed. This maximizes the retention of independent and effective features, avoids feature collinearity issues interfering with subsequent model training, and improves the overall effectiveness of the feature set.

[0082] The third stage involves core feature optimization. Based on the LASSO regression algorithm with minimum absolute shrinkage and selection operator, and combined with a ten-fold cross-validation mechanism, high-precision feature selection is achieved. Leveraging the variable compression and feature selection advantages of LASSO regression, independent features are iteratively optimized in terms of weights. Simultaneously, ten-fold cross-validation avoids the randomness of a single selection, preventing overfitting and underfitting issues. Ultimately, core features with statistically significant correlations and the highest predictive value for the efficacy of neoadjuvant chemotherapy in breast cancer are selected. This results in a dual-modal radiomics feature subset with optimal dimensionality, strongest correlation, and best stability, which serves as the core input data for subsequent model training.

[0083] In one embodiment of this application, such as Figure 3 As shown, step S5, the training and construction process of the neoadjuvant chemotherapy efficacy prediction model for breast cancer includes:

[0084] Training, validation, and test datasets were constructed. The dataset samples were all breast cancer patients who received standard neoadjuvant chemotherapy. All samples retained pre-chemotherapy breast MR and breast-specific PET dual-modal imaging data and postoperative pathological efficacy gold standard results.

[0085] All samples in the dataset are processed according to steps S2-S4 to obtain the optimal bimodal image omics feature subset for each sample;

[0086] Using the optimal bimodal radiomics feature subset as input and the chemotherapy efficacy results of the gold standard in pathology as labels, an initial prediction model is constructed based on random forest, support vector machine or deep learning network algorithm;

[0087] The model hyperparameters were iteratively optimized using a validation dataset. The model's accuracy, AUC value, specificity, and sensitivity were validated based on a test dataset. The model with the best performance was selected as the final predictive model for the efficacy of neoadjuvant chemotherapy in breast cancer.

[0088] It should be noted that, firstly, training datasets, validation datasets, and test datasets were constructed according to the standard proportions. The samples in all three datasets were strictly selected from breast cancer patients diagnosed with the clinically standardized neoadjuvant chemotherapy regimen. All enrolled samples had complete original breast MR and breast-specific PET dual-modal imaging data before chemotherapy, and all had postoperative pathological examination results. The postoperative pathological results were used as the gold standard label for evaluating the efficacy of chemotherapy.

[0089] Secondly, all patient samples in the dataset are processed uniformly throughout the entire process according to the standardized procedures of steps S2 to S4 above, to obtain the optimal dual-modal radiomics feature subset corresponding to each sample, thereby achieving standardization and homogenization of the features of the entire dataset and eliminating sample data processing bias.

[0090] Then, using the optimal subset of bimodal radiomics features as the input feature variables for the model, and the efficacy of neoadjuvant chemotherapy as determined by the postoperative pathological gold standard as the model classification label, initial prediction models were constructed in parallel using artificial intelligence algorithms. The random forest algorithm demonstrates strong resistance to overfitting and is well-suited for high-dimensional feature data; the support vector machine algorithm excels at classifying small sample sizes and nonlinear data; and deep learning networks can uncover deep hidden correlations within features. By comparing and selecting the best model from multiple models, the predictive performance of the final model was improved.

[0091] Finally, the hyperparameters of each initial model were iteratively debugged and optimized using the validation dataset. Core hyperparameters such as the learning rate, number of iterations, and regularization coefficient were continuously corrected to improve the model fitting effect. Simultaneously, the optimized models were comprehensively validated using independent test datasets. Evaluation metrics for each model were accurately calculated and compared to fully assess the model's prediction accuracy, negative discrimination ability, and positive discrimination ability, resulting in a predictive model for the efficacy of neoadjuvant chemotherapy in breast cancer.

[0092] In one embodiment of this application, the dataset sample covers breast cancer patients with different molecular subtypes and different tumor stages, and the pathological gold standard results are determined based on the pathological evaluation criteria for neoadjuvant chemotherapy in breast cancer to determine pathological complete remission or non-pathological complete remission.

[0093] It should be noted that the dataset included comprehensive case stratification coverage, covering breast cancer patients with different molecular subtypes such as Luminal, HER2-positive, and triple-negative, as well as clinical cases of different tumor stages such as stage I, II, and III. The efficacy outcomes were divided into two categories: pathological complete remission and non-pathological complete remission, which effectively ensured the accuracy of the model labels and improved the clinical adaptability of the model.

[0094] In one embodiment of this application, the method further includes: fusing patient clinicopathological features, immunohistochemical indicators, and the optimal dual-modal radiomics feature subset to construct an image-clinical joint prediction model, thereby achieving accurate prediction of the efficacy of neoadjuvant chemotherapy through multi-dimensional fusion.

[0095] It should be noted that this model integrates the microscopic heterogeneity characteristics of imaging lesions, the macroscopic pathological characteristics of clinical lesions, and the molecular immune characteristics, achieving multi-dimensional joint prediction from microscopic imaging characteristics to macroscopic clinical characteristics. This significantly improves the accuracy and reliability of predicting the efficacy of neoadjuvant chemotherapy for breast cancer, and better meets the application needs of individualized clinical diagnosis and treatment.

[0096] The present application and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present application. The actual structure is not limited to this. In conclusion, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present application, such design should fall within the protection scope of the present application.

