Image classification method and device, electronic equipment and storage medium

Abstract

The present disclosure relates to the field of medical artificial intelligence and computer-aided diagnosis, and particularly relates to an image classification method and device, an electronic device and a storage medium, comprising: obtaining an image to be classified, performing automatic region segmentation on the image to be classified to obtain a target region, the image to be classified comprising an image of a non-functioning pituitary adenoma; extracting radiomics features of the target region based on a feature extraction model; performing dimension reduction processing on the extracted radiomics features based on a preset feature selection rule to obtain a first radiomics feature set, the first radiomics feature set being used to quantify the invasiveness of the structure of the target region; and inputting the first radiomics feature set into a classification model to obtain a classification result. By using the method of the present disclosure, the non-functioning pituitary adenoma can be classified according to the degree of invasiveness, and a high accuracy can be achieved, thereby assisting doctors to more quickly and accurately diagnose the type of non-functioning pituitary adenoma, and reducing the treatment burden of patients and the occurrence of postoperative adverse conditions.

Description

Technical Field

[0001] This disclosure relates to the fields of medical artificial intelligence and computer-aided diagnosis, and in particular to an image classification method and apparatus, electronic device and storage medium. Background Technology

[0002] Pituitary adenoma (PA) accounts for 10% of all intracranial masses, and in autopsy studies, it accounts for approximately 15% of all cases. PAs are classified according to size into microadenomas (diameter <10 mm) and macroadenomas (diameter ≥10 mm), or according to their hormone secretion into functional pituitary adenoma (FPA), such as prolactinoma, adrenocorticotropic hormone adenoma, growth hormone adenoma, and thyroid-stimulating hormone adenoma, and non-functional pituitary adenoma (NFPA), such as silent gonadotropin-secreting cell adenoma and silent adrenocorticotropic hormone-secreting cell adenoma. Except for patients with prolactinomas who typically respond well to dopamine agonists, surgical treatment is suitable for patients with neurological dysfunction due to mass effects or those with hypersecretion syndrome.

[0003] Different subtypes of pituitary adenomas have different diagnostic and treatment methods. Dopamine agonists (DA) and somatostatin analogs (SSA) are key treatment options for prolactinomas and growth hormone adenomas, respectively. However, the therapeutic value of DA and SSA for non-functioning pituitary adenomas is unclear. Therefore, the identification of non-functioning pituitary adenomas is an important task for clinicians. Currently, the diagnosis of NFPA is made only after other subtypes of PA have been ruled out. Furthermore, although most PAs are benign, a large number of PAs have an aggressive clinical course. As the tumor grows, PAs invade surrounding structures, leading to incomplete resection and a poor prognosis. Unlike FPA, due to the difficulty in diagnosing non-functioning pituitary adenomas, NFPA often presents as an aggressive macroadenoma at the time of diagnosis, making its treatment more complex, its prognosis worse, and its medical costs significantly increased.

[0004] In recent years, with the development of radiomics technology, computers can analyze high-throughput image features to identify subtle differences between various pituitary adenomas, providing new possibilities for accurate diagnosis. This artificial intelligence-based technology can not only extract information that is difficult to detect in traditional medical imaging, but also build predictive models through big data analysis, further optimizing the diagnostic process. However, currently, the classification of non-functional pituitary adenomas in clinical practice mainly relies on the manual judgment of pathologists. This process requires pathological staining of tumor tissue combined with microscopic observation. Although this method has high accuracy, it is time-consuming, usually requiring a week after surgery to reach a definitive conclusion. This delay may not only affect patients' treatment decisions but also cause certain difficulties in subsequent clinical management. Summary of the Invention

[0005] In view of this, this disclosure proposes an image classification scheme.

[0006] According to one aspect of this disclosure, an image classification method is provided, comprising:

[0007] An image to be classified is acquired, and automated region segmentation is performed on the image to be classified to obtain a target region. The image to be classified includes images of non-functional pituitary adenomas. Based on a feature extraction model, radiomics features of the target region are extracted. Based on a preset feature selection rule, the extracted radiomics features are dimensionality reduced to obtain a first radiomics feature set. The first radiomics feature set is used to quantify the invasiveness of the target region structure. The first radiomics feature set is input into a classification model to obtain a classification result.

[0008] In one possible implementation, the method further includes: acquiring a first image set, the first image set including a first full-size image set and a first target region image set, the first target region image set being obtained by region segmentation of the first full-size image set; extracting radiomics features of the first image set based on a feature extraction model to obtain a first feature dataset, the first feature dataset including a first full-size feature dataset and a first target region feature dataset; and performing dimensionality reduction processing on the first feature dataset to obtain a second radiomics feature set.

[0009] In one possible implementation, the step of performing dimensionality reduction processing on the extracted radiomics features based on preset feature selection rules to obtain a first radiomics feature set includes: selecting features from the second radiomics feature set among the radiomics features extracted from the image to be classified to obtain the first radiomics feature set.

[0010] In one possible implementation, the method further includes: dividing the second radiomics feature set into a training set and a test set; training the training set using multiple first models; evaluating the training results based on the test set; and obtaining the classification model based on the evaluation results. The step of training the training set using multiple first models includes: inputting the fused second radiomics feature set into the first model for training, wherein the second radiomics feature set incorporates coronal and sagittal features; and determining the classification model based on the influence of coronal and sagittal features on the classification performance of the multiple first models.

[0011] In one possible implementation, the first model includes a random forest model, a support vector machine model, a logistic regression model, and a deep learning model; obtaining the classification model based on the evaluation results includes: selecting one model from the first models based on the evaluation results to obtain the classification model; or fusing at least two first models to obtain the classification model.

[0012] In one possible implementation, before extracting the radiomics features of the first image set based on the feature extraction model, the method further includes: performing standardization preprocessing on the first image set to obtain a first image set with preset numerical distribution characteristics; and performing wavelet transform on the first image set to extract multi-scale feature information.

[0013] In one possible implementation, the image type to be classified includes magnetic resonance imaging, and the image to be classified includes images of silent gonadotropin-releasing tumors (GRT-RT) and silent adrenocorticotropic hormone-secreting tumors (ACT-HT-MT), wherein the invasiveness of silent GRT-RT-RT and ACT-HT-MT differs; the first radiomics feature set is input into a classification model to obtain a classification result, including: the classification model classifies the image to be classified as an image of a silent GRT-RT or an image of a silent ACTO based on the difference in invasiveness.

[0014] According to another aspect of this disclosure, an image classification apparatus is provided, comprising: a target region acquisition module, configured to acquire an image to be classified, perform automated region segmentation on the image to be classified to obtain a target region, wherein the image to be classified includes an image of a non-functional pituitary adenoma;

[0015] The radiomics feature extraction module is used to extract radiomics features of the target region based on a feature extraction model.

[0016] The first radiomics feature set acquisition module is used to perform dimensionality reduction processing on the extracted radiomics features based on preset feature selection rules to obtain the first radiomics feature set, which is used to quantify the invasiveness of the target region structure.

[0017] The classification result acquisition module is used to input the first image omics feature set into the classification model to obtain the classification result.

[0018] In one possible implementation, the apparatus further includes: a first image set acquisition module, configured to acquire a first image set, the first image set including a first full-size image set and a first target region image set, the first target region image set being obtained by region segmentation of the first full-size image set; a first feature dataset acquisition module, configured to extract radiomics features of the first image set based on a feature extraction model to obtain a first feature dataset, the first feature dataset including a first full-size feature dataset and a first target region feature dataset; and a second radiomics feature set acquisition module, configured to perform dimensionality reduction processing on the first feature dataset to obtain a second radiomics feature set.

[0019] In one possible implementation, the first radiomics feature set acquisition module is used to select features from the second radiomics feature set among the radiomics features extracted from the image to be classified, to obtain the first radiomics feature set.

[0020] In one possible implementation, the apparatus further includes: a second radiomics feature set partitioning module for partitioning the second radiomics feature set into a training set and a test set; a training set training module for training the training set using multiple first models; an evaluation module for evaluating the training results based on the test set; a classification model acquisition module for obtaining the classification model based on the evaluation results; the training set training module for inputting the fused second radiomics feature set into the first model for training, wherein the second radiomics feature set incorporates coronal and sagittal features; and determining the classification model based on the influence of coronal and sagittal features on the classification performance of multiple first models.

[0021] In one possible implementation, the first model includes a random forest model, a support vector machine model, a logistic regression model, and a deep learning model; the classification model acquisition module is used to select one model from the first models based on the evaluation results to obtain the classification model; or to fuse at least two first models to obtain the classification model.

[0022] In one possible implementation, the device further includes: a preprocessing module for performing standardized preprocessing on the first image set to obtain a first image set with preset numerical distribution characteristics; and a wavelet transform module for performing wavelet transform on the first image set to extract multi-scale feature information.

[0023] In one possible implementation, the image type to be classified includes magnetic resonance imaging, and the image to be classified includes images of silent gonadotropin-secreting cell adenomas and images of silent adrenocorticotropic hormone-secreting cell adenomas, wherein the invasiveness of silent gonadotropin-secreting cell adenomas and silent adrenocorticotropic hormone-secreting cell adenomas differs; the classification result acquisition module is used by the classification model to classify the image to be classified as an image of a silent gonadotropin-secreting cell adenoma or an image of a silent adrenocorticotropic hormone-secreting cell adenoma based on the difference in invasiveness.

[0024] According to another aspect of this disclosure, a non-volatile computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the above-described method.

