A method for detecting ovarian cancer peritoneal metastasis lesions

Abstract

The application relates to an ovarian cancer abdominal cavity metastasis lesion detection method and belongs to the technical field of lesion detection. The method solves the problem that the prior art cannot accurately detect ovarian cancer abdominal cavity metastasis lesions. The method comprises the following steps: acquiring a CT image of a patient with lesion metastasis and a lesion segmentation label to construct an initial sample set; acquiring a CT image of a patient without lesion metastasis; taking a sample in the initial sample set as a positive sample, taking a sample of the CT image of the patient without lesion metastasis as a negative sample, generating a simulated real sample based on the positive sample and the negative sample, expanding the initial sample set to obtain an expanded sample set; training a deep learning network model for lesion segmentation based on the expanded sample set to obtain a lesion segmentation model; and inputting a CT image of a patient to be detected into the lesion segmentation model to obtain a lesion metastasis detection result. The metastasis lesion detection is quickly and accurately obtained.

Description

Technical Field

[0001] This invention relates to the field of lesion detection technology, and in particular to a method for detecting peritoneal metastatic lesions of ovarian cancer. Background Technology

[0002] Ovarian cancer is one of the most aggressive and deadliest malignant tumors of the female reproductive system. Advanced-stage patients typically exhibit widespread peritoneal metastasis, with common sites including the peritoneum, greater omentum, liver, spleen, mesentery, and lymph nodes. Abdominal and pelvic CT imaging is a crucial tool for clinical assessment and diagnosis of peritoneal dissemination of ovarian cancer, significantly impacting decisions regarding tumor burden assessment and surgical or neoadjuvant chemotherapy. However, the application of computer-aided diagnostic (CAD) and deep learning models in ovarian cancer image analysis still faces several challenges: First, ovarian cancer lesion annotation is extremely difficult: Advanced ovarian cancer lesions are distributed throughout the entire abdominal cavity, exhibiting complex morphologies, indistinct borders, and varying sizes. Manually annotating each lesion is extremely labor-intensive, and inconsistencies in annotation standards between different physicians can lead to poor consistency. Second, ovarian cancer lesions exhibit diverse characteristics: metastatic lesions present differently in different anatomical locations, making it difficult for existing methods to uniformly detect lesions throughout the entire abdominal cavity. Third, data access is limited: sharing of medical imaging data is limited, especially abdominal and pelvic CT images with detailed lesion annotations, which are usually only available to a few research institutions, making it difficult to train deep learning models.

[0003] To overcome these challenges, weakly supervised learning and self-supervised learning methods have been gradually introduced into the field of medical image analysis in recent years. However, these methods can only learn the global features of images and cannot distinguish between lesion areas and normal tissue. Furthermore, data augmentation methods are mostly limited to simple transformations such as rotation, flipping, and contrast adjustment, and cannot create new lesion data. Radiomics-based methods are limited by the ability to manually design features and also heavily rely on manual delineation. Currently, quantitative studies of ovarian cancer images still heavily depend on manually labeled data, which is not only time-consuming and labor-intensive but also suffers from inconsistent labeling standards and poor reproducibility.

[0004] Accurate lesion segmentation requires high-quality annotated data. Large-scale, high-quality lesion annotation data, including primary and metastatic lesions, is needed to ensure the model's generalization ability can adapt to the complexity of peritoneal disseminated lesions. Current lesion detection methods based on CNNs or Transformers mainly rely on end-to-end supervised learning; however, these methods heavily depend on large-scale annotated data, making it difficult to address issues such as high annotation costs, insufficient data, and poor model generalization. Summary of the Invention

[0005] Based on the above analysis, the present invention aims to provide a method for detecting peritoneal metastatic lesions of ovarian cancer, in order to solve the problem that existing methods cannot accurately detect peritoneal metastatic lesions of ovarian cancer.

