Three-dimensional medical image anomaly detection method and system based on self-supervised learning

Abstract

A three-dimensional medical image anomaly detection method and system based on self-supervised learning, the method comprising the following steps: constructing three independent feature embedding networks based on three original feature maps of different scales; converting the three original feature maps into one-dimensional embedding vectors using the feature embedding network; calculating the sum of the mean squared errors between the embedding vectors of the pseudo-abnormal image and the normal image at each scale; updating the parameters of the feature embedding network using the Adam optimizer; constructing a decoder symmetrical to the encoder structure; expanding the three embedding vectors using a fully connected layer and reshaping them respectively to generate three low-resolution feature maps, the spatial resolution of the low-resolution feature maps being the same as that of the layer3 feature map; fusing the three low-resolution feature maps to obtain a layer3 reconstructed feature map; and an upsampling module of the decoder cascade-reconstructing the layer3 reconstructed feature map layer by layer to obtain a layer1 reconstructed feature map and a layer2 reconstructed feature map that are consistent with the resolutions of the layer1 feature map and the layer2 feature map, respectively.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image analysis, and in particular to a three-dimensional medical image anomaly detection method and system based on self-supervised learning. Background Art

[0002] In the field of medical image analysis, anomaly detection technology is a current research and application focus, widely used for early disease detection and auxiliary diagnosis. This technology aims to identify abnormal regions that differ significantly from normal anatomical structures in three-dimensional medical images such as CT and MRI. Because abnormal regions are often sparsely distributed, have diverse morphologies, and have blurred boundaries, existing supervised learning-based methods often rely on large amounts of high-quality annotated data. However, in real clinical settings, abnormal samples are difficult to obtain and manual annotation is expensive, making such methods difficult to apply in unlabeled or weakly labeled conditions.

[0003] In recent years, self-supervised learning methods have received increasing attention in the field of anomaly detection. These methods construct prior tasks that do not rely on manual annotation, such as reconstruction, contrastive learning, or contextual prediction, enabling the model to automatically learn the underlying structural features of the data itself. In medical image analysis scenarios, self-supervised learning can leverage consistent features between normal samples, enabling the model to identify potential abnormal regions that deviate from the normal distribution during the inference phase. This unsupervised modeling approach has significant advantages in the absence of abnormal samples and is particularly suitable for real-world clinical scenarios where lesion annotations are scarce.

[0004] With the advancement of 3D imaging technology, 3D convolutional neural networks (3D CNNs) have become the mainstream architecture for medical volume data processing, thanks to their ability to simultaneously model both spatial and depth features. Compared to traditional 2D slice-based methods, 3D CNNs maintain contextual structural continuity and have demonstrated greater accuracy and stability in lesion detection in 3D structures such as the lungs and brain.

[0005] However, current 3D anomaly detection methods in medical imaging are still mostly implemented in a supervised manner. Anomaly localization in 3D space usually requires pixel-level mask annotation as a training target, which is a huge workload in the actual data preparation process and limits the applicability of the model under unlabeled conditions. To reduce label dependence, existing technologies have introduced mechanisms such as mask modeling and regional occlusion reconstruction. By constructing pseudo-labels or local masking to simulate abnormal areas, the model is guided to learn the difference characteristics between abnormal and normal structures. However, most of these methods still need to rely on a certain proportion of abnormal areas or auxiliary supervisory signals. When samples are scarce or the distribution is complex, the model performance is still at risk of instability.