[0025] According to another aspect of this disclosure, a computer program product is provided, including a computer program or a non-volatile computer-readable storage medium carrying the computer program, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0026] In this embodiment of the disclosure, an image to be classified is acquired, and the image to be classified is automatically segmented to obtain a target region. The image to be classified includes an image of a non-functional pituitary adenoma. Based on a feature extraction model, radiomics features of the target region are extracted. Based on a preset feature selection rule, the extracted radiomics features are dimensionality reduced to obtain a first radiomics feature set. The first radiomics feature set is used to quantify the invasiveness of the target region structure. The first radiomics feature set is input into a classification model to obtain a classification result. By automating region segmentation of images of non-functioning pituitary adenomas, features can be extracted only from the region of interest after automated segmentation, avoiding the noise influence of full-image features and improving the accuracy of classification tasks. Radiomic features of the target region are extracted and then dimensionality-reduced. Since different types of non-functioning pituitary adenomas exhibit varying degrees of invasiveness—for example, silent gonadotropin-secreting adenomas are often more invasive, while silent adrenocorticotropic hormone-secreting adenomas are mostly non-invasive—the dimensionality-reduced radiomic feature set quantifies the invasiveness of the region of interest structure in images of non-functioning pituitary adenomas, enabling accurate classification of non-functioning pituitary adenomas and providing reliable radiographic evidence for preoperative assessment of tumor biological behavior. Simultaneously, by inputting the radiomic feature set into a trained classification model, the model classifies non-functioning pituitary adenomas based on the learned feature patterns. Therefore, the embodiments of this disclosure can classify non-functional pituitary adenomas according to whether they are invasive or not, and achieve a high degree of accuracy, thereby assisting doctors to diagnose the type of non-functional pituitary adenoma more quickly and accurately, reducing the treatment burden on patients and the occurrence of postoperative adverse events.

[0027] Other features and aspects of this disclosure will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0028] The accompanying drawings, which are included in and form part of this specification, illustrate exemplary embodiments, features, and aspects of this disclosure together with the specification and serve to explain the principles of this disclosure.

[0029] Figure 1 This image shows a schematic MRI image of a silent adrenocorticotropic hormone cell adenoma according to an embodiment of the present disclosure.

[0030] Figure 2 A flowchart illustrating an image classification method according to an embodiment of the present disclosure is shown.

[0031] Figure 3 A flowchart illustrating the training of a segmentation model according to an embodiment of this disclosure is shown.

[0032] Figure 4A schematic diagram showing the localization of a pituitary adenoma using a coronal image of a raw MRI according to an embodiment of the present disclosure.

[0033] Figure 5 A schematic diagram showing a coronal image of a raw MRI according to an embodiment of the present disclosure.

[0034] Figure 6 A schematic diagram showing a manually annotated coronal image of the original MRI according to an embodiment of the present disclosure.

[0035] Figure 7 This diagram illustrates the image fusion of a coronal image of the original MRI and a manually annotated coronal image of the original MRI, according to an embodiment of the present disclosure.

[0036] Figure 8 This diagram illustrates a region of interest (ROI) segmentation of a coronal image of an original MRI image according to an embodiment of the present disclosure.

[0037] Figure 9 This diagram illustrates a manually annotated coronal image of an original MRI scan according to an embodiment of the present disclosure.

[0038] Figure 10 This diagram illustrates an overlapping image of a manually annotated coronal image of an original MRI image after ROI region segmentation according to an embodiment of the present disclosure.

[0039] Figure 11 A schematic diagram of normalized preprocessing of a coronal image of a raw MRI according to an embodiment of the present disclosure is shown.

[0040] Figure 12 A schematic diagram of the standardized preprocessing of a raw coronal MRI image according to an embodiment of the present disclosure, after manual annotation.

[0041] Figure 13 A schematic diagram illustrating the training process of the nnU-Net model according to an embodiment of the present disclosure is shown.

[0042] Figures 14A-14F The diagrams show a comparison between automated segmentation and manual segmentation of different layers using the nnU-Net model according to embodiments of this disclosure.

[0043] Figures 15A-15E The diagrams show a comparison between automated segmentation and manual segmentation of different layers using the nnU-Net model, according to another embodiment of the present disclosure.

[0044] Figure 16 A schematic diagram showing the first twenty features after dimensionality reduction according to an embodiment of the present disclosure is provided.

[0045] Figure 17A schematic diagram showing a heatmap of the first twenty features after dimensionality reduction according to an embodiment of the present disclosure.

[0046] Figure 18 A schematic diagram of the ROC curve of a full-size feature dataset according to an embodiment of the present disclosure is shown.

[0047] Figure 19 A schematic diagram of the ROC curve of a target region feature dataset according to an embodiment of the present disclosure is shown.

[0048] Figure 20 This diagram illustrates a full-process model training flowchart according to an embodiment of the present disclosure.

[0049] Figure 21 A schematic flowchart of a full-process model test according to an embodiment of the present disclosure is shown.

[0050] Figure 22 A schematic diagram of an image classification processing flow according to an embodiment of the present disclosure is shown.

[0051] Figure 23 A schematic diagram illustrating the ROC results of imaging and clinical information of a pituitary adenoma patient according to an embodiment of the present disclosure.

[0052] Figure 24 A block diagram of an image classification apparatus according to an embodiment of the present disclosure is shown.

[0053] Figure 25 A block diagram of an electronic device 1900 according to an embodiment of the present disclosure is shown. Detailed Implementation

[0054] Various exemplary embodiments, features, and aspects of this disclosure will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0055] As used herein, the terms “comprising,” “including,” “having,” or variations thereof are open-ended and include one or more of the stated features, integrals, elements, steps, components, or functions, but do not exclude the presence or addition of one or more other features, integrals, elements, steps, components, functions, or groups thereof.

[0056] When an element is referred to as “connected,” “coupled,” “responding,” or a variation thereof relative to another element, it may be directly connected, coupled, or responding to another element, or there may be an intermediate element present.

[0057] Although the terms first, second, third, etc., may be used herein to describe various elements / operations, these elements / operations should not be limited by these terms. These terms are only used to distinguish one element / operation from another. Therefore, without departing from the teachings of the inventive concept, a first element / operation in some embodiments may be referred to as a second element / operation in other embodiments.

[0058] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0059] Furthermore, to better illustrate this disclosure, numerous specific details are set forth in the following detailed description. Those skilled in the art will understand that this disclosure can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art have not been described in detail in order to highlight the main points of this disclosure.

[0060] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant regions.

[0061] Pituitary adenoma (PA) accounts for 10% of all intracranial masses, and in autopsy studies, it accounts for approximately 15% of all cases. PAs are classified according to size into microadenomas (diameter <10 mm) and macroadenomas (diameter ≥10 mm), or according to their hormone secretion into functional pituitary adenoma (FPA), such as prolactinoma, adrenocorticotropic hormone adenoma, growth hormone adenoma, and thyroid-stimulating hormone adenoma, and non-functional pituitary adenoma (NFPA), such as silent gonadotropin-secreting cell adenoma and silent adrenocorticotropic hormone-secreting cell adenoma. Figure 1 This diagram illustrates an MRI image of a silent adrenocorticotropic hormone-secreting cell adenoma according to an embodiment of the present disclosure. T1W represents T1-weighted imaging, IR represents inversion recovery sequence, and +C represents contrast-enhanced scanning. Figure 1 As can be seen, the tumor grows aggressively and surrounds the internal carotid artery. However, apart from those patients with prolactinomas who usually respond well to dopamine agonists, surgical treatment is suitable for other patients who experience neurological dysfunction due to the mass effect or have hypersecretion syndrome.

[0062] Different subtypes of pituitary adenomas have different diagnostic and treatment methods. Dopamine agonists (DA) and somatostatin analogs (SSA) are key treatment options for prolactinomas and growth hormone adenomas, respectively. However, the therapeutic value of DA and SSA for non-functioning pituitary adenomas is unclear. Therefore, the identification of non-functioning pituitary adenomas is an important task for clinicians. Currently, the diagnosis of NFPA is made only after other subtypes of PA have been ruled out. Furthermore, although most PAs are benign, a large number of PAs have an aggressive clinical course. As the tumor grows, PAs invade surrounding structures, leading to incomplete resection and a poor prognosis. Unlike FPA, due to the difficulty in diagnosing non-functioning pituitary adenomas, NFPA often presents as an aggressive macroadenoma at the time of diagnosis, making its treatment more complex, its prognosis worse, and its medical costs significantly increased.

[0063] In recent years, with the development of radiomics technology, computers can analyze high-throughput image features to identify subtle differences between various pituitary adenomas, providing new possibilities for accurate diagnosis. This artificial intelligence-based technology can not only extract information that is difficult to detect in traditional medical imaging, but also build predictive models through big data analysis, further optimizing the diagnostic process. However, currently, the classification of non-functional pituitary adenomas in clinical practice mainly relies on the manual judgment of pathologists. This process requires pathological staining of tumor tissue combined with microscopic observation. Although this method has high accuracy, it is time-consuming, usually requiring a week after surgery to reach a definitive conclusion. This delay may not only affect patients' treatment decisions but also cause certain difficulties in subsequent clinical management.

[0064] In this embodiment, an image to be classified is acquired, and automated region segmentation is performed on the image to be classified to obtain a target region. The image to be classified includes an image of a non-functional pituitary adenoma. Based on a feature extraction model, radiomics features of the target region are extracted. Based on a preset feature selection rule, the extracted radiomics features are subjected to dimensionality reduction processing to obtain a first radiomics feature set. The first radiomics feature set is used to quantify the invasiveness of the target region structure. The first radiomics feature set is input into a classification model to obtain a classification result. Therefore, non-functional pituitary adenomas can be classified according to their degree of invasiveness with high accuracy, thereby assisting doctors in diagnosing the type of non-functional pituitary adenoma more quickly and accurately, reducing the treatment burden on patients and the occurrence of postoperative adverse events.

[0065] Figure 2 A flowchart illustrating an image classification method according to an embodiment of this disclosure is shown. Figure 2As shown, this image classification method includes:

[0066] S11. Obtain the image to be classified, perform automated region segmentation on the image to be classified to obtain the target region, wherein the image to be classified includes images of non-functional pituitary adenomas.

[0067] In this embodiment of the disclosure, a non-functioning pituitary adenoma can be a type of benign pituitary tumor that does not present with clinical manifestations of hyperhormone secretion but causes neurological symptoms and / or pituitary hypofunction through tumor space-occupying effects. Non-functioning pituitary adenomas can include silent gonadotropin-secreting cell adenomas, silent adrenocorticotropic hormone-secreting cell adenomas, silent growth hormone-secreting cell adenomas, silent prolactin-secreting cell adenomas, etc.

[0068] The images to be classified may include images of non-functional pituitary adenomas. The source type (i.e., data format) of the images to be classified may be computed tomography (CT), magnetic resonance imaging (MRI), positron emission tomography (PET), etc., and this disclosure does not limit this. In one possible implementation, the type (i.e., data format) of the images to be classified includes magnetic resonance imaging (MRI). MRI can reveal the characteristics of tumors from multiple angles using different scanning sequences (e.g., T1-weighted, T2-weighted, enhanced T1-weighted, diffusion-weighted imaging, etc.), and can directly obtain coronal, sagittal, and axial images without reconstruction, thus providing important basis for the assessment and classification of tumor invasiveness.