[0006] On one hand, embodiments of the present invention provide a method for detecting peritoneal metastatic lesions of ovarian cancer, comprising the following steps:

[0007] CT images and lesion segmentation labels of patients with metastatic lesions were obtained to construct an initial sample set;

[0008] CT images of patients without lesion metastasis are acquired. Samples from the initial sample set are used as positive samples, and samples from the CT images of patients without lesion metastasis are used as negative samples. Simulated positive samples are generated based on the positive and negative samples, and the initial sample set is expanded to obtain an expanded sample set.

[0009] A lesion segmentation model is obtained by training a deep learning network model for lesion segmentation based on an expanded sample set;

[0010] The CT images of the patient to be tested are input into the lesion segmentation model to obtain the lesion metastasis detection results.

[0011] Based on further improvements to the above method, the following approach is used to generate positive simulation samples:

[0012] For each positive sample and each negative sample, perform abdominal and pelvic organ segmentation to obtain the sub-organ mask for each positive sample and each negative sample;

[0013] For each positive sample, similar negative samples are selected as paired samples based on the sub-organ mask;

[0014] Based on the positive sample, each paired sample of the positive sample is dissected and aligned, and the aligned paired samples and the positive sample are used as positive and negative sample pairs in the second sample set; construct the second sample set;

[0015] A simulated sample generation model is obtained by training a generative adversarial network based on the second sample set;

[0016] The simulated positive sample corresponding to each negative sample in the second sample set is obtained based on the simulation sample generation model.

[0017] Based on the above method, a further improvement is made: for each positive sample, similar negative samples are selected as paired samples for that positive sample based on the sub-organ mask, including:

[0018] The similarity between the positive sample and each negative sample is obtained based on the similarity between the sub-organ mask of the positive sample and the sub-organ mask of each negative sample.

[0019] Negative samples with a similarity greater than a first threshold are selected as paired samples for the positive sample.

[0020] Based on the further improvement of the above method, the similarity between the positive sample and each negative sample is calculated using the following formula:

[0021]

[0022] in, This represents the CT image of the i-th positive sample. Let K represent the CT image of the j-th negative sample, and K represent the number of sub-organ masks. This represents the mask of the k-th sub-organ of the i-th positive sample. This represents the mask of the k-th sub-organ of the j-th negative sample. Let |·| represent the similarity between the i-th positive sample and the j-th negative sample, and |·| represent the number of voxels in the mask.

[0023] Based on a further improvement to the above method, each paired sample of the positive sample is dissected and aligned, and the aligned paired samples and the positive sample are used as positive and negative sample pairs in a second sample set, including:

[0024] Using the CT image of the positive sample as the target image, histogram matching is performed on the CT images of each paired sample to obtain the first transformed CT image of each paired sample;

[0025] With the goal of minimizing the overlap error between sub-organ masks, rigid registration is performed on the CT image of the positive sample and the first transformed CT image of each paired sample to obtain the second transformed CT image of each paired sample.

[0026] With the goal of maximizing the similarity between sub-organs, non-rigid registration is performed on the CT image of the positive sample and the second transformed CT image of each paired sample to obtain the aligned paired samples.

[0027] Based on a further improvement of the above method, the objective function of the rigid registration is:

[0028]

[0029] Among them, T affine This represents the transformation matrix for rigid registration, where K represents the number of sub-organ masks. This represents the mask of the k-th sub-organ of a positive sample. This represents the mask of the k-th sub-organ of the paired sample.

[0030] Based on a further improvement of the above method, the objective function of the non-rigid registration is:

[0031]

[0032] in, p(a,b) represents the joint probability distribution of gray value a in image A and gray value b in image B, P(a) represents the probability of gray value a in image A, P(b) represents the probability of gray value b in image B, and K represents the number of sub-organ masks. This represents the mask of the k-th sub-organ of a positive sample. This represents the second conversion of paired sample CT images. T represents the mask of the k-th sub-organ after rigid registration transformation of paired samples. B-spline Let β represent the non-rigid registration transformation matrix, and let β represent the weighting parameter. This represents the 2-norm of a matrix.

[0033] Based on a further improvement of the above method, the generative adversarial network is a conditional generative adversarial network;

[0034] For each positive and negative sample pair in the second sample set, the lesion morphology and location information of the positive sample are used as the generation conditions corresponding to the positive and negative sample pair. The conditional generative adversarial network is trained based on the second sample set to obtain the simulation sample generation model.

