Gastric cancer metastasis risk assessment method and device

Abstract

The invention provides a gastric cancer metastasis risk assessment method and device which can be applied to the technical field of artificial intelligence and medical image analysis. The method comprises the following steps: acquiring a first initial point, a second initial point and a third initial point in an image of a target object; according to a region segmentation iterative algorithm and a preset distance threshold value, determining a gastric cancer focus target region, a peritoneal middle region and an organ middle region respectively corresponding to the first initial point, the second initial point and the third initial point from the image; processing the gastric cancer focus target region, the peritoneal middle region and the visceral organ middle region according to a region recognition algorithm to obtain a peritoneal target region; voxel blocks corresponding to the gastric cancer focus target area and the peritoneum target area are processed according to the prediction model, and a gastric cancer metastasis risk assessment result is obtained.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of artificial intelligence and medical image analysis, in particular to a gastric cancer metastasis risk assessment method and device. BACKGROUND

[0002] Early gastric cancer often has no obvious symptoms, resulting in many patients being diagnosed with advanced gastric cancer. Peritoneal metastasis is a common symptom of advanced gastric cancer, and about half of patients with distal metastasis of gastric cancer will develop peritoneal metastasis. Peritoneal metastasis is considered incurable, and the median survival time of patients is usually only 3 to 6 months. Therefore, timely and accurate assessment of the risk of peritoneal metastasis of gastric cancer is important for determining the treatment method and improving the prognosis of patients.

[0003] In the field of medical image analysis, existing imageomics methods require manual segmentation of the boundaries of gastric cancer lesions or peritoneum and extraction of artificial defined features, which has high requirements for the experience and time of operators, high cost, poor stability, and low-dimensional semantic information reduces the accuracy of model evaluation. SUMMARY

[0004] In view of the above problems, the present disclosure provides a gastric cancer metastasis risk assessment method and device.

[0005] According to a first aspect of the present disclosure, a gastric cancer metastasis risk assessment method is provided, which is executed by an electronic device and includes:

[0006] Obtaining a first initial point, a second initial point and a third initial point in an image image of a target object, wherein the first initial point represents a position marker point in an initial region of a gastric cancer lesion in the image image, the second initial point represents a position marker point in an initial region of peritoneum in the image image, and the third initial point represents a position marker point in an initial region of an organ in the image image; determining a gastric cancer lesion target region, a peritoneum intermediate region and an organ intermediate region corresponding to the first initial point, the second initial point and the third initial point, respectively, from the image image according to a region segmentation iterative algorithm and a preset distance threshold; processing the gastric cancer lesion target region, the peritoneum intermediate region and the organ intermediate region according to a region recognition algorithm to obtain a peritoneum target region; and processing voxel blocks corresponding to the gastric cancer lesion target region and the peritoneum target region, respectively, according to a prediction model to obtain a gastric cancer metastasis risk assessment result.

[0007] According to an embodiment of the present disclosure, the gastric cancer lesion target region, the peritoneum intermediate region and the organ intermediate region corresponding to the first initial point, the second initial point and the third initial point, respectively, are determined from the image image according to the region segmentation iterative algorithm and the preset distance threshold, including:

[0008] Based on the region segmentation iterative algorithm and a preset distance threshold, a first candidate point set corresponding to a first initial point is determined from the pixels in the image, wherein the first candidate point set includes the first initial point and at least one first pixel associated with the first initial point; based on the region segmentation iterative algorithm and the preset distance threshold, a second candidate point set corresponding to a second initial point is determined from the pixels in the image, wherein the first candidate point set includes the second initial point and at least one second pixel associated with the second initial point; based on the region segmentation iterative algorithm and the preset distance threshold, a third candidate point set corresponding to a third initial point is determined from the pixels in the image, wherein the first candidate point set includes the third initial point and at least one third pixel associated with the third initial point; based on the first candidate point set, the second candidate point set, and the third candidate point set, the target region of the gastric cancer lesion, the intermediate peritoneum region, and the intermediate organ region corresponding to the first candidate point set, the second candidate point set, and the third candidate point set, respectively, are determined.

[0009] According to embodiments of this disclosure, a first candidate point set is determined from pixels in an image based on a region segmentation iterative algorithm and a preset distance threshold, including:

[0010] Multiple rounds of filtering operations are performed on the pixels in the image to determine the first candidate point set. The i-th round of filtering operations includes:

[0011] The process involves determining the i-th unselected pixel from the unselected pixels in an image, where the distance between the i-th unselected pixel and at least one pixel in the (i-1)-th first candidate point set is equal to a preset distance threshold. The (i-1)-th first candidate point set represents the first candidate point set determined in the (i-1)-th round of selection, where i is a positive integer. Size extraction is performed on the pixels in the (i-1)-th first candidate point set and the i-th unselected pixel, respectively, to obtain at least one (i-1)-th candidate voxel block and one i-th unselected voxel block. A deep learning discriminator is used to process at least one (i-1)-th candidate voxel block and one i-th unselected voxel block to obtain the association result between the i-th unselected voxel block and each (i-1)-th candidate voxel block. If the number of successfully associated results exceeds a preset threshold, the i-th unselected pixel is added to the (i-1)-th first candidate point set, resulting in the i-th first candidate point set.

[0012] According to embodiments of this disclosure, the deep learning discriminator includes a feature extraction coding layer and a multilayer perceptron layer; wherein, by processing at least one (i-1)th candidate voxel block and an i-th voxel block to be screened by the deep learning discriminator, the association result between the i-th voxel block to be screened and each (i-1)th candidate voxel block is obtained, including:

[0013] The (i-1)th candidate voxel block and the i-th voxel block to be screened are respectively input into the feature extraction coding layer to obtain the i-th initial candidate feature and the i-th initial feature to be screened; the i-th initial candidate feature and the i-th initial feature to be screened are respectively input into the multilayer perceptron layer to obtain the (i-1)th target candidate feature and the i-th target feature to be screened; when the similarity between the (i-1)th target candidate feature and the i-th target feature to be screened is greater than the preset similarity threshold, the association result of the representation association between the i-th voxel block to be screened and the (i-1)th candidate voxel block is obtained.

[0014] According to embodiments of this disclosure, the training method for a deep learning discriminator includes the following operations:

[0015] The first training sample is obtained, which includes the initial regions of the gastric cancer lesion, peritoneum, and organs after image segmentation. A predetermined number of pixels are selected from each of these regions for size extraction, resulting in sample pixel blocks and labels. The labels represent the region type corresponding to the pixel blocks. The pixel blocks are input to a feature extraction encoding layer to obtain initial pixel features. These initial pixel features are then input to a multilayer perceptron layer to obtain target pixel features. Based on the labels, the positive and negative pixel features corresponding to the target pixel features are determined. The similarity between the target pixel features and positive pixel features, and between the target pixel features and negative pixel features, is processed using an unsupervised learning loss function to obtain loss values. Based on these loss values, a deep learning discriminator is trained to obtain the trained deep learning discriminator.

[0016] According to embodiments of this disclosure, a peritoneal target region is obtained by processing the gastric cancer lesion target region, the intermediate peritoneal region, and the intermediate organ region using a region recognition algorithm, including:

[0017] The overlapping regions in the intermediate peritoneum are trimmed to obtain the remaining trimmed peritoneum region. The overlapping region represents the overlap between the intermediate peritoneum region and the target gastric cancer lesion region, or the overlap between the intermediate peritoneum region and the intermediate region of an organ. If the volume of a connected region in the trimmed peritoneum region is greater than or equal to a first preset volume threshold, the connected region is defined as a peritoneal connected region. A connected region represents a set of adjacent pixels with the same pixel value in the trimmed peritoneum region. If the volume of the peritoneal connected region is greater than or equal to a second preset volume threshold, the peritoneal connected region is contracted according to the target gastric cancer lesion region to obtain the target peritoneum region. If the volume of the peritoneal connected region is less than the second preset volume threshold, the peritoneal connected region is defined as the target peritoneum region.