[0069] The target region can be a region of interest (ROI). In this embodiment of the disclosure, the target region can be a tumor portion of an image of a non-functional pituitary adenoma.

[0070] Automated segmentation can use computer models to automatically analyze image data (such as CT, MRI, PET) with little or no human intervention, distinguishing and outlining the region of interest in the image from the background or other tissue structures.

[0071] In one possible implementation, before performing automated region segmentation on the image to be classified, the method further includes preprocessing the image to be classified. Preprocessing may involve performing Gaussian blur denoising on the image to be classified. Gaussian blur denoising can reduce image noise and improve the signal-to-noise ratio. For images of non-functional pituitary adenomas, appropriate denoising before feature extraction can prevent noise itself from being mistakenly identified as important texture information by the model, leading to unreliable features.

[0072] After Gaussian blur denoising, contrast enhancement can be used to highlight tissue boundaries, enhance local contrast, make tumor boundaries clearer, and significantly enhance local details of non-functional pituitary adenoma images (such as irregular edges of the tumor and fine internal structures), making the contrast with surrounding normal tissue more obvious, which greatly facilitates subsequent segmentation operations.

[0073] Preprocessing may further include pixel value normalization and size standardization of the images to be classified. Pixel value normalization can employ Z-Score normalization, Min-Max normalization, etc., and this disclosure does not limit the specific methods used. Pixel value normalization can unify the pixel value distribution of the images, ensuring the comparability of texture, intensity, and other features extracted from the collected images. Size standardization can adjust the image resolution, giving the features extracted from the images (such as texture and volume) real physical meaning, allowing for direct comparison and ensuring data dimensionality consistency.

[0074] After preprocessing the image to be classified, it can be input into a pre-trained segmentation model to perform automated region segmentation. Figure 3 A flowchart illustrating the training of a segmentation model according to an embodiment of this disclosure is shown. Figure 3 As shown, in one possible implementation, the segmentation model training process may include:

[0075] S11-1. Obtain the first segmentation training dataset and label the first segmentation training dataset;

[0076] The segmentation model in this embodiment can be a machine learning model, such as a support vector machine or random forest, or a deep learning model, such as a fully convolutional network or U-Net model. This embodiment does not limit the specific implementation. In one possible implementation, the segmentation model can be an nnU-Net (No-new-Net) model.

[0077] The segmentation dataset can be an image dataset, which may include coronal and / or sagittal data of images of non-functional pituitary adenomas. Taking the nnU-Net model as an example, coronal and sagittal images can be viewed as two modalities of the same structure (T1 and T2 modalities). By organizing them according to modal dimensions, nnU-Net automatically learns complementary features of different perspectives without the need for manual design of complex networks. In this embodiment, by fusing coronal and sagittal image data of non-functional pituitary adenomas, the information limitations of a single perspective can be overcome, and the difficulty of designing a fusion strategy is reduced by using the nnU-Net model.

[0078] When labeling the first segmentation training dataset, it can be done manually. The coronal and / or sagittal data of the non-functional pituitary adenoma image are labeled according to preset standards. The preset standards can be tumor density, signal characteristics, morphological appearance, etc., and this embodiment does not limit this.

[0079] For example, Figure 4 A schematic diagram showing the localization of a pituitary adenoma using a coronal image of a raw MRI according to an embodiment of the present disclosure. Figure 5 A schematic diagram showing a coronal image of a raw MRI according to an embodiment of the present disclosure. Figure 6 A schematic diagram showing a manually annotated coronal image of the original MRI according to an embodiment of the present disclosure. Figure 7 This diagram illustrates the image fusion of a coronal image of the original MRI and a manually annotated coronal image of the original MRI, according to an embodiment of the present disclosure. Figure 8 This diagram illustrates a region of interest (ROI) segmentation of a coronal image of an original MRI image according to an embodiment of the present disclosure. Figure 9 This diagram illustrates a manually annotated coronal image of an original MRI scan according to an embodiment of the present disclosure. Figure 10 This diagram illustrates an overlapping image of a manually annotated coronal image of an original MRI image after ROI region segmentation according to an embodiment of the present disclosure. Figure 11 A schematic diagram of normalized preprocessing of a coronal image of a raw MRI according to an embodiment of the present disclosure is shown. Figure 12 This diagram illustrates a standardized preprocessing of a raw coronal MRI image according to an embodiment of the present disclosure, after manual annotation. Figures 4-12 The layer in the middle represents a (imaging) slice of a tumor on a coronal MRI. Figures 4-12 (All are coronal). From Figures 4-12 The preprocessing and segmentation process can be seen.

[0080] S11-2. In the first segmentation training dataset, images that meet the training requirements of the segmentation model are selected to obtain the second segmentation training dataset.

[0081] Meeting the requirements for model training can include screening images for quality, such as excluding images with severe artifacts, images with extremely poor quality or low signal-to-noise ratio, and screening images for clinical consistency, such as excluding images that have undergone postoperative follow-up, to ensure that the labeled data matches the clinical scenario of the model training target.

[0082] S11-3. Organize the second segmentation training dataset according to the preset data format;

[0083] The preset data format can be the format required by the segmentation model. In the nnU-Net model, it can be the coronal / sagittal images organized into a form that the model can recognize, according to the fixed folder structure, naming rules, and metadata format specified by the nnU-Net model.

[0084] S11-4. Input the second segmentation training dataset into the segmentation model for training;

[0085] After formatting the second segmentation training dataset, it can be input into the segmentation model for training. The nnU-Net model allows for automated training. It automatically divides the training dataset into several disjoint subsets. Several rounds of training and validation are performed sequentially. In each round, a majority of the data (i.e., the second segmentation training dataset in this embodiment) is selected as the training set to update the model weights; the remaining small subset serves as the validation set to monitor model performance and prevent overfitting. During this process, the nnU-Net model automatically calculates the most suitable hyperparameter combination, completing all rounds of training and validation, thus achieving automatic segmentation without any manual intervention.

[0086] For example, Figure 13 A schematic diagram illustrating the training process of the nnU-Net model according to an embodiment of this disclosure is shown. See also... Figure 13 Here, the training loss (loss_tr) is the loss function value calculated by the model on the current training batch. The pseudo-Dice coefficient is the Dice similarity coefficient calculated on the training set. The validation loss (loss_val) is the loss function value calculated by the model on the validation set. The pseudo-Dice coefficient (moving average) is the moving average of the pseudo-Dice values ​​on the validation set. Figure 13As can be seen, after multiple rounds of training, the changes in parameters such as loss_tr, pseudo dice, loss_val, and pseudo dice(mov.avg.) all tend to stabilize.

[0087] In this embodiment of the disclosure, after the second segmentation training dataset is input into the segmentation model for training, the training results can be evaluated. The untrained dataset can be input into the trained segmentation model for automated region segmentation. The segmented target region can be compared with the manually segmented target region to verify the accuracy of the trained segmentation model.

[0088] For example, Figures 14A-14F The diagrams show a comparison between automated segmentation and manual segmentation of different layers using the nnU-Net model according to embodiments of this disclosure. Figures 14A-14F The overlap results of automated segmentation using the nnU-Net model and manual segmentation are as follows: the average Dice coefficient for all slices is 0.9561, the average overlap rate is 95.61%, the overall 3D Dice coefficient is 0.8289, and the overall overlap rate is 82.89%.

[0089] Figures 15A-15E The diagrams show a comparison between automated segmentation and manual segmentation of different layers using the nnU-Net model, according to another embodiment of the present disclosure. Figures 15A-15E The overlap results of automated segmentation using the nnU-Net model and manual segmentation are as follows: the average Dice coefficient for all slices is 0.7600, the average overlap rate is 76.00%, the overall 3D Dice coefficient is 0.8316, and the overall overlap rate is 83.16%. (See also...) Figures 14A-14F and Figures 15A-15E , Figures 14A-14F Two of the images do not contain non-functional pituitary adenomas. Figures 15A-15EThe image in the middle contains all images containing non-functional pituitary adenomas. As can be seen from the average overlap rate and the overall overlap rate in the image, the automated segmentation of this embodiment has a high accuracy rate regardless of whether the image to be classified contains non-functional pituitary adenomas (the overall overlap rate in the image is all above 80%).

[0090] In this embodiment of the disclosure, by performing multimodal training on the segmentation model, the segmentation accuracy and robustness of the model can be significantly improved. By performing automated training in the nnU-Net model, the optimal network structure and training strategy can be automatically adjusted, and multiple image data such as MRI and CT can be processed simultaneously, significantly improving the effect and efficiency of image segmentation of non-functional pituitary adenomas.

[0091] S12. Based on the feature extraction model, extract the image omics features of the target region;

[0092] A feature extraction model can be a predefined, computable algorithmic framework or set of functions that automatically quantifies and describes the visual attributes of image data within a target region. These attributes include, but are not limited to, intensity statistics, spatial texture patterns, morphological geometry, and higher-order transform domain properties. Feature extraction models can provide standardized, reproducible quantitative input data (i.e., feature vectors) for the subsequent construction of diagnostic, prognostic, or predictive machine learning models.

[0093] Radiomics features can refer to quantitative, high-dimensional data descriptors extracted from medical imaging data through reproducible and automated computational methods.

[0094] The feature extraction model can be a morphological feature model, including volume, surface area, surface area-to-volume ratio, etc.; it can be a texture feature model, including gray-level co-occurrence matrix model, gray-level run-length matrix model, gray-level size region matrix model, neighborhood gray-level difference matrix model, etc.; or it can be a multi-scale filtering feature model, including wavelet transform model, etc. The embodiments disclosed herein do not limit this.

[0095] In one possible implementation, before extracting the radiomics features of the first image set based on the feature extraction model, the method further includes: performing standardization preprocessing on the first image set to obtain a first image set with preset numerical distribution characteristics; and performing wavelet transform on the first image set to extract multi-scale feature information.

