A two-stage radiomic lesion identification and localization method and apparatus

Abstract

The application discloses a two-stage radiomics lesion identification and positioning method and device, adopts a multilayer perceptron network to analyze multi-modal image data, detects FCD by extracting features taking the gray matter as a region of interest by using a radiomics method, and the features combine shape, first-order statistics and texture features from multi-modal and wavelet images. In addition, the application also introduces asymmetric features of left and right hemispheres, avoids potential interference in the contralateral area of the unilateral FCD patient caused by the compensatory mechanism of the left and right hemispheres of epilepsy. According to the rich high-dimensional features and asymmetric features of radiomics, the sensitive features of FCD are fully explored, the two-stage detection method is combined to identify FCD abnormalities, the extracted features are more perfect, false positive results are avoided, and the accuracy and sensitivity of detecting FCD are improved from different scales from coarse to fine granularity.

Description

A two-stage radiomics method and device for lesion identification and localization Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a two-stage radiomics lesion identification and localization method and apparatus. Background Technology

[0002] Epilepsy is considered a disturbance at the level of the entire brain network. Frontal lobe epilepsy is a common type of epilepsy that can be treated surgically, and it is more common in children. Focal cortical dysplasia (FCD) is the most common pathological cause of epilepsy in children. Previous studies have shown that preoperative diagnosis of lesions can improve postoperative outcomes; therefore, accurate detection and localization of epileptogenic lesions are crucial for resection planning and surgical effectiveness. Some studies have assessed changes in cortical thickness and volume in children with frontal lobe epilepsy. These studies have found changes in the frontal lobe cortex compared to controls, but specific lesion localization studies are lacking. Furthermore, studies have found compensatory mechanisms in the contralateral hemisphere in patients with unilateral frontal lobe epilepsy (Swartz et al., 1996; Widjaja et al., 2014), which could lead to potential disturbances in the contralateral region in patients with unilateral FCD. Moreover, FCD lesions in pediatric patients are often large, and surgical interventions usually focus on resecting the epileptogenic zone rather than all abnormal areas, which also leads to a decrease in the accuracy of epileptogenic zone detection.

[0003] Recent studies have shown promising performance in the research of type 2 diabetes and breast cancer, as well as excellent results in diagnosing temporal lobe epilepsy using MRI and PET images. In radiomics studies focusing on pediatric FCD, a two-stage Bayesian classifier was employed (Kulaseharan et al., 2019). First, voxel classification was performed based on cortical thickness and gray-white matter boundary ambiguity. Subsequently, voxels classified as lesions were reclassified using texture features; however, this method only utilizes a subset of texture features and does not fully leverage the rich high-dimensional features of radiomics. Furthermore, some studies extract morphological features from pixels on the entire cortical surface and predict lesion location through clustering, which suffers from a high false-positive rate. While MRI and wavelet images (Cheong et al., 2021) are currently used to obtain rich high-dimensional radiomic features in temporal lobe epilepsy, their application in FCD research has not been widely explored. Therefore, this invention aims to explore radiomics methods for identifying FCD abnormalities. Summary of the Invention

[0004] This invention addresses the problems of imperfect feature extraction and high false positive rates in existing technologies by proposing a two-stage radiomics lesion identification and localization method and device. It utilizes two types of features (radiomics features and asymmetric features) and a two-stage detection method (coarse localization and fine localization) to achieve higher accuracy and sensitivity in FCD lesion detection.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] On the one hand, this invention provides a two-stage radiomics lesion identification and localization method, comprising two stages: frontal lobe detection and frontal lobe region detection, wherein:

[0007] The first phase includes the following steps:

[0008] S11. Preprocess the original image and perform cortical reconstruction to obtain a multimodal image;

[0009] S12. Based on the preprocessed multimodal image and the wavelet transform image decomposed by low-pass and high-pass filters, the left and right frontal lobes of the brain are segmented and the gray matter regions of interest are extracted.

[0010] S13. Extract radiomics features from the region of interest in gray matter, calculate the asymmetric features on the left and right sides of the extracted radiomics features, and then standardize and filter the radiomics features and asymmetric features.

[0011] S14. Input the filtered features into an MLP classifier for classification and output the diseased frontal lobe.

