Training method of classification model, image classification method, computing device, storage medium and program product

Abstract

Embodiments of the present application provide a classification model training method, an image classification method, a computing device, a computer readable storage medium and a computer program product. The classification model training method comprises: obtaining a pre-training image dataset; pre-training a shallow network and a deep network of a to-be-trained classification model using the pre-training image dataset to obtain a first classification model after pre-training; obtaining a SPECT bone imaging dataset and generating an expanded dataset based on the SPECT bone imaging dataset using an image generation technology; inputting the SPECT bone imaging dataset into the first classification model to train the deep network of the first classification model using the SPECT bone imaging dataset to obtain a second classification model; combining the SPECT bone imaging dataset and the expanded dataset into a synthetic dataset, and adjusting network parameters of the second classification model using the synthetic dataset to obtain a target classification model after training. The technical solution provided by the embodiments of the present application improves the classification accuracy.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method for training a classification model, an image classification method, a computing device, a computer-readable storage medium, and a computer program product. Background Technology

[0002] Single-photon emission computed tomography (SPECT), as a functional imaging technique, plays an important role in modern medical diagnostics. In the diagnosis of bone diseases, SPECT bone imaging is widely used, clearly displaying functional changes in the entire skeleton, and is crucial for the early detection of various diseases such as bone metastases, bone infections, and fractures.

[0003] Currently, SPECT bone scans are primarily classified based on the experience and visual perception of radiologists. Differences in experience level, fatigue level, and subjective perception among doctors can all influence the classification results. Summary of the Invention

[0004] This application provides a method for training a classification model, an image classification method, a computing device, a computer-readable storage medium, and a computer program product.

[0005] Firstly, this application provides a method for training a classification model, comprising: Obtain the pre-trained image dataset; Using the pre-trained image dataset, the shallow and deep networks of the classification model to be trained are pre-trained to obtain the first classification model after pre-training. Obtain the SPECT bone scintigraphy dataset and use image generation technology to generate an expanded dataset based on the SPECT bone scintigraphy dataset; The SPECT bone scintigraphy dataset is input into the first classification model to train the deep network of the first classification model using the SPECT bone scintigraphy dataset, thereby obtaining the second classification model. The SPECT bone imaging dataset and the augmented dataset are merged into a synthetic dataset, and the network parameters of the second classification model are adjusted using the synthetic dataset to obtain the trained target classification model.

[0006] Secondly, this application provides an image classification method, including: Obtain the SPECT bone scintigraphy data to be classified; A classification model is loaded and trained through the following operations: A pre-training image dataset is acquired; the shallow and deep networks of the classification model to be trained are pre-trained using the pre-training image dataset to obtain a pre-trained first classification model; a SPECT bone scintigraphy dataset is acquired, and an expanded dataset is generated based on the SPECT bone scintigraphy dataset using image generation technology; the SPECT bone scintigraphy dataset is input into the first classification model to train the deep network of the first classification model using the SPECT bone scintigraphy dataset to obtain a second classification model; the SPECT bone scintigraphy dataset and the expanded dataset are merged into a synthetic dataset, and the network parameters of the second classification model are adjusted using the synthetic dataset to obtain a trained classification model. The SPECT bone scintigraphy data to be classified is input into the classification model, and the classification result output by the classification model corresponding to the SPECT bone scintigraphy data to be classified is obtained.

[0007] Thirdly, this application provides a training apparatus for a classification model, comprising: The first acquisition module is used to acquire the pre-trained image dataset; The pre-training module is used to pre-train the shallow and deep networks of the classification model to be trained using the pre-training image dataset, so as to obtain the first classification model after pre-training. The second acquisition module is used to acquire the SPECT bone imaging dataset and generate an expanded dataset based on the SPECT bone imaging dataset using image generation technology; The first training module is used to input the SPECT bone scintigraphy dataset into the first classification model, so as to train the deep network of the first classification model using the SPECT bone scintigraphy dataset to obtain the second classification model. The second training module is used to merge the SPECT bone imaging dataset and the augmented dataset into a synthetic dataset, and to adjust the network parameters of the second classification model using the synthetic dataset to obtain the trained target classification model.

[0008] Fourthly, this application provides an image classification device, comprising: The third acquisition module is used to acquire the SPECT bone scintigraphy data to be classified; The model loading module is used to load a classification model, which is trained and generated through the following operations: acquiring a pre-training image dataset; using the pre-training image dataset to pre-train the shallow and deep networks of the classification model to be trained, obtaining a pre-trained first classification model; acquiring a SPECT bone scintigraphy dataset, and using image generation technology to generate an expanded dataset based on the SPECT bone scintigraphy dataset; inputting the SPECT bone scintigraphy dataset into the first classification model to train the deep network of the first classification model using the SPECT bone scintigraphy dataset, obtaining a second classification model; merging the SPECT bone scintigraphy dataset and the expanded dataset into a synthetic dataset, and using the synthetic dataset to adjust the network parameters of the second classification model, obtaining a trained classification model; The input module is used to input the SPECT bone scintigraphy data to be classified into the classification model and obtain the classification result output by the classification model corresponding to the SPECT bone scintigraphy data to be classified.

[0009] Fifthly, this application provides a computing device, including a processing component and a storage component; The storage component stores a computer program; the computer program is invoked and executed by the processing component to implement the training method of the classification model provided in the embodiments of this application, or to implement the image classification method provided in the embodiments of this application.

[0010] Sixthly, this application provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processing component, it implements the training method of the classification model provided in this application embodiment, or implements the image classification method provided in this application embodiment.

[0011] In a seventh aspect, this application provides a computer program product, including a computer program or instructions, which, when executed by a processing component, implement the training method of the classification model provided in this application embodiment, or implement the image classification method provided in this application embodiment.