[0096] Before feature extraction, the first image set can be preprocessed by standardization. Similarly, Z-Score standardization can be used for processing. The standardization process can be found in formula (1):

[0097]

[0098] in, This represents the intensity value of a single pixel in the original image. This is the average of the intensity values ​​of all pixels in the original image. The standard deviation of the intensity values ​​of all pixels in the original image. This is the intensity value of the pixel after standardization.

[0099] Preset numerical distribution characteristics can adjust the pixel intensity of the original image to a preset range (e.g., mean of 0 and variance of 1), thereby reducing model performance fluctuations caused by differences in brightness or contrast.

[0100] After standardization, wavelet transform can be applied to the first image set to extract multi-scale feature information. Wavelet transform can decompose an image into spaces with different frequencies and scales, thereby extracting multi-resolution texture information.

[0101] In this embodiment, by performing standardized preprocessing on the first image set followed by wavelet transform, many biological characteristics of non-functional pituitary adenomas may be most pronounced at specific scales. Therefore, wavelet transform, through multi-scale decomposition, can make explicit these texture patterns, which are difficult to detect in the original images and are strongly correlated with invasiveness, thereby achieving multi-scale, multi-resolution texture analysis capabilities. Simultaneously, since wavelet features are calculated at multiple scales, they also exhibit stronger robustness to unavoidable minor noise, resolution differences, or changes in reconstruction parameters during image acquisition.

[0102] In one possible implementation, the pyradiomics library can be used to extract image omics features of the target region. Pyradiomics can be used to extract multiple high-dimensional image features, such as texture, morphology, and statistics.

[0103] In this embodiment of the disclosure, by extracting high-dimensional image features, a comprehensive and in-depth digital profile of non-functional pituitary adenomas can be constructed, providing a solid data foundation for subsequent image classification.

[0104] In one possible implementation, the method further includes: acquiring a first image set, the first image set including a first full-size image set and a first target region image set, the first target region image set being obtained by region segmentation of the first full-size image set; extracting radiomics features of the first image set based on a feature extraction model to obtain a first feature dataset, the first feature dataset including a first full-size feature dataset and a first target region feature dataset; and performing dimensionality reduction processing on the first feature dataset to obtain a second radiomics feature set.

[0105] In one possible implementation, extracting radiomics features from the first image set based on the feature extraction model may include: extracting features from the first full-size image set and the first target region image set, respectively, based on the feature extraction model. That is, feature extraction is performed on the four image sets respectively, and the resulting first feature dataset may include four feature datasets: a first full-size coronal feature dataset, a first full-size sagittal feature dataset, a first target region coronal feature dataset, and a first target region sagittal feature dataset.

[0106] When extracting features, the feature extraction methods mentioned above can be used, which will not be elaborated on in this embodiment.

[0107] In this embodiment of the disclosure, by simultaneously extracting features from the entire image and the region of interest, richer and more multi-dimensional feature information can be obtained from the image data. This allows for the simultaneous capture of the characteristics of the lesion area itself as well as the background information of the surrounding tissue, thereby increasing the range of feature extraction and improving feature diversity.

[0108] For example, suppose the four feature datasets generated after extracting radiomics features from full-size MRI images and ROI images are as follows:

[0109] 1. full_image_features represents the full-size coronal feature set, with 558 sample data.

[0110] 2. roi_image_features represents the coronal feature set of the first target region, consisting of 558 sample data.

[0111] 3. full_image_features_T2 represents the full-size sagittal (T2) feature set, with 180 sample data.

[0112] 4. roi_image_features_T2 represents the sagittal (T2) feature set of the first target region, consisting of 180 sample data.

[0113] After feature extraction, the feature dimensions of the four feature datasets can be consistent.

[0114] In one possible implementation, after obtaining the four feature datasets mentioned above, feature fusion processing can be performed on them. Specifically, the feature vectors from the four original feature sets can be horizontally concatenated to form a fused ultra-high-dimensional feature vector. Then, this ultra-high-dimensional feature vector is reduced in dimensionality.

[0115] Dimensionality reduction can begin with variance thresholding. Applying a variance thresholding method removes features with excessively low variance from the ultra-high-dimensional feature matrix. Subsequent feature processing allows for global feature selection or extraction, achieving both dimensionality reduction and filtering. This step can employ Principal Component Analysis (PCA) to perform PCA on the fused features, retaining principal components with over 95% cumulative variance contribution and extracting the most representative information globally. Alternatively, Linear Discriminant Analysis (LDA) can be used. Based on sample labels, LDA is performed on the fused features to find the projection direction that best distinguishes different categories, taking into account the combined effects of all planar and regional features. Finally, SelectKBest can be used. SelectKBest, based on the F-test (f_classif) of the analysis of variance, directly selects the k most discriminative features from the fused ultra-high-dimensional feature set. Through these steps, the final dimensionality-reduced features are obtained. The second radiomics feature set obtained in this way includes the second full-size coronal radiomics feature set, the second full-size sagittal radiomics feature set, the second target region coronal radiomics feature set, and the second target region sagittal radiomics feature set. The feature dimensions of these four radiomics feature sets are consistent.

[0116] In this embodiment, the feature datasets are first fused, and then dimensionality reduction is performed, which can preserve the correlation information of the features more completely. By using dimensionality reduction methods such as PCA, interference can be removed, redundant features can be reduced, the risk of overfitting can be lowered, and thus the classification accuracy can be improved.

[0117] In one possible implementation, after obtaining the four feature datasets mentioned above, dimensionality reduction can be performed on the four feature datasets. Feature fusion can then be performed on the four dimensionality-reduced feature datasets. Specifically, dimensionality reduction can be performed on each of the four feature datasets separately, using the methods described above, which will not be elaborated upon in this embodiment. After dimensionality reduction, the dimensionality-reduced features can be horizontally concatenated to form a fused feature set. The second radiomics feature set obtained in this way includes a fused feature set resulting from the fusion of the following four feature sets: the second full-size coronal radiomics feature set, the second full-size sagittal radiomics feature set, the second target region coronal radiomics feature set, and the second target region sagittal radiomics feature set.

[0118] In this embodiment of the disclosure, the feature dataset is first reduced in dimensionality and then fused. This modular design can not only significantly improve processing efficiency, but also optimize the dimensionality reduction algorithm for different feature sets, thereby enhancing the robustness and discriminativeness of feature representation.

[0119] S13. Based on preset feature selection rules, the extracted radiomics features are subjected to dimensionality reduction processing to obtain a first radiomics feature set, which is used to quantify the invasiveness of the target region structure.

[0120] After extracting the radiomics features of the target region, the extracted radiomics features can be dimensionality-reduced based on a preset feature selection rule to obtain a first radiomics feature set. Specifically, the preset feature selection rule can be to select features that are strongly correlated with invasiveness from the extracted features to obtain the first radiomics feature set. Feature selection can be performed using methods such as analysis of variance (ANOVA) and the Mann-Whitney U test to calculate the significance of the distribution difference of each feature among different invasiveness groups, and retain features with a difference threshold less than a preset threshold. Other methods can also be used for dimensionality reduction. The dimensionality-reduced feature set only needs to be able to quantitatively characterize the invasiveness of the target region structure. This embodiment of the present disclosure does not impose any limitations on this.

[0121] In one possible implementation, the step of performing dimensionality reduction processing on the extracted radiomics features based on preset feature selection rules to obtain a first radiomics feature set includes: selecting features from the second radiomics feature set among the radiomics features extracted from the image to be classified to obtain the first radiomics feature set.

[0122] After obtaining the second radiomics feature set, features from the second radiomics feature set can be selected to directly obtain the first radiomics feature set. That is, the features in the first radiomics feature set and the second radiomics feature set are the same, or the first feature set is a subset of the second feature set.

[0123] For example, Figure 16 A schematic diagram showing the first twenty features after dimensionality reduction according to an embodiment of the present disclosure is provided, wherein 1-20 in the diagram are numbered as follows:

[0124] First-order statistics include: 1. Ibp-2D_firstorder_Entropy, which is the first-order entropy (after applying an Ibp-2D filter); Shape features include: 4. original_shape_MajorAxisLength, which is the maximum axis length (on the original image); 5. original_shape_Maximum3Ddiameter, which is the maximum three-dimensional diameter (on the original image); 8. original_shape_SurfaceArea, which is the surface area (on the original image); 11. original_shape_Maximum2DDiameterSlice, which is the maximum two-dimensional diameter (within the slice) (on the original image); 12. original_shape_Maximum2DdiameterColumn, which is the maximum two-dimensional diameter (column direction) (on the original image).

[0125] The characteristics of a gray-level run-length matrix (GLRLM) are as follows: 2. logarithm_glrlm_GrayLevelNonUniformity (after applying a logarithmic filter) is gray-level non-uniformity; 6. wavelet-LLH_glrlm_RunLengthNonUniformity (after applying an LLH wavelet filter) is run-length non-uniformity; 17. wavelet-HHL_glrlm_RunLengthNonUniformity (after applying an HHL wavelet filter) is run-length non-uniformity; 16. square_glrlm_GrayLevelNonUniformity (after applying a square filter) is gray-level non-uniformity; 19. original_glrlm_GrayLevelNonUniformity (on the original image) is gray-level non-uniformity; 9. bp-2D_glrlm_GrayLevelNonUniformity (after applying an Ibp-2D filter) is gray-level non-uniformity; 10. squareroot_glrlm_GrayLevelNonUniformity refers to the gray level non-uniformity (after applying the square root filter), and wavelet-LLH_glrlm_GrayLevelNonUniformity refers to the gray level non-uniformity (after applying the LLH wavelet filter).

[0126] The characteristics of the gray-level size region matrix (GLSZM) are: 3. gradient_glszm_GrayLevelNonUniformity (after applying a gradient filter), 7. wavelet-HLL_glszm_GrayLevelNonUniformity (after applying an HLL wavelet filter), 15. logarithm_glszm_GrayLevelNonUniformity (after applying a logarithmic filter), and 13. wavelet-LHL_glszm_GrayLevelNonUniformity (after applying an LHL wavelet filter).

[0127] The characteristics of the gray level dependency matrix (GLDM) are: 18. exponential_gldm_GrayLevelNonUniformity, i.e., gray level non-uniformity (after applying an exponential filter); 20. wavelet-LHL_gldm_DependenceNonUniformity, i.e., dependency non-uniformity (after applying an LHL wavelet filter).