[0006] In particular, research on self-supervised learning mechanisms and label-free modeling in the field of 3D medical image anomaly detection remains limited. One of the key challenges facing this area of technology is how to build a stable self-supervised learning framework for 3D volume data without requiring any manual annotation, while effectively improving the model's responsiveness to anomalies and its detection accuracy. Summary of the Invention

[0007] In view of the above-mentioned shortcomings of the prior art, the present invention provides a three-dimensional medical image anomaly detection method and system based on self-supervised learning. By introducing an unlabeled self-supervised training strategy and utilizing the internal consistency of normal samples and a multi-scale reconstruction mechanism, accurate detection of abnormal areas in three-dimensional medical images is achieved, solving the problems of traditional methods' heavy reliance on labeled data and difficulty in obtaining abnormal samples.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is:

[0009] First, a 3D medical image anomaly detection method based on self-supervised learning includes the following steps:

[0010] Step S1, acquiring and preprocessing three-dimensional medical image data to generate a normal image; adding noise to the normal image to generate a pseudo abnormal image;

[0011] Step S2: inputting the normal image and the pseudo abnormal image into a three-dimensional convolutional neural network encoder for multi-scale feature extraction to obtain three original feature maps of different scales, wherein the original feature maps are sorted from high to low in terms of resolution into a layer 1 feature map, a layer 2 feature map, and a layer 3 feature map;

[0012] Step S3: constructing three independent feature embedding networks based on the original feature maps at three different scales; using the feature embedding networks, converting the three original feature maps into one-dimensional embedding vectors; each feature embedding network processes the original feature map at a corresponding scale; calculating the sum of the mean squared errors between the embedding vectors of the pseudo-abnormal image and the normal image at each scale; and updating the parameters of the feature embedding networks using an Adam optimizer;

[0013] Step S4: construct a decoder symmetrical to the encoder structure; use a fully connected layer to expand the three embedding vectors and reshape them respectively to generate three low-resolution feature maps, where the spatial resolution of the low-resolution feature maps is the same as that of the layer3 feature map;

[0014] The three low-resolution feature maps are fused to obtain a layer3 reconstructed feature map; the upsampling module of the decoder cascades and reconstructs the layer3 reconstructed feature map layer by layer to obtain a layer1 reconstructed feature map and a layer2 reconstructed feature map with the same resolution as the layer1 feature map and the layer2 feature map respectively.

[0015] Preferably, in step S1, the preprocessing includes performing size normalization, pixel intensity normalization, spatial resampling and denoising on the three-dimensional medical image data.

[0016] Preferably, in step S1, in the process of adding noise to the normal image, Perlin noise and Worley noise are mixed and superimposed to generate a disturbed texture; in the noise fusion process, weighted parameters are used to control the contribution of the two noises to the texture and boundary.

[0017] Preferably, in step S2, the network structure of the three-dimensional convolutional neural network adopts a 7×7×7 three-dimensional convolutional layer in the first layer with a step size of 2 for extracting low-level texture features; the 7×7×7 three-dimensional convolutional layer is connected to a batch normalization layer and a ReLU activation function, and then a 3×3×3 three-dimensional maximum pooling layer is introduced.

[0018] Preferably, in step S3, each of the feature embedding networks includes: a group of 1×1×1 three-dimensional convolutional layers, a spatial pooling layer connected after the 1×1×1 three-dimensional convolutional layer, and a fully connected layer connected at the end.

[0019] Preferably, in step S4, the Huber loss function is used to maintain the structural consistency of the reconstruction process by minimizing the structural difference between the reconstructed feature map and the original feature map at each scale.

[0020] Preferably, the method further comprises step S5 of respectively calculating the cosine similarity between the reconstructed feature map at each scale and the original feature map, and fusing the two maps to generate an abnormality score map.

[0021] In a second aspect, a three-dimensional medical image anomaly detection system based on self-supervised learning includes: an acquisition and preprocessing module, a pseudo-anomaly generation module, a three-dimensional feature extraction module, a feature embedding module, and a symmetric decoding and reconstruction module; the system is used to implement the steps of the method described in the first aspect;

[0022] The acquisition and preprocessing module is used to acquire and preprocess the three-dimensional medical image to generate the normal image; the pseudo-abnormal generation module is used to generate the pseudo-abnormal image based on the normal image; the three-dimensional feature extraction module inputs the normal image and the pseudo-abnormal image into the encoder of the three-dimensional convolutional neural network for multi-scale feature extraction to obtain three original feature maps of different scales; the feature embedding module constructs three independent feature embedding networks based on the three original feature maps of different scales and uses the feature embedding networks to convert the three original feature maps into one-dimensional embedding vectors; the symmetric decoding and reconstruction module is used to construct the decoder that is symmetrical with the encoder structure; the decoder is used to generate the low-resolution feature map and then generate the reconstructed feature map.

