Ovarian cancer peritoneal metastatic focus detection method

Abstract

The invention relates to an ovarian cancer peritoneal metastatic focus detection method, belongs to the technical field of focus detection, and solves the problem that the ovarian cancer peritoneal metastatic focus cannot be accurately detected in the prior art. The method comprises the following steps: acquiring a CT image of a patient with focus metastasis and a focus 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 simulation positive sample based on the positive sample and the negative sample, and expanding the initial sample set to obtain an expanded sample set; training a deep learning network model for focus segmentation based on the expanded sample set to obtain a focus segmentation model; and inputting a CT image of a patient to be detected into the focus segmentation model to obtain a focus metastasis detection result. And the metastatic focus detection can be quickly and accurately obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of lesion detection, and in particular, to a method for detecting peritoneal metastasis lesions of ovarian cancer. Background Art

[0002] Ovarian cancer is one of the most invasive and highly lethal malignant tumors in the female reproductive system. Advanced patients usually show extensive peritoneal metastasis, and common metastatic sites include peritoneum, greater omentum, liver, spleen, mesentery, lymph nodes, etc. Abdominopelvic CT images are important tools for clinical evaluation and diagnosis of peritoneal dissemination of ovarian cancer, and are of great significance for implementing tumor burden assessment and making decisions on surgery or neoadjuvant chemotherapy. However, the current application of computer-aided diagnosis (CAD) and deep learning models in ovarian cancer image analysis still faces many challenges: First, it is extremely difficult to label ovarian cancer lesions: Since advanced ovarian cancer lesions are spread throughout the abdominal cavity, with complex shapes, blurred boundaries, and different sizes, the workload of manually labeling lesions one by one is extremely large, and there may be differences in labeling criteria among different doctors, resulting in poor consistency. Second, the features of ovarian cancer lesions are diverse: Metastatic lesions show different manifestations in different anatomical locations, and it is difficult for existing methods to uniformly detect lesions throughout the abdominal cavity. Third, data acquisition is limited: The sharing of medical image data is restricted, especially abdominopelvic CT image data with detailed lesion labels, which are usually only available to a few research institutions, making it difficult to train deep learning models.

[0003] In order to overcome the above challenges, in recent years, methods such as weakly supervised learning and self-supervised learning have gradually been introduced into the field of medical image analysis, but these methods can only learn the global features of images and cannot distinguish between lesion regions and normal tissues. In addition, most data augmentation methods are limited to simple transformations such as rotation, flipping, and contrast adjustment, and cannot create new lesion data. And radiomics-based methods are limited by the ability of manually designed features and also highly rely on manual delineation. Currently, the image quantitative research on ovarian cancer still highly depends on manually labeled data, which is not only time-consuming and laborious, but also has inconsistent labeling standards and poor repeatability of research.

[0004] A prerequisite for accurate lesion segmentation is high-quality annotation data. Large-scale and high-quality lesion label data, including primary lesions and metastatic lesions, are required to ensure the generalization ability of the model to adapt to the complexity of peritoneal dissemination lesions. Currently, lesion detection methods based on CNN or Transformer mainly rely on end-to-end supervised learning, but these methods highly depend on large-scale labeled data and are difficult to solve problems such as high labeling cost, insufficient data, and poor model generalization. Summary of the Invention

[0005] In view of the above analysis, the embodiments of the present invention aim to provide a method for detecting peritoneal metastatic lesions of ovarian cancer to solve the problem that existing methods cannot accurately detect peritoneal metastatic lesions of ovarian cancer.

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

[0007] Obtain the CT images of patients with lesion metastasis and lesion segmentation labels to construct an initial sample set;

[0008] Obtain the CT images of patients without lesion metastasis. Using the samples in the initial sample set as positive samples and the samples of the CT images of patients without lesion metastasis as negative samples, generate simulated positive samples based on the positive and negative samples, expand the initial sample set, and obtain an expanded sample set;

[0009] Train a deep learning network model for lesion segmentation based on the expanded sample set to obtain a lesion segmentation model;

[0010] Input the CT image of the patient to be detected into the lesion segmentation model to obtain a lesion metastasis detection result.