[0128] Among these, the non-uniformity features such as logarithm_glrlm_GrayLevelNonUniformity, wavelet-HHL_glrlm_RunLengthNonUniformity, exponential_gldm_GrayLevelNonUniformity, and wavelet-LHL_gldm_DependenceNonUniformity have higher non-uniformity values, indicating more uneven and chaotic texture within the tumor, and thus stronger invasiveness. Ibp-2D_firstorder_Entropy measures the randomness or unpredictability of image information. Higher entropy values ​​indicate a more chaotic and complex grayscale distribution within the tumor, directly reflecting the disordered internal structure and the presence of multiple components (such as water, necrotic tissue, active cells, and hemorrhage), and are strongly correlated with high invasiveness. The morphological characteristics of original_shape_Maximum3Ddiameter, original_shape_MajorAxisLength, original_shape_SurfaceArea, and original_shape_Maximum2DdiameterSlice may also affect the invasiveness of tumors.

[0129] Figure 17A schematic diagram showing a heatmap of the first twenty features after dimensionality reduction according to an embodiment of the present disclosure. Wherein, Figure 17 The meaning of the number in the middle and Figure 16 Consistent. From Figure 17 As can be seen, the dependencies between the features after dimensionality reduction are not strong, indicating that the features after dimensionality reduction can effectively reflect the feature distribution characteristics of the original data. There is no obvious redundancy among the features, and they have strong independence and distinguishability, which can effectively quantify the invasiveness of the target region structure.

[0130] In this embodiment of the disclosure, the invasiveness of the region of interest structure in the image of a non-functional pituitary adenoma is quantified by using a dimensionality-reduced radiomics feature set, which enables accurate classification of non-functional pituitary adenomas and provides a reliable imaging basis for preoperative assessment of the biological behavior of the tumor.

[0131] In one possible implementation, after dimensionality reduction, a Synthetic Minority Over-sampling Technique (SMOTE) can be used to generate minority class samples to balance the class ratio. Specifically, SMOTE increases the number of minority class samples by interpolating between them to generate new synthetic samples, thus balancing the ratio of minority to majority classes (e.g., 1:1 in binary classification).

[0132] In this embodiment, by generating minority class samples and balancing the class ratio, SMOTE generates new synthetic samples through linear interpolation, which enriches the feature space of the minority class and fills the data distribution gaps between existing minority class samples. When classifying invasive subtypes of non-functional pituitary adenomas, without class balancing, restricted samples (majority class) may dominate, leading to model bias and misclassification of invasive subtypes (minority class) as "restricted," resulting in a high false negative rate. SMOTE, however, synthesizes samples of invasive subtypes, allowing the model to learn from a sufficient number of positive examples during training, significantly improving the sensitivity of invasive subtype identification and directly reducing the risk of missed diagnoses in clinical practice.

[0133] S14. Input the first image omics feature set into the classification model to obtain the classification result.

[0134] The classification model can be a pre-determined classification model. After obtaining the first impact omics feature set, the first impact omics feature set can be input into the pre-determined classification model to obtain the classification result.

[0135] In one possible implementation, the method further includes: dividing the second radiomics feature set into a training set and a test set; training the training set using multiple first models; evaluating the training results based on the test set; and obtaining the classification model based on the evaluation results.

[0136] After obtaining the second radiomics feature set, the second radiomics feature set can be divided into a training set and a test set. The training set and the test set can be divided according to a ratio, such as 80% training set and 20% test set, or 70% training set and 30% test set, etc. The embodiments disclosed herein do not limit this.

[0137] In one possible implementation, training the training set using multiple first models includes: inputting a fused second radiomics feature set into the first model for training, wherein the second radiomics feature set integrates coronal and sagittal features; and determining a classification model based on the influence of coronal and sagittal features on the classification performance of the multiple first models.

[0138] The second radiomics feature set obtained above includes the second full-size coronal radiomics feature set, the second full-size sagittal radiomics feature set, the second target region coronal radiomics feature set, and the second target region sagittal radiomics feature set; or a fused feature set obtained by fusing the four feature sets of the second full-size coronal radiomics feature set, the second full-size sagittal radiomics feature set, the second target region coronal radiomics feature set, and the second target region sagittal radiomics feature set.

[0139] The second radiomics feature set can be input into the first model for training. Since the second radiomics feature set integrates coronal and sagittal features, multiple first models can be trained using coronal and sagittal features respectively. This allows for simultaneous testing of coronal and sagittal features. Subsequently, the classification model can be determined by evaluating the training effects of different first models on coronal and sagittal features.

[0140] For example, taking the four feature datasets mentioned above as an example, the resulting second radiomics feature set can still be:

[0141] 1. full_image_features represents the full-size coronal feature set, with 558 sample data.

[0142] 2. roi_image_features represents the coronal feature set of the first target region, consisting of 558 sample data.

[0143] 3. full_image_features_T2 represents the full-size sagittal (T2) feature set, with 180 sample data.

[0144] 4. roi_image_features_T2 represents the sagittal (T2) feature set of the first target region, consisting of 180 sample data.

[0145] However, the feature dimensions have changed. Assuming there were 500 feature dimensions before dimensionality reduction, after the dimensionality reduction process described above, it can be reduced to 100 feature dimensions. Since these four feature datasets all have 100 feature dimensions, these four feature datasets from the second radiomics feature set can be directly input into the first model for training.

[0146] The resulting second image omics feature set can also be a large feature dataset with the number of samples being the sum of the number of samples in the four feature sets mentioned above (558+558+180+180). Assuming there are 500 feature dimensions before dimensionality reduction, after the dimensionality reduction process mentioned above, it can be reduced to 100 feature dimensions. This large feature dataset can be directly input into the first model for training.

[0147] The first model can be a traditional machine learning model, such as linear models, logistic regression, support vector machine (SVM), or a tree-based model, such as decision tree, random forest, gradient boosting machine, or a deep learning model, such as convolutional neural network (CNN), recurrent neural network (RNN), or any other applicable classification or regression algorithm. This disclosure does not limit the specific model.

[0148] When training the classification model, it can be trained according to actual needs. For example, suppose we need to classify non-functioning pituitary adenomas into silent gonadotropin-releasing adenomas, silent adrenocorticotropic hormone-secreting adenomas, and silent growth hormone-secreting adenomas. That is, the second radiomics feature set includes these four types of non-functioning pituitary adenomas. Then, we can input the dimensionality-reduced training set features into various first models for training to obtain the classification results of these four types of non-functioning pituitary adenomas.

[0149] In one possible implementation, the image to be classified includes images of silent gonadotropin-releasing tumors (GRT-RT) and silent adrenocorticotropic hormone-secreting tumors (ACT-HT-RT); the invasiveness of silent GRT-RT and silent ACTC-RT differs; the first radiomics feature set is input into a classification model to obtain a classification result, including: the classification model classifies the image to be classified as an image of a silent GRT-RT or an image of a silent ACTC-RT based on the difference in invasiveness.

[0150] Silent gonadotropin-secreting adenomas often exhibit higher invasiveness, while silent adrenocorticotropic hormone-secreting adenomas are mostly non-invasive. Based on the invasiveness of these two types of tumors (target regions) quantified using the first radiomics feature set, inputting the first radiomics feature set into the classification model yields the classification results for these two types of images. That is, the image to be classified can be categorized as an image of a silent gonadotropin-secreting adenoma or an image of a silent adrenocorticotropic hormone-secreting adenoma. In clinical practice, these two types of non-functional pituitary adenomas account for approximately 90% to 95% of all non-functional pituitary adenomas. Due to their more prominent invasiveness and significant clinical implications, binary classification training on these two types of images can not only improve classification accuracy but also significantly enhance the applicability of the model in real clinical scenarios.

[0151] In one possible implementation, the first model includes a random forest model, a support vector machine model, a logistic regression model, and a deep learning model.

[0152] Random forests, being an ensemble method based on multiple decision trees, offer high stability. Support vector machines are suitable for high-dimensional feature data and effectively handle small sample problems. Logistic regression provides interpretable decision boundaries in binary classification tasks.

[0153] Deep learning models can be constructed using TensorFlow / Keras to build fully connected neural networks (MLPs). The input layer can adjust the number of neurons based on the feature dimension, which can be a dimensionality-reduced feature dimension. Multiple fully connected layers (dense layers) can be added as hidden layers after the input layer. These fully connected layers can introduce non-linearity by modifying the linear units (ReLU activation function). To avoid overfitting, dropout layers can be added between the hidden layers, improving the model's generalization ability. The output layer can use a normalized exponential function (Softmax activation function) for classification prediction; this function is suitable for multi-class classification problems.

[0154] After training a deep learning model, it can be optimized. For example, the Adam optimizer can be used to improve gradient update efficiency and accelerate convergence, and / or the optimal model checkpoint can be used to automatically save the best performing model weights, and / or the early stopping mechanism can be used to monitor the validation set loss and prevent overfitting.

[0155] After training with multiple primary models, the test set can be input into each model, and the models can be evaluated based on their classification results. During evaluation, since the secondary radiomics features integrate coronal and sagittal features, the training effectiveness of different primary models for these features can be assessed. This allows for the selection of models that effectively integrate dual-view features, thereby improving classification accuracy and robustness.

[0156] Evaluation can be conducted using classification metrics such as accuracy, precision, recall, F1 score, and ROC curve (Receiver Operating Characteristic Curve). The ROC curve can calculate the false positive rate (FPR) and true positive rate (TPR), which can visualize the model's discriminative ability. Additionally, the AUC value (Area Under Curve) measures the classifier's ability to distinguish between positive and negative samples. When the AUC value is greater than a certain preset threshold, it indicates excellent classification performance.

[0157] When evaluating the training results, you can choose the core metrics according to your needs. For example, if you need to balance precision and recall, you can prioritize F1-score. If you need to evaluate the model's overall ability to distinguish between positive and negative samples, you can prioritize AUC. You can also combine multiple metrics for analysis. For example, in the case of imbalanced class distribution, avoid relying solely on accuracy and consider parameters such as precision, recall, and F1-score simultaneously.