[0035] Based on the above method, a further improvement is made, and the training loss of the conditional generative adversarial network is calculated using the following formula:

[0036] L cgan =L G +L D +λL C

[0037] Among them, L G L represents the generator loss. D L represents the discriminator loss. C λ represents the contrast loss, and λ represents the weighting coefficient.

[0038] Based on the further improvement of the above method, the contrast loss is calculated using the following formula:

[0039]

[0040] in, This represents the CT image of the k-th simulated positive sample. This represents the CT image of the p-th simulated positive sample, which was generated under the same conditions as the k-th simulated positive sample. p n represents the number of simulated positive samples whose generation conditions are the same as those of the k-th simulated positive sample's CT image. n This represents the number of simulated positive samples whose generation conditions differ from those of the k-th simulated positive sample's CT image. Let q represent the simulated positive sample whose generation conditions are different from those of the k-th simulated positive sample's CT image. N represents the number of samples in the current training batch. τ represents the temperature coefficient used to control the sharpness of the sample distribution. sim(·,·) represents the similarity calculation function.

[0041] Compared with existing technologies, this invention constructs an initial sample set by acquiring CT images of patients with lesion metastasis and lesion segmentation labels, and then acquires CT images of patients without lesion metastasis. Using samples from the initial sample set as positive samples and samples from CT images of patients without lesion metastasis as negative samples, simulated positive samples are generated based on the positive and negative samples. Image generation is then performed to generate new lesion data, thereby expanding the positive samples, reducing the annotation burden, and solving the problems of difficult annotation and insufficient data for abdominal metastatic lesions in CT images of advanced ovarian cancer. By training a deep learning network model based on the expanded sample set, the generalization ability of the trained model is improved, making the model's detection accuracy more accurate. For patients to be detected, their CT images can be input into the lesion segmentation model to quickly and accurately obtain lesion metastasis detection results.

[0042] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0043] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0044] Figure 1 This is a flowchart of a method for detecting peritoneal metastatic lesions of ovarian cancer according to an embodiment of the present invention. Detailed Implementation

[0045] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0046] A specific embodiment of the present invention discloses a method for detecting peritoneal metastatic lesions of ovarian cancer, such as... Figure 1 As shown, it includes the following steps:

[0047] S1. Obtain CT images and lesion segmentation labels of patients with metastatic lesions to construct an initial sample set; obtain CT images of patients without metastatic lesions as negative samples;

[0048] S2. Obtain CT images of patients without lesion metastasis; use samples from the initial sample set as positive samples and samples from CT images of patients without lesion metastasis as negative samples, generate simulated positive samples based on the positive and negative samples, and expand the initial sample set to obtain an expanded sample set.

[0049] S3. A lesion segmentation model is obtained by training a deep learning network model for lesion segmentation based on an expanded sample set;

[0050] S4. Input the CT image of the patient to be tested into the lesion segmentation model to obtain the lesion metastasis detection result.

[0051] Compared with existing technologies, the ovarian cancer peritoneal metastasis detection method provided in this embodiment constructs an initial sample set by acquiring CT images of patients with metastatic lesions and lesion segmentation labels, and then acquires CT images of patients without metastatic lesions. Using samples from the initial sample set as positive samples and samples from CT images of patients without metastatic lesions as negative samples, simulated positive samples are generated based on the positive and negative samples. Image generation is then performed to generate new lesion data, thereby expanding the positive samples, reducing the annotation burden, and solving the problems of difficult annotation and insufficient data for peritoneal metastatic lesions in CT images of advanced ovarian cancer. By training a deep learning network model based on the expanded sample set, the generalization ability of the trained model is improved, making the detection accuracy of the model more accurate. For the patient to be detected, inputting their CT image into the lesion segmentation model can quickly and accurately obtain the lesion metastasis detection result.

[0052] In practice, the lesion segmentation label is the lesion mask of the CT image.