Claims

1. A method for predicting the efficacy of neoadjuvant chemotherapy for breast cancer based on radiomics, characterized in that, Includes the following steps: S1. Obtain bimodal raw image data of the breast cancer patient to be predicted before chemotherapy, wherein the bimodal raw image data includes breast MR image data and breast-specific PET image data; S2. The breast MR image data and breast-specific PET image data are preprocessed respectively to obtain standardized dual-modal image data after registration and alignment; S3. Based on the preprocessed standardized bimodal image data, the region of interest of breast cancer lesions is delineated, and breast MR radiomics features and breast-specific PET radiomics features are extracted to form a bimodal original radiomics feature set. S4. Perform feature filtering and optimization on the original bimodal image omics feature set, remove redundant and invalid features, and obtain the optimal bimodal image omics feature subset; S5. Input the optimal bimodal radiomics feature subset into the pre-trained neoadjuvant chemotherapy efficacy prediction model for breast cancer, and output the neoadjuvant chemotherapy efficacy prediction results for the patient. The efficacy prediction results include two types of evaluation results: pathological complete remission and non-pathological complete remission.

2. The method for predicting the efficacy of neoadjuvant chemotherapy for breast cancer based on radiomics according to claim 1, characterized in that, In step S1, the breast MR imaging data is multi-parameter breast MR imaging data, including dynamic contrast-enhanced DCE-MRI, diffusion-weighted DWI-MRI, and apparent diffusion coefficient ADC imaging data; the breast-specific PET imaging data is breast-specific positron emission tomography glucose metabolism PET imaging data, including the maximum standardized uptake value of the lesion and the original imaging parameters of metabolic volume.

3. The method for predicting the efficacy of neoadjuvant chemotherapy for breast cancer based on radiomics according to claim 1, characterized in that, The preprocessing process in step S2 specifically includes: Denoising, grayscale normalization, and slice interpolation resampling were performed on breast MR images and breast-specific PET images respectively to unify the spatial resolution, pixel spacing, and image dimensions of the two types of images. Using breast MR images as the reference images, a rigid registration combined with a local elastic registration algorithm is used to perform pixel-level precise registration between breast-specific PET images and breast MR images, eliminating spatial positional deviations between the two modal images. By using image thresholding to remove chest wall, breast fat, and blood vessel-irrelevant tissue areas, pure breast lesion image areas are preserved, thus completing the standardized preprocessing of dual-modal images.

4. The method for predicting the efficacy of neoadjuvant chemotherapy for breast cancer based on radiomics according to claim 1, characterized in that, In step S3, the region of interest is the solid area of ​​the breast cancer tumor obtained by combining artificial and intelligent segmentation, excluding the normal breast tissue and edema area surrounding the tumor. The breast MR radiomics features include first-order statistical features, second-order texture features, morphological features, and higher-order wavelet transform features; the breast-specific PET radiomics features include metabolic intensity features, metabolic heterogeneity texture features, and lesion metabolic distribution morphology features.

5. The method for predicting the efficacy of neoadjuvant chemotherapy for breast cancer based on radiomics according to claim 1, characterized in that, The feature selection and optimization in step S4 specifically includes: The variance thresholding method is used to remove low-discrimination features with variances below a preset threshold, thus completing the initial screening. The Pearson correlation coefficient method was used to remove highly correlated redundant features and retain independent features. Based on the LASSO regression algorithm with minimum absolute contraction and selection operator, combined with ten-fold cross-validation, features highly correlated with the efficacy of neoadjuvant chemotherapy for breast cancer were screened out, forming a subset of bimodal radiomics features.

6. The method for predicting the efficacy of neoadjuvant chemotherapy for breast cancer based on radiomics according to claim 1, characterized in that, In step S5, the training and construction process of the neoadjuvant chemotherapy efficacy prediction model for breast cancer includes: Training, validation, and test datasets were constructed. The dataset samples were all breast cancer patients who received standard neoadjuvant chemotherapy. All samples retained pre-chemotherapy breast MR and breast-specific PET dual-modal imaging data and postoperative pathological efficacy gold standard results. All samples in the dataset are processed according to steps S2-S4 to obtain the optimal bimodal image omics feature subset for each sample; Using the optimal bimodal radiomics feature subset as input and the chemotherapy efficacy results of the gold standard in pathology as labels, an initial prediction model is constructed based on random forest, support vector machine or deep learning network algorithm; The model hyperparameters were iteratively optimized using a validation dataset. The model's accuracy, AUC value, specificity, and sensitivity were validated based on a test dataset. The model with the best performance was selected as the final predictive model for the efficacy of neoadjuvant chemotherapy in breast cancer.

7. The method for predicting the efficacy of neoadjuvant chemotherapy for breast cancer based on radiomics according to claim 6, characterized in that, The dataset samples cover breast cancer patients with different molecular subtypes and tumor stages. The pathological gold standard results are used to determine pathological complete remission or non-pathological complete remission based on the pathological evaluation criteria for neoadjuvant chemotherapy in breast cancer.

8. The method for predicting the efficacy of neoadjuvant chemotherapy for breast cancer based on radiomics according to claim 1, characterized in that, Also includes: By fusing the patient's clinicopathological features, immunohistochemical indicators, and the optimal dual-modal radiomics feature subset, an imaging-clinical joint prediction model is constructed to achieve accurate prediction of the efficacy of neoadjuvant chemotherapy through multi-dimensional fusion.

Images