[0012] This embodiment first pre-trains the shallow and deep networks of the classification model using a pre-trained image dataset, thus providing a rich foundation of general visual features for the first classification model. Then, an expanded dataset is generated based on the SPECT bone scintigraphy dataset to significantly increase the number and diversity of samples. Next, the deep network of the first classification model is specifically trained using the SPECT bone scintigraphy dataset, enabling the deep network to adapt to the domain-specific abstract features of medical images. Finally, the SPECT bone scintigraphy dataset and the expanded dataset are merged into a synthetic dataset, and the parameters of the second classification model are adjusted based on this synthetic dataset to obtain the target classification model, thereby achieving effective training for small-sample SPECT bone scintigraphy classification. This ensures that the classification model retains general features while effectively incorporating the diversity of generated samples and the clinical accuracy of real data. Through this training method that combines hierarchical transfer learning with generative data augmentation, refined optimization of feature extraction is achieved. This suppresses overfitting and negative transfer problems caused by data scarcity and domain differences from the training source. As a result, the trained target classification model can more accurately identify bone metastases, benign lesions, and healthy categories on low-resolution, noisy SPECT bone scintigraphy, effectively reducing the risk of missing small lesions and reliance on physician subjective judgment. At the same time, it significantly reduces the need for a large amount of high-quality labeled data, thereby improving classification accuracy.

[0013] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description

[0014] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart of an embodiment of a training method for a classification model provided in this application; Figure 2 A flowchart of an embodiment of an image classification method provided in this application; Figure 3 A block diagram of one embodiment of a training apparatus for a classification model provided in this application; Figure 4 A block diagram of one embodiment of an image classification device provided in this application; Figure 5 A block diagram of one embodiment of a computing device provided in this application. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0016] It should be noted that, in the cases involving user information in the embodiments of this application, the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse. In addition, the various models involved in this application (including but not limited to language models or large models) comply with relevant laws and standards.

[0017] In recent years, the rapid development of deep learning technology, especially the application of convolutional neural networks (CNNs) in medical image processing, has provided new methods for the automated analysis and lesion classification of SPECT bone scans. Researchers have attempted to use deep learning models to identify and classify lesions in SPECT bone scans to assist radiologists in improving diagnostic efficiency and objectivity. However, due to the inherent characteristics of the SPECT imaging principle and limitations in clinical data acquisition, existing automated classification techniques for SPECT bone scans still have several prominent problems.

[0018] First, SPECT images have relatively low spatial resolution and are significantly affected by scattering, attenuation, and noise, making it easy to miss or misdiagnose small lesions or early abnormalities. Second, acquiring high-quality annotated SPECT bone imaging data is difficult and costly. Clinically available annotated datasets for training are typically small (often only a few hundred cases) and suffer from class imbalance (e.g., a relatively large number of bone metastases compared to fewer cases of certain benign lesions). Directly training deep learning models on small datasets is prone to overfitting, resulting in insufficient model generalization ability. Furthermore, traditional feature extraction methods rely on manual design or shallow learning, failing to fully exploit the deep abstract features in SPECT images, and deep models struggle to converge effectively when data is insufficient. Simultaneously, current diagnostic methods still heavily rely on the experience and subjective visual judgment of radiologists, are easily influenced by individual differences, and lack standardization and reproducibility.

[0019] To alleviate the aforementioned problems, the inventors attempted to propose several data augmentation and model optimization schemes. For example, traditional data augmentation methods such as rotation, flipping, translation, and brightness adjustment were used to expand the training samples, or generative adversarial networks (GANs) were introduced to synthesize new samples, increasing the amount of data and improving class balance. Furthermore, a transfer learning strategy was employed, using models pre-trained on large-scale natural image datasets such as ImageNet as initialization, and fine-tuning them on the SPECT dataset to accelerate convergence and improve performance. However, the inventors found that these methods still have limitations in practical applications: traditional data augmentation methods generate samples with limited diversity, failing to effectively simulate the complex noise and lesion variations in SPECT images; samples synthesized by a single generative model may introduce artifacts or distribution biases, affecting model robustness; direct transfer learning is prone to negative transfer due to the significant differences between natural and medical image domains, especially in the deep feature extraction stage. Existing schemes still have room for further improvement in accuracy, recall, and generalization ability in small-sample SPECT bone imaging classification tasks.

[0020] To address the technical problem of improving the accuracy of small-sample SPECT bone scintigraphy classification tasks, this application provides a solution. The basic idea is as follows: First, a pre-trained image dataset is used to fully pre-train the shallow and deep networks of the classification model, providing a rich foundation of general visual features for the first classification model. Then, an expanded dataset is generated based on the SPECT bone scintigraphy dataset to significantly increase the number and diversity of samples. Next, the deep network of the first classification model is specifically trained using the SPECT bone scintigraphy dataset, adapting the deep network to the domain-specific abstract features of medical images. Finally, the SPECT bone scintigraphy dataset and the expanded dataset are merged into a synthetic dataset, and the parameters of the second classification model are adjusted based on this synthetic dataset to obtain the target classification model, thereby achieving effective training for small-sample SPECT bone scintigraphy classification. This ensures that the classification model retains general features while effectively incorporating the diversity of generated samples and the clinical accuracy of real data. Through this training method that combines hierarchical transfer learning with generative data augmentation, refined optimization of feature extraction is achieved. This suppresses overfitting and negative transfer problems caused by data scarcity and domain differences from the training source. As a result, the trained target classification model can more accurately identify bone metastases, benign lesions, and healthy categories on low-resolution, noisy SPECT bone scintigraphy, effectively reducing the risk of missing small lesions and reliance on physician subjective judgment. At the same time, it significantly reduces the need for a large amount of high-quality labeled data, thereby improving classification accuracy.

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] The implementation details of the technical solutions in the embodiments of this application are described in detail below.

[0023] Figure 1 A flowchart of an embodiment of a training method for a classification model provided in this application.