[0053] During implementation, abdominal and pelvic CT images of patients with metastatic lesions are obtained. Senior gynecological radiologists can manually annotate ovarian cancer lesions to obtain lesion segmentation labels, i.e., lesion masks. The annotations include metastatic sites such as the peritoneum, greater omentum, liver, spleen, mesentery, and lymph nodes. The abdominal and pelvic CT images of patients with metastatic lesions and their corresponding lesion labels are used as one sample in the initial sample set to construct the initial sample set.

[0054] Since the initial sample set contains a small number of samples, it is necessary to generate simulation samples.

[0055] Acquire abdominal and pelvic CT images of patients without lesion metastasis. Samples from the initial sample set are used as the generation target; therefore, samples from the initial sample set are treated as positive samples, and abdominal and pelvic CT images of patients without lesion metastasis are treated as negative samples. Simulated positive samples are generated based on the positive and negative samples to expand the initial sample set.

[0056] Specifically, the following method is used to generate simulated positive samples:

[0057] S21. Perform abdominal and pelvic organ segmentation on each positive and negative sample to obtain the sub-organ mask for each positive and negative sample.

[0058] S22. For each positive sample, select similar negative samples as paired samples based on the sub-organ mask.

[0059] S23. Based on the positive sample, dissect and align each paired sample of the positive sample, and use the aligned paired samples and the positive sample as positive and negative sample pairs in the second sample set; construct the second sample set;

[0060] S24. A simulated sample generation model is obtained by training a generative adversarial network based on the second sample set;

[0061] S25. Based on the simulation sample generation model, obtain the simulation positive sample corresponding to each negative sample in the second sample set.

[0062] In practice, to ensure that the generated simulation samples conform to the real anatomical environment, this invention introduces an anatomical alignment strategy, so that the generated simulation positive samples have similar anatomical structural features to the original positive samples.

[0063] Specifically, the main organs of the abdominal and pelvic cavity, such as the liver, spleen, intestines, and bladder, are segmented for each positive and negative sample to obtain a mask for each sub-organ, which is used for anatomical region alignment.

[0064] During implementation, the images with and without lesions with the most similar anatomical backgrounds on the original images are initially screened to select paired samples for each positive sample.

[0065] Specifically, for each positive sample, similar negative samples are selected as paired samples based on the sub-organ mask, including:

[0066] The similarity between the positive sample and each negative sample is obtained based on the similarity between the sub-organ mask of the positive sample and the sub-organ mask of each negative sample.

[0067] Negative samples with a similarity greater than a first threshold are selected as paired samples for the positive sample.

[0068] During implementation, based on the masks of each sub-organ in the abdominal and pelvic cavity, the volume, location, and spatial layout information of each sub-organ are obtained. The differences in organ volume and relative spatial location between the two images are calculated to obtain the similarity.

[0069] Specifically, the similarity between the positive sample and each negative sample is calculated using the following formula:

[0070]

[0071] in, This represents the CT image of the i-th positive sample. Let K represent the CT image of the j-th negative sample, and K represent the number of sub-organ masks. This represents the mask of the k-th sub-organ of the i-th positive sample. This represents the mask of the k-th sub-organ of the j-th negative sample. Let |·| represent the similarity between the i-th positive sample and the j-th negative sample, and |·| represent the number of voxels in the mask.

[0072] For the i-th positive sample, calculate its similarity with each negative sample, take the n negative samples with high similarity as the paired samples of the positive sample, and then perform dissection and alignment on each paired sample of the positive sample based on the positive sample. The aligned paired samples and the positive sample are used as positive and negative sample pairs in the second sample set, thereby constructing the second sample set.

[0073] Specifically, based on the positive sample, each paired sample of the positive sample is dissected and aligned, and the aligned paired samples and the positive sample are used as positive and negative sample pairs in the second sample set, including:

[0074] Using the CT image of the positive sample as the target image, histogram matching is performed on the CT images of each paired sample to obtain the first transformed CT image of each paired sample;

[0075] With the goal of minimizing the overlap error between sub-organ masks, rigid registration is performed on the CT image of the positive sample and the first transformed CT image of each paired sample to obtain the second transformed CT image of each paired sample.