[0023] Compared with the prior art, the beneficial effects of the present invention are embodied in:

[0024] Different from the existing technology that only samples a single noise or local occlusion to construct pseudo-anomalies and reconstructs mostly single-scale technologies, the present invention proposes to fuse Perlin noise and Worley noise to generate highly realistic pseudo-anomaly areas with delicate textures and natural boundaries; at the same time, it constructs a three-scale decoding path that is completely symmetrical with the 3D ResNet-50 encoder structure, realizes cross-level structural alignment, and significantly enhances the model's ability to distinguish multi-morphological anomalies.

[0025] Different from the existing technical solutions of simple splicing or single mapping on multi-scale features, the present invention designs a "multi-scale embedding + distillation" self-supervised framework: for the three-layer features of layer1, layer2, and layer3, feature embedding networks are constructed respectively to generate compact embedding vectors, and a robust optimization based on Huber loss and a feature distillation mechanism with frozen encoder features as teachers and decoded outputs as students are introduced to ensure cross-scale semantic consistency in the reconstruction process, greatly improving the detection accuracy of hidden or complex boundary lesions.

[0026] Different from the existing technical solutions that rely on empirical thresholds or manual parameter adjustment, the present invention proposes a cosine similarity distribution based on multi-scale reconstruction errors and an F1-driven cross-validation strategy to automatically determine the voxel-level and image-level classification boundaries, and realize threshold self-calibration under unsupervised conditions, greatly enhancing the system's versatility and clinical feasibility in different patient and equipment environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is an overall architecture diagram of a three-dimensional medical image anomaly detection method based on self-supervised learning according to Example 1 of the present invention;

[0028] Figure 2 This is a diagram of the overall architecture of 3D ResNet-50 according to Example 1 of the present invention;

[0029] Figure 3 This is an architectural diagram of the bottleneck layer in 3D ResNet-50 according to Example 1 of the present invention. DETAILED DESCRIPTION

[0030] In order to make the technical means, creative features, objectives and effects of the invention easier to understand, the present invention is further described with reference to specific figures. However, the present invention is not limited to the following implementation cases.

[0031] It should be noted that the structures, proportions, sizes, etc. illustrated in the drawings in this specification are only used to match the contents disclosed in the specification so that people familiar with this technology can understand and read them. They are not used to limit the conditions under which the present invention can be implemented. Therefore, they have no substantive technical significance. Any structural modification, change in proportional relationship or adjustment of size should still fall within the scope of the technical content disclosed in the present invention without affecting the efficacy and purpose that can be achieved by the present invention. Example

[0032] like Figure 1 A method for detecting anomalies in three-dimensional medical images based on self-supervised learning is shown, comprising the following steps:

[0033] Step S1: Acquire and preprocess three-dimensional medical image data to generate a normal image; add noise to the normal image to generate a pseudo abnormal image.

[0034] Step S1-1: Acquire three-dimensional medical image data and perform preprocessing.

[0035] First, 3D medical image data is obtained from public medical imaging databases or hospital imaging systems. This image data may include formats such as MRI (magnetic resonance imaging) and CT (computed tomography). The raw image data is then normalized, or preprocessed, to generate a normalized image. Preprocessing includes size normalization, pixel intensity normalization, and spatial resampling. Size normalization adjusts all images to a uniform spatial resolution to eliminate variations caused by different equipment or imaging conditions. Pixel intensity normalization uses linear stretching or Z-score normalization to adjust image grayscale values to a uniform range for subsequent model training. Spatial resampling interpolates the image to a fixed voxel spacing to ensure consistency of 3D structural features. Furthermore, denoising is optionally performed on the image, using algorithms such as Gaussian filtering and median filtering, to improve image quality. The preprocessed 3D medical image, or the normalized image, serves as input for the subsequent self-supervised training phase.