[0024] Figure 1 The training method for the classification model shown may include the following steps: 101: Obtain the pre-trained image dataset.

[0025] In this embodiment, a pre-trained image dataset can be obtained first. This pre-trained image dataset is preferably a large-scale natural image dataset, such as the ImageNet dataset, which contains millions of labeled images covering multiple categories and can provide rich, general visual knowledge for subsequent feature extraction.

[0026] 102: Using a pre-trained image dataset, the shallow and deep networks of the classification model to be trained are pre-trained to obtain the first classification model after pre-training. The classification model to be trained can be implemented as a deep learning model based on convolutional neural networks, such as ResNet50, VGG, or EfficientNet. During pre-training, a pre-training image dataset can be input into the classification model to be trained, and all network layers of the model can be trained through supervised learning, including shallow networks for extracting basic features such as edges and textures, and deep networks for extracting higher-level abstract features.

[0027] In the embodiments of this application, pre-training can employ standard optimization algorithms, such as stochastic gradient descent or Adam optimizer, setting appropriate learning rates and training epochs until the model's loss on the pre-training dataset converges, thereby obtaining a first classification model with strong general feature extraction capabilities.

[0028] 103: Obtain the SPECT bone scintigraphy dataset and use image generation techniques to generate an expanded dataset based on the SPECT bone scintigraphy dataset.

[0029] The SPECT bone imaging dataset can consist of multiple clinically acquired whole-body skeletal SPECT images. Each SPECT image can be preprocessed and labeled with a corresponding category, such as healthy, bone metastases, or benign lesions.

[0030] After obtaining the SPECT bone scintigraphy dataset, image generation techniques can be used to generate an expanded dataset based on the SPECT bone scintigraphy dataset.

[0031] In embodiments of this application, image generation techniques may include, for example, generative models, such as generative adversarial networks, diffusion models, or other models capable of generating new samples with a distribution similar to the original data. By inputting the SPECT bone scan dataset as a condition or seed into the generative model, a large number of new SPECT bone scan samples are generated. These samples are highly similar to the original data in visual features and statistical distribution, thus forming an expanded dataset. The expanded dataset is significantly larger than the original SPECT bone scan dataset, used to further enhance the diversity and quantity of the data.

[0032] In another embodiment of this application, the image generation technique may include, for example, image blending techniques. Image blending techniques are an augmentation method for creating new samples by linearly or regionally fusing multiple images. The specific process is as follows: In one implementation, two (or more) images of different or the same category are randomly selected from the SPECT bone imaging dataset, such as a healthy bone imaging image and a bone metastasis bone imaging image. Then, the pixel values ​​of the two images can be weighted linearly interpolated, and the category labels can be fused using soft labels to generate a new blended image. Specifically, during the pixel value blending process, the pixel values ​​of the new image can be weighted linearly interpolated according to the weight λ, that is, the new image x' = λ× x_A + (1-λ) × x_B, where λ ∈ [0,1]. Accordingly, the category labels are fused using the same weight λ for weighted averaging, i.e., the new label y' = λ × y_A + (1-λ) × y_B, resulting in a soft label vector. The values ​​of each element in this vector range from [0,1], and the sum is 1, representing the probability of the mixed image belonging to each category. For example, when selecting an image of the healthy category (label y_A = [1,0,0]) and an image of the bone metastasis tumor category (label y_B = [0,1,0]), if the sampled weight λ = 0.7, then the category label y' of the generated mixed image is 0.7 × [1,0,0] + 0.3 × [0,1,0] = [0.7, 0.3, 0], indicating that the mixed image belongs to the healthy category with a 70% probability, the bone metastasis tumor category with a 30% probability, and the benign lesion category with a 0% probability. During training, the model's loss function (such as cross-entropy loss) will be calculated based on the soft label instead of the traditional hard label, thereby encouraging the model to learn a smoother feature space and avoid overfitting to hard decision boundaries.

[0033] In another implementation, a local region (such as the pelvis or spine) can be randomly cropped from one image and pasted into the corresponding position in another image, while adjusting the label according to the cropped area ratio. This process can be performed in batches, generating a large number of mixed samples to form an augmented dataset. These new samples visually present gradient or spliced ​​lesion features, helping the model learn smoother decision boundaries.

[0034] 104: Input the SPECT bone scintigraphy dataset into the first classification model to train the deep network of the first classification model using the SPECT bone scintigraphy dataset, and obtain the second classification model.

[0035] In the embodiments of this application, a first classification model can be trained under supervision using a SPECT bone scintigraphy dataset. The first classification model can output classification predictions based on the input SPECT bone scintigraphy image data, and update network parameters through backpropagation by calculating the loss between the predictions and the true labels (such as cross-entropy loss). Here, the training focuses on deep networks to enable them to adapt to the specific domain characteristics of SPECT bone scintigraphy, such as abstract patterns of bone metabolic abnormalities, thereby improving the model's ability to specifically represent medical images while retaining the general features of shallow layers.

[0036] After multiple training iterations, a second classification model can be obtained, which has already demonstrated preliminary classification performance on the SPECT bone imaging task.

[0037] 105: The SPECT bone imaging dataset and the augmented dataset were merged into a synthetic dataset, and the network parameters of the second classification model were adjusted using the synthetic dataset to obtain the trained target classification model.

[0038] Finally, the SPECT bone scan dataset and the augmented dataset can be merged into a synthetic dataset, and the network parameters of the second classification model can be adjusted using the synthetic dataset to obtain the trained target classification model.

[0039] The merged synthetic dataset combines the accuracy of real clinical data with the diversity of generated samples, effectively mitigating class imbalance and further reducing the risk of overfitting. During parameter tuning, the synthetic dataset is input into the secondary classification model, and fine-tuning is performed using a small learning rate or partial layer unfreezing to optimize model parameters on a larger dataset.