[0018] According to embodiments of this disclosure, the method for assessing the risk of gastric cancer metastasis further includes:

[0019] Acquire the initial image of the target object; process the initial image using an image resampling algorithm to obtain an enhanced image; process the enhanced image using a normalization algorithm to obtain the final image.

[0020] According to embodiments of this disclosure, the prediction model includes a classification sub-model. Based on the prediction model, voxel blocks corresponding to the target regions of gastric cancer lesions and peritoneal target regions are processed to obtain gastric cancer metastasis risk assessment results, including:

[0021] Feature extraction is performed on the voxel blocks corresponding to the target regions of gastric cancer lesions and peritoneum, respectively, to obtain the features of the gastric cancer lesion region and the peritoneum region. The gastric cancer lesion region features and peritoneum region features are then processed according to a classification sub-model to obtain the gastric cancer metastasis risk assessment results. The training method for the classification sub-model includes:

[0022] A second training sample is obtained, which includes gastric cancer lesion region features, peritoneal region features, and classification labels. The classification labels represent the gastric cancer metastasis results in the sample images. Feature filtering is performed on the gastric cancer lesion region features and peritoneal region features to obtain a target feature candidate set. Based on a stepwise search strategy, the candidate features in the target feature candidate set are stepwise input into the classification sub-model, and the gastric cancer metastasis risk assessment result of the sample is output. Based on the gastric cancer metastasis risk assessment result and classification labels, the classification sub-model is trained to obtain the trained classification sub-model.

[0023] According to embodiments of this disclosure, feature screening is performed on the regional features of gastric cancer lesions and the regional features of peritoneum in the sample to obtain a candidate set of target features, including:

[0024] Based on the correlation between the gastric cancer lesion region features and peritoneal region features of the sample and the classification label, multiple first candidate features and multiple second candidate features are determined. Among them, the correlation between the first candidate features and the classification label is greater than a preset correlation threshold, and the correlation between the second candidate features and the classification label is less than or equal to the preset correlation threshold. Based on the multiple first candidate features, an initial feature candidate set is determined. The multiple second candidate features are then filtered to determine a third candidate feature. Among them, the correlation coefficient between the third candidate feature and the first candidate feature is less than a preset coefficient threshold. Based on the third candidate feature and the initial feature candidate set, a target feature candidate set is determined.

[0025] A second aspect of this disclosure provides a device for assessing the risk of gastric cancer metastasis, comprising:

[0026] The acquisition module is used to acquire a first initial point, a second initial point, and a third initial point in the image of the target object. The first initial point represents a location marker point in the initial region of the gastric cancer lesion in the image, the second initial point represents a location marker point in the initial region of the peritoneum in the image, and the third initial point represents a location marker point in the initial region of the organ in the image.

[0027] The determination module is used to determine the target regions of gastric cancer lesions, the intermediate peritoneal region, and the intermediate organ region corresponding to the first initial point, the second initial point, and the third initial point, respectively, from the image based on the region segmentation iterative algorithm and the preset distance threshold.

[0028] The generation module is used to process the target region of gastric cancer lesions, the intermediate region of peritoneum, and the intermediate region of organs according to the region recognition algorithm to obtain the target region of peritoneum.

[0029] The assessment module is used to process voxel blocks corresponding to the target areas of gastric cancer lesions and peritoneal target areas according to the prediction model, and obtain the gastric cancer metastasis risk assessment results.

[0030] According to the gastric cancer metastasis risk assessment method and apparatus provided in this disclosure, the gastric cancer lesion target region, peritoneal intermediate region, and organ intermediate region corresponding to the first initial point, second initial point, and third initial point respectively are determined from the image by using a region segmentation iterative algorithm and a preset distance threshold. The gastric cancer lesion target region, peritoneal intermediate region, and organ intermediate region are processed according to a region recognition algorithm to obtain the peritoneal target region. The voxel blocks corresponding to the gastric cancer lesion target region and the peritoneal target region are processed according to a prediction model to obtain the gastric cancer metastasis risk assessment result. Since the gastric cancer lesion target region and the peritoneal target region are quickly segmented based on some points such as the first initial point, second initial point, and third initial point, the accuracy and stability of the segmentation are improved, and the reliance on human experience and human intervention is significantly reduced while ensuring segmentation precision. Furthermore, the prediction model automatically analyzes the voxel blocks corresponding to the gastric cancer lesion target region and the peritoneal target region, thereby identifying and extracting more complex features highly correlated with the risk of peritoneal metastasis of gastric cancer, and more accurately assessing the risk of peritoneal metastasis in preoperative gastric cancer patients. Attached Figure Description

[0031] The above and other objects, features and advantages of this disclosure will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0032] Figure 1 A flowchart of a gastric cancer metastasis risk assessment method according to an embodiment of the present disclosure is shown;

[0033] Figure 2A flowchart illustrating the generation of a target region for a gastric cancer lesion, a mid-peritoneal region, and a mid-organ region according to an embodiment of the present disclosure is shown.

[0034] Figure 3 A flowchart illustrating the generation of a peritoneal target region according to an embodiment of the present disclosure is shown;

[0035] Figure 4 A structural block diagram of a gastric cancer metastasis risk assessment device according to an embodiment of the present disclosure is shown. Detailed Implementation

[0036] The embodiments according to this disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this disclosure for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concepts of this disclosure.

[0037] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. The terms “comprising,” “including,” etc., as used herein indicate the presence of features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0038] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0039] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0040] In the process of realizing this disclosure, it was found that in the field of medical image analysis, existing radiomics methods require manual segmentation of the boundaries of gastric cancer lesions or peritoneum and extraction of manually defined features, which requires high experience and time from the operator, is costly, has poor stability, and low-dimensional semantic information reduces the accuracy of model evaluation.

[0041] In view of the above, an embodiment of this disclosure provides a method and apparatus for assessing the risk of gastric cancer metastasis. The method includes: acquiring a first initial point, a second initial point, and a third initial point in an image of a target object; determining, based on a region segmentation iterative algorithm and a preset distance threshold, a gastric cancer lesion target region, a peritoneal intermediate region, and an organ intermediate region corresponding to each of the first, second, and third initial points in the image; processing the gastric cancer lesion target region, the peritoneal intermediate region, and the organ intermediate region using a region recognition algorithm to obtain a peritoneal target region; and processing the voxel blocks corresponding to the gastric cancer lesion target region and the peritoneal target region using a prediction model to obtain a gastric cancer metastasis risk assessment result.

[0042] In the technical solution disclosed herein, the user information (including but not limited to user personal information, user image information, user device information, such as location information) and data (including but not limited to data used for analysis, stored data, and displayed data) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with relevant laws, regulations, and standards, necessary confidentiality measures have been taken, and they do not violate public order and good morals. Corresponding operation entry points are provided for users to choose to authorize or refuse.

[0043] It should be noted that the sequence numbers of the operations in the following methods are for descriptive purposes only and should not be considered as indicating the execution order of the operations. Unless explicitly stated otherwise, the method does not need to be executed in the exact order shown.

[0044] Figure 1 A flowchart of a gastric cancer metastasis risk assessment method according to an embodiment of the present disclosure is shown.

[0045] like Figure 1 As shown, the method 100 includes operations S110 to S140.

[0046] In operation S110, the first initial point, the second initial point, and the third initial point in the image of the target object are acquired.

