Prostate mr volume automatic calculation method and system based on generative adversarial network, electronic equipment and storage medium

Abstract

The application provides a prostate MRI volume automatic calculation method and system based on a generative adversarial network, an electronic device and a storage medium, relates to the technical field of generative adversarial networks, and obtains a first type of image reflecting an anatomical structure and a second type of image reflecting a blood vessel distribution, aligns and integrates voxels of the two types of images to form a multi-modal image set. Through a generation unit of an attention generative adversarial network, boundary features of a peripheral zone and a central zone of the prostate in the multi-modal image set are extracted, and an initial image containing a prostate region is generated; then, through an evaluation of a discrimination unit, a probability graph representing whether each voxel belongs to the prostate region is obtained. Through a convolution network, the probability graph is converted into an initial voxel-level segmentation result, morphological processing is performed on the voxel-level segmentation result, a target voxel-level segmentation result from which isolated noise points are removed is obtained, and then the prostate MRI volume is calculated, so that the accuracy and reliability of the prostate MRI volume calculation can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of generative adversarial networks, and in particular to a prostate MRI volume automatic calculation method and system based on a generative adversarial network, an electronic device, and a storage medium. BACKGROUND

[0002] In the clinical diagnosis and efficacy monitoring of prostate diseases, the prostate magnetic resonance imaging (MRI) volume is a key quantitative indicator for evaluating lesion progression and judging treatment response. With the development of multi-modal MRI technology, image data containing information such as anatomical structures and blood vessel distribution are increasingly rich, and there is an urgent need in the clinic for automatic and high-precision calculation of prostate MRI volume based on these multi-modal data, which requires not only efficient integration of multi-source image features to improve segmentation accuracy, but also intelligent algorithms to reduce errors caused by manual intervention and provide reliable data support for early disease screening and disease classification.

[0003] Currently, there is a scheme based on convolutional neural networks (CNN) for automatic calculation of prostate MRI volume. This scheme extracts prostate region features from single-modal MRI images through CNN, performs voxel-level segmentation on the images, and labels the voxels belonging to the prostate region, and then calculates the volume in combination with the physical size of the voxels; wherein the segmentation process mainly relies on the gray scale and texture features of a single modality, and the region division is achieved through multi-layer convolution operations.

[0004] However, the existing scheme has obvious defects: on the one hand, it only relies on single-modal image features, making it difficult to fully integrate multi-dimensional information such as anatomical structures and blood vessel distribution, resulting in blurred boundary recognition of the peripheral zone and central zone of the prostate, and the segmentation result is prone to edge shift; on the other hand, the CNN has limited feature capture capability for small glandular regions, and is prone to misjudgment of isolated noise voxels as prostate regions, resulting in decreased voxel labeling accuracy, and ultimately leading to deviations in the volume calculation result from the actual value, making it difficult to meet the clinical requirements for high-precision quantitative indicators. SUMMARY

[0005] The purpose of the present application is to provide a prostate MRI volume automatic calculation method and system based on a generative adversarial network, an electronic device, and a storage medium, to solve the problem of low accuracy and reliability of prostate MRI volume calculation caused by insufficient single-modal fusion, weak feature capture of small glands, and inaccurate voxel labeling in the prior art.

[0006] To solve the above technical problems, in a first aspect, the present application provides a prostate MRI volume automatic calculation method based on a generative adversarial network, comprising:

[0007] acquire a first type of image reflecting an anatomical structure and a second type of image reflecting a blood vessel distribution of a target object, and align and integrate voxels of the first type of image and the second type of image to form a multi-modal image set;

[0008] extract boundary features of a peripheral zone and a central zone of the prostate in the multi-modal image set by a generation unit of an attention generative adversarial network, generate an initial image containing a prostate region, and evaluate the initial image by a discrimination unit of the attention generative adversarial network to obtain a probability map representing whether each voxel belongs to the prostate region;

[0009] convert the probability map into an initial voxel-level segmentation result by a convolutional network, and perform morphological processing on the initial voxel-level segmentation result to obtain a target voxel-level segmentation result in which isolated noise points are removed;

[0010] calculate a prostate MRI volume based on the target voxel-level segmentation result.