[0040] After training, the target classification model can achieve high accuracy and robustness in SPECT bone scintigraphy classification tasks and can be directly used for clinical auxiliary diagnosis, such as outputting the corresponding class probability after inputting a new SPECT bone scintigraphy image.

[0041] This embodiment first pre-trains the shallow and deep networks of the classification model using a pre-trained image dataset, thus providing a rich foundation of general visual features for the first classification model. Then, an expanded dataset is generated based on the SPECT bone scintigraphy dataset to significantly increase the number and diversity of samples. Next, the deep network of the first classification model is specifically trained using the SPECT bone scintigraphy dataset, enabling the deep network to adapt to the domain-specific abstract features of medical images. Finally, the SPECT bone scintigraphy dataset and the expanded dataset are merged into a synthetic dataset, and the parameters of the second classification model are adjusted based on this synthetic dataset to obtain the target classification model, thereby achieving effective training for small-sample SPECT bone scintigraphy classification. This ensures that the classification model retains general features while effectively incorporating the diversity of generated samples and the clinical accuracy of real data. Through this training method that combines hierarchical transfer learning with generative data augmentation, refined optimization of feature extraction is achieved. This suppresses overfitting and negative transfer problems caused by data scarcity and domain differences from the training source. As a result, the trained target classification model can more accurately identify bone metastases, benign lesions, and healthy categories on low-resolution, noisy SPECT bone scintigraphy, effectively reducing the risk of missing small lesions and reliance on physician subjective judgment. At the same time, it significantly reduces the need for a large amount of high-quality labeled data, thereby improving classification accuracy.

[0042] To ensure that the images in the dataset focus on the most relevant skeletal sites for clinical diagnosis, thereby improving the training efficiency and accuracy of subsequent classification models, in some embodiments, obtaining the SPECT bone imaging dataset can be specifically achieved as follows: Acquire multiple whole-body skeletal SPECT image data; For any whole-body skeletal SPECT image data, multiple bone regions are identified in the whole-body skeletal SPECT image data using image recognition algorithms; Based on multiple bone regions, SPECT image data of the whole skeleton is cropped to obtain image data of multiple bone regions; A SPECT bone imaging dataset was generated based on multiple bone region image data corresponding to multiple whole-body skeletal SPECT image data.

[0043] First, multiple whole-body skeletal SPECT image data can be collected from clinical sources. These whole-body skeletal SPECT image data are usually stored in DICOM format. The whole-body skeletal SPECT image data for each patient includes two projection images, a frontal view and a back view, covering the complete skeletal range from head to feet. The images show the uptake and distribution of the radioactive tracer in the bone tissue.

[0044] For any full-body skeletal SPECT image, image recognition algorithms can be used to identify multiple bone regions within the image. These image recognition algorithms can employ threshold-based segmentation methods, edge detection algorithms (such as the Canny operator), or deep learning-based semantic segmentation models (such as U-Net or its variants). These algorithms can automatically identify and locate key bone regions, such as the spine, pelvis, ribs, long bones of the limbs, and regions of interest in the skull, based on the grayscale distribution and anatomical features of the SPECT image, thereby generating corresponding bone region masks or bounding boxes.

[0045] Furthermore, based on multiple identified bone regions, the whole-body skeletal SPECT image data can be cropped. That is, according to the boundary or mask of each bone region, the corresponding sub-images are extracted from the original whole-body image to obtain multiple independent bone region image data. These sub-images retain local details at the original resolution while removing irrelevant background areas to reduce noise interference and computational redundancy.

[0046] Finally, multiple bone region image data corresponding to all whole-body skeletal SPECT image data are aggregated and combined with category labels provided by clinicians (such as healthy, bone metastases, or benign lesions) to generate a SPECT bone imaging dataset. Each bone region image data in this dataset has a standardized size (e.g., adjusted to a uniform resolution, such as 256×256 pixels, through a resize operation) and can be further normalized to meet the input requirements of subsequent models.

[0047] The SPECT bone imaging dataset obtained in this way not only preserves the local features of bone lesions, but also significantly increases the number of effective samples (multiple local images can be generated from each whole-body image), which helps to alleviate the problem of small sample size and improve the model's sensitivity to local lesions.

[0048] It should be noted that those skilled in the art can adjust the number of bone regions or the parameters of the recognition algorithm according to specific application requirements.

[0049] In some embodiments, each bone region image data in the SPECT bone imaging dataset has a specific category label.

[0050] In the embodiments of this application, each bone region image in the SPECT bone scintigraphy dataset has a specific category label. These labels can be annotated by experienced radiologists based on clinical diagnostic criteria and image features. For example, images can be classified into healthy categories (uniform bone uptake, no abnormal hot spots), bone metastatic tumor categories (multiple, irregular high-uptake areas, often accompanied by multiple lesions throughout the body), or benign lesion categories (abnormal uptake in a single or specific location, usually associated with degeneration or inflammation). This category labeling ensures the supervised learning properties of the dataset, enabling subsequent training processes to optimize parameters based on real clinical judgment, thereby improving the diagnostic consistency and reliability of the model.

[0051] In some embodiments, the SPECT bone scintigraphy dataset is input into the first classification model to train the deep network of the first classification model using the SPECT bone scintigraphy dataset, thereby obtaining the second classification model. Specifically, this can be implemented as follows: Input the SPECT bone scintigraphy dataset into the first classification model, so that the first classification model outputs the classification result of each bone region image data respectively; Based on the differences between the classification results and the category labels, the network parameters of the deep network of the first classification model are updated to obtain the second classification model.