[0047] According to embodiments of this disclosure, the target object is characterized as a gastric cancer patient. An image of the target object is acquired using an acquisition device. The image can be a computed tomography (CT) scan. The acquisition device can be a CT scanner. The conventional parameters can be set as follows: tube voltage 120V, tube current 300mA, slice thickness 1.25mm, and field of view 512mm*512mm.

[0048] According to embodiments of this disclosure, one or more locations are marked in the three-dimensional image of the gastric cancer lesion, peritoneum, and surrounding organs such as the liver, spleen, pancreas, kidney, rectum, and colon, serving as initial points for segmenting the corresponding regions. The first initial point represents a location marker within the initial region of the gastric cancer lesion in the image, the second initial point represents a location marker within the initial region of the peritoneum in the image, and the third initial point represents a location marker within the initial region of the organs in the image.

[0049] In operation S120, based on the region segmentation iterative algorithm and the preset distance threshold, the target regions of gastric cancer lesions, the intermediate peritoneal region, and the intermediate organ region corresponding to the first initial point, the second initial point, and the third initial point are determined from the image.

[0050] According to embodiments of this disclosure, a first initial point is used as the initial iteration point. Pixels in the image that satisfy a preset distance threshold with respect to the initial iteration point are selected. The initial iteration point and the pixels satisfying the preset distance threshold are used as the next iteration point. Selection continues according to the region segmentation iteration algorithm and the preset distance threshold until no pixels in the image satisfy the preset distance threshold are found, at which point the iteration stops. In each iteration, multiple target pixels are selected from the initial iteration point and the pixels satisfying the preset distance threshold based on the physiological characteristics of the pixels. A circular region centered on the target pixel and with a radius equal to the preset distance threshold is determined. Finally, the circular regions corresponding to all target pixels are merged to determine the target region of the gastric cancer lesion.

[0051] According to an embodiment of this disclosure, taking the above-described determination of the gastric cancer lesion target region corresponding to the first initial point as an example, the peritoneal intermediate region corresponding to the second initial point is determined from the image image based on the region segmentation iterative algorithm and the preset distance threshold; the organ intermediate region corresponding to the third initial point is determined from the image image based on the region segmentation iterative algorithm and the preset distance threshold.

[0052] In operation S130, the target area of ​​the gastric cancer lesion, the intermediate area of ​​the peritoneum, and the intermediate area of ​​the organs are processed according to the region recognition algorithm to obtain the target area of ​​the peritoneum.

[0053] According to embodiments of this disclosure, the peritoneal target region characterizes an area highly associated with the risk of peritoneal metastasis from gastric cancer. The peritoneal target region is obtained based on the segmentation of the central peritoneum region adjacent to the target region of the gastric cancer lesion.

[0054] According to the embodiments of this disclosure, the regions in the middle peritoneum that overlap with the target region of the gastric cancer lesion and the middle region of the organ are first excluded, and then the volume of the remaining region after exclusion is detected. Based on a preset threshold and the distance between the pixels in the region and the target region of the gastric cancer lesion, the target peritoneum region is determined.

[0055] In operation S140, the voxel blocks corresponding to the target areas of gastric cancer lesions and peritoneum are processed according to the prediction model to obtain the gastric cancer metastasis risk assessment results.

[0056] According to embodiments of this disclosure, the prediction model can be constructed based on a machine learning network, and the voxel block of the gastric cancer lesion target region can be extracted based on pixels in the gastric cancer lesion target region with a size of L. The voxel block corresponding to the peritoneal target region can be extracted based on pixels in the peritoneal target region with a size of L.

[0057] According to embodiments of this disclosure, voxel blocks corresponding to the target regions of gastric cancer lesions and peritoneum are input into a prediction model, and the model outputs a gastric cancer metastasis risk assessment result. The gastric cancer metastasis risk assessment result characterizes the probability of peritoneal metastasis from gastric cancer.

[0058] According to embodiments of this disclosure, the gastric cancer metastasis risk assessment result serves as intermediate result information for determining health status, and all the above steps are performed by electronic devices.

[0059] According to embodiments of this disclosure, by using partial initial points such as the first initial point, the second initial point, and the third initial point as initial iteration points, the target region of gastric cancer lesions and the target region of peritoneum are quickly segmented, improving the accuracy and stability of segmentation. While ensuring segmentation precision, the reliance on human experience and human intervention is significantly reduced. Furthermore, based on the prediction model, the voxel blocks corresponding to the target regions of gastric cancer lesions and peritoneum are automatically analyzed, thereby identifying and extracting more complex features that are highly correlated with the risk of peritoneal metastasis of gastric cancer, and more accurately assessing the risk of peritoneal metastasis in preoperative gastric cancer patients.

[0060] Figure 2 A flowchart illustrating the generation of the target region of a gastric cancer lesion, the intermediate region of the peritoneum, and the intermediate region of the organs according to embodiments of the present disclosure is shown.

[0061] like Figure 2 As shown, based on the region segmentation iterative algorithm and the preset distance threshold, the target regions of gastric cancer lesions, the intermediate peritoneal region, and the intermediate organ region corresponding to the first initial point, the second initial point, and the third initial point are determined from the image, including operations S210 to S240.

[0062] In operation S210, based on the region segmentation iterative algorithm and the preset distance threshold, a first candidate point set corresponding to the first initial point is determined from the pixels in the image.

[0063] According to embodiments of this disclosure, the first candidate point set includes a first initial point and at least one first pixel point associated with the first initial point.

[0064] According to embodiments of this disclosure, the region segmentation iterative algorithm can be a region growing segmentation algorithm. Based on the region segmentation iterative algorithm and a preset distance threshold, multiple rounds of filtering operations are performed on the pixels in the image, with a first initial point as the initial iteration point, to obtain at least one first pixel associated with the first initial point. A first candidate point set is obtained based on the first initial point and the at least one first pixel associated with the first initial point.

[0065] For example, a preset distance threshold d is used. The first initial point is used as the initial iteration point. The first pixel point that satisfies the preset distance threshold d with the initial iteration point is selected from the image. The first iteration point and the first pixel point that satisfies the preset distance threshold are used as the next iteration point. The selection continues according to the region segmentation iteration algorithm and the preset distance threshold until there are no first pixels in the image that satisfy the preset distance threshold. The iteration stops. The first candidate point set is formed based on the first initial point and the selected first pixels.

[0066] In operation S220, based on the region segmentation iterative algorithm and the preset distance threshold, a set of second candidate points associated with the second initial point is determined from the pixels in the image.

[0067] According to embodiments of this disclosure, the second candidate point set includes a second initial point and at least one second pixel point associated with the second initial point.

[0068] According to embodiments of this disclosure, based on a region segmentation iterative algorithm and a preset distance threshold, multiple rounds of filtering operations are performed on pixels in the image image using a second initial point as the initial iteration point, until no second pixel in the image image meets the preset distance threshold, at which point the iteration stops. At least one second pixel associated with the second initial point is obtained, and a second candidate point set is obtained based on the second initial point and the at least one second pixel associated with the second initial point.

[0069] In operation S230, based on the region segmentation iterative algorithm and the preset distance threshold, a set of third candidate points corresponding to the third initial point is determined from the pixels in the image.

[0070] According to embodiments of this disclosure, the third candidate point set includes a third initial point and at least one third pixel point associated with the third initial point.

[0071] According to embodiments of this disclosure, based on a region segmentation iterative algorithm and a preset distance threshold, multiple rounds of filtering operations are performed on pixels in the image image using a third initial point as the initial iteration point, until no third pixel in the image image satisfies the preset distance threshold, at which point the iteration stops. At least one third pixel associated with the third initial point is obtained, and a third candidate point set is obtained based on the third initial point and the at least one third pixel associated with the third initial point.