[0158] For example, Figure 18 This diagram illustrates the ROC curve of a full-size feature dataset according to an embodiment of the present disclosure. Figure 19 A schematic diagram of the ROC curve of a target region feature dataset according to an embodiment of the present disclosure is shown. Figure 18 and Figure 19 The images show models trained using Random Forest, Logistic Regression, and Support Vector Machine (SVM), respectively. It can be seen that... Figure 18 The random forest model performed well. Figure 19 The logistic regression model performed well, and the classification AUC value after region segmentation was generally higher than that of the original full-size image. Therefore, the accuracy of classification by the model after region segmentation will be significantly improved.

[0159] After evaluating the training results, the best model can be selected as the classification model, or multiple well-performing models can be selected as classification models. The first radiomics feature set can be input into multiple classification models to obtain multiple classification results for doctors to refer to.

[0160] In this embodiment of the disclosure, by inputting radiomics features into multiple models, doctors can be assisted in a better and more comprehensive assessment of the patient's condition. By analyzing and comparing the results of different classification models, the accuracy of diagnosing non-functional pituitary adenomas can be further improved.

[0161] In one possible implementation, obtaining the classification model based on the evaluation results includes: selecting one model from the first models based on the evaluation results to obtain the classification model; or fusing at least two first models to obtain the classification model.

[0162] After evaluating the training results, at least two first models can be fused to obtain the classification model. Specifically, a stacking generalization model can be used for model fusion. Multiple first models can be selected as base models, and then an untrained model can be selected as a meta-model to learn the optimal combination of the prediction results of the base models. After training, the trained meta-model can be used as the classification model.

[0163] In this embodiment of the disclosure, by fusing multiple first models and performing cross-validation, the final classification model exhibits more stable and reliable performance than a single model when facing unknown data, thus avoiding performance fluctuations caused by relying on a single model.

[0164] In this embodiment, an image to be classified is acquired, and automated region segmentation is performed on the image to be classified to obtain a target region. The image to be classified includes an image of a non-functional pituitary adenoma. Based on a feature extraction model, radiomics features of the target region are extracted. Based on a preset feature selection rule, the extracted radiomics features are subjected to dimensionality reduction processing to obtain a first radiomics feature set. The first radiomics feature set is used to quantify the invasiveness of the target region structure. The first radiomics feature set is input into a classification model to obtain a classification result. Therefore, non-functional pituitary adenomas can be classified according to their degree of invasiveness with high accuracy, thereby assisting doctors in diagnosing the type of non-functional pituitary adenoma more quickly and accurately, reducing the treatment burden on patients and the occurrence of postoperative adverse events.

[0165] Application scenario examples

[0166] Figure 20 This diagram illustrates a full-process model training flowchart according to an embodiment of the present disclosure. Figure 21 A schematic flowchart of a full-process model test according to an embodiment of the present disclosure is shown.

[0167] like Figure 20 and Figure 21 As shown, the overall process is divided into two core parts:

[0168] 1. Train the nnU-Net segmentation model to accurately segment the tumor region.

[0169] 2. Training the classification model: Extract key features based on the segmentation results and train a classifier to distinguish between silent ACTH-type and SF-type pituitary adenomas.

[0170] Training the nnU-Net segmentation model

[0171] First, data preparation was carried out, including 746 coronal MRI images and 198 sagittal MRI images, all of which were manually labeled by professional physicians.

[0172] The data is divided into raw MRI images and corresponding tumor-labeled images.

[0173] After screening, approximately 559 sagittal data points that meet the training requirements of nnU-Net were selected as the training set.

[0174] Training of segmentation models

[0175] According to the nnU-Net specification, the data is organized into the format expected by nnU-Net.

[0176] We employ the nnU-Net automated training process to optimize segmentation performance.

[0177] After training, the segmentation effect was evaluated and hyperparameters were adjusted to improve segmentation accuracy, with the final segmentation overlap rate reaching over 80%.

[0178] Training on pituitary tumor classification models

[0179] First, a classification task analysis was conducted. Research shows that pituitary tumor classification primarily depends on the characteristics of the tumor itself and its surrounding area. Therefore, two datasets can be constructed:

[0180] 1. Full-size MRI images (containing complete anatomical information).

[0181] 2. Image of the ROI of the tumor region (focusing only on the segmented tumor and its surrounding area).

[0182] During classifier training, the two sets of features were extracted separately and then integrated. Simultaneously, the impact of fusing coronal and sagittal features on classification performance was tested to ultimately select the optimal classification model. Finally, 558 cases were selected from 746 coronal cases, and 180 cases were selected from 198 sagittal cases for training and validation. In addition, a dataset of 154 patients sharing both coronal and sagittal features was also included.

[0183] Data preprocessing

[0184] 1. Image quality optimization

[0185] Gaussian blur denoising can reduce image noise and improve the signal-to-noise ratio.

[0186] Contrast enhancement can improve local contrast and make tumor boundaries clearer.

[0187] Pixel value normalization can unify the pixel value distribution of MRI images.

[0188] Size standardization allows for adjustment of image resolution, ensuring consistency in data dimensions.

[0189] 2. Pyradiomics Feature Extraction

[0190] Extract radiomics features from full-size MRI images and ROI images. Generate a feature dataset:

[0191] (1). full_image_features.csv (full image features, 558 data points).

[0192] (2). roi_image_features.csv (ROI features, 558 data points).

[0193] (3). full_image_features_T2.csv (T2 modality full image features, 180 data points).

[0194] (4). roi_image_features_T2.csv (T2 modal ROI features, 180 data points).

[0195] 3. Feature Standardization and Selection

[0196] Z-score normalization (StandardScaler) can standardize the eigenvalues ​​to a mean of 0 and a standard deviation of 1.

[0197] Variance threshold can be used to remove features with low variance.

[0198] Principal component analysis (PCA) can retain more than 95% of the key information of the cumulative variance contribution rate.

[0199] Linear discriminant analysis (LDA) can enhance class differentiation.

[0200] SelectKBest can select the k most discriminative features based on f_classif.

[0201] Category balance:

[0202] SMOTE (Synthetic Minority Over-sampling Technique) can be used to generate minority class samples to balance the class ratio.

[0203] Training of classification models

[0204] 1. Dataset partitioning

[0205] Training set (80%)

[0206] Test set (20%)

[0207] 2. Machine Learning Models

[0208] Random Forest (RandomForestClassifier) ​​is an ensemble method based on multiple decision trees, and it has high stability.

[0209] Support Vector Machines (SVM, kernel='linear') are suitable for high-dimensional feature data and can effectively handle small sample problems.

[0210] Logistic regression can provide interpretable decision boundaries in binary classification tasks.

[0211] 3. Deep Learning Models

[0212] Fully connected neural networks (MLPs) can be built using TensorFlow / Keras:

[0213] Input layer: Adjust the number of neurons according to the feature dimension.

[0214] Hidden layers: Add multiple Dense layers with ReLU activation function.

[0215] Dropout layer: Prevents overfitting and improves the model's generalization ability.

[0216] Output layer: Softmax activation function is used for classification prediction.

[0217] 4. Training optimization strategies

[0218] Adam optimizer: Improves gradient update efficiency and accelerates convergence.

[0219] ModelCheckpoint: Automatically saves the weights of the best-performing model.

[0220] EarlyStopping: Monitors validation set loss to prevent overfitting.

[0221] Regarding model evaluation

[0222] 1. Classification index analysis

[0223] Accuracy, Precision, Recall, and F1-score (considering both Precision and Recall)

[0224] 2. ROC curve & AUC value

[0225] ROC curve (Receiver Operating Characteristic Curve): Calculates the false positive rate (FPR) and true positive rate (TPR), visualizing the model's ability to distinguish between different targets.

[0226] AUC (Area Under Curve): Measures the ability of a classifier to distinguish between positive and negative samples. An AUC > 0.8 indicates excellent classification performance.

[0227] Figure 22 A schematic diagram of an image classification processing flow according to an embodiment of the present disclosure is shown. Figure 22 As shown, the process may include the following steps: when new patient MRI images are available, data preprocessing is performed first, then the ROI region is cut out using the nnUNet segmentation model, then feature extraction is performed using Pyradiomics, followed by data processing such as dimensionality reduction, and finally the trained classification model is applied to output the classification result.

[0228] In this embodiment, automated medical image segmentation based on nnU-Net improves the accuracy of tumor region extraction. Using nnU-Net for automatic tumor region segmentation eliminates the need for manual hyperparameter adjustment, adapts to medical image data with different resolutions and scanning protocols, and improves segmentation quality and generalization ability. Automated segmentation allows feature extraction only from the lesion region, avoiding the noise impact of full-image features and improving the accuracy of classification tasks. The overlap rate between manual and automatic segmentation reaches over 80%.

[0229] In this embodiment, Pyradiomics is used to extract image features, enhancing the information representation capability of the classification model. Pyradiomics is employed to extract high-dimensional image features such as texture, morphology, and statistics, and features of the entire image and the Region of Interest (ROI) (regions segmented by nnU-Net) are calculated separately to provide richer classification information. Comparing the classification performance of the entire image features and the ROI features, in the test set of this embodiment, the ROC of the ROI features is improved by more than 5%, reaching over 80%, indicating that the segmented features are more discriminative.

[0230] In this embodiment, data standardization and dimensionality reduction improve model training efficiency. Z-score normalization is used to unify the feature scale, preventing inconsistent feature value ranges from affecting classifier training. Dimensionality reduction is achieved using PCA (Principal Component Analysis) and T-tests, reducing the feature dimension from over 500 to less than 100 while retaining principal components, thus removing interference and improving classification accuracy.

[0231] In this embodiment of the disclosure, SMOTE (Synthetic Minority Oversampling) is used to generate synthetic samples to solve the class imbalance of medical image data (e.g., if the ratio of SF to ACTH is 3:1, it can be processed to 1:1), which can handle class imbalance and improve the recognition rate of minority classes.

[0232] In the embodiments of this disclosure, multiple classifiers and model fusion are employed to improve classification stability. Random forest (RF), support vector machine (SVM), logistic regression (LR), and deep learning-based (DNN) classification methods are used, and a stacking model fusion strategy is employed to reduce the risk of overfitting from a single model and improve generalization ability.

[0233] In this embodiment, an end-to-end automated process can be achieved, improving the reproducibility of the model. The entire process is automated through programming, including data preprocessing, nnU-Net segmentation, Pyradiomics feature extraction, dimensionality reduction, classification training, and model evaluation, ensuring efficient process and stable results. The tumor classification accuracy can reach over 80%.