[0052] First, bone region images from the SPECT bone scintigraphy dataset can be input batch by batch or image by image into a pre-trained first classification model. Upon receiving the input, this model outputs a classification result for each bone region image, typically represented as a probability distribution. For example, a softmax function can be used to convert the output into probability values ​​for each category (e.g., healthy, bone metastases, benign lesions). Next, based on the difference between the model's output classification result and the specific category label corresponding to each bone region image, a loss value is calculated. This difference is usually quantified using a loss function, such as the cross-entropy loss function or other loss metrics suitable for multi-class tasks. This loss value reflects the degree of deviation between the model's current prediction and the true label. Subsequently, the backpropagation algorithm is used to update the deep network parameters of the first classification model based on the calculated loss value. Specifically, an optimizer (e.g., stochastic gradient descent, Adam, or variants) can be used to adjust the weights and biases of the deep network, allowing it to gradually learn the abstract features and patterns related to diseases in SPECT bone scintigraphy. Meanwhile, the parameters of the shallow network remain relatively stable to retain general feature extraction capabilities. After multiple training rounds of iterative optimization until the loss converges or the preset stopping condition is met, a second classification model is obtained. The deep network of this model has been significantly adapted to the specific task requirements of SPECT bone imaging, showing stronger domain-specific classification performance.

[0053] The above training method not only makes full use of the supervision signals provided by the category labels, but also achieves targeted optimization of deep networks, effectively reducing the risk of negative transfer in the process of transferring from general pre-training to medical image tasks, and improving the model's generalization ability and classification accuracy on small sample datasets.

[0054] During the training process described above, the specific dataset size or computational resources, batch size, learning rate, or training epochs, and other hyperparameters can be flexibly selected by those skilled in the art according to actual application needs.

[0055] In some embodiments, the SPECT bone scan dataset and the augmented dataset are merged into a synthetic dataset, and the network parameters of the second classification model are adjusted using the synthetic dataset to obtain the trained target classification model. Specifically, this can be implemented as follows: The SPECT bone imaging dataset and the augmented dataset were merged into a synthetic dataset, and the parameters of the fully connected layer network of the second classification model were adjusted using the synthetic dataset to obtain the trained target classification model.

[0056] In the embodiments of this application, firstly, the SPECT bone scintigraphy dataset can be merged with a previously generated augmented dataset to form a larger synthetic dataset with a more balanced class distribution. In one possible implementation, the generated samples in the augmented dataset inherit or match the class labels of the original SPECT bone scintigraphy dataset, and a filtering operation can be performed during the merging process, such as removing low-quality generated images (e.g., samples with obvious artifacts or excessive deviation from the true distribution), thereby ensuring the overall quality and clinical relevance of the synthetic dataset. This synthetic dataset combines accurate annotation of real clinical images with the diversity of generated samples, effectively expanding the number of training samples and alleviating the class imbalance problem.

[0057] Next, the parameters of the fully connected layers in the secondary classification model can be tuned using a synthetic dataset. In one possible implementation, image data from the synthetic dataset can be batch-input into the secondary classification model. During forward propagation, the network parameters of the feature extraction parts, such as convolutional layers, are kept fixed, allowing only the weights and biases of the fully connected layers (usually located at the end of the model, used for classification decisions) to participate in the update. By calculating the loss value (e.g., cross-entropy loss) between the model's output classification result and the corresponding class label, the parameters of the fully connected layers are iteratively optimized using a backpropagation algorithm and an optimizer (e.g., Adam or stochastic gradient descent) with a small learning rate. This tuning process typically uses fewer training epochs to avoid disrupting previous pre-training and deep transfer learning, while allowing the fully connected layers to quickly adapt to the comprehensive feature distribution of the synthetic dataset. After parameter tuning, a trained target classification model is obtained. The fully connected layers of this model have fully integrated the knowledge from both real and generated data, exhibiting higher classification accuracy and robustness, especially suitable for fine differentiation between bone metastases, benign lesions, and healthy categories in SPECT bone scintigraphy.

[0058] By fine-tuning the parameters of only the fully connected layers, the final optimization can be completed efficiently with limited computing resources, while retaining the stable feature extraction capabilities of both shallow and deep layers of the model, further enhancing its application value on small clinical datasets.

[0059] In some embodiments, generating an expanded dataset based on a SPECT bone scan dataset using image generation techniques can be specifically implemented as follows: An augmented dataset is generated based on the SPECT bone scintigraphy dataset using a diffusion model; or, Generative adversarial networks are used to generate augmented datasets based on SPECT bone scintigraphy datasets; or, An expanded dataset was generated based on the SPECT bone scintigraphy dataset using a natural regression model.

[0060] In one embodiment of this application, an augmented dataset is generated based on a SPECT bone scintigraphy dataset using a diffusion model. The data distribution can be learned by simulating the process of forward noise addition and reverse denoising recovery. In one possible implementation, real images from the SPECT bone scintigraphy dataset are first forward diffused, i.e., Gaussian noise is progressively added until the image approximates a pure noise distribution. Simultaneously, a neural network (such as a denoising network based on the U-Net architecture) is trained to learn the reverse diffusion step. In this step, conditional information, such as category labels (healthy, bone metastases, or benign lesions) or textual prompts describing specific skeletal abnormalities, can be introduced to guide the generation process. After training, only the reverse process is performed, progressively denoising from random noise, generating a large number of new SPECT bone scintigraphy samples. These samples have high fidelity and diversity in skeletal anatomy and lesion uptake patterns. After quality screening (e.g., removing abnormal samples based on structural similarity indices), the augmented dataset is formed. This approach is particularly suitable for the noise characteristics of SPECT images, producing high-quality, diverse samples and improving the model's robustness to real-world clinical variations.