[0072] In operation S240, based on the first candidate point set, the second candidate point set, and the third candidate point set, the target region of the gastric cancer lesion, the intermediate region of the peritoneum, and the intermediate region of the organ corresponding to each of the first candidate point set, the second candidate point set, and the third candidate point set are determined.

[0073] According to embodiments of this disclosure, a gastric cancer lesion target region corresponding to the first candidate point set is determined based on a preset distance threshold and a first candidate point set; a peritoneal intermediate region corresponding to the second candidate point set is determined based on a preset distance threshold and a second candidate point set; and an organ intermediate region corresponding to the third candidate point set is determined based on a preset distance threshold and a third candidate point set.

[0074] According to embodiments of this disclosure, determining the target region of a gastric cancer lesion corresponding to the first candidate point set based on a preset distance threshold and the first candidate point set may include: determining circular regions for the first initial point and the first pixel point respectively, with the preset distance threshold as the radius and the first initial point and the first pixel point in the first candidate point set as the center, and merging all circular regions to obtain the target region of the gastric cancer lesion. The process of determining the intermediate peritoneal region and the intermediate organ region is similar to the process of determining the target region of the gastric cancer lesion described above, and will not be repeated here.

[0075] For example, taking the determination of the target region of a gastric cancer lesion using the first candidate point set as an example, the first candidate point set includes the first initial point. First pixel First pixel If the preset distance threshold is d, then the first initial point is determined. The circular region S1, the first pixel The circular region S2, the first pixel The circular region S3 is used to determine the target area of ​​the gastric cancer lesion based on the circular regions S1, S2, and S3. d can be 3 mm.

[0076] According to embodiments of this disclosure, a first candidate point set is determined from pixels in an image based on a region segmentation iterative algorithm and a preset distance threshold, including:

[0077] Multiple rounds of filtering operations are performed on the pixels in the image to determine the first candidate point set. The i-th round of filtering operations includes:

[0078] The process involves determining the i-th unselected pixel from the unselected pixels in an image, where the distance between the i-th unselected pixel and at least one pixel in the (i-1)-th first candidate point set is equal to a preset distance threshold. The (i-1)-th first candidate point set represents the first candidate point set determined in the (i-1)-th round of selection, where i is a positive integer. Size extraction is performed on the pixels in the (i-1)-th first candidate point set and the i-th unselected pixel, respectively, to obtain at least one (i-1)-th candidate voxel block and one i-th unselected voxel block. A deep learning discriminator is used to process at least one (i-1)-th candidate voxel block and one i-th unselected voxel block to obtain the association result between the i-th unselected voxel block and each (i-1)-th candidate voxel block. If the number of successfully associated results exceeds a preset threshold, the i-th unselected pixel is added to the (i-1)-th first candidate point set, resulting in the i-th first candidate point set.

[0079] According to embodiments of this disclosure, one pixel is selected in each round of filtering. The i-th pixel to be filtered represents the pixel to be filtered in the i-th round of filtering. In the i-th round of filtering, at least one pixel is determined from the pixels that were not filtered in the previous i-1 rounds of filtering, and the distance between it and at least one pixel in the (i-1)-th first candidate point set is equal to a preset distance threshold. Then, an i-th pixel to be filtered is randomly selected from the at least one pixel that is equal to the preset distance threshold.

[0080] For example, the (i-1)th first candidate point set includes pixels. Pixel In the image, determine the pixels that were not filtered in the first i-1 rounds of filtering. Or pixels At least one pixel whose distance to the other is equal to a preset distance threshold d is selected, and then a random pixel to be selected is determined from the at least one pixel whose distance to the other is equal to the preset distance threshold d.

[0081] According to embodiments of this disclosure, the pixels in the (i-1)th first candidate point set include a first initial point and the first pixels selected in the first (i-1)th rounds of filtering operations. In the image, pixels in the (i-1)th first candidate point set are extracted with a size of L centered on each pixel to obtain the (i-1)th candidate voxel block corresponding to each pixel in the (i-1)th first candidate point set; in the image, the i-th pixel to be filtered is extracted with a size of L centered on the i-th pixel to be filtered to obtain the i-th voxel block to be filtered corresponding to the i-th pixel to be filtered. For example, the size L can be 32 mm.

[0082] According to embodiments of this disclosure, the deep learning discriminator can be constructed from a convolutional neural network. The deep learning discriminator processes each (i-1)th candidate voxel block and the ith voxel block to be screened, respectively, to obtain the association results between the ith voxel block to be screened and each (i-1)th candidate voxel block.

[0083] According to embodiments of this disclosure, the association result characterizes the similarity of pixel features; the larger the association result, the greater the likelihood that the pixels belong to the same physiological structure. When the association result is greater than an association threshold, an association result indicating successful association is obtained.

[0084] According to embodiments of this disclosure, the preset quantity threshold can be 50% of the total number of associated results.

[0085] For example, the first candidate point set of the (i-1)th mentioned above includes pixels. Pixel For example, pixels. The corresponding (i-1)th candidate voxel block Pixel The corresponding (i-1)th candidate voxel block The (i-1)th candidate voxel block is processed according to the deep learning discriminator. The association result D1 is obtained by comparing the i-th candidate voxel block with the i-th candidate voxel block; the (i-1)-th candidate voxel block is then processed using a deep learning discriminator. The association result D2 is obtained by comparing the i-th voxel block to be screened.

[0086] According to embodiments of this disclosure, there are two association results, D1 and D2. Therefore, the preset quantity threshold can be 1. Each association result is compared with the association threshold. If the number of association results indicating successful association is greater than the preset quantity threshold, the i-th pixel to be screened is added to the (i-1)-th first candidate point set to obtain the i-th first candidate point set.

[0087] According to embodiments of this disclosure, the deep learning discriminator includes a feature extraction coding layer and a multilayer perceptron layer; wherein, by processing at least one (i-1)th candidate voxel block and an i-th voxel block to be screened by the deep learning discriminator, the association result between the i-th voxel block to be screened and each (i-1)th candidate voxel block is obtained, including:

[0088] The (i-1)th candidate voxel block and the i-th voxel block to be screened are respectively input into the feature extraction coding layer to obtain the i-th initial candidate feature and the i-th initial feature to be screened; the i-th initial candidate feature and the i-th initial feature to be screened are respectively input into the multilayer perceptron layer to obtain the (i-1)th target candidate feature and the i-th target feature to be screened; when the similarity between the (i-1)th target candidate feature and the i-th target feature to be screened is greater than the preset similarity threshold, the association result of the representation association between the i-th voxel block to be screened and the (i-1)th candidate voxel block is obtained.

[0089] According to embodiments of this disclosure, the feature extraction coding layer can be constructed based on the PyTorch deep learning framework. The feature extraction coding layer can be a three-dimensional convolutional neural network with a global average pooling layer at the end for outputting features. The three-dimensional convolutional neural network can be constructed based on a deep residual network (3D ResNet34), with an output feature dimension of 1024, and weights initialized using a depth initialization algorithm (Xavier).

[0090] According to embodiments of this disclosure, the multilayer perceptron layer maps features from an initial dimension to a higher-dimensional space through a linear layer, with an output dimension of 2048. A batch normalization layer and a linear rectification function (ReLU) are added, and finally, another linear layer reduces the dimensionality of the features to a lower dimension, with an output dimension of 512.

[0091] According to embodiments of this disclosure, the (i-1)th candidate voxel block to be compared is... and the i-th voxel block to be screened The inputs are fed into the feature extraction coding layer to obtain the i-th initial candidate features. and the i-th initial feature to be screened .