[0036] Step S1-2: Apply a forged abnormal data enhancement strategy based on Perlin noise and Worley noise to the data.

[0037] To achieve the goal of self-supervised learning, controllable artificial abnormal regions are introduced into normal images, allowing the model to learn to recover the original normal anatomical structure from the interference areas, thereby improving its ability to identify true abnormal regions during the inference phase. In this embodiment, a fusion generation mechanism of Perlin noise and Worley noise is used to enhance the image to generate pseudo-abnormal regions with complex textures and natural boundaries.

[0038] During the implementation, several local voxel regions are first selected from the normal image as pseudo-abnormality injection areas based on random sampling. Within each selected area, a perturbed texture is generated by mixing and superimposing Perlin noise and Worley noise. Perlin noise is used to construct a multi-scale, continuously changing texture structure that can simulate the delicate and dense tissue features present in real lesions; Worley noise is used to introduce spatial distribution unevenness, generate blurred boundaries or local void features, and enhance the uncertainty of the lesion's appearance. By introducing weighting parameters in the noise fusion process, the contribution of the two noises to the texture and boundaries is controlled, achieving a perturbation effect that is closer to the appearance of a real abnormality.

[0039] The fused noise is superimposed on the normal image to form a pseudo-abnormal image. Finally, the pseudo-abnormal image and its corresponding normal image are input into the model, allowing the model to learn the prior characteristics of normal tissue structure during the reconstruction task, so as to accurately detect the real abnormal area during the inference stage.

[0040] Step S2: Use a three-dimensional convolutional neural network to extract multi-scale features.

[0041] To effectively extract structural semantic information from 3D medical images, the present invention uses a 3D convolutional neural network as an encoder to perform multi-scale feature extraction on the input image. Preferably, this embodiment uses a 3D ResNet-50 network as the 3D convolutional neural network. Figure 2 The overall architecture diagram of 3D ResNet-50 in an embodiment of the present invention includes 3D convolutional layers, batch normalization, ReLU activation layers, and bottleneck layers. Figure 3 This figure shows the bottleneck layer architecture of the 3D ResNet-50 model in an embodiment of the present invention. The 3D ResNet-50 model adapts to the dimensional characteristics of 3D medical imaging by introducing 3D convolution operations on top of the standard 2D ResNet structure. Its network weights remain frozen during training and do not participate in gradient updates, serving only as a feature extractor.

[0042] The network structure begins with a 7×7×7 3D convolutional layer with a stride of 2 to extract low-level texture features. This convolutional layer is followed by a batch normalization layer and a ReLU activation function to enhance the network's nonlinear representation capabilities while mitigating the vanishing gradient problem. This layer is followed by a 3×3×3 3D max pooling layer for initial downsampling, reducing spatial resolution and expanding the receptive field.

[0043] The encoder backbone consists of four stages of residual modules (Stage 1 to Stage 4). Each stage is composed of multiple stacked 3D residual units, using a cross-layer connection mechanism to achieve layer-by-layer feature accumulation and semantic fusion. In Stages 1 to 3, the network outputs three encoded feature maps of different scales: high-resolution feature map layer 1, medium-resolution feature map layer 2, and low-resolution feature map layer 3, corresponding to structural and semantic information at different spatial resolutions.

[0044] In this example, a normal image and a pseudo-abnormal image are fed into the shared-weight 3D ResNet-50 encoder, generating three sets of corresponding multi-scale feature maps. Since the encoder does not participate in parameter updates, the extracted feature representations are passed as static input to a subsequent trainable feature embedding module for feature alignment and reconstruction.

[0045] Step S3: embedding features of normal images and pseudo-abnormal images and optimizing multi-scale feature loss.

[0046] In this embodiment, to restore the structure of the disturbed regions in the pseudo-anomaly image, the multi-scale features output by the encoder are further subjected to trainable feature mapping. The present invention constructs corresponding feature embedding networks for each of the three feature maps output by the encoder: high-resolution feature map layer 1, medium-resolution feature map layer 2, and low-resolution feature map layer 3. These networks are used to map the features into compact embedding vectors.