[0011] Based on a further improvement of the above method, the following method is used to generate simulated positive samples:

[0012] Perform abdominal and pelvic organ segmentation on each positive sample and each negative sample to obtain the sub-organ masks of each positive sample and each negative sample;

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

[0014] Based on the positive sample, perform anatomical alignment on each paired sample of the positive sample, and use the aligned paired sample and the positive sample as the positive and negative sample pairs in the second sample set; construct the second sample set;

[0015] Train a generative adversarial network based on the second sample set to obtain a simulated sample generation model;

[0016] Based on the simulated sample generation model, obtain the simulated positive samples corresponding to each negative sample in the second sample set.

[0017] Based on a further improvement of the above method, for each positive sample, screening similar negative samples based on the sub-organ mask as the paired samples of the positive sample includes:

[0018] Obtain the similarity between the positive sample and each negative sample based on the similarity between each sub-organ mask of the positive sample and the sub-organ mask of each negative sample;

[0019] Screen negative samples with a similarity greater than the first threshold as the paired samples of 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] where represents the CT image of the i-th positive sample, represents the CT image of the j-th negative sample, K represents the number of sub-organ masks, represents the k-th sub-organ mask of the i-th positive sample, represents the k-th sub-organ mask of the j-th negative sample, represents the similarity between the i-th positive sample and the j-th negative sample, and |·| represents the number of voxels in the mask.

[0023] Based on the further improvement of the above method, for each paired sample of the positive sample, anatomical alignment is performed based on the positive sample, and the aligned paired sample and the positive sample are used as the positive and negative sample pairs in the second sample set, including:

[0024] Performing histogram matching on the CT image of each paired sample with the CT image of the positive sample as the target image to obtain the first transformed CT image of each paired sample;

[0025] Taking the minimum overlap error between sub-organ masks as the optimization objective, performing rigid registration 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] Taking the maximum similarity between sub-organs as the optimization objective, performing non-rigid registration on the CT image of the positive sample and the second transformed CT image of each paired sample to obtain the aligned paired sample.

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

[0028]

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

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

[0031]

[0032] Among them, p(a, b) represents the joint probability distribution where the gray value in image A is a and the gray value in image B is b, P(a) represents the probability that the gray value in image A is a, P(b) represents the probability that the gray value in image B is b, and K represents the number of sub-organ masks. represents the k-th sub-organ mask of the positive sample. represents the CT image of the paired sample after the second transformation. represents the k-th sub-organ mask of the paired sample after rigid registration transformation, T B-spline represents the non-rigid registration transformation matrix, and β represents the weight parameter. represents the 2-norm of the matrix.

[0033] Based on the 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, using the lesion morphology and location information of the positive sample as the generation condition corresponding to this positive and negative sample pair, train the conditional generative adversarial network based on the second sample set to obtain a simulation sample generation model.

[0035] Based on the further improvement of the above method, the following formula is used to calculate the training loss of the conditional generative adversarial network:

[0036] L cgan = L G + L D + λL C

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

[0038] Based on the further improvement of the above method, the following formula is used to calculate the contrast loss:

[0039]

[0040] Among them, represents the CT image of the k-th simulated positive sample. represents the CT image of the p-th simulated positive sample with the same generation condition as the CT image of the k-th simulated positive sample, n p represents the number of simulated positive samples with the same generation condition as the CT image of the k-th simulated positive sample, n n represents the number of simulated positive samples with different generation conditions from the CT image of the k-th simulated positive sample. Denote the q-th simulated real sample with different generation conditions from the CT image of the k-th simulated real sample, N denote the number of samples in the current training batch, τ denote the temperature coefficient used to control the sharpness of the sample distribution, and sim(·,·) denote the similarity calculation function.

[0041] Compared with the prior art, the present invention constructs an initial sample set by obtaining CT images and lesion segmentation labels of patients with lesion metastasis, and obtains CT images of patients without lesion metastasis; uses the samples in the initial sample set as positive samples and the samples of CT images of patients without lesion metastasis as negative samples, generates simulated real samples based on the positive and negative samples, performs image generation, generates new lesion data, thereby expanding the positive samples, reducing the annotation burden, solving the problems of difficult annotation and insufficient data of abdominal cavity metastasis lesions in CT images of advanced ovarian cancer, and training a deep learning network model based on the expanded sample set, thereby improving the generalization ability of the trained model, making the detection results 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.