[0061] In another embodiment of this application, an expanded dataset is generated based on a SPECT bone scintigraphy dataset using a generative adversarial network (GAN). Image generation can be achieved through an adversarial training mechanism between the generator and the discriminator. In one possible implementation, a conditional GAN ​​structure can be chosen. Images from the SPECT bone scintigraphy dataset are used as real samples input to the discriminator, while random noise vectors combined with conditional information (such as class labels) are input to the generator. The generator is responsible for generating fake SPECT bone scintigraphy images, while the discriminator attempts to distinguish between real and fake images. By alternately optimizing the generator (making the generated images harder to be identified as fake) and the discriminator (improving the accuracy of distinguishing between real and fake images), the model gradually captures the latent distribution features of SPECT images. After training convergence, the generator generates new samples in batches based on random input or specified conditions. These samples are similar to the original data in lesion details and overall uptake distribution. After evaluation and screening, an expanded dataset is formed. This method has high training efficiency and generates images with rich details, which helps to simulate the complex multifocal lesion patterns in SPECT bone scintigraphy.

[0062] In another embodiment of this application, an augmented dataset can also be generated based on a SPECT bone scintigraphy dataset using a natural regression model. In one possible implementation, the images in the SPECT bone scintigraphy dataset are first converted into sequence representations (e.g., through rasterization scanning or block processing). Then, an autoregressive network (e.g., a Transformer-based model) is trained to learn dependencies between pixels. Conditional inputs, such as category labels or masks to guide anomalous behavior in specific bone regions, can be incorporated during generation. Starting from the beginning of the image, the model sequentially predicts the value of the next pixel or region until the entire image is completed. This process can be parallelized for efficiency, and the generated samples inherit the statistical properties of the original dataset, with diversity controlled through techniques such as temperature sampling. After post-processing and quality control, these samples form the augmented dataset. This approach excels in maintaining local image consistency and is suitable for scenarios in SPECT bone scintigraphy where accurate simulation of continuous skeletal structures is required.

[0063] In some embodiments, before inputting the SPECT bone scintigraphy dataset into a first classification model to train a deep network of the first classification model using the SPECT bone scintigraphy dataset to obtain a second classification model, the method may further include: Fix the network parameters of the shallow network of the first classification model.

[0064] After initial pre-training using a pre-trained image dataset (such as a large-scale natural image dataset), the first classification model's shallow network has sufficiently learned general basic visual features, such as edges, textures, and simple shapes. These features have high general applicability between natural and medical images (such as SPECT bone scans). By fixing the network parameters of the shallow network—that is, prohibiting updates to the weights and biases of the shallow network during subsequent training (e.g., by setting the learning rate of the corresponding layers to zero, freezing gradient calculations, or marking those layers as untrainable in the deep learning framework)—it is possible to ensure that these general features are fully preserved. At the same time, the training focus is concentrated on the deep network, making it specifically adapted to the domain-specific abstract features of SPECT bone scans, such as complex patterns of bone metabolic abnormalities, distribution of multiple lesions, or higher-order representations under noise interference. This fixed operation can be implemented in the training framework, such as in platforms like PyTorch or TensorFlow, through parameter grouping or layer attribute settings. This allows only the parameters of deep networks (such as the last few convolutional blocks and some fully connected layers) to be iteratively updated through backpropagation and the optimizer when supervised training is performed on the input SPECT bone imaging dataset.

[0065] This strategy of fixing the parameters of the shallow network can significantly reduce the risk of negative transfer during the migration from the general pre-training domain to the medical image-specific domain. Specifically, it avoids the over-modification or destruction of shallow general features due to the limited samples in the SPECT bone scintigraphy dataset, thereby improving the overall stability and domain adaptability of the secondary classification model. Especially in clinical scenarios with limited data, this strategy helps improve classification accuracy, recall, and enhance the model's generalization performance. Those skilled in the art can flexibly determine the boundary between the shallow and deep networks based on the specific architecture of the selected classification model (such as different deep variants of the ResNet series). For example, the first few residual blocks can be considered shallow, while subsequent blocks can be considered deep, and the fixed range can be adjusted accordingly.

[0066] Figure 2 A flowchart of an embodiment of an image classification method provided in this application.

[0067] Figure 2 The image classification method shown may include the following steps: 201: Obtain the SPECT bone scintigraphy data to be classified; 202: Load the classification model. The classification model is trained and generated through the following operations: Obtain a pre-training image dataset; use the pre-training image dataset to pre-train the shallow and deep networks of the classification model to be trained, obtaining the pre-trained first classification model; obtain the SPECT bone scintigraphy dataset, and use image generation technology to generate an expanded dataset based on the SPECT bone scintigraphy dataset; input the SPECT bone scintigraphy dataset into the first classification model to train the deep network of the first classification model using the SPECT bone scintigraphy dataset, obtaining the second classification model; merge the SPECT bone scintigraphy dataset and the expanded dataset into a synthetic dataset, and use the synthetic dataset to adjust the network parameters of the second classification model, obtaining the trained classification model. 203: Input the SPECT bone scintigraphy data to be classified into the classification model, and obtain the classification result output by the classification model corresponding to the SPECT bone scintigraphy data to be classified.

[0068] The detailed implementation methods and beneficial effects of each step in this embodiment have been described in detail in the foregoing embodiments, and will not be elaborated here.

[0069] Figure 3 A block diagram of one embodiment of a training device for a classification model provided in this application.

[0070] Figure 3 The training apparatus for the classification model shown may include: The first acquisition module 301 is used to acquire a pre-trained image dataset; The pre-training module 302 is used to pre-train the shallow and deep networks of the classification model to be trained using a pre-training image dataset, so as to obtain the first classification model after pre-training. The second acquisition module 303 is used to acquire the SPECT bone imaging dataset and generate an extended dataset based on the SPECT bone imaging dataset using image generation technology. The first training module 304 is used to input the SPECT bone scintigraphy dataset into the first classification model, so as to train the deep network of the first classification model using the SPECT bone scintigraphy dataset to obtain the second classification model. The second training module 305 is used to merge the SPECT bone imaging dataset and the augmented dataset into a synthetic dataset, and to adjust the network parameters of the second classification model using the synthetic dataset to obtain the trained target classification model.