[0076] With the goal of maximizing the similarity between sub-organs, non-rigid registration is performed on the CT image of the positive sample and the second transformed CT image of each paired sample to obtain the aligned paired samples.

[0077] In practice, the CT image of the positive sample is first used as the target image. Histogram matching is used to map the contrast, grayscale range, texture and other features of the CT image of each paired sample to the distribution space of the positive sample.

[0078] It should be noted that the sub-organ mask of the first converted CT image of the paired sample is the same as the sub-organ mask of the original CT image of the paired sample.

[0079] In implementation, the HistMatch histogram matching function can be used to match samples. The statistical distribution should be as close as possible to that of the positive sample. The first converted CT image of each paired sample is obtained and labeled as

[0080] To achieve accurate anatomical alignment, rigid registration is performed first, followed by non-rigid registration.

[0081] In practice, the optimization objective of rigid registration is to minimize the overlap error of the mask.

[0082] Specifically, the objective function for rigid registration is:

[0083]

[0084] Among them, T affine This represents the transformation matrix for rigid registration, where K represents the number of sub-organ masks. This represents the mask of the k-th sub-organ of a positive sample. This represents the mask of the k-th sub-organ of the paired sample.

[0085] In implementation, the Affine affine transformation method is used for rigid matching, with positive samples... For reference, a rigid registration method (Affine affine transformation) is used to register the paired samples. Initially aligned to the anatomical space of the positive sample's CT image, a second transformed CT image is obtained for each paired sample, labeled as...

[0086] After obtaining the transformation matrix for rigid registration, the paired samples will be... Multiply by the rigid registration matrix, Second-transformed CT images of paired samples were obtained. Similarly, each sub-organ mask of the paired samples The corresponding transformation is performed based on the rigid registration transformation matrix, i.e., multiplied by the rigid registration matrix, to obtain the sub-organ mask after rigid registration transformation.

[0087] During implementation, the optimization objective is to maximize the similarity between sub-organs, with positive samples... For reference, the B-spline method was used to process the second-transformed CT images. and Non-rigid registration is performed to further refine local differences in the images, resulting in aligned CT images. adapt The anatomical spatial structure is highly matched, thus supporting the subsequent simulation of lesions to be realistically superimposed on reasonable anatomical locations.

[0088] Specifically, the objective function for non-rigid registration is:

[0089]

[0090] in, p(a,b) represents the joint probability distribution of gray value a in image A and gray value b in image B, P(a) represents the probability of gray value a in image A, P(b) represents the probability of gray value b in image B, and K represents the number of sub-organ masks. This represents the mask of the k-th sub-organ of a positive sample. This represents the second conversion of paired sample CT images. T represents the mask of the k-th sub-organ after rigid registration transformation of paired samples. B-spline Let β represent the non-rigid registration transformation matrix, and let β represent the weighting parameter. This represents the 2-norm of a matrix.

[0091] During implementation, regular expressions are added. To prevent overfitting, the similarity between each sub-organ is calculated, resulting in smaller local differences in CT images. The aligned paired samples and the positive castration version are highly matched in anatomical space structure, thus supporting the subsequent simulation of positive lesions to be realistically superimposed on reasonable anatomical locations.

[0092] After obtaining the transformation matrix for non-rigid registration, the paired samples... Multiply with the non-rigid registration matrix, Obtain the third transformed CT image of the paired sample. Aligned CT images The aligned paired samples are obtained.

[0093] Positive samples and aligned paired samples constitute a positive-negative sample pair in the second sample set, where the CT images of the positive samples are... CT images of negative samples are

[0094] In practice, to make the generated positive simulation samples more realistic, the generative adversarial network adopts a conditional generative adversarial network.

[0095] During implementation, for each positive sample, the morphological and location information of the lesion is extracted as a generation condition. The morphological and location information of the lesion includes: lesion type (such as peritoneal metastasis, greater omentum metastasis, etc.), volume, maximum diameter, sphericity, density information (mean and standard deviation of HU value), and relative position of the lesion.