[0042] In the present invention, the above technical solutions can also be combined with each other to achieve more preferred combination schemes. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can be made obvious from the description, or understood by implementing the present invention. The objectives and other advantages of the present invention can be realized and obtained from the content specifically pointed out in the description and the drawings. Description of the Drawings

[0043] The drawings are only for the purpose of showing specific embodiments and are not considered as limiting the present invention. Throughout the drawings, the same reference signs denote the same components;

[0044] Figure 1 It is a flowchart of the method for detecting abdominal cavity metastasis lesions of ovarian cancer in an embodiment of the present invention. Detailed Embodiments

[0045] The following will specifically describe the preferred embodiments of the present invention with reference to the drawings. The drawings form a part of this application and are used together with the embodiments of the present invention to explain the principles of the present invention, rather than to limit the scope of the present invention.

[0046] A specific embodiment of the present invention discloses a method for detecting abdominal cavity metastasis lesions of ovarian cancer, as Figure 1 shown, including the following steps:

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

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

[0049] S3. Train a deep learning network model for lesion segmentation based on the expanded sample set to obtain a lesion segmentation model;

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

[0051] Compared with the prior art, the ovarian cancer peritoneal metastasis lesion detection method provided in this embodiment constructs an initial sample set by obtaining the CT images and lesion segmentation labels of patients with lesion metastasis, and obtains the CT images of patients without lesion metastasis; uses the samples in the initial sample set as positive samples, and the samples of the CT images of patients without lesion metastasis as negative samples. Generate simulated positive samples based on the positive and negative samples for image generation to generate new lesion data, thereby expanding the positive samples and reducing the annotation burden, solving the problems of difficult annotation and insufficient data of peritoneal metastasis 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 result of the model more accurate. For the patient to be detected, inputting its CT image into the lesion segmentation model can quickly and accurately obtain the lesion metastasis detection result.

[0052] During implementation, the lesion segmentation label is the lesion mask of the CT image.

[0053] During implementation, obtain the abdominal and pelvic CT images of patients with lesion metastasis. A senior gynecological specialist radiologist can manually annotate the ovarian cancer lesions to obtain their lesion segmentation labels, that is, lesion masks. The annotation content includes: metastatic sites such as peritoneum, greater omentum, liver, spleen, mesentery, lymph nodes, etc. The abdominal and pelvic CT images of patients with lesion metastasis and the corresponding lesion labels are used as a sample in the initial sample set to construct the initial sample set.

[0054] The number of samples in the initial sample set is small, so simulated samples need to be generated.

[0055] Obtain the abdominal and pelvic CT images of patients without lesion metastasis. Since the samples in the initial sample set are the generation targets, the samples in the initial sample set are used as positive samples, and the abdominal and pelvic CT images of patients without lesion metastasis are used as negative samples. Generate simulated positive samples 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 sample and each negative sample to obtain the sub-organ masks of each positive sample and each negative sample;

[0058] S22. For each positive sample, screen for similar negative samples based on the sub-organ mask as the paired samples of this positive sample;

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

[0060] S24. Train a generative adversarial network based on the second sample set to obtain a simulation sample generation model;

[0061] S25. Obtain the simulated positive samples corresponding to each negative sample in the second sample set based on the simulation sample generation model.

[0062] During implementation, in order to ensure that the generated simulation samples conform to the real anatomical environment, the present invention introduces an anatomical alignment strategy, so that the generated simulated positive samples have similar anatomical structure features to the original positive samples.

[0063] Specifically, first perform abdominal and pelvic major organ segmentation on each positive and negative sample, such as the liver, spleen, intestine, bladder, etc., to obtain each sub-organ mask for anatomical region alignment.

[0064] During implementation, initially screen for the lesion-free images and the images with lesions on the original image that have the most similar anatomical background, so as to screen out the paired samples of each positive sample.