[0092] According to embodiments of this disclosure, the i-th initial candidate feature is processed according to the multilayer perceptron layer. and the i-th initial feature to be screened The i-1th target candidate features are obtained. and the i-th target features to be screened .

[0093] According to embodiments of this disclosure, the preset similarity threshold can be 0.5. In the (i-1)th target candidate feature... and the i-th target features to be screened If the similarity between the two voxels is greater than the preset similarity threshold, the association result of the characterization association between the i-th voxel block to be screened and the (i-1)-th candidate voxel block is obtained.

[0094] In one embodiment, similarity As shown in formula (1):

[0095] (1);

[0096] in, Characterize the features of the (i-1)th target candidate. Characterize the features to be selected for the i-th target. Characterization The Euclidean length, Characterization The Euclidean length.

[0097] According to embodiments of this disclosure, the training method for a deep learning discriminator includes the following operations:

[0098] The first training sample is obtained, which includes the initial regions of the gastric cancer lesion, peritoneum, and organs after image segmentation. A predetermined number of pixels are selected from each of these regions for size extraction, resulting in sample pixel blocks and labels. The labels represent the region type corresponding to the pixel blocks. The pixel blocks are input to a feature extraction encoding layer to obtain initial pixel features. These initial pixel features are then input to a multilayer perceptron layer to obtain target pixel features. Based on the labels, the positive and negative pixel features corresponding to the target pixel features are determined. The similarity between the target pixel features and positive pixel features, and between the target pixel features and negative pixel features, is processed using an unsupervised learning loss function to obtain loss values. Based on these loss values, a deep learning discriminator is trained to obtain the trained deep learning discriminator.

[0099] According to an embodiment of this disclosure, the first training sample is derived from sample image images of 200 gastric cancer patients. The first training sample is preprocessed, and then 20 preprocessed first training samples are randomly selected. The regions of the gastric cancer lesion, peritoneum, and surrounding organs such as liver, spleen, pancreas, kidney, rectum, and colon are initially segmented to obtain the initial region of the gastric cancer lesion, the initial region of the peritoneum, and the initial region of the organs.

[0100] According to embodiments of this disclosure, five pixels can be selected from each of the initial regions of the gastric cancer lesion, the peritoneum, and the organ. These pixels are then extracted from the image, centered on the selected pixels and with a size of 32 mm, to obtain sample pixel blocks for each pixel. For example, pixel 1 is selected, and its label is 1 to indicate that it originates from the initial region of the gastric cancer lesion.

[0101] According to embodiments of this disclosure, sample pixel blocks are input to a feature extraction coding layer to obtain initial pixel features of the sample; the initial pixel features of the sample are input to a multilayer perceptron layer to obtain target pixel features of the sample.

[0102] According to embodiments of this disclosure, positive and negative sample pairs are determined for each sample target pixel feature based on the tag corresponding to each pixel. A positive sample pair includes a sample target pixel feature and positive sample pixel features, where the sample target pixel feature and positive sample pixel features originate from the same region, and the sample target pixel feature and negative sample pixel features originate from different regions. The sample pixel feature corresponding to the sample target pixel feature includes one positive sample pixel feature and the remaining negative sample pixel features.

[0103] In one embodiment, the loss value As shown in formula (2):

[0104] (2);

[0106] in, Characterizes the target pixel features of the sample. Characterizes the positive sample pixel features corresponding to the target pixel features of the sample. Characterizes the sample pixel features corresponding to the target pixel features of the sample. It is a parameter characterizing positive temperature, and can be 0.5. Characterization The number of samples.

[0107] According to embodiments of this disclosure, a deep learning discriminator is trained iteratively based on the loss value, and an optimizer (adaptive moment estimation, Adam) is used to adjust the parameters of the feature extraction coding layer and the multilayer perceptron layer until the loss value meets the training termination condition.

[0108] Figure 3 A flowchart illustrating the generation of a peritoneal target region according to an embodiment of this disclosure is shown.

[0109] like Figure 3 As shown, the target region of gastric cancer lesion, the intermediate region of peritoneum, and the intermediate region of organs are processed according to the region recognition algorithm to obtain the target region of peritoneum, including operations S310 to S340.

[0110] In operation S310, the overlapping area in the middle region of the peritoneum is trimmed to obtain the remaining trimmed peritoneum area.

[0111] According to embodiments of this disclosure, the overlapping region represents the overlapping area between the intermediate peritoneum and the target area of ​​the gastric cancer lesion, or the overlapping area between the intermediate peritoneum and the intermediate area of ​​an organ.

[0112] According to embodiments of this disclosure, the remaining peritoneal cut area after cutting represents the area after excluding the overlapping area from the middle peritoneal region.

[0113] In operation S320, if the volume of the connected region in the peritoneal cutting area is greater than or equal to a first preset volume threshold, the connected region is determined as a peritoneal connected region.

[0114] According to embodiments of this disclosure, a connected region represents a set of pixels with the same pixel value in a peritoneal cutting region.

[0115] According to embodiments of this disclosure, the volume of a connected region is determined based on the number of pixels and the pixel size within the connected region.

[0116] According to embodiments of this disclosure, the first preset volume threshold can be 10. 3 mm 3， The volume of the connected regions in the peritoneal trimming area is detected. If the volume of the connected regions in the peritoneal trimming area is less than a first preset volume threshold, the connected regions are excluded. If the volume of the connected regions in the peritoneal trimming area is greater than or equal to the first preset volume threshold, the connected regions are identified as peritoneal connected regions.

[0117] According to embodiments of this disclosure, if there are multiple connected regions with a volume greater than or equal to a first preset volume threshold in the peritoneal cutting area, only the connected region closest to the gastric cancer lesion target area is retained as the peritoneal connecting region.

[0118] In operation S330, when the volume of the peritoneal connecting region is greater than or equal to the second preset volume threshold, the peritoneal connecting region is contracted according to the target region of the gastric cancer lesion to obtain the target peritoneal region.

[0119] According to embodiments of this disclosure, the volume of the peritoneal communication region is detected, and the second preset volume threshold can be 10. 6 mm 3 .

[0120] According to embodiments of this disclosure, when the volume of the peritoneal connecting region is greater than or equal to a second preset volume threshold, shrinking the peritoneal connecting region to obtain the peritoneal target region based on the gastric cancer lesion target region may include: first calculating the distance from each pixel in the peritoneal connecting region to the gastric cancer lesion target region, then sequentially including each pixel in the peritoneal connecting region into the candidate region in ascending order of distance, until the volume of the candidate region is equal to the second preset volume threshold, stopping the shrinkage, and using the shrunken region as the peritoneal target region.

[0121] In operation S340, if the volume of the peritoneal connection region is less than the second preset volume threshold, the peritoneal connection region is determined as the peritoneal target region.

[0122] According to embodiments of this disclosure, when the volume of the peritoneal connecting region is less than a second preset volume threshold, the peritoneal connecting region is directly determined as the peritoneal target region.

[0123] According to embodiments of this disclosure, the method for assessing the risk of gastric cancer metastasis further includes:

[0124] Acquire the initial image of the target object; process the initial image using an image resampling algorithm to obtain an enhanced image; process the enhanced image using a normalization algorithm to obtain the final image.

[0125] According to an embodiment of this disclosure, the target is a gastric cancer patient, and an initial image of the target is acquired using an acquisition device.

[0126] According to an embodiment of this disclosure, the initial image is a three-dimensional image. The initial image is processed according to an image resampling algorithm to obtain an enhanced image, in which the pixel size is uniform.

[0127] For example, a bicubic interpolation algorithm can be used to adjust the pixel size of the initial image in the x, y, and z axes to a fixed pixel value. The fixed pixel value can be 1 mm pixels.