[0012] The second phase includes the following steps:

[0013] S21. Obtain multimodal images of the diseased frontal lobe region; the frontal lobe region includes the upper part of the middle frontal gyrus, the frontal pole, the lateral orbitofrontal region, the medial orbitofrontal region, the posterior part of the inferior frontal gyrus, the orbital part of the inferior frontal gyrus, the triangular part of the inferior frontal gyrus, the precentral gyrus, the lower part of the middle frontal gyrus, and the superior frontal gyrus.

[0014] S22. Based on the multimodal image of the frontal lobe and the wavelet transform image decomposed by low-pass and high-pass filters, the frontal lobe is segmented and the gray matter region of interest is extracted.

[0015] S23. Extract radiomics features from the region of interest in gray matter, calculate the asymmetric features on the left and right sides of the extracted radiomics features, and then standardize and filter the radiomics features and asymmetric features.

[0016] S24. Input the filtered features of each sub-region into a separate MLP classifier for classification to obtain the diseased sub-regions predicted by the model for each sub-region; finally, superimpose all the predicted diseased sub-regions to obtain the predicted lesion location.

[0017] Furthermore, in steps S11 and S21, the multimodal images include T1 images, FLAIR images, and PET images, and step S11 uses the FreeSurfer tool to complete cortical reconstruction.

[0018] Further, steps S12 and S22 use the FreeSurfer tool to segment and extract the gray matter region of interest.

[0019] Further, steps S12 and S22 use the PyRadiomic tool to decompose the wavelet-transformed image.

[0020] Furthermore, in steps S13 and S23, the PyRadiomic tool is used to extract radiomics features, which include 14 shape features, 18 first-order features, 24 gray-level co-occurrence matrix features, 16 gray-level run-length matrix features, 16 gray-level size region matrix features, 5 neighborhood gray-level difference matrix features, and 14 gray-level dependence matrix features.

[0021] Furthermore, the asymmetric features of the left and right sides calculated in steps S13 and S23 are obtained by subtracting the features of the left and right frontal lobes or subregions, as shown in the following formula:

[0022] asymmetru=2×(f left -f right ) / (f left +f right )

[0023] Among them, f left and f right These represent features of the left and right brain regions. Asymmetry is defined as the asymmetric feature value of the left brain region, and -asymmetry is defined as the asymmetric feature value of the right brain region.

[0024] Furthermore, the standardization process for radiomics features and asymmetric features in steps S13 and S23 includes:

[0025] The first step was to standardize the features using the patient's internal z-score for patients and the control group's internal z-score for normal controls.

[0026] The second step involves assigning z-scores to the patients from the first step, using the mean and standard deviation of the control group.

[0027] Furthermore, steps S13 and S23 employ a random forest classifier to filter the standardized features.

[0028] Furthermore, the MLP classifier in step S14 or step S24 has two hidden layers and two output nodes, with the two hidden layers containing 40 and 10 nodes respectively, and uses dropout of 0.4 on the input layer, while using the Adam optimization algorithm and employing the cross-entropy loss function.

[0029] On the other hand, the present invention provides a two-stage radiomics lesion identification and localization device, comprising the following modules to implement the two-stage radiomics lesion identification and localization method described in any of the above claims:

[0030] The multimodal image processing module is used to preprocess the original image and reconstruct the cortex, including registration, resampling, skull peeling, head motion correction, and intensity normalization steps to obtain multimodal images;

[0031] The wavelet image conversion module is used to decompose the preprocessed multimodal image into a wavelet converted image through low-pass and high-pass filters;

[0032] The gray matter region of interest segmentation module, based on the preprocessed multimodal image and the wavelet transform image decomposed by low-pass and high-pass filters, segments the left and right frontal lobes and their frontal lobe regions of the brain, and extracts the gray matter region of interest.

[0033] The radiomics feature extraction module is used to extract radiomics features from regions of interest in gray matter.

[0034] The asymmetric feature calculation module is used to calculate the asymmetric features on the left and right sides of the extracted radiomics features;

[0035] The feature selection module is used for standardization and feature selection of radiomics features and asymmetric features;

[0036] An MLP classifier is used to classify the input filtering features and output the diseased frontal lobe or diseased frontal lobe region.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] This invention provides a two-stage radiomics-based method and apparatus for lesion identification and localization. It employs a multi-layer perceptron (MLP) network to analyze multimodal (T1, FLAIR, PET) image data. Radiomics methods are used to extract features from gray matter as regions of interest to detect focal discrepancies (FCDs). These features combine shape, first-order statistical, and texture features from multimodal and wavelet images. Furthermore, this invention introduces asymmetric features between the left and right hemispheres to avoid potential interference from the contralateral region in unilateral FCD patients due to compensatory mechanisms in the left and right hemispheres of epilepsy. By fully exploring the sensitive features of FCDs based on the rich high-dimensional and asymmetric features of radiomics, and combining this with a two-stage detection method, FCD abnormalities are identified, feature extraction is more comprehensive, false positives are avoided, and the accuracy and sensitivity of FCD detection are improved from coarse to fine granularity at different scales.