[0065] Specifically, for each positive sample, screening for similar negative samples based on the sub-organ mask as the paired samples of this positive sample includes:

[0066] Obtain the similarity between this positive sample and each negative sample based on the similarity between each sub-organ mask of this positive sample and the sub-organ masks of each negative sample;

[0067] Screen for negative samples with a similarity greater than the first threshold as the paired samples of this positive sample.

[0068] During implementation, based on the sub-organ masks of the abdominal and pelvic regions, obtain the volume, position, and spatial layout information of each sub-organ, calculate the organ volume difference and relative spatial position difference between the two images, and obtain the similarity.

[0069] Specifically, use the following formula to calculate the similarity between this positive sample and each negative sample:

[0070]

[0071] Where, denotes the CT image of the \(i\)-th positive sample, denotes the CT image of the \(j\)-th negative sample, and \(K\) denotes the number of sub-organ masks. denotes the \(k\)-th sub-organ mask of the \(i\)-th positive sample, denotes the \(k\)-th sub-organ mask of the \(j\)-th negative sample, denotes the similarity between the \(i\)-th positive sample and the \(j\)-th negative sample, and \(|\cdot|\) denotes the number of voxels in the mask.

[0072] For the \(i\)-th positive sample, calculate its similarity with each negative sample, select the \(n\) negative samples with higher similarity as the paired samples of this positive sample, and then perform anatomical alignment on each paired sample of this positive sample based on this positive sample. The aligned paired samples and this positive sample are used as the positive and negative sample pairs in the second sample set, thereby constructing the second sample set.

[0073] Specifically, performing anatomical alignment on each paired sample of this positive sample based on this positive sample, and using the aligned paired samples and this positive sample as the positive and negative sample pairs in the second sample set includes:

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

[0075] Taking the minimum overlap error between sub-organ masks as the optimization objective, perform rigid registration on the CT image of this positive sample and the first transformed CT image of each paired sample to obtain the second transformed CT image of each paired sample;

[0076] Taking the maximum similarity between sub-organs as the optimization objective, perform non-rigid registration on the CT image of this positive sample and the second transformed CT image of each paired sample to obtain the aligned paired samples.

[0077] During implementation, first, using the CT image of this positive sample as the target image, and adopting histogram matching, map the features such as contrast, gray level range, and texture of the CT image of each of its paired samples into the distribution space of this positive sample.

[0078] It should be noted that the sub-organ masks of the first transformed CT image of the paired sample are the same as those of the original CT image of the paired sample.

[0079] During implementation, the HistMatch histogram matching function can be used to make the statistical distribution of as close as possible to that of the positive sample to obtain the first transformed CT image of each paired sample, denoted as

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

[0081] During implementation, the optimization goal of rigid registration is to minimize the overlap error of the masks.

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

[0083]

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

[0085] During implementation, the Affine affine transformation method is used for rigid matching. Taking the of the positive sample as a reference, the of the paired sample is preliminarily aligned into the anatomical space of the CT image of the positive sample using the rigid registration method (Affine affine transformation) to obtain the second transformed CT image of each paired sample, labeled as

[0086] After obtaining the transformation matrix of rigid registration, multiply the of the paired sample by the rigid registration matrix, to obtain the second transformed CT image of the paired sample Similarly, each sub-organ mask of the paired sample is correspondingly transformed according to the transformation matrix of rigid registration, that is, multiplied by the rigid registration matrix, to obtain the sub-organ mask after rigid registration transformation

[0087] During implementation, with the maximum similarity between sub-organs as the optimization goal, taking the of the positive sample as a reference, the B-spline method is used to perform non-rigid registration on the second transformed CT image and to further finely adjust the local result differences of the images and obtain the aligned CT image Adapting to be highly matched in the anatomical space structure, thus supporting subsequent simulated lesions to be truly superimposed on reasonable anatomical positions.

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

[0089]

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

[0091] During implementation, overfitting is prevented by adding a regularization term By calculating the similarity between each sub-organ, the local differences of the CT images are made smaller, and the aligned paired samples and the positive castrated version are highly matched in the anatomical space structure, thus supporting that the subsequent simulated true lesions can be truly superimposed on the reasonable anatomical positions.