[0128] According to embodiments of this disclosure, processing an enhanced image image using a normalization algorithm to obtain an image image may include: truncating the pixel intensity in the enhanced image image according to the window level and window width of the abdominal window, and normalizing the pixel intensity to a specified range using a normalization algorithm to obtain the image image. For example, the specified range may be (-1, 1).

[0129] According to embodiments of this disclosure, the initial image is enhanced and normalized to obtain an image with uniform pixel size, which enhances the contrast of each region and accelerates the convergence of the model.

[0130] According to embodiments of this disclosure, the prediction model includes a classification sub-model. Based on the prediction model, voxel blocks corresponding to the target regions of gastric cancer lesions and peritoneal target regions are processed to obtain gastric cancer metastasis risk assessment results, including:

[0131] Feature extraction was performed on the voxel blocks corresponding to the target regions of gastric cancer lesions and peritoneum to obtain the features of the gastric cancer lesion region and the peritoneum region.

[0132] According to embodiments of this disclosure, the pixel values ​​of non-gastric cancer lesion target regions and non-peritoneal target regions in the image can be set to zero. Based on the feature extraction coding layer, feature extraction is performed on the voxel blocks corresponding to the gastric cancer lesion target region and the peritoneal target region respectively to obtain gastric cancer lesion region features and peritoneal region features.

[0133] According to embodiments of this disclosure, gastric cancer lesion regional features characterize the depth features within the target region of the gastric cancer lesion, and peritoneal regional features characterize the depth features within the target region of the peritoneum.

[0134] According to embodiments of this disclosure, gastric cancer lesion region features and peritoneal region features are processed using a classification sub-model to obtain gastric cancer metastasis risk assessment results. The training method for the classification sub-model includes:

[0135] A second training sample is obtained, which includes gastric cancer lesion region features, peritoneal region features, and classification labels. The classification labels represent the gastric cancer metastasis results in the sample images. Feature filtering is performed on the gastric cancer lesion region features and peritoneal region features to obtain a target feature candidate set. Based on a stepwise search strategy, the candidate features in the target feature candidate set are stepwise input into the classification sub-model, and the gastric cancer metastasis risk assessment result of the sample is output. Based on the gastric cancer metastasis risk assessment result and classification labels, the classification sub-model is trained to obtain the trained classification sub-model.

[0136] According to embodiments of this disclosure, the classification sub-model can be constructed based on a machine learning model. The machine learning model can be derived from logistic regression, Naive Bayes classification, support vector machine, k-nearest neighbor, decision tree, random forest, etc.

[0137] According to embodiments of this disclosure, the second training sample includes sample image images, and the gastric cancer lesion region features and peritoneal region features of the sample are obtained from the sample image images.

[0138] According to embodiments of this disclosure, feature selection is performed on the gastric cancer lesion region features and peritoneal region features of the sample based on classification labels. A target feature candidate set is obtained based on the selected features, and the candidate features in the target feature candidate set have a high degree of correlation with the classification labels.

[0139] According to embodiments of this disclosure, candidate features in the target feature candidate set are sorted from smallest to largest based on their correlation, and then, based on an upward stepwise search strategy, the candidate features in the target feature candidate set are progressively input into the classification sub-model. After each candidate feature is input into the classification sub-model, a gastric cancer metastasis risk assessment result for the sample is generated.

[0140] According to embodiments of this disclosure, the average accuracy of the classification sub-model is obtained based on the gastric cancer metastasis risk assessment results and classification labels of the samples. The classification sub-model is then iteratively trained until the average accuracy of the classification sub-model meets a preset accuracy threshold.

[0141] According to embodiments of this disclosure, feature screening is performed on the regional features of gastric cancer lesions and the regional features of peritoneum in the sample to obtain a candidate set of target features, including:

[0142] Based on the correlation between the gastric cancer lesion region features and peritoneal region features of the sample and the classification label, multiple first candidate features and multiple second candidate features are determined. Among them, the correlation between the first candidate features and the classification label is greater than a preset correlation threshold, and the correlation between the second candidate features and the classification label is less than or equal to the preset correlation threshold. Based on the multiple first candidate features, an initial feature candidate set is determined. The multiple second candidate features are then filtered to determine a third candidate feature. Among them, the correlation coefficient between the third candidate feature and the first candidate feature is less than a preset coefficient threshold. Based on the third candidate feature and the initial feature candidate set, a target feature candidate set is determined.

[0143] According to embodiments of this disclosure, the correlation between the gastric cancer lesion region features and the classification label of the sample is tested using the Mann-Whitney U statistical algorithm, and the correlation between the peritoneal region features and the classification label of the sample is tested. If the correlation is greater than a preset correlation threshold, a first candidate feature is determined; if the correlation is less than or equal to the preset correlation threshold, it is determined as a second candidate feature.

[0144] According to embodiments of this disclosure, the initial feature candidate set includes multiple first candidate features. The feature screening of multiple second candidate features to determine the third candidate feature may include: for a second candidate feature, calculating the correlation between the second candidate feature and multiple first candidate features in the initial feature candidate set to obtain multiple correlation results. When the number of correlation results representing low correlation meets a preset number, the second candidate feature is determined as the third candidate feature.

[0145] According to an embodiment of this disclosure, the preset coefficient threshold can be 0.5. When the Spearman correlation coefficient between the second candidate feature and the first candidate feature is less than the preset coefficient threshold, a correlation result representing low correlation is obtained.

[0146] According to embodiments of this disclosure, feature filtering is performed on multiple second candidate features to determine at least one third candidate feature, and the at least one third candidate feature is included in an initial feature candidate set to obtain a target feature candidate set.

[0147] Based on the above-mentioned method for assessing the risk of gastric cancer metastasis, this disclosure also provides a device for assessing the risk of gastric cancer metastasis. The following will be combined with... Figure 4 The device is described in detail.

[0148] Figure 4 A structural block diagram of a gastric cancer metastasis risk assessment device according to an embodiment of the present disclosure is shown.

[0149] like Figure 4 As shown, the gastric cancer metastasis risk assessment device 400 of this embodiment includes an acquisition module 410, a determination module 420, a generation module 430, and an assessment module 440.

[0150] The acquisition module 410 is used to acquire a first initial point, a second initial point, and a third initial point in the image of the target object. The first initial point represents a location marker within the initial region of a gastric cancer lesion in the image, the second initial point represents a location marker within the initial region of the peritoneum in the image, and the third initial point represents a location marker within the initial region of an organ in the image. In one embodiment, the acquisition module 410 can be used to perform the operation S110 described above, which will not be repeated here.

[0151] The determination module 420 is used to determine, based on a region segmentation iterative algorithm and a preset distance threshold, the target region of the gastric cancer lesion, the intermediate peritoneal region, and the intermediate organ region corresponding to the first initial point, the second initial point, and the third initial point, respectively, from the image. In one embodiment, the determination module 420 can be used to perform the operation S120 described above, which will not be repeated here.

[0152] The generation module 430 is used to process the target region of the gastric cancer lesion, the intermediate region of the peritoneum, and the intermediate region of the organ according to the region recognition algorithm to obtain the target region of the peritoneum. In one embodiment, the generation module 430 can be used to perform the operation S130 described above, which will not be repeated here.

[0153] The evaluation module 440 is used to process the voxel blocks corresponding to the target areas of gastric cancer lesions and peritoneal target areas according to the prediction model to obtain the gastric cancer metastasis risk assessment results. In one embodiment, the evaluation module 440 can be used to perform the operation S140 described above, which will not be repeated here.

[0154] According to embodiments of this disclosure, the determining module 420 includes a first determining submodule, a second determining submodule, a third determining submodule, and a fourth determining submodule.