[0039] The method and apparatus of the present invention are applicable not only to childhood epilepsy, but also to the detection of lesions in adult epilepsy; not only to the detection of epileptic lesions in the frontal lobe, but also to the detection of lesions in other brain regions (such as the temporal lobe, parietal lobe, etc.); not only to the detection of epilepsy, but also to the detection of lesions in other brain diseases. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0041] Figure 1 is a flowchart of the two-stage radiomics lesion identification and localization method provided by the present invention.

[0042] Figure 2 is a functional flowchart of the two-stage radiomics lesion identification and localization device module provided by the present invention. Detailed Implementation

[0043] To improve the accuracy of epileptogenic zone detection and reduce potential interference from contralateral regions in patients with unilateral frontal lobe disease (FCD), this invention designs a two-stage detection method for frontal lobe epilepsy, from the frontal lobe to sub-regions. First, the affected frontal lobe is identified in both left and right frontal lobe regions. Then, all sub-regions of the affected frontal lobe are analyzed to determine the specific location of the FCD. This two-stage method offers the advantage of a progressively refined detection process. In the first stage, abnormalities within the cerebral hemispheres can be more easily identified, thus narrowing the search area for the lesion and detecting the affected frontal lobe (coarse localization). In the second stage, the focus can be concentrated on detecting abnormal regions within the affected frontal lobe (fine localization), thereby improving the accuracy of FCD detection.

[0044] To better understand this technical solution, the method of the present invention will be described in detail below with reference to the accompanying drawings.

[0045] The two-stage radiomics lesion identification and localization method provided by this invention, as shown in Figure 1, includes two stages: frontal lobe detection and frontal lobe region detection, wherein:

[0046] The first phase includes the following steps:

[0047] S11. Preprocess the original image and perform cortical reconstruction to obtain a multimodal image;

[0048] S12. Based on the preprocessed multimodal image and the wavelet transform image decomposed by low-pass and high-pass filters, the left and right frontal lobes of the brain are segmented and the gray matter regions of interest are extracted.

[0049] S13. Extract radiomics features from the region of interest in gray matter, calculate the asymmetric features on the left and right sides of the extracted radiomics features, and then standardize and filter the radiomics features and asymmetric features.

[0050] S14. Input the filtered features into an MLP classifier for classification and output the diseased frontal lobe.

[0051] The second phase includes the following steps:

[0052] S21. Obtain multimodal images of the diseased frontal lobe region; the frontal lobe region includes the upper part of the middle frontal gyrus, the frontal pole, the lateral orbitofrontal region, the medial orbitofrontal region, the posterior part of the inferior frontal gyrus, the orbital part of the inferior frontal gyrus, the triangular part of the inferior frontal gyrus, the precentral gyrus, the lower part of the middle frontal gyrus, and the superior frontal gyrus.

[0053] S22. Based on the multimodal image of the frontal lobe and the wavelet transform image decomposed by low-pass and high-pass filters, the frontal lobe is segmented and the gray matter region of interest is extracted.

[0054] S23. Extract radiomics features from the region of interest in gray matter, calculate the asymmetric features on the left and right sides of the extracted radiomics features, and then standardize and filter the radiomics features and asymmetric features.

[0055] S24. Input the filtered features of each sub-region into a separate MLP classifier for classification to obtain the diseased sub-regions predicted by the model for each sub-region; finally, superimpose all the predicted diseased sub-regions to obtain the predicted lesion location.