[0096] For each positive and negative sample pair in the second sample set, the morphology and location information of the lesion in the positive sample are used as the generation conditions corresponding to the positive and negative sample pair. The conditional generative adversarial network is trained based on the second sample set to obtain the simulation sample generation model.

[0097] In practice, a conditional generative adversarial network consists of a generator and a discriminator.

[0098] The generator is used to generate simulated positive samples based on the negative samples in the positive-negative sample pair and the generation conditions;

[0099] The discriminator is used to distinguish between genuine and fake positive samples and simulated positive samples.

[0100] In practice, the generator's processing includes conditional fusion, an encoder, and a decoder. The initial input is CT images of negative samples from a pair of positive and negative samples. The generation condition c is combined with the CT image using a feature concatenation and fusion approach, and then input into the network: the generation condition is converted into a numerical vector, and then a multilayer perceptron (MLP) is used to map the numerical vector of the generation condition to a high-dimensional feature space to obtain the conditional feature representation; then the conditional feature representation is concatenated with the CT image along the channel dimension and input into the subsequent encoder module.

[0101] The encoder uses a convolutional neural network to extract spatial features from the fused input data step by step: First, shallow structural information in the image is extracted through a series of convolutions, batch normalization, ReLU layers or LReLU layers; second, deeper semantic information of the image is obtained through progressive downsampling; at each encoder stage, feature maps are preserved for subsequent skip connections.

[0102] The decoder progressively upsamples and reconstructs the deep feature representation obtained from the encoder through a deconvolutional network. First, the decoder takes the deep feature representation from the encoder as input, progressively upsamples it, and restores the spatial resolution. Second, it employs the classic skip connections in the U-Net structure; after each deconvolution process, the encoder uses the extracted feature map and its corresponding mirror-image convolutional layer feature map in a skip connection, directly stacking the corresponding feature maps before feeding them into the next layer. Finally, a 1×1 convolutional layer converts the multi-channel feature mapping into a single-channel image, generating a CT image of a simulated positive sample containing simulated lesions.

[0103] The loss function of the generator is expressed as:

[0104]

[0105] Where D(·) represents the discriminator and G(·) represents the generator. This represents the CT image of the negative sample of the k-th positive-negative sample pair in the current training batch, where c represents the generation condition and N represents the number of samples in the current training batch.

[0106] The discriminator's input consists of two parts: one is the CT image of the positive sample in the positive-negative sample pair of the second sample set and the generation conditions, which constitute the real image pair. Secondly, the CT images of simulated positive samples generated by the generator. And the generation conditions, constituting the simulated image pair

[0107] After receiving the input, the discriminator extracts features from real image pairs and simulated image pairs using a convolutional neural network, and outputs the probability of the real image. The specific loss function is as follows:

[0108]

[0109] Where D(·) is the discriminator and G(·) is the generator. This represents the CT image of the positive sample in the k-th positive-negative sample pair of the current training batch. This represents the CT image of the negative sample of the k-th positive-negative sample pair in the current training batch, where c represents the generation condition and N represents the number of samples in the current training batch.

[0110] The discriminator targets real images Output a probability close to 1 for simulated image pairs. Output the probability that is close to 0.

[0111] The training process employs a classic generative adversarial optimization strategy. First, the generator G is fixed, and the discriminator D is trained to accurately distinguish between real and simulated images. Then, the discriminator D is fixed, and the generator G is trained to generate simulated images as realistic as possible to deceive the discriminator. The two processes alternate iteratively until, at the end of training, the generator can stably output realistic simulated lesion images.

[0112] A generative adversarial network (GAN) is trained based on a second sample set to obtain a simulated sample generation model. CT images of negative samples from the second sample set and their corresponding generation conditions are input into the simulated sample generation model to generate simulated positive samples. The difference between the positive samples and the simulated positive samples is used for binarization with the Otsu thresholding method to extract the lesion region (ROI) and generate the lesion mask for the simulated positive samples, i.e., the lesion segmentation label.

[0113] During implementation, a graphics processor can be used to generate the image.