[0047] Specifically, the feature embedding network consists of the following structural units: a set of 1×1×1 three-dimensional convolutional layers, which compress the channel dimension while preserving the spatial structure; a subsequent spatial pooling operation to obtain a global feature response for the entire image; and finally, a fully connected layer to map the pooled result into a fixed-length one-dimensional embedding vector. These three sub-networks have independent parameters and act on feature maps of different scales to preserve the distribution of semantic features at each scale.

[0048] During the training phase, a normal image and its corresponding pseudo-abnormal image are fed into the encoder, generating three sets of feature maps with consistent scales. Embedding vectors for layer 1, layer 2, and layer 3 are then extracted through their respective feature embedding networks. Based on these embedding vectors, the mean square error (MSE) between the embedding vectors of the pseudo-abnormal image and the normal image at each scale is calculated as the feature reconstruction error at that scale.

[0049] This method uses the sum of the multi-scale MSEs as the training loss function and employs the Adam optimizer to update the parameters of the feature embedding network. By minimizing the distance between normal images and pseudo-anomaly images in feature space, the model learns to recover normal structure from perturbed regions. With repeated training, the embedding network gradually improves its ability to recover perturbed features, thereby enhancing the accuracy and robustness of the overall model in identifying true anomaly regions during the inference phase.

[0050] Step S4: 3D feature reconstruction and structure restoration based on the symmetric decoding structure.

[0051] In this embodiment, to achieve feature reconstruction of normal images, a three-dimensional decoder network is constructed, symmetrical to the encoder structure, to reconstruct the multi-scale feature vectors embedded in the image layer by layer. First, the embedded vectors are expanded and reshaped to obtain their own low-resolution feature maps. These are then fused to obtain the fused low-resolution feature maps, which serve as the initial input to the decoder. Subsequently, the decoder gradually restores the spatial resolution and reconstructs the image's structural details through a cascade of upsampling modules.

[0052] Specifically, the fully connected layer is used to expand and reshape the three embedding vectors output from layer 1, layer 2, and layer 3, obtaining a feature map with the same spatial resolution as the encoder's low-resolution feature map layer 3. Furthermore, the three expanded and reshaped feature maps are fused, either using average fusion or weighted average fusion, to obtain a reconstructed feature map for layer 3, which serves as the input to the decoder. The decoder includes multiple hierarchical decoding units, each of which consists of an upsampling operation and a three-dimensional convolutional layer to restore the spatial structure and extract layer-by-layer semantic features. Upsampling can be achieved using deconvolution, trilinear interpolation plus convolution, or other spatial scale recovery methods. During the decoding process, a reconstructed feature map for layer 2 and a reconstructed feature map for layer 1 that are symmetrical with those of the encoder are obtained, ultimately resulting in three reconstructed feature maps with the same resolution as the encoder output.

[0053] To improve the structural consistency of decoder learning, a feature distillation mechanism is introduced to guide the reconstruction process. Specifically, the original feature maps of layers 1, 2, and 3 extracted by the frozen encoder network serve as the teacher signal; the reconstructed feature maps output by the decoder serve as the student predictions; and the Huber loss function is used for supervised optimization between the two. By minimizing the structural difference between the reconstructed and original features at each scale, the decoder is guided to learn the normal representations modeled by the encoder while preserving the spatial structure, thereby enhancing the model's ability to recover normal features from pseudo-anomalous images.

[0054] Step S5: Anomaly scoring mechanism and discrimination strategy based on multi-scale structural error.

[0055] During the inference phase, the encoder, along with a trained feature embedding network and decoder, performs feature extraction, embedding mapping, and structural reconstruction on the 3D medical images under test. Anomaly scoring is then performed based on the reconstruction error. The input image is first fed into the encoder, where its multi-scale feature representation is extracted. The feature embedding module then generates a corresponding multi-scale embedding vector. This embedding vector is then expanded and reshaped before being fed into the decoder network, resulting in three reconstructed feature maps corresponding to the spatial resolutions of the original layers 1, 2, and 3.