[0155] The first determining submodule is used to determine a first candidate point set corresponding to a first initial point from the pixels in the image based on a region segmentation iterative algorithm and a preset distance threshold. The first candidate point set includes the first initial point and at least one first pixel associated with the first initial point.

[0156] The second determining submodule is used to determine a second candidate point set associated with the second initial point from the pixels in the image based on the region segmentation iterative algorithm and a preset distance threshold, wherein the first candidate point set includes the second initial point and at least one second pixel associated with the second initial point.

[0157] The third determination submodule is used to determine a set of third candidate points corresponding to the third initial point from the pixels in the image based on the region segmentation iterative algorithm and a preset distance threshold. The set of first candidate points includes the third initial point and at least one third pixel point associated with the third initial point.

[0158] The fourth determination submodule is used to determine the target area of ​​gastric cancer lesions, the intermediate peritoneal region, and the intermediate organ region corresponding to the first candidate point set, the second candidate point set, and the third candidate point set, respectively.

[0159] According to embodiments of this disclosure, the first determining submodule includes a filtering unit.

[0160] The filtering unit is used to perform multiple rounds of filtering operations on the pixels in the image to determine the first candidate point set.

[0161] According to embodiments of this disclosure, the filtering unit includes a first filtering subunit, a second filtering subunit, a third filtering subunit, and a fourth filtering subunit.

[0162] The first filtering subunit is used to determine the i-th pixel to be filtered from the unfiltered pixels in the image, wherein the distance between the i-th pixel to be filtered and at least one pixel in the (i-1)-th first candidate point set is equal to a preset distance threshold, and the (i-1)-th first candidate point set represents the first candidate point set determined in the (i-1)-th round of filtering operation, where i is a positive integer.

[0163] The second filtering subunit is used to extract the size of the pixels in the (i-1)th first candidate point set and the i-th pixel to be filtered, respectively, to obtain at least one (i-1)th candidate voxel block and the i-th voxel block to be filtered.

[0164] The third filtering subunit is used to process at least one (i-1)th candidate voxel block and the i-th voxel block to be filtered according to the deep learning discriminator, and to obtain the association result between the i-th voxel block to be filtered and each (i-1)th candidate voxel block.

[0165] The fourth filtering subunit is used to add the i-th pixel to be filtered into the (i-1)-th first candidate point set when the number of successfully associated results is greater than a preset threshold, thus obtaining the i-th first candidate point set.

[0166] According to embodiments of this disclosure, the third screening subunit is configured to process at least one (i-1)th candidate voxel block and the i-th voxel block to be screened using a deep learning discriminator, to obtain the association results between the i-th voxel block to be screened and each (i-1)th candidate voxel block, including:

[0167] The (i-1)th candidate voxel block and the i-th voxel block to be screened are respectively input into the feature extraction coding layer to obtain the i-th initial candidate feature and the i-th initial feature to be screened; the i-th initial candidate feature and the i-th initial feature to be screened are respectively input into the multilayer perceptron layer to obtain the (i-1)th target candidate feature and the i-th target feature to be screened; when the similarity between the (i-1)th target candidate feature and the i-th target feature to be screened is greater than the preset similarity threshold, the association result of the representation association between the i-th voxel block to be screened and the (i-1)th candidate voxel block is obtained.

[0168] According to embodiments of this disclosure, the generation module 430 includes a first generation submodule, a second generation submodule, a third generation submodule, and a fourth generation submodule.

[0169] The first generation submodule is used to trim the overlapping area in the intermediate peritoneum region to obtain the remaining trimmed peritoneum region. The overlapping area represents the overlapping area between the intermediate peritoneum region and the target area of ​​the gastric cancer lesion or the overlapping area between the intermediate peritoneum region and the intermediate area of ​​the organ.

[0170] The second generation submodule is used to determine the connected region as a peritoneal connected region when the volume of the connected region in the peritoneal cutting region is greater than or equal to a first preset volume threshold. The connected region represents a set of pixels with the same pixel value in the peritoneal cutting region.

[0171] The third generation submodule is used to shrink the peritoneal connecting region according to the gastric cancer lesion target region when the volume of the peritoneal connecting region is greater than or equal to the second preset volume threshold, so as to obtain the peritoneal target region.

[0172] The fourth generation submodule is used to determine the peritoneal connecting region as the peritoneal target region when the volume of the peritoneal connecting region is less than the second preset volume threshold.

[0173] According to embodiments of this disclosure, any plurality of modules, including modules, submodules, units, and subunits, can be combined into one module for implementation, or any one of these modules can be split into multiple modules. Alternatively, at least a portion of the functionality of one or more of these modules can be combined with at least a portion of the functionality of other modules and implemented in one module. According to embodiments of this disclosure, at least one of modules, submodules, units, and subunits can be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or implemented in hardware or firmware by any other reasonable means of integrating or packaging circuitry, or implemented in any one of software, hardware, and firmware methods, or in a suitable combination of any of these. Alternatively, at least one of modules, submodules, units, and subunits can be at least partially implemented as a computer program module, which, when run, can perform corresponding functions.

[0174] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions. Those skilled in the art will understand that the features recited in the various embodiments and / or claims of this disclosure can be combined and / or combined in various ways, even if such combinations or combinations are not expressly described in this disclosure. In particular, the features described in the various embodiments and / or claims of this disclosure may be combined and / or combined in various ways without departing from the spirit and teachings of this disclosure. All such combinations and / or combinations fall within the scope of this disclosure.

[0175] The embodiments according to this disclosure have been described above. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of this disclosure. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. The scope of this disclosure is defined by the appended claims and their equivalents. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of this disclosure, and all such substitutions and modifications should fall within the scope of this disclosure.

Claims

1. A method for assessing the risk of gastric cancer metastasis, characterized in that, The method is performed by an electronic device, and the method includes: Acquire a first initial point, a second initial point, and a third initial point in an image of a target object, wherein the first initial point represents a location marker within the initial region of a gastric cancer lesion in the image, the second initial point represents a location marker within the initial region of the peritoneum in the image, and the third initial point represents a location marker within the initial region of an organ in the image. Based on the region segmentation iterative algorithm and the preset distance threshold, the target regions of gastric cancer lesions, the intermediate peritoneal regions, and the intermediate organ regions corresponding to the first initial point, the second initial point, and the third initial point are determined from the image. The target peritoneal region is obtained by processing the gastric cancer lesion target region, the intermediate peritoneal region, and the intermediate organ region according to the region recognition algorithm. Based on the prediction model, voxel blocks corresponding to the target regions of the gastric cancer lesions and the peritoneum are processed to obtain the gastric cancer metastasis risk assessment results.

2. The method according to claim 1, characterized in that, Based on a region segmentation iterative algorithm and a preset distance threshold, the target regions of gastric cancer lesions, the intermediate peritoneal region, and the intermediate organ region corresponding to the first initial point, the second initial point, and the third initial point, respectively, are determined from the image, including: Based on the region segmentation iterative algorithm and a preset distance threshold, a first candidate point set corresponding to the first initial point is determined from the pixels in the image, wherein the first candidate point set includes the first initial point and at least one first pixel associated with the first initial point; Based on the region segmentation iterative algorithm and the preset distance threshold, a second candidate point set associated with the second initial point is determined from the pixels in the image, wherein the first candidate point set includes the second initial point and at least one second pixel associated with the second initial point; Based on the region segmentation iterative algorithm and the preset distance threshold, a third candidate point set corresponding to the third initial point is determined from the pixels in the image, wherein the first candidate point set includes the third initial point and at least one third pixel associated with the third initial point; Based on the first candidate point set, the second candidate point set, and the third candidate point set, the target region of the gastric cancer lesion, the intermediate peritoneal region, and the intermediate organ region corresponding to the first candidate point set, the second candidate point set, and the third candidate point set, respectively, are determined.