[0071] In some embodiments, the first acquisition module 301 is specifically used for: Acquire multiple whole-body skeletal SPECT image data; For any whole-body skeletal SPECT image data, multiple bone regions are identified in the whole-body skeletal SPECT image data using image recognition algorithms; Based on multiple bone regions, SPECT image data of the whole skeleton is cropped to obtain image data of multiple bone regions; A SPECT bone imaging dataset was generated based on multiple bone region image data corresponding to multiple whole-body skeletal SPECT image data.

[0072] In some embodiments, each bone region image data in the SPECT bone imaging dataset has a specific category label; In some embodiments, the first training module 304 is specifically used for: Input the SPECT bone scintigraphy dataset into the first classification model, so that the first classification model outputs the classification result of each bone region image data respectively; Based on the differences between the classification results and the category labels, the network parameters of the deep network of the first classification model are updated to obtain the second classification model.

[0073] In some embodiments, the second training module 305 is specifically used for: The SPECT bone imaging dataset and the augmented dataset were merged into a synthetic dataset, and the parameters of the fully connected layer network of the second classification model were adjusted using the synthetic dataset to obtain the trained target classification model.

[0074] In some embodiments, the second acquisition module 303 is specifically used for: An augmented dataset is generated based on the SPECT bone scintigraphy dataset using a diffusion model; or, Generative adversarial networks are used to generate augmented datasets based on SPECT bone scintigraphy datasets; or, An expanded dataset was generated based on the SPECT bone scintigraphy dataset using a natural regression model.

[0075] In some embodiments, the device may further include: The parameter fixing module is used to fix the network parameters of the shallow network of the first classification model.

[0076] The detailed implementation methods and beneficial effects of each step in the device of this embodiment have been described in detail in the foregoing embodiments, and will not be elaborated here.

[0077] Figure 3 The training device for the classification model can perform... Figure 1 The implementation principle and technical effects of the training method for the classification model described in the illustrated embodiments will not be repeated here. The specific methods by which each module and unit of the training device for the classification model in the above embodiments perform operations have been described in detail in the embodiments related to this method, and will not be elaborated upon here.

[0078] Figure 4 A block diagram of one embodiment of an image classification device provided in this application.

[0079] Figure 4 The image classification device shown may include: The third acquisition module 401 is used to acquire the SPECT bone scintigraphy data to be classified; The model loading module 402 is used to load the classification model. The classification model is trained and generated through the following operations: acquiring a pre-training image dataset; using the pre-training image dataset, pre-training the shallow and deep networks of the classification model to be trained to obtain a pre-trained first classification model; acquiring a SPECT bone scintigraphy dataset and using image generation technology to generate an expanded dataset based on the SPECT bone scintigraphy dataset; inputting the SPECT bone scintigraphy dataset into the first classification model to train the deep network of the first classification model using the SPECT bone scintigraphy dataset to obtain a second classification model; merging the SPECT bone scintigraphy dataset and the expanded dataset into a synthetic dataset, and using the synthetic dataset to adjust the network parameters of the second classification model to obtain the trained classification model. The input module 403 is used to input the SPECT bone scintigraphy data to be classified into the classification model and obtain the classification result output by the classification model corresponding to the SPECT bone scintigraphy data to be classified.

[0080] Figure 4 The image classification device described above can perform... Figure 2The implementation principle and technical effects of the image classification method described in the illustrated embodiments will not be repeated here. The specific methods by which each module and unit of the image classification device in the above embodiments performs its operations have been described in detail in the embodiments related to this method, and will not be elaborated upon here.

[0081] It should be noted that some processes described in the above embodiments and accompanying drawings include multiple operations appearing in a specific order. However, it should be clearly understood that these operations may not be executed in the order they appear in this document, or they may be executed in parallel. The operation numbers, such as 101, 102, etc., are merely used to distinguish different operations and do not represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should also be noted that the descriptions such as "first" and "second" in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0082] Figure 5 This is a schematic diagram of the structure of one embodiment of a computing device provided in this application. Figure 5 As shown, in practice, the computing device may include a storage component 501 and a processing component 502.

[0083] Storage component 501 is used to store computer programs and can be configured to store various other data to support operation on a computing device. Examples of this data include instructions for any application or method used to operate on the computing device, data structures, contact data, phone book data, messages, pictures, videos, etc.

[0084] Processing component 502, coupled to storage component 501, is used to execute computer programs in storage component 501 for implementing, etc. Figure 1 The training method or implementation of the classification model shown. Figure 2 The classification method shown.

[0085] Furthermore, such as Figure 5 As shown, the computing device may also include other components such as a communication component 503, a display component 504, a power supply component 505, and an audio component 506. Figure 5 The diagram only shows some components and does not mean that the device includes only these components. Figure 5 The components shown. Additionally... Figure 5The components within the dashed box are optional, not mandatory, and their specific requirements depend on the product form of the computing device. The computing device in this embodiment can be a terminal device such as a desktop computer, laptop computer, smartphone, or IoT (Internet of Things) device, or a server-side device such as a conventional server, cloud server, or server array. If the computing device in this embodiment is implemented as a terminal device such as a desktop computer, laptop computer, or smartphone, it may include... Figure 5 The components within the dashed box; if the computing device in this embodiment is implemented as a conventional server, cloud server, or server array, etc., then it may not include... Figure 5 The component within the dashed box.

[0086] The processing component described above includes one or more processors to execute computer instructions to complete all or part of the steps in the method described above. Alternatively, the processing component may be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the method described above.

[0087] The aforementioned storage components can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0088] The aforementioned communication component is configured to facilitate wired or wireless communication between the device housing the communication component and other devices. The device housing the communication component can access wireless networks based on communication standards, such as mobile communication networks, or combinations thereof. In one exemplary embodiment, the communication component receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel.