[0114] The generated simulated positive samples are added to the initial sample set to expand it, resulting in an expanded sample set. A deep learning network model is then trained based on this expanded sample set to obtain a lesion segmentation model, thereby improving the model's generalization ability.

[0115] Because advanced ovarian cancer lesions exhibit diverse morphologies and often metastasize to multiple sites throughout the abdominal cavity, it is necessary to focus on both fine local features and the ability to capture global contextual information. The deep learning network model of this invention employs the SwinUNETR architecture, which combines the advantages of the Swin Transformer and U-Net architectures. The training loss function is the Dice loss function, and the output is the lesion segmentation result.

[0116] First, the input sample is a 512×512×N 3D CT image. Using the self-attention mechanism of the Swin Transformer, the input data is segmented into multiple local windows (4×4×4 pixel cube windows). The Swin Transformer consists of two components: a window-based multi-head self-attention module (W-MSA) that rearranges the input features in a non-overlapping manner to generate small patches, and then performs self-attention computation within each patch; and a shifted window-based multi-head self-attention (SW-MSA) module that moves the features horizontally and vertically, performing self-attention computation on non-overlapping patches. Through a multi-layered stacked structure (4 Swin Transformer blocks), deep-level global semantic features are gradually extracted, ultimately outputting a global high-dimensional feature representation.

[0117] The U-Net decoder receives the global high-dimensional feature representation output by the Transformer and performs progressive feature upsampling to gradually restore the original image size. At each upsampling stage, multi-scale feature maps (local feature maps) saved in the Swin Transformer encoder stage are fused through skip connections. By progressive feature fusion and layer-by-layer restoration of spatial resolution, the decoder can accurately capture the fine local boundaries and morphological features of lesions, obtaining high-precision lesion segmentation results and achieving accurate image semantic segmentation.

[0118] The model outputs a binary segmentation mask of the same size as the CT image, where 1 indicates that the pixel belongs to the lesion region and 0 indicates that it does not belong to the lesion region. The training process uses the Dice loss function to optimize the network parameters. The specific formula for calculating the loss for a single sample is as follows:

[0119]

[0120] Among them, y s This represents the true value of the s-th voxel in the lesion segmentation mask of the sample. V represents the predicted value of the s-th voxel in the lesion segmentation mask of the sample, V represents the number of voxels in the sample CT image, and ∈ is a smoothing constant to avoid a denominator of 0.

[0121] Once the preset loss accuracy or number of iterations is reached, training is complete, and a lesion segmentation model is obtained. Inputting the CT images of the patient to be detected into the lesion segmentation model allows for rapid and accurate detection of lesion metastasis.

[0122] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0123] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting peritoneal metastatic lesions of ovarian cancer, characterized in that, Includes the following steps: CT images and lesion segmentation labels of patients with metastatic lesions were obtained to construct an initial sample set; CT images of patients without lesion metastasis are acquired. Samples from the initial sample set are used as positive samples, and samples from the CT images of patients without lesion metastasis are used as negative samples. Simulated positive samples are generated based on the positive and negative samples, and the initial sample set is expanded to obtain an expanded sample set. A lesion segmentation model is obtained by training a deep learning network model for lesion segmentation based on an expanded sample set; The CT images of the patient to be tested are input into the lesion segmentation model to obtain the lesion metastasis detection results; The following methods are used to generate positive simulation samples: For each positive sample and each negative sample, perform abdominal and pelvic organ segmentation to obtain the sub-organ mask for each positive sample and each negative sample; For each positive sample, similar negative samples are selected as paired samples based on the sub-organ mask; Based on the positive sample, each paired sample of the positive sample is dissected and aligned, and the aligned paired samples and the positive sample are used as positive and negative sample pairs in the second sample set; construct the second sample set; A simulated sample generation model is obtained by training a generative adversarial network based on the second sample set; The simulated positive sample corresponding to each negative sample in the second sample set is obtained based on the simulation sample generation model.