[0056] Both stages involve feature extraction and preprocessing, standardization, and filtering of gray matter regions of interest (ROIs), as shown in Figure 2. First, the original images (T1, FLAIR, PET) undergo preprocessing and cortical reconstruction, including registration, resampling, skull stripping, head motion correction, and intensity normalization. Cortical reconstruction is primarily performed using FreeSurfer. Then, the left and right frontal lobes and their subregions are segmented and their gray matter ROIs extracted. Radiomics features are extracted from these ROIs, and asymmetric features between the left and right hemispheres are calculated based on the extracted features. Standardization and feature filtering are then performed based on the original and asymmetric features. The filtered features are input into an MLP network for classification, ultimately achieving lesion detection. Specifically:

[0057] Based on preprocessed multimodal images (such as T1, FLAIR, and PET), the FreeSurfer tool was used to segment gray matter regions of interest. Based on these gray matter regions of interest, radiomics features were extracted from the T1 and wavelet-transformed images, FLAIR images, and PET images. A total of 107 radiomics features were extracted from each image, including 14 shape features, 18 first-order features, 24 gray-level co-occurrence matrix (GLCM) features, 16 gray-level run-length matrix (GLRLM) features, 16 gray-level size zone matrix (GLSZM) features, 5 neighborhood gray-tone difference matrix (NGTDM) features, and 14 gray-level dependence matrix (GLDM) features. In wavelet transform images with eight decompositions, where H represents a high-pass filter and L represents a low-pass filter: LLL, LLH, LHL, LHH, HLL, HLH, HHL, HHH; 744 wavelet features were computed, including first-order, GLCM, GLRLM, GLSZM, NGTDM, and GLDM features. For each region of interest for each participant, a total of 1037 features were merged from T1, FLAIR, and PET. Decomposition of wavelet transform images and radiomics feature extraction were processed using the PyRadiomic tool (Van Griethuysen et al., 2017).

[0058] Asymmetric features are obtained by subtracting features from the left and right frontal lobes and their subregions, using the following formula:

[0059] asymmetry=2×(fleft -f right ) / (f left +f right )

[0060] Among them, f left and f right Representing the features of the left and right brain regions, we define asymmetry as the asymmetric feature value of the left brain region and -asymmetry as the asymmetric feature value of the right brain region.

[0061] In each participant, these extracted features underwent two standardized processing procedures:

[0062] (1) For patients, features were standardized using the patient’s internal z-score. For normal controls, features were standardized using the control group’s internal z-score.

[0063] (2) Patients in (1) were scored using z-scores based on the mean and standard deviation of the control group.

[0064] Standardized features were selected using a random forest classifier and then classified using an MLP model. To determine the optimal data processing and network parameters, we conducted a series of experiments using 5-fold cross-validation on experimental data. Hyperparameter selection was based on the performance metrics of each 5-fold cross-validation model on its respective validation set. The MLP model (Spitzer et al., 2022) had two hidden layers with 40 and 10 nodes respectively, two output nodes, and used a dropout of 0.4 on the input layer to learn more robust representations. The Adam optimization algorithm was employed, along with a cross-entropy loss function. Classification performance was evaluated using accuracy, specificity, sensitivity, and area under the curve. Furthermore, lesion detection performance in the frontal lobe was evaluated by measuring the overlap between the predicted region and the ground truth label.

[0065] On the other hand, the present invention provides a two-stage radiomics lesion identification and localization device, comprising the following modules to implement the two-stage radiomics lesion identification and localization method described in any of the above claims:

[0066] The multimodal image processing module is used to preprocess the original image and reconstruct the cortex, including registration, resampling, skull peeling, head motion correction, and intensity normalization steps to obtain multimodal images;

[0067] The wavelet image conversion module is used to decompose the preprocessed multimodal image into a wavelet converted image through low-pass and high-pass filters;

[0068] The gray matter region of interest segmentation module, based on the preprocessed multimodal image and the wavelet transform image decomposed by low-pass and high-pass filters, segments the left and right frontal lobes and their frontal lobe regions of the brain, and extracts the gray matter region of interest.

[0069] The radiomics feature extraction module is used to extract radiomics features from regions of interest in gray matter.

[0070] The asymmetric feature calculation module is used to calculate the asymmetric features on the left and right sides of the extracted radiomics features;

[0071] The feature selection module is used for standardization and feature selection of radiomics features and asymmetric features;

[0072] An MLP classifier is used to classify the input filtering features and output the diseased frontal lobe or diseased frontal lobe region.