[0234] This invention utilizes a classifier built with machine learning and other algorithms to automatically segment and distinguish between silent ACTH and SF expression types of non-functional pituitary adenomas using preoperative enhanced MRI images, achieving high accuracy. This assists neurosurgeons in determining treatment plans more quickly and effectively, reducing the treatment burden on patients and the occurrence of postoperative adverse events.

[0235] Application scenario examples

[0236] This disclosure provides a radiomics method for automatically identifying pituitary adenoma regions and predicting the pathological grade of pituitary adenomas based on enhanced MRI images. Specifically, this method utilizes enhanced MRI images to identify and segment pituitary adenoma regions, and combines quantitative indicators to predict the pathological grade of pituitary adenomas. This method can be used for preoperative prediction of the pathological grade of pituitary adenomas, which is of great significance for guiding treatment strategies and prognostic assessments for patients with pituitary adenomas. The non-invasive imaging features of MRI play an important role in the diagnosis of pituitary adenomas, but inter-observer variability may affect the reliability of quantitative indicators. Radiomics, as an emerging approach, can convert images into high-dimensional, mineable data, providing biomarkers for the diagnosis and pathological grading of pituitary adenomas. By combining the features of enhanced MRI images with radiomics analysis methods, automatic identification of pituitary adenoma regions and accurate prediction of pathological grades can be achieved, providing clinicians with more accurate diagnostic information, thereby guiding personalized treatment plans and prognostic assessments.

[0237] This disclosure correlates high-throughput radiomics features with pituitary adenoma grading information, representing an upgrade from traditional radiomics methods. It further explores and investigates the interpretability of radiomics features. By ensuring an understanding of the contribution of each radiomics type to predicting the pathological grade of pituitary adenoma, it provides important reference for future radiomics research.

[0238] This disclosure provides an image processing method for automatically identifying pituitary adenoma regions and predicting the pathological grade of pituitary adenomas based on enhanced MRI images. The specific implementation process is as follows:

[0239] S1: Delineate the ROI region, i.e. the pituitary adenoma lesion area, on enhanced MRI images of patients with pituitary adenoma confirmed by surgical pathology, and train the model using nnUNet.

[0240] Specifically, in this embodiment of the disclosure, the pituitary adenoma lesion area can be delineated by a doctor based on clinical experience in the enhanced MRI images of patients with pituitary adenoma confirmed by surgical pathology. In order to reduce personal subjectivity, it can be jointly decided by 2-3 doctors with 5-10 years of experience.

[0241] nnUNet is an open-source deep learning framework for medical image segmentation, built on top of the PyTorch deep learning library. nnUNet aims to provide an end-to-end solution, including data preprocessing, model training, and inference, to make medical image segmentation tasks easier.

[0242] S2: Based on the characteristic images of the pituitary adenoma, obtain the radiomics features of the pituitary adenoma for subsequent classifier training;

[0243] The method for extracting radiomics features in this embodiment is based on the segmentation results described above. Software can be used to read each bimodal image slice of each patient's Mr sequence and integrate them into a three-dimensional raw, almost raw raster data (NRRD) image. The same process is performed on each image slice using a ROI mask to generate a three-dimensional labeled NRRD image. The original image can then be normalized (see formula (1)) and subjected to wavelet transform. Subsequently, pyradiomics is used to extract radiomics features. Eighteen first-order statistical features are extracted from the original image, including 14 shape features, 22 gray-level co-occurrence matrix (GLCM) features, 16 gray-level run-length matrix (GLRLM) features, 16 gray-level size region matrix (GLSZM) features, and 14 gray-level dependency matrix (GLDM) features.

[0244] S3: Multiple methods, including independent samples t-test, multicollinearity analysis, and principal component analysis, were used to reduce the dimensionality of the extracted features and obtain representative features of the radiomics characteristics of pituitary adenomas.

[0245] Because high-dimensional radiometric features may contain redundant and irrelevant information, which can lead to overfitting and reduced classifier performance, dimensionality reduction is necessary. To achieve the best dimensionality reduction effect and effectively avoid overfitting, this embodiment uses independent sample t-test, multicollinearity detection, and principal component analysis (PCA) for dimensionality reduction. After comparison, the top-ranked radiomics features are: large-area low-gray-level emphasis in the original image gray-level co-occurrence matrix (GLSZM), correlation of the wavelet transform-LHL direction gray-level co-occurrence matrix (GLCM), cluster shadow in the wavelet transform-LHH direction gray-level co-occurrence matrix (GLCM), information measurement correlation coefficient 2 in the wavelet transform-LHH direction gray-level co-occurrence matrix (GLCM), maximum correlation coefficient in the wavelet transform-LHH direction gray-level co-occurrence matrix (GLCM), high gray-level emphasis in large and small areas in the wavelet transform-LHH direction gray-level co-occurrence matrix (GLCM), and gray-level emphasis in the wavelet transform-LHH direction gray-level co-occurrence matrix (GLCM). Emphasis on low gray levels in regions of varying sizes; prominence of gray-level co-occurrence matrix (GLCM) clusters in the wavelet transform-HLH direction; proportion of gray-level regions in the wavelet transform-HLH direction; maximum value of first-order statistical features in the wavelet transform-HHL direction; median of first-order statistical features in the wavelet transform-HHL direction; shading of gray-level co-occurrence matrix (GLCM) clusters in the wavelet transform-HHL direction; non-uniformity of gray-level regions in the wavelet transform-HHL direction; gray-level size dependence emphasis in the wavelet transform-HHH direction; non-uniformity of gray-level regions in the wavelet transform-HHH direction; normalization of size non-uniformity of gray-level regions in the wavelet transform-HHH direction; small area emphasis in gray-level regions in the wavelet transform-HHH direction; small area emphasis with low gray levels in gray-level regions in the wavelet transform-HHH direction.

[0246] S4: Use machine learning and other algorithms to train a classifier, combine it with pathological test results, and use dimensionality reduction features to predict the pathological subtype of pituitary adenoma images.

[0247] Different subtypes of pituitary adenomas vary in tumor morphology, size, and relationship with surrounding blood vessels, leading to differences in postoperative outcomes and tumor recurrence rates among patients with different subtypes. Utilizing imaging data to predict the pathological subtype of pituitary adenomas in advance can assist neurosurgeons in developing surgical plans.

[0248] Application scenario examples

[0249] Imaging and clinical information of 44 patients with pituitary adenomas were collected, as follows:

[0250] From left to right: Serial number, gender, age, pathology.

[0251] A 63-year-old female patient with a pituitary adenoma showed occasional positive results for adrenocorticotropic hormone (ACTH) and positive results for transcription factor T-PIT on immunohistochemical staining.

[0252] Two males, aged 44, had pituitary adenomas. Immunohistochemical staining showed positive results for luteinizing hormone (LH) and follicle-stimulating hormone (FSH).

[0253] Three women, 71 years old, had pituitary adenomas. Immunohistochemical staining occasionally showed positive results for luteinizing hormone (LH).

[0254] Four males, 48 ​​years old, had pituitary adenomas and were positive for luteinizing hormone (LH) and follicle-stimulating hormone (FSH) by immunohistochemistry.

[0255] Five women, 37 years old, had pituitary adenomas. Immunohistochemical staining showed negative results for all six pituitary hormones, but positive results for T-PIT.

[0256] Six males, 35 years old, had pituitary adenomas and strokes. Immunohistochemical staining occasionally showed positive results for luteinizing hormone (LH).

[0257] Seven men, 51 of whom had pituitary adenomas, were positive for luteinizing hormone (LH) and follicle-stimulating hormone (FSH) by immunohistochemical staining.

[0258] Eight women, 50 years old, had pituitary adenomas, and immunohistochemical staining showed scattered positive results for adrenocorticotropic hormone (ACTH).

[0259] Nine males, 68 years old, had pituitary adenomas. Immunohistochemical staining showed positive results for follicle-stimulating hormone (FSH) and luteinizing hormone (LH).

[0260] Ten males, 62 years old, had pituitary adenomas and were positive for luteinizing hormone (LH) on immunohistochemical staining.

[0261] 11 females, 42 years old, with pituitary adenoma; immunohistochemical staining showed positive follicle-stimulating hormone (FSH).

[0262] 12 males, 64 years old, with pituitary adenomas. Immunohistochemical staining showed focal positive results for luteinizing hormone (LH) and follicle-stimulating hormone (FSH).

[0263] A 60-year-old male patient with a pituitary adenoma tested positive for luteinizing hormone (LH) and follicle-stimulating hormone (FSH) on immunohistochemical staining.

[0264] A 14-year-old male with a 63-year-old pituitary adenoma was found to be positive for luteinizing hormone (LH) on immunohistochemical staining.

[0265] A 15-year-old male with a pituitary adenoma tested positive for luteinizing hormone (LH) and follicle-stimulating hormone (FSH) on immunohistochemical staining.

[0266] A 16-year-old female patient with a pituitary adenoma tested positive for follicle-stimulating hormone (FSH) and luteinizing hormone (LH) on immunohistochemical staining.

[0267] A 17-year-old female with a 37-year-old pituitary adenoma, positive for adrenocorticotropic hormone (ACTH) by immunohistochemistry.

[0268] An 18-year-old male with a 55-year-old pituitary adenoma was diagnosed with positive luteinizing hormone (LH) and follicle-stimulating hormone (FSH) staining by immunohistochemistry.

[0269] A 19-year-old female with a 38-year-old pituitary adenoma. Immunohistochemical staining showed negative results for all six pituitary hormones, but positive for T-PIT.

[0270] A 20-year-old female with a 34-year-old pituitary adenoma, positive for adrenocorticotropic hormone (ACTH) on immunohistochemical staining, and a focally elevated Ki-67 proliferation index. Regular follow-up examinations are recommended.

[0271] A 21-year-old male, aged 59, has a pituitary adenoma that infiltrates bone tissue. Immunohistochemical staining shows positive results for follicle-stimulating hormone (FSH). Follow-up examination is recommended.

[0272] A 22-year-old female with a 49-year-old pituitary adenoma showed positive immunohistochemical staining for adrenocorticotropic hormone (ACTH) and scattered positive luteinizing hormone (LH).

[0273] A 23-year-old female with a 68-year-old pituitary adenoma, and immunohistochemical staining showed positive results for follicle-stimulating hormone (FSH).