3. The method according to claim 2, characterized in that, Based on a region segmentation iterative algorithm and a preset distance threshold, a first candidate point set is determined from the pixels in the image, including: Multiple rounds of filtering operations are performed on the pixels in the image to determine a first candidate point set, wherein the i-th round of filtering operations includes: The i-th pixel to be filtered is determined from the unfiltered pixels in the image, wherein the distance between the i-th pixel to be filtered and at least one pixel in the (i-1)-th first candidate point set is equal to the preset distance threshold, and the (i-1)-th first candidate point set represents the first candidate point set determined in the (i-1)-th round of filtering operation, where i is a positive integer; Size extraction is performed on the pixels in the (i-1)th first candidate point set and the i-th pixel to be screened, respectively, to obtain at least one (i-1)th candidate voxel block and one i-th voxel block to be screened. The deep learning discriminator processes at least one of the (i-1)th candidate voxel blocks and the i-th voxel block to be screened to obtain the association result between the i-th voxel block to be screened and each of the (i-1)th candidate voxel blocks; If the number of successfully associated results is greater than a preset threshold, the i-th pixel to be screened is added to the (i-1)-th first candidate point set to obtain the i-th first candidate point set.

4. The method according to claim 3, characterized in that, The deep learning discriminator includes a feature extraction coding layer and a multi-layer perceptron layer; Specifically, the process involves processing at least one of the (i-1)th candidate voxel blocks and the i-th voxel block to be screened using a deep learning discriminator to obtain the association results between the i-th voxel block to be screened and each of the (i-1)th candidate voxel blocks, including: The (i-1)th candidate voxel block and the i-th voxel block to be screened are respectively input into the feature extraction coding layer to obtain the i-th initial candidate feature and the i-th initial feature to be screened. The i-th initial candidate feature and the i-th initial feature to be screened are respectively input into the multilayer perceptron layer to obtain the (i-1)-th target candidate feature and the i-th target feature to be screened. If the similarity between the (i-1)th target candidate feature and the i-th target feature to be screened is greater than a preset similarity threshold, the association result of the characterization association between the i-th voxel block to be screened and the (i-1)th candidate voxel block is obtained.

5. The method according to claim 4, characterized in that, The training method for the deep learning discriminator includes the following operations: Obtain the first training sample, which includes the initial region of the gastric cancer lesion, the initial region of the peritoneum, and the initial region of the organ after the segmentation of the sample image. A preset number of pixels are selected from the initial region of the gastric cancer lesion, the initial region of the peritoneum, and the initial region of the organ in the sample for size extraction to obtain sample pixel blocks and labels, wherein the labels represent the region type corresponding to the sample pixel blocks; The sample pixel block is input into the feature extraction coding layer to obtain the initial pixel features of the sample; The initial pixel features of the sample are input into the multilayer perceptron layer to obtain the target pixel features of the sample; Based on the label, determine the positive and negative sample pixel features corresponding to the target pixel features of the sample; The similarity between the sample target pixel features and the positive sample pixel features, and the similarity between the sample target pixel features and the negative sample pixel features are processed according to the unsupervised learning loss function to obtain the loss value; Based on the loss value, the deep learning discriminator is trained to obtain the trained deep learning discriminator.

6. The method according to claim 1, characterized in that, The target peritoneal region is obtained by processing the gastric cancer lesion target region, the intermediate peritoneal region, and the intermediate organ region according to the region recognition algorithm, including: The overlapping area in the intermediate peritoneum is trimmed to obtain the remaining trimmed peritoneum area, wherein the overlapping area represents the overlapping area between the intermediate peritoneum and the target area of ​​the gastric cancer lesion or the overlapping area between the intermediate peritoneum and the intermediate area of ​​the organ. If the volume of a connected region in the peritoneal cutting area is greater than or equal to a first preset volume threshold, the connected region is defined as a peritoneal connected region, wherein the connected region represents a set of adjacent pixels with the same pixel value in the peritoneal cutting area. If the volume of the peritoneal connecting region is greater than or equal to a second preset volume threshold, the peritoneal connecting region is contracted according to the target region of the gastric cancer lesion to obtain the target peritoneal region. If the volume of the peritoneal connection region is less than the second preset volume threshold, the peritoneal connection region is determined as the peritoneal target region.

7. The method according to claim 1, characterized in that, The gastric cancer metastasis risk assessment method also includes: Obtain the initial image of the target object; The initial image is processed according to an image resampling algorithm to obtain an enhanced image; The enhanced image is obtained by processing the image according to a standardization algorithm.

8. The method according to claim 1, characterized in that, The prediction model includes a classification sub-model. Based on the prediction model, voxel blocks corresponding to the target region of the gastric cancer lesion and the target region of the peritoneum are processed to obtain gastric cancer metastasis risk assessment results, including: Feature extraction is performed on the voxel blocks corresponding to the gastric cancer lesion target region and the peritoneum target region, respectively, to obtain the gastric cancer lesion region features and the peritoneum region features; The gastric cancer lesion region features and peritoneal region features are processed according to the classification sub-model to obtain the gastric cancer metastasis risk assessment result. The training method of the classification sub-model includes: Obtain a second training sample, which includes gastric cancer lesion region features and peritoneal region features, and classification labels, wherein the classification labels characterize the gastric cancer metastasis results in the sample image. Feature screening is performed on the gastric cancer lesion region features and the peritoneal region features of the sample to obtain a candidate set of target features; Based on a stepwise search strategy, candidate features from the target feature candidate set are stepwise input into the classification sub-model, and the gastric cancer metastasis risk assessment result of the sample is output. Based on the gastric cancer metastasis risk assessment results of the samples and the classification labels, a classification sub-model is trained to obtain the trained classification sub-model.

9. The method according to claim 8, characterized in that, Feature screening was performed on the gastric cancer lesion region features and the peritoneal region features of the samples to obtain a candidate set of target features, including: Based on the correlation between the gastric cancer lesion region features and the peritoneal region features of the sample and the classification label, multiple first candidate features and multiple second candidate features are determined, wherein the correlation between the first candidate features and the classification label is greater than a preset correlation threshold, and the correlation between the second candidate features and the classification label is less than or equal to the preset correlation threshold. Based on multiple first candidate features, an initial feature candidate set is determined; Feature filtering is performed on multiple second candidate features to determine a third candidate feature, wherein the correlation coefficient between the third candidate feature and the first candidate feature is less than a preset coefficient threshold. The target feature candidate set is determined based on the third candidate feature and the initial feature candidate set.

10. A device for assessing the risk of gastric cancer metastasis, the device comprising: The acquisition module is used to acquire a first initial point, a second initial point, and a third initial point in an image of a target object, wherein the first initial point represents a location marker point within the initial region of a gastric cancer lesion in the image, the second initial point represents a location marker point within the initial region of the peritoneum in the image, and the third initial point represents a location marker point within the initial region of an organ in the image. The determination module is used to determine, based on the region segmentation iterative algorithm and a preset distance threshold, the target region of the gastric cancer lesion, the intermediate region of the peritoneum and the intermediate region of the organ corresponding to the first initial point, the second initial point and the third initial point respectively from the image; The generation module is used to process the target region of the gastric cancer lesion, the intermediate region of the peritoneum, and the intermediate region of the organ according to the region recognition algorithm to obtain the target region of the peritoneum. The assessment module is used to process the voxel blocks corresponding to the gastric cancer lesion target region and the peritoneal target region according to the prediction model to obtain the gastric cancer metastasis risk assessment result.

Images