[0056] For each scale, the structural error between the reconstructed feature map output by the decoder and the original feature map at the corresponding level of the encoder is calculated. This error reflects the degree of structural deviation in the embedding space through cosine similarity. These multi-scale errors are additively fused to generate an anomaly score map, which can be used to reflect the anomaly probability distribution at the voxel level. In practical applications, the score map can also be statistically aggregated to generate image-level or sample-level anomaly score indicators for overall risk assessment.

[0057] To determine the discrimination threshold, we used cross-validation to traverse the candidate threshold range on the validation set. Using the F1-score as the evaluation metric, we selected the optimal threshold as the final classification standard. During the inference process, when the reconstruction error of a region in the image exceeds the threshold, the region is considered a possible anomaly. When the global score of the entire image exceeds the threshold, the sample is considered to have a structural anomaly.

[0058] Step 6: System implementation of model deployment optimization and efficient inference process.

[0059] In order to facilitate the deployment and operation of the method of the present invention in a practical environment, a variety of optimization strategies are adopted.

[0060] During the training phase, using the mixed precision training function in the PyTorch framework can reduce video memory usage and speed up training.

[0061] To adapt to different types of devices, the model supports channel pruning and quantization, reducing model size and computing resource requirements, making it easier to run on embedded devices or edge computing devices.

[0062] During the inference phase, a sliding window approach is used to process large images, dividing the image into blocks and feeding them into the model to improve operational efficiency. Furthermore, asynchronous processing between the GPU and CPU is supported, allowing data reading, model calculation, and result output to proceed simultaneously, further accelerating processing speed.

[0063] To validate the effectiveness of the proposed self-supervised learning-based 3D medical image anomaly detection method, a retrospective dataset of 580 brain CT images was collected as an experimental dataset, consisting of 290 normal samples and 290 samples diagnosed as abnormal by professional physicians. All image data underwent a unified preprocessing process, including size normalization, pixel intensity standardization, and spatial resampling, to ensure consistency in spatial resolution and grayscale distribution.

[0064] The dataset was then divided into a training set and a validation set in a ratio of 8:2. The training set contains only normal samples, which are used for learning the model in the self-supervised training phase of the present invention; while the validation set contains both normal and abnormal samples, which are used to evaluate the model's ability to recognize abnormal areas without label supervision. Since there are relatively few existing unsupervised or self-supervised anomaly detection methods for three-dimensional medical images, several representative three-dimensional models are introduced in this experimental design as comparison baselines to comprehensively evaluate the performance advantages of the method of the present invention. Among the three-dimensional models introduced in this study, 3D ResNet-18 adopts a standard supervised training process, and 3D STFPM and SSMCTB adopt a self-supervised learning method.

[0065] The experiment uses three indicators: accuracy, precision, and F1-score to comprehensively evaluate the detection effect of each model on the validation set. The experimental results are shown in Table 1:

[0066] Table 1 Experimental results

[0067]

[0068] Experimental results show that our method outperforms other control methods in three core metrics: accuracy, precision, and F1 score. In particular, the F1 score reaches 0.7097, which is approximately 3.4% higher than the traditional supervised method 3D ResNet-18, demonstrating improved anomaly detection performance. Notably, our method achieves superior performance to supervised methods without relying on anomalous samples for training, fully validating the effectiveness and robustness of the proposed self-supervised multi-scale structure alignment mechanism.

[0069] In summary, the method of the present invention not only performs superior in accuracy, but also has stronger generalization ability and practicality, and is suitable for promotion and use in real clinical scenarios.

[0070] The above description is merely illustrative of certain exemplary embodiments of the present invention. It goes without saying that those skilled in the art will be able to modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and description are illustrative in nature and should not be construed as limiting the scope of protection of the claims.