[0274] A 24-year-old male with a 42-year-old pituitary adenoma showed scattered positive results for follicle-stimulating hormone (FSH) and luteinizing hormone (LH) on immunohistochemical staining.

[0275] A 25-year-old female with a 54-year-old pituitary adenoma, infiltrating the submucosal and bone tissue, showed scattered positive results for luteinizing hormone (LH) and adrenocorticotropic hormone (ACTH) on immunohistochemical staining.

[0276] A 26-year-old female with a 32-year-old pituitary adenoma, positive for adrenocorticotropic hormone (ACTH) by immunohistochemical staining.

[0277] A 27-year-old male with a pituitary adenoma, whose immunohistochemical staining showed positive results for luteinizing hormone (LH) and follicle-stimulating hormone (FSH).

[0278] A 28-year-old male with a 43-year-old pituitary adenoma showed focal positive immunohistochemical staining for follicle-stimulating hormone (FSH).

[0279] A 29-year-old male with a 68-year-old pituitary adenoma, and positive for follicle-stimulating hormone (FSH) on immunohistochemical staining.

[0280] A 30-year-old male, aged 65, was diagnosed with a pituitary neuroendocrine tumor (PitNET) / adenoma. He tested positive for gonadotropins (LH and FSH) via immunohistochemical staining, and was also positive for transcription factor SF-1.

[0281] 31-year-old female, 63 years old, with adrenocorticotropic hormone-secreting pituitary neuroendocrine tumor (PitNET) / adenoma, sparsely granular type.

[0282] A 32-year-old male with a pituitary adenoma, positive for luteinizing hormone (LH) and follicle-stimulating hormone (FSH) on immunohistochemical staining, consistent with a gonadotropin-secreting pituitary neuroendocrine tumor (PitNET) / adenoma.

[0283] 33-year-old female, 60 years old, pituitary neuroendocrine tumor (PitNET) / adenoma, sparse granular type.

[0284] A 34-year-old male with a 55-year-old pituitary adenoma, with positive immunohistochemical staining for luteinizing hormone (LH) and follicle-stimulating hormone (FSH).

[0285] A 35-year-old female with a 44-year-old pituitary adenoma. Immunohistochemical staining showed negative results for all six pituitary hormones, but positive for transcription factor T-PIT.

[0286] A 36-year-old male with a 69-year-old pituitary adenoma showed positive immunohistochemical staining for follicle-stimulating hormone (FSH) and focal positive luteinizing hormone (LH).

[0287] A 37-year-old female patient with a 47-year-old pituitary adenoma showed occasional positive results for adrenocorticotropic hormone (ACTH) and positive results for transcription factor T-PIT on immunohistochemical staining.

[0288] A 38-year-old male with a 53-year-old pituitary adenoma, positive for luteinizing hormone (LH) and follicle-stimulating hormone (FSH) on immunohistochemical staining.

[0289] A 39-year-old male with a 44-year-old pituitary adenoma that infiltrates bone tissue. Immunohistochemistry shows positive follicle-stimulating hormone (FSH) and scattered positive luteinizing hormone (LH). Regular follow-up examinations are recommended.

[0290] A 40-year-old woman with a 60-year-old pituitary adenoma showed focal positive results for adrenocorticotropic hormone (ACTH) on immunohistochemical staining.

[0291] A 41-year-old male with a 60-year-old pituitary adenoma showed positive immunohistochemical staining for follicle-stimulating hormone (FSH) and scattered positive luteinizing hormone (LH).

[0292] A 42-year-old female with a 59-year-old pituitary adenoma, positive for immunohistochemical transcription factor T-PIT.

[0293] A 43-year-old male with a 66-year-old pituitary adenoma tested positive for luteinizing hormone (LH) and follicle-stimulating hormone (FSH) on immunohistochemistry.

[0294] The patient was 44 years old and 46 years old, with pituitary adenoma. Immunohistochemical staining showed occasional positive results for adrenocorticotropic hormone (ACTH) and positive results for transcription factor T-PIT.

[0295] Figure 23 A schematic diagram illustrating the ROC results of imaging and clinical information of a pituitary adenoma patient according to an embodiment of this disclosure is shown. Figure 23 As shown, the AUC values ​​of both Logistic Regression and Support Vector Machine (SVM) are 0.75, indicating good classification performance.

[0296] Figure 24 A block diagram of an image classification apparatus according to an embodiment of the present disclosure is shown, such as Figure 24 As shown, the device includes:

[0297] The target region acquisition module 31 is used to acquire the image to be classified, perform automatic region segmentation on the image to be classified, and obtain the target region. The image to be classified includes images of non-functional pituitary adenomas.

[0298] The radiomics feature extraction module 32 is used to extract radiomics features of the target region based on the feature extraction model;

[0299] The first radiomics feature set acquisition module 33 is used to perform dimensionality reduction processing on the extracted radiomics features based on preset feature selection rules to obtain the first radiomics feature set, which is used to quantify the invasiveness of the target region structure.

[0300] The classification result acquisition module 34 is used to input the first image omics feature set into the classification model to obtain the classification result.

[0301] This disclosure also provides a non-volatile computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

[0302] This disclosure also provides a computer program product, including a computer program or a non-volatile computer-readable storage medium carrying the computer program, wherein the computer program, when executed by a processor, implements the steps of the above method.

[0303] Figure 25 A block diagram of an electronic device 1900 according to an embodiment of the present disclosure is shown. For example, the electronic device 1900 may be provided as a server or a terminal device. (Refer to...) Figure 25The electronic device 1900 includes a processing component 1922, which further includes one or more processors, and a memory resource represented by a memory 1932 for storing instructions, such as application programs, that can be executed by the processing component 1922. The application programs stored in the memory 1932 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 1922 is configured to execute instructions to perform the methods described above.

[0304] Electronic device 1900 may also include a power supply component 1926 configured to perform power management of electronic device 1900, a wired or wireless network interface 1950 configured to connect electronic device 1900 to a network, and an input / output interface 1958 (I / O interface). Electronic device 1900 can operate on an operating system, such as Windows Server, stored in memory 1932. TM Mac OS X TM Unix TM Linux TM FreeBSD TM Or similar.

[0305] In an exemplary embodiment, a non-volatile computer-readable storage medium is also provided, such as a memory 1932 including computer program instructions that can be executed by a processing component 1922 of an electronic device 1900 to perform the above-described method.

[0306] Computer-readable storage media can be tangible devices capable of holding and storing programs / instructions used by instruction execution devices. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0307] The computer program (or computer-readable program instructions) described herein can be downloaded from a computer-readable storage medium to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage medium in the respective computing / processing device.

[0308] The computer program (or computer program instructions) used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing state information from the computer-readable program instructions to implement various aspects of this disclosure.

[0309] Various aspects of this disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0310] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0311] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0312] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0313] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. An image classification method, characterized in that, The method includes: Acquire an image to be classified, perform automated region segmentation on the image to be classified to obtain a target region, wherein the image to be classified includes an image of a non-functional pituitary adenoma; Based on the feature extraction model, the image omics features of the target region are extracted; Based on preset feature selection rules, the extracted radiomics features are dimensionality reduced to obtain a first radiomics feature set, which is used to quantify the invasiveness of the target region structure. The first image omics feature set is input into the classification model to obtain the classification result.

2. The method according to claim 1, characterized in that, The method further includes: A first image set is obtained, which includes a first full-size image set and a first target region image set. The first target region image set is obtained by region segmentation of the first full-size image set. Based on the feature extraction model, the radiomics features of the first image set are extracted to obtain the first feature dataset, which includes a first full-size feature dataset and a first target region feature dataset. The first feature dataset is subjected to dimensionality reduction processing to obtain the second image omics feature set.

3. The method according to claim 2, characterized in that, The extracted radiomics features are dimensionality-reduced based on preset feature selection rules to obtain a first radiomics feature set, including: From the radiomics features extracted from the image to be classified, features from the second radiomics feature set are selected to obtain the first radiomics feature set.

4. The method according to claim 2, characterized in that, The method further includes: The second image omics feature set is divided into a training set and a test set; The training set is trained using multiple first models; The training results are evaluated based on the test set. Based on the evaluation results, the classification model is obtained; The training of the training set using multiple first models includes: The fused second radiomics feature set is input into the first model for training. The second radiomics feature set is a fusion of coronal and sagittal features. The classification model was determined based on the influence of coronal and sagittal features on the classification performance of various first-order models.

5. The method according to claim 4, characterized in that, The first model includes a random forest model, a support vector machine model, a logistic regression model, and a deep learning model; the classification model obtained based on the evaluation results includes: Based on the evaluation results, one model is selected from the first models to obtain the classification model; or at least two first models are fused to obtain the classification model.

6. The method according to claim 2, characterized in that, Before extracting the radiomics features of the first image set based on the feature extraction model, the method further includes: The first image set is preprocessed by standardization to obtain a first image set with a preset numerical distribution characteristic; Wavelet transform is performed on the first image set to extract multi-scale feature information.

7. The method according to any one of claims 1-6, characterized in that, The image types to be classified include magnetic resonance imaging, and the images to be classified include images of silent gonadotropin-secreting cell adenomas and images of silent adrenocorticotropic hormone-secreting cell adenomas. There are differences in the invasiveness of silent gonadotropin-secreting cell adenomas and silent adrenocorticotropic hormone-secreting cell adenomas. The first radiomics feature set is input into the classification model to obtain the classification result, including: the classification model classifies the image to be classified as an image of a silent gonadotropin-releasing adenoma or an image of a silent adrenocorticotropic hormone-secreting adenoma based on the difference in invasiveness.

8. An image classification device, characterized in that, The device includes: The target region acquisition module is used to acquire the image to be classified, perform automated region segmentation on the image to be classified, and obtain the target region. The image to be classified includes images of non-functional pituitary adenomas. The radiomics feature extraction module is used to extract radiomics features of the target region based on a feature extraction model. The first radiomics feature set acquisition module is used to perform dimensionality reduction processing on the extracted radiomics features based on preset feature selection rules to obtain the first radiomics feature set, which is used to quantify the invasiveness of the target region structure. The classification result acquisition module is used to input the first image omics feature set into the classification model to obtain the classification result.

9. A non-volatile computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images