[0073] For the first stage of frontal lobe detection:

[0074] For each subject, the brain was divided into two hemispheres, serving as two sets of data for analysis. To detect whether patients had frontal lobes with focal discrepancy (FCD), labels for the left and right hemispheres were determined based on postoperative images and the side of surgery. First, based on the preprocessed multimodal images (T1, FLAIR, PET) of the left and right frontal lobes, the FreeSurfer tool was used to segment the gray matter regions of interest (ROIs). Based on the ROIs, radiomics features were extracted from the multimodal images and wavelet images decomposed by low-pass and high-pass filters. Left and right asymmetric features were calculated from the original features. All original and asymmetric features were standardized using z-scores for both patient-internal and control-internal analysis. The standardized features were then filtered using a random forest classifier, and a five-fold cross-validation was used to search for the optimal screening threshold from 0 to the maximum feature importance. Important features were selected and retained based on the optimal threshold. The filtered features were then input into an MLP classifier for classification. The MLP classifier parameters were selected based on the performance metrics of each five-fold cross-validation model on its respective validation set, while dropout was used to prevent overfitting. Finally, the model outputs whether the left and right frontal lobes were diseased, thus detecting the diseased frontal lobes.

[0075] For the second-stage frontal lobe region detection:

[0076] Based on the identification of the left and right hemispheres of the frontal lobe, to further accurately detect lesion areas within the patient's frontal lobe, a separate MLP classifier was trained for each sub-region of the frontal lobe. The prediction results of all sub-regions were then combined to determine the predicted lesion location for the patient. The manually created lesion labels were based on the postoperative resection area. The precise label of the sub-region was determined by calculating whether it overlapped with the manually created label. The detailed detection process for sub-regions is basically the same as that for the frontal lobe. First, the 10 sub-regions of the frontal lobe identified in the first stage are analyzed separately, including the upper part of the middle frontal gyrus, the frontal pole, the lateral orbitofrontal region, the medial orbitofrontal region, the posterior part of the inferior frontal gyrus, the orbital part of the inferior frontal gyrus, the triangular part of the inferior frontal gyrus, the precentral gyrus, the lower part of the middle frontal gyrus, and the superior frontal gyrus. Regions of interest in gray matter are segmented. Radiomic features are extracted from multimodal images and wavelet images. Left and right asymmetric features are calculated. All features are standardized using z-scores. A random forest classifier is used for feature selection. The selected features of each sub-region are input into a separate MLP for classification. The diseased sub-regions predicted by the model are obtained for each sub-region. Finally, all the predicted diseased sub-regions of the patient are superimposed to obtain the predicted lesion location for the patient.

[0077] To verify the effectiveness of this method, all experiments used 5-fold cross-validation to verify detection performance. The proposed method introduces asymmetric features into the frontal lobe and its subregions. The performance of the proposed method is evaluated by comparing the results with the original feature results. Experimental results are shown in Tables 1 and 2. Compared to the original features, the method with asymmetric features achieved an accuracy of 93.0% and a sensitivity of 89.2% in the first-stage frontal lobe detection task; in the frontal lobe region, the accuracy was 85.1% and the sensitivity was 85.6%. Combining the two-stage comprehensive localization results of the frontal lobe and its subregions, the overlap rate between the model-predicted region and the actual lesion label was calculated. Our method achieved an overlap rate of 55.1%, demonstrating its ability to effectively detect not only positive patients but also identify FCD abnormalities in negative patients.

[0078] Table 1. Results of radiomics detection

[0079]

[0080] Table 2. Results of the Two-Stage Integrated Positioning

[0081]

[0082] The experimental data of this invention are FCD multimodal data, and are also applicable to single-modal data. The method of this invention is applicable not only to lesion detection in childhood epilepsy but also to adult epilepsy; not only to lesion detection in the frontal lobe but also to lesion detection in other brain regions (such as the temporal lobe and parietal lobe); not only to the detection of epilepsy but also to the detection of lesions in other brain diseases. Furthermore, the two-stage lesion detection method of this invention uses radiomics to extract features, but can also be applied to morphological methods for feature extraction; the two-stage lesion detection method of this invention uses an MLP classifier for prediction, but can also be applied to other classifiers for feature prediction.