[0089] The aforementioned display components may include a screen, which may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen can be implemented as a touchscreen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors can sense not only the boundaries of touch or swipe actions but also the duration and pressure associated with the touch or swipe operation.

[0090] The aforementioned power supply components provide power to various components within the device in which they reside. These power supply components may include a power management system, one or more power sources, and other components associated with generating, managing, and distributing power to the device in which they reside.

[0091] The aforementioned audio component can be configured to output and / or input audio signals. For example, the audio component includes a microphone (MIC) configured to receive external audio signals when the device containing the audio component is in an operating mode, such as call mode, recording mode, or voice recognition mode. The received audio signals can be further stored in memory or transmitted via a communication component. In some embodiments, the audio component also includes a speaker for outputting audio signals.

[0092] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to implement the steps in the above-described method embodiments. The computer-readable storage medium includes volatile or non-volatile components, or a combination thereof, and can be removable or non-removable. Examples of computer-readable storage media include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random-access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), flash memory or other memory technologies, CD-ROM, Digital Video Disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium. Accordingly, this application also provides a computer program product, which includes a computer program or instructions that, when executed by a processor, cause the processor to implement the steps in the above method embodiments. It should be understood that each step or combination of steps in the above method flow can be implemented by the computer program or instructions. Furthermore, these computer programs or instructions can be applied to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device, enabling the processor of the general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to function as an apparatus for implementing the corresponding functions in the above method embodiments.

[0093] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0094] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0095] Finally, it should be noted that the above are merely embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for training a classification model, characterized in that, include: Obtain the pre-trained image dataset; Using the pre-trained image dataset, the shallow and deep networks of the classification model to be trained are pre-trained to obtain the first classification model after pre-training. Obtain the SPECT bone scintigraphy dataset and generate an expanded dataset based on the SPECT bone scintigraphy dataset using image generation technology; The SPECT bone scintigraphy dataset is input into the first classification model to train the deep network of the first classification model using the SPECT bone scintigraphy dataset, thereby obtaining the second classification model. The SPECT bone imaging dataset and the augmented dataset are merged into a synthetic dataset, and the network parameters of the second classification model are adjusted using the synthetic dataset to obtain the trained target classification model.

2. The method according to claim 1, characterized in that, The acquisition of the SPECT bone scintigraphy dataset includes: Acquire multiple whole-body skeletal SPECT image data; For any whole-body skeletal SPECT image data, multiple bone regions are identified in the whole-body skeletal SPECT image data using an image recognition algorithm; Based on the multiple bone regions, the whole-body skeletal SPECT image data is cropped to obtain multiple bone region image data; The SPECT bone imaging dataset is generated based on multiple bone region image data corresponding to the multiple whole-body skeletal SPECT image data.

3. The method according to claim 2, characterized in that, Each bone region image in the SPECT bone imaging dataset has a specific category label; The step of inputting the SPECT bone scintigraphy dataset into the first classification model to train the deep network of the first classification model using the SPECT bone scintigraphy dataset to obtain the second classification model includes: The SPECT bone imaging dataset is input into the first classification model, so that the first classification model outputs the classification result of each bone region image data respectively; Based on the difference information between the classification results and the category labels, the network parameters of the deep network of the first classification model are updated to obtain the second classification model.

4. The method according to claim 1, characterized in that, The step of merging the SPECT bone scan dataset and the augmented dataset into a synthetic dataset, and using the synthetic dataset to adjust the network parameters of the second classification model to obtain the trained target classification model includes: The SPECT bone imaging dataset and the augmented dataset are merged into a synthetic dataset, and the fully connected layer network parameters of the second classification model are adjusted using the synthetic dataset to obtain the trained target classification model.

5. The method according to claim 1, characterized in that, The process of generating an expanded dataset based on the SPECT bone scan dataset using image generation technology includes: An expanded dataset is generated based on the SPECT bone scintigraphy dataset using a diffusion model; or, An augmented dataset is generated based on the SPECT bone scintigraphy dataset using a generative adversarial network; or, An expanded dataset was generated based on the SPECT bone scintigraphy dataset using a natural regression model.

6. The method according to claim 1, characterized in that, Before inputting the SPECT bone scintigraphy dataset into the first classification model to train the deep network of the first classification model using the SPECT bone scintigraphy dataset to obtain the second classification model, the method further includes: The network parameters of the shallow network of the first classification model are fixed.

7. An image classification method, characterized in that, include: Obtain the SPECT bone scintigraphy data to be classified; Load the classification model, which is trained and generated through the following operation: obtain a pre-trained image dataset; Using the pre-trained image dataset, the shallow and deep networks of the classification model to be trained are pre-trained to obtain the first classification model after pre-training. Obtain the SPECT bone scintigraphy dataset and use image generation technology to generate an expanded dataset based on the SPECT bone scintigraphy dataset; The SPECT bone imaging dataset is input into the first classification model to train the deep network of the first classification model using the SPECT bone imaging dataset, thereby obtaining the second classification model; the SPECT bone imaging dataset and the augmented dataset are merged into a synthetic dataset, and the network parameters of the second classification model are adjusted using the synthetic dataset to obtain the trained classification model; The SPECT bone scintigraphy data to be classified is input into the classification model, and the classification result output by the classification model corresponding to the SPECT bone scintigraphy data to be classified is obtained.

8. A computing device, characterized in that, This includes processing components and storage components; The storage component stores a computer program; the computer program is invoked and executed by the processing component to implement the training method of the classification model as described in any one of claims 1 to 6, or to implement the image classification method as described in claim 7.

9. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processing component, implements the training method of the classification model as described in any one of claims 1 to 6, or implements the image classification method as described in claim 7.

10. A computer program product, characterized in that, It includes a computer program or instructions that, when executed by a processing component, implement the training method of the classification model as described in any one of claims 1 to 6, or implement the image classification method as described in claim 7.

Images