Claims

1. A 3D medical image anomaly detection method based on self-supervised learning, characterized in that: The steps include: Step S1, acquiring and preprocessing three-dimensional medical image data to generate a normal image; adding noise to the normal image to generate a pseudo abnormal image; Step S2: inputting the normal image and the pseudo abnormal image into a three-dimensional convolutional neural network encoder for multi-scale feature extraction, respectively, to obtain three corresponding original feature maps of different scales, and sorting the original feature maps from high to low resolution into layer 1 feature map, layer 2 feature map and layer 3 feature map; Step S3: constructing three independent feature embedding networks based on the original feature maps at three different scales; using the feature embedding networks, converting the three original feature maps into one-dimensional embedding vectors; each feature embedding network processes the original feature map at a corresponding scale; calculating the sum of the mean squared errors between the embedding vectors of the pseudo-abnormal image and the normal image at each scale; and updating the parameters of the feature embedding networks using an Adam optimizer; Step S4: construct a decoder symmetrical to the encoder structure; use a fully connected layer to expand the three embedding vectors and reshape them respectively to generate three low-resolution feature maps, where the spatial resolution of the low-resolution feature maps is the same as that of the layer3 feature map; The three low-resolution feature maps are fused to obtain a layer3 reconstructed feature map; the upsampling module of the decoder cascades and reconstructs the layer3 reconstructed feature map layer by layer to obtain a layer1 reconstructed feature map and a layer2 reconstructed feature map that have the same resolution as the layer1 feature map and the layer2 feature map, respectively; The preprocessing includes performing size normalization, pixel intensity normalization, spatial resampling processing and denoising processing on the three-dimensional medical image data; Each of the feature embedding networks includes: a group of 1×1×1 three-dimensional convolutional layers, a spatial pooling layer connected to the 1×1×1 three-dimensional convolutional layer, and a fully connected layer connected at the end.

2. The method for detecting anomalies in three-dimensional medical images based on self-supervised learning according to claim 1, wherein: In the step S1, in the process of adding noise to the normal image, Perlin noise and Worley noise are mixed and superimposed to generate a disturbance texture; In the noise fusion process, the contribution of the two noises to the texture and boundary is controlled by weighting parameters.

3. The method for detecting anomalies in three-dimensional medical images based on self-supervised learning according to claim 1, wherein: In step S2, the network structure of the three-dimensional convolutional neural network adopts a 7×7×7 three-dimensional convolution layer in the first layer with a step size of 2 for extracting low-level texture features; the 7×7×7 three-dimensional convolution layer is connected to a batch normalization layer and a ReLU activation function, and then a 3×3×3 three-dimensional maximum pooling layer is introduced.

4. The method for detecting anomalies in three-dimensional medical images based on self-supervised learning according to claim 1, wherein: In step S4, the Huber loss function is used to maintain the structural consistency of the reconstruction process by minimizing the structural difference between the reconstructed feature map and the original feature map at each scale.

5. The method for detecting anomalies in three-dimensional medical images based on self-supervised learning according to claim 4, wherein: The method further comprises step S5 of respectively calculating the cosine similarity between the reconstructed feature map at each scale and each original feature map and fusing them to generate an abnormality score map.

6. A 3D medical image anomaly detection system based on self-supervised learning, characterized by: include: Acquisition and preprocessing module, pseudo-anomaly generation module, 3D feature extraction module, feature embedding module and symmetric decoding and reconstruction module; The system is used to implement the steps of the method according to claim 1; The acquisition and preprocessing module is used to acquire and preprocess the three-dimensional medical image to generate the normal image; The pseudo-abnormality generating module is used to generate the pseudo-abnormality image according to the normal image; The three-dimensional feature extraction module inputs the normal image and the pseudo-abnormal image into the encoder of the three-dimensional convolutional neural network for multi-scale feature extraction to obtain the three original feature maps of different scales; the feature embedding module constructs three independent feature embedding networks according to the three original feature maps of different scales and uses the feature embedding networks to convert the three original feature maps into one-dimensional embedding vectors; the symmetric decoding and reconstruction module is used to construct the decoder that is symmetrical with the encoder structure; the decoder is used to generate the low-resolution feature map and then generate the reconstructed feature map.

Images