[0092] After obtaining the non-rigid registration transformation matrix, multiply the of the paired sample by the non-rigid registration matrix, to obtain the third transformed CT image of the paired sample i.e., the aligned CT image to obtain the aligned paired sample.

[0093] The positive sample and the aligned paired sample form a positive and negative sample pair in the second sample set. Among them, the CT image of the positive sample is The CT image of the negative sample is

[0094] During implementation, in order to make the generated simulated true samples more realistic, the generative adversarial network adopts a conditional generative adversarial network.

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

[0096] For each positive and negative sample pair in the second sample set, using the morphological and positional information of the lesion of the positive sample as the generation condition corresponding to the positive and negative sample pair, the conditional generative adversarial network is trained based on the second sample set to obtain a simulation sample generation model.

[0097] During implementation, the conditional generative adversarial network includes: a generator and a discriminator;

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

[0099] The discriminator is used to distinguish the authenticity of the positive samples and the simulated real samples.

[0100] During implementation, the processing process of the generator includes condition fusion, an encoder, and a decoder. First, the input is the CT image of the negative sample in the positive and negative sample pairs and the generation condition c. In the way of feature splicing fusion, the two are jointly input into the network: the generation condition is converted into a numerical vector, and then the numerical vector of the generation condition is mapped to a high-dimensional feature space using a multi-layer perceptron (MLP) to obtain the conditional feature representation; then the conditional feature representation and the CT image are spliced in the channel dimension and jointly input into the subsequent encoder module.

[0101] The encoder uses a convolutional neural network to gradually extract the spatial features of the fused input data: First, through a series of convolutional, batch normalization, ReLU layer or LReLU layer, the shallow structure information in the image is extracted; Second, through gradual downsampling, deeper semantic information of the image is obtained; At each encoder stage, the feature map is retained for subsequent skip connections.

[0102] The decoder gradually upsamples and reconstructs the deep feature representation obtained by the encoder through a transposed convolutional network. First, the decoder takes the deep feature representation of the encoder as the input, gradually upsamples and restores the spatial resolution; Second, the classic skip connection in the U-Net structure is adopted. After each transposed convolution process of the encoder, the extracted feature map and the feature map of the convolutional layer corresponding to it in a mirror image relationship are used for skip connection, and the corresponding feature maps are directly superimposed and then passed to the next layer. Finally, through a 1×1 convolutional layer, the multi-channel feature map is converted into a single-channel image to generate the CT image of the simulated real 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, represents the CT image of the negative sample of the kth positive and negative sample pair in the current training batch, c represents the generation condition, and N represents the number of samples in the current training batch.

[0106] The input of the discriminator includes two parts. One is the CT image of the positive sample in the positive and negative sample pairs of the second sample set and the generation condition, forming a real image pair The other is the CT image of the simulated real sample generated by the generator and the generation condition, forming a simulated image pair

[0107] After receiving the input, the discriminator extracts features from the real image pairs and simulated image pairs through a convolutional neural network, and the output is 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. represents the CT image of the positive sample of the k-th positive and negative sample pair in the current training batch. represents the CT image of the negative sample of the k-th positive and negative sample pair in the current training batch, c represents the generation condition, and N represents the number of samples in the current training batch.

[0110] The discriminator outputs a probability close to 1 for the real image pair and outputs a probability close to 0 for the simulated image pair and outputs a probability close to 0.

[0111] The training process adopts the classic alternating optimization strategy of generative adversarial networks. First, the generator G is fixed, and the discriminator D is trained to enable the discriminator to accurately distinguish real images and simulated images. Then, the discriminator D is fixed, and the generator G is trained to make the simulated images generated by the generator as realistic as possible to deceive the discriminator. The two are alternately iterated, and when the training ends, the generator can stably output realistic simulated lesion images.

[0112] Based on the second sample set, a simulation sample generation model is obtained by training a generative adversarial network. The CT images of the negative samples in the second sample set and the corresponding generation conditions are input into the simulation sample generation model to generate simulated positive samples. Through the difference between the positive samples and the simulated positive samples, binaryzation processing is performed using the Otsu threshold method to extract the lesion region ROI, and the lesion mask of the simulated positive sample, that is, the lesion segmentation label, is generated.