[0083] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. However, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A two-stage radiomics method for lesion identification and localization, characterized in that, The study comprises two stages: frontal lobe detection and frontal lobe region detection. The first stage includes the following steps: S11, preprocessing and cortical reconstruction of the original image, including registration, resampling, skull stripping, head motion correction, and intensity normalization, to obtain a multimodal image; S12, based on the preprocessed multimodal image and its wavelet transform image obtained through low-pass and high-pass filters, segmenting the left and right frontal lobes of the brain and extracting regions of interest (ROIs) from the gray matter; S13, extracting radiomics features from the ROIs, calculating asymmetric features between the left and right sides of the extracted radiomics features, and then standardizing and filtering the radiomics features and asymmetric features; S14, inputting the filtered features into an MLP classifier for classification, outputting the diseased frontal lobe. The second stage includes the following steps: S21, obtaining the diseased frontal lobe... Multimodal images of the frontal lobe region; the frontal lobe region includes the upper part of the middle frontal gyrus, the frontal pole, the lateral orbitofrontal region, the medial orbitofrontal region, the posterior part of the inferior frontal gyrus, the orbital part of the inferior frontal gyrus, the triangular part of the inferior frontal gyrus, the precentral gyrus, the lower part of the middle frontal gyrus, and the superior frontal gyrus; S22, based on the multimodal images of the frontal lobe region and the wavelet transform images decomposed by low-pass and high-pass filters, the frontal lobe region is segmented, and gray matter regions of interest are extracted; S23, radiomics features are extracted for the gray matter regions of interest, and the left and right asymmetric features are calculated for the extracted radiomics features, and then the radiomics features and asymmetric features are standardized and feature filtered; S24, the filtered features of each sub-region are input into a separate MLP classifier for classification to obtain the diseased sub-regions predicted by the model for each sub-region; finally, all the predicted diseased sub-regions are superimposed to obtain the predicted lesion location.

2. The two-stage radiomics lesion identification and localization method according to claim 1, characterized in that, In steps S11 and S21, the multimodal images include T1 images, FLAIR images, and PET images, wherein step S11 uses the FreeSurfer tool to complete cortical reconstruction.

3. The two-stage radiomics lesion identification and localization method according to claim 1, characterized in that, Steps S12 and S22 use the FreeSurfer tool to segment and extract the gray matter region of interest.

4. The two-stage radiomics lesion identification and localization method according to claim 1, characterized in that, Steps S12 and S22 use the PyRadiomic tool to decompose the wavelet-transformed image.

5. The two-stage radiomics lesion identification and localization method according to claim 1, characterized in that, In steps S13 and S23, the PyRadiomic tool is used to extract radiomics features, which include 14 shape features, 18 first-order features, 24 gray-level co-occurrence matrix features, 16 gray-level run-length matrix features, 16 gray-level size region matrix features, 5 neighborhood gray-level difference matrix features, and 14 gray-level dependence matrix features.

6. The two-stage radiomics lesion identification and localization method according to claim 1, characterized in that, In steps S13 and S23, the asymmetric features of the left and right sides are calculated by subtracting the features of the left and right frontal lobes or subregions. The specific formulas are as follows: ,in, and Features representing the left and right brain regions, Defined as asymmetric feature values ​​of the left brain region. Defined as the asymmetric feature value of the right brain region.

7. The two-stage radiomics lesion identification and localization method according to claim 1, characterized in that, The standardization process for radiomics features and asymmetric features in steps S13 and S23 includes: First, for patients, features are standardized using the patient's internal z-score; for normal controls, features are standardized using the control group's internal z-score; Second, the patients in step one are z-scored according to the control group's mean and standard deviation.

8. The two-stage radiomics lesion identification and localization method according to claim 1, characterized in that, Steps S13 and S23 use a random forest classifier to filter the standardized features.

9. The two-stage radiomics lesion identification and localization method according to claim 1, characterized in that, The MLP classifier in step S14 or step S24 has two hidden layers and two output nodes. The two hidden layers contain 40 and 10 nodes respectively, and dropout of 0.4 is used on the input layer. The Adam optimization algorithm is used, and the cross-entropy loss function is employed.

10. A two-stage radiomics lesion identification and localization device, characterized in that, The method comprises the following modules to implement the two-stage radiomics lesion identification and localization method according to any one of claims 1-9: a multimodal image processing module for preprocessing and cortical reconstruction of the original image, including registration, resampling, craniotomy, head motion correction, and intensity normalization steps to obtain a multimodal image; a wavelet image conversion module for decomposing the preprocessed multimodal image into wavelet-converted images using low-pass and high-pass filters; a gray matter region of interest segmentation module for segmenting the left and right frontal lobes and their respective frontal lobe regions based on the preprocessed multimodal image and the wavelet-converted images decomposed by low-pass and high-pass filters, and extracting gray matter regions of interest; a radiomics feature extraction module for extracting radiomics features from the gray matter regions of interest; and an asymmetric feature calculation module for calculating left and right asymmetric features from the extracted radiomics features. The feature selection module is used to standardize and select radiomics features and asymmetric features; the MLP classifier is used to classify the input selected features and output the diseased frontal lobe or diseased frontal lobe region.

Images