2. The method for detecting peritoneal metastatic lesions of ovarian cancer according to claim 1, characterized in that, For each positive sample, similar negative samples are selected as paired samples based on the sub-organ mask, including: The similarity between the positive sample and each negative sample is obtained based on the similarity between the sub-organ mask of the positive sample and the sub-organ mask of each negative sample. Negative samples with a similarity greater than a first threshold are selected as paired samples for the positive sample.

3. The method for detecting peritoneal metastatic lesions of ovarian cancer according to claim 2, characterized in that, The similarity between the positive sample and each negative sample is calculated using the following formula: ； ； in, This represents the CT image of the i-th positive sample. Let K represent the CT image of the j-th negative sample, and K represent the number of sub-organ masks. This represents the mask of the k-th sub-organ of the i-th positive sample. This represents the mask of the k-th sub-organ of the j-th negative sample. Let |·| represent the similarity between the i-th positive sample and the j-th negative sample, and |·| represent the number of voxels in the mask.

4. The method for detecting peritoneal metastatic lesions of ovarian cancer according to claim 1, characterized in that, Based on the positive sample, each paired sample of the positive sample is dissected and aligned. The aligned paired samples and the positive sample are then used as positive and negative sample pairs in the second sample set, including: Using the CT image of the positive sample as the target image, histogram matching is performed on the CT images of each paired sample to obtain the first transformed CT image of each paired sample; With the goal of minimizing the overlap error between sub-organ masks, rigid registration is performed on the CT image of the positive sample and the first transformed CT image of each paired sample to obtain the second transformed CT image of each paired sample. With the goal of maximizing the similarity between sub-organs, non-rigid registration is performed on the CT image of the positive sample and the second transformed CT image of each paired sample to obtain the aligned paired samples.

5. The method for detecting peritoneal metastatic lesions of ovarian cancer according to claim 4, characterized in that, The objective function for rigid registration is: ； in, This represents the transformation matrix for rigid registration, where K represents the number of sub-organ masks. This represents the mask of the k-th sub-organ of the i-th positive sample. This represents the mask of the k-th sub-organ of the j-th paired sample.

6. The method for detecting peritoneal metastatic lesions of ovarian cancer according to claim 4, characterized in that, The objective function for the non-rigid registration is: ； in, , This indicates that the grayscale value in image A is... And the grayscale value in image B is The joint probability distribution, This indicates that the grayscale value in image A is... The probability, This indicates that the grayscale value in image B is... The probability, where K represents the number of sub-organ masks, This represents the mask of the k-th sub-organ of the i-th positive sample. This represents the second conversion of paired sample CT images. This represents the mask of the k-th sub-organ after rigid registration transformation of paired samples. Let β represent the non-rigid registration transformation matrix, and let β represent the weighting parameter. This represents the 2-norm of a matrix.

7. The method for detecting peritoneal metastatic lesions of ovarian cancer according to claim 1, characterized in that, The generative adversarial network is a conditional generative adversarial network; For each positive and negative sample pair in the second sample set, the lesion morphology and location information of the positive sample are used as the generation conditions corresponding to the positive and negative sample pair. The conditional generative adversarial network is trained based on the second sample set to obtain the simulation sample generation model.

8. The method for detecting peritoneal metastatic lesions of ovarian cancer according to claim 7, characterized in that, The training loss of the conditional generative adversarial network is calculated using the following formula: ； in, Indicates generator loss. Indicates discriminator loss. λ represents the contrast loss, and λ represents the weighting coefficient.

9. The method for detecting peritoneal metastatic lesions of ovarian cancer according to claim 8, characterized in that, The contrast loss is calculated using the following formula: ； in, This represents the CT image of the k-th simulated positive sample. This represents the CT image of the p-th simulated positive sample, which was generated under the same conditions as the k-th simulated positive sample. This represents the number of simulated positive samples that share the same generation conditions as the CT image of the k-th simulated positive sample. This represents the number of simulated positive samples whose generation conditions differ from those of the k-th simulated positive sample's CT image. Let q represent the simulated positive sample whose generation conditions differ from those of the k-th simulated positive sample, and N represent the number of samples in the current training batch. The temperature coefficient is used to control the sharpness of the sample distribution, and sim(·,·) represents the similarity calculation function.

Images