[0113] During implementation, a graphics processor can be used for image generation.

[0114] The generated simulated positive samples are added to the initial sample set to expand the initial sample set. Based on the expanded sample set, a deep learning network model is trained to obtain a lesion segmentation model, thereby improving the generalization ability of the model.

[0115] Since the lesions of advanced ovarian cancer have diverse morphologies and mostly metastasize to multiple parts of the whole abdomen, it is necessary to pay attention to both local fine features and the ability to capture global context information. The deep learning network model of the present invention adopts the SwinUNETR structure, which combines the advantages of the Swin Transformer and the U-Net architecture. The training loss function is the Dice loss function, and the lesion segmentation result is output.

[0116] First, input the three-dimensional CT image (512×512×N) of the sample. Using the self-attention mechanism of Swin Transformer, the input data is segmented into multiple local windows (cubic windows with a size of 4×4×4 pixels). Swin Transformer contains two components. One is the window-based multi-head self-attention module (Windows Multi-Head Self Attention, W-MSA), which rearranges the input features in a non-overlapping manner to generate patches, and then performs self-attention calculations within each patch. The other is the shifted window-based multi-head self-attention mechanism module (Shifted Windows Multi-head Self-Attention, SW-MSA) module, which moves the features in the horizontal and vertical directions and performs self-attention calculations on non-overlapping patches. Through a multi-layer stacked structure (4 Swin Transformer blocks), deep global semantic features are gradually extracted, and finally a global high-dimensional feature representation is output.

[0117] The U-Net decoder is used to accept the global high-dimensional feature representation output by the Transformer and perform step-by-step feature upsampling operations to gradually restore it to 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. Through gradual feature fusion and layer-by-layer restoration of the spatial resolution, the decoder can accurately capture the local fine boundaries and morphological features of the lesions, obtain high-precision lesion segmentation results, and achieve accurate image semantic segmentation.

[0118] The model output is a binary segmentation mask with the same size as the CT image. 1 indicates that the pixel belongs to the lesion area, and 0 indicates the non-lesion area. During the training process, the Dice loss function is used to optimize the network parameters. The specific formula for calculating the loss of a sample is as follows:

[0119]

[0120] where y s represents the true value of the s-th voxel of the lesion segmentation mask of the sample, represents the predicted value of the s-th voxel of 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 the denominator being 0.

[0121] When the preset loss accuracy or the number of iterations is reached, the training is completed, and a lesion segmentation model is obtained. Input the CT image of the patient to be detected into the lesion segmentation model, and the lesion metastasis detection result can be quickly and accurately obtained.

[0122] Those skilled in the art can understand that all or part of the processes for implementing the methods of the above embodiments can be completed by instructing relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. Among them, the computer-readable storage medium is a disk, an optical disc, a read-only memory or a random access memory, etc.

[0123] As described above, the above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

Claims

1. A method for detecting peritoneal metastasis lesions of ovarian cancer, characterized in that, It includes the following steps: Obtain the CT images of patients with lesion metastasis and lesion segmentation labels to construct an initial sample set; Obtain the CT images of patients without lesion metastasis. Use the samples in the initial sample set as positive samples, and the samples of the CT images of patients without lesion metastasis as negative samples. Generate simulated positive samples based on the positive samples and negative samples, and expand the initial sample set to obtain an expanded sample set; Train a deep learning network model for lesion segmentation based on the expanded sample set to obtain a lesion segmentation model; Input the CT image of the patient to be detected into the lesion segmentation model to obtain a lesion metastasis detection result.

2. The method for detecting peritoneal metastasis lesions of ovarian cancer according to claim 1, wherein Generate simulated positive samples in the following way: Perform abdominal and pelvic organ segmentation on each positive sample and each negative sample to obtain the sub-organ masks of each positive sample and each negative sample; For each positive sample, screen similar negative samples based on the sub-organ masks as the paired samples of this positive sample; Perform anatomical alignment on each paired sample of this positive sample based on this positive sample, and use the aligned paired sample and this positive sample as the positive and negative sample pairs in the second sample set; construct the second sample set; Train a generative adversarial network based on the second sample set to obtain a simulated sample generation model; Obtain the simulated positive samples corresponding to each negative sample in the second sample set based on the simulated sample generation model.

3. The method for detecting peritoneal metastatic lesions of ovarian cancer according to claim 2, wherein For each positive sample, screening similar negative samples based on the sub-organ masks as the paired samples of this positive sample includes: Obtain the similarity between this positive sample and each negative sample based on the similarity between each sub-organ mask of this positive sample and the sub-organ mask of each negative sample; Screen the negative samples with a similarity greater than the first threshold as the paired samples of this positive sample.

4. The method for detecting peritoneal metastatic lesions of ovarian cancer according to claim 3, wherein Use the following formula to calculate the similarity between this positive sample and each negative sample: Among them, represents the CT image of the i-th positive sample, represents the CT image of the j-th negative sample, and K represents the number of sub-organ masks, represents the k-th sub-organ mask of the i-th positive sample, represents the k-th sub-organ mask of the j-th negative sample, represents the similarity between the i-th positive sample and the j-th negative sample, and |·| represents the number of voxels in the mask.

5. The method for detecting peritoneal metastasis lesions of ovarian cancer according to claim 2, wherein Performing anatomical alignment on each paired sample of this positive sample based on this positive sample, and using the aligned paired sample and this positive sample as the positive and negative sample pairs in the second sample set includes: Perform histogram matching on the CT image of each paired sample with the CT image of this positive sample as the target image to obtain the first transformed CT image of each paired sample; Taking the minimum overlap error between sub-organ masks as the optimization target, perform rigid registration on the CT image of this positive sample and the first transformed CT image of each paired sample to obtain the second transformed CT image of each paired sample; Taking the maximum similarity between sub-organs as the optimization target, perform non-rigid registration on the CT image of this positive sample and the second transformed CT image of each paired sample to obtain the aligned paired sample.

6. The method for detecting peritoneal metastasis lesions of ovarian cancer according to claim 5, wherein The objective function of the rigid registration is: Among them, T affine represents the transformation matrix for rigid registration, K represents the number of sub-organ masks, represents the k-th sub-organ mask of the positive sample, represents the k-th sub-organ mask of the paired sample.

7. The method for detecting peritoneal metastasis lesions of ovarian cancer according to claim 5, wherein The objective function of the non-rigid registration is: Among them, p(a, b) represents the joint probability distribution of the gray value a in image A and the gray value b in image B, P(a) represents the probability of the gray value a in image A, P(b) represents the probability of the gray value b in image B, and K represents the number of sub-organ masks. represents the k-th sub-organ mask of the positive sample. represents the second converted CT image of the paired sample. represents the k-th sub-organ mask after the rigid registration transformation of the paired sample, T B-spline represents the non-rigid registration transformation matrix, and β represents the weight parameter. represents the 2-norm of the matrix.

8. The method for detecting peritoneal metastatic lesions of ovarian cancer according to claim 2, wherein The generative adversarial network is a conditional generative adversarial network; For each positive and negative sample pair in the second sample set, use the lesion morphology and location information of the positive sample as the generation condition corresponding to this positive and negative sample pair, and train the conditional generative adversarial network based on the second sample set to obtain a simulated sample generation model.

9. The method for detecting peritoneal metastatic lesions of ovarian cancer according to claim 8, wherein, Use the following formula to calculate the training loss of the conditional generative adversarial network: L cgan = L G + L D + λL C Among them, L G represents the generator loss, and L D represents the discriminator loss, and L C represents the contrastive loss, and λ represents the weight coefficient.

10. The method for detecting peritoneal metastatic lesions of ovarian cancer according to claim 9, characterized in that, Use the following formula to calculate the contrast loss: Among them, represents the CT image of the k-th simulated real sample, represents the CT image of the p-th simulated real sample with the same generation conditions as the CT image of the k-th simulated real sample, n p represents the number of simulated real samples with the same generation conditions as the CT image of the k-th simulated real sample, n n represents the number of simulated real samples with different generation conditions from the CT image of the k-th simulated real sample, represents the q-th simulated real sample with different generation conditions from the CT image of the k-th simulated real sample, 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, and sim(·,·) represents the similarity calculation function.

Images