[0011] Optionally, the extracting boundary features of a peripheral zone and a central zone of the prostate in the multi-modal image set by a generation unit of an attention generative adversarial network, generating an initial image containing a prostate region, and evaluating the initial image by a discrimination unit of the attention generative adversarial network to obtain a probability map representing whether each voxel belongs to the prostate region comprises:

[0012] preprocess the multi-modal image set by a generation unit of an attention generative adversarial network, extract boundary features of a peripheral zone and a central zone of the prostate from the preprocessing result, fuse the boundary features with structural features in the first type of image and blood vessel features in the second type of image, and generate an initial image containing a prostate region;

[0013] compare the initial image with gland features in the multi-modal image set by a discrimination unit of the attention generative adversarial network, identify micro glands in the prostate region based on a comparison result, and output difference information between the initial image and the gland features;

[0014] enhance a focusing degree on the boundary features according to the difference information by the generation unit, adjust a fusion ratio of the structural features and the blood vessel features based on the enhanced focusing degree, and generate a target image;

[0015] recompare the target image by the discrimination unit and output new difference information, and repeat a processing procedure of the generation unit and the discrimination unit until a change amplitude of the difference information is less than a preset difference threshold;

[0016] After the repeated processing process ends, a probability distribution of each voxel belonging to the prostate region is obtained based on the target image of the last generation, and a probability map is generated based on the probability distribution.

[0017] Optionally, the focusing degree for the boundary feature is enhanced according to the difference information, the fusion ratio of the structure feature and the blood vessel feature is adjusted based on the enhanced focusing degree, and a target image is generated, including:

[0018] The difference information is analyzed to determine a first associated region corresponding to both the prostate peripheral zone and the central zone boundary, and a second associated region corresponding to the microgland.

[0019] For the first associated region, the proportion of the boundary feature in feature fusion is increased to enhance the focusing degree.

[0020] For the second associated region, the proportion of the blood vessel feature in feature fusion is increased according to the enhanced focusing degree.

[0021] For the irrelevant region in the initial image, the basic proportion of the structure feature in feature fusion is maintained.

[0022] The boundary feature and the structure feature after adjusting the proportion, and the blood vessel feature maintaining the basic proportion are fused to generate a target image.

[0023] Optionally, the probability map is converted into an initial voxel-level segmentation result by the convolutional network, and morphological processing is performed on the initial voxel-level segmentation result to obtain a target voxel-level segmentation result in which isolated noise points are removed, including:

[0024] The probability values of each voxel in the probability map are extracted, and the voxels belonging to the prostate region are divided according to the probability values of each voxel by the convolutional network to form an initial voxel-level segmentation result.

[0025] The voxel groups with continuous positions in the initial voxel-level segmentation result are checked, and the number of voxels contained in each voxel group is determined.

[0026] The voxel groups with a number of voxels less than a preset number threshold are marked as isolated redundant voxel groups.

[0027] The redundant voxel groups are removed from the initial voxel-level segmentation result, the remaining voxel groups are retained and integrated to form a target voxel-level segmentation result.

[0028] Optionally, the probability values of each voxel in the probability map are extracted, and the voxels belonging to the prostate region are divided according to the probability values of each voxel by the convolutional network to form an initial voxel-level segmentation result, including:

[0029] The probability values of the voxels are processed layer by layer by using a convolutional network, and a reference value for judging the attribution of the voxels is set;

[0030] The probability values of the voxels are compared with the reference value, a first label indicating belonging to the prostate region is given to voxels whose probability values are not lower than the reference value, and a second label indicating not belonging to the prostate region is given to voxels whose probability values are lower than the reference value;

[0031] All voxels with the first label and the second label are fused to form an initial voxel-level segmentation result.

[0032] Optionally, the spatial coordinates of the voxels of the first type of image and the second type of image are aligned and integrated to form a multi-modal image set, including:

[0033] The spatial coordinates of the first type of image are determined as reference coordinates, and the spatial positions of the voxels in the second type of image are adjusted according to the reference coordinates, so that the spatial positions match the corresponding positions in the reference coordinates;

[0034] The voxels of the first type of image and the voxels of the second type of image with matched spatial positions are determined as corresponding voxel pairs;

[0035] The image information of the first type of image and the image information of the second type of image in the corresponding voxel pairs are associated;

[0036] All associated corresponding voxel pairs are combined to form a multi-modal image set.

[0037] Optionally, the prostate MRI volume is calculated based on the target voxel-level segmentation result, including:

[0038] The total number of voxels belonging to the prostate region in the target voxel-level segmentation result is counted;

[0039] The actual length, the actual width and the actual height of a single voxel in an actual scanning space are determined, and the volume of a single voxel is calculated by using the actual length, the actual width and the actual height;

[0040] The total number of voxels is multiplied by the volume of a single voxel to obtain the volume of the prostate MRI.

[0041] In a second aspect, the present application provides a prostate MRI volume automatic calculation system based on a generative adversarial network, including:

[0042] The acquisition module is configured to acquire a first type of image reflecting an anatomical structure and a second type of image reflecting a blood vessel distribution of a target object, and align and integrate voxels of the first type of image and the second type of image to form a multi-modal image set.

[0043] The generation module is configured to extract boundary features of a peripheral zone and a central zone of the prostate in the multi-modal image set by a generation unit of an attention generative adversarial network, generate an initial image containing a prostate region, and evaluate the initial image by a discrimination unit of the attention generative adversarial network to obtain a probability map representing whether each voxel belongs to the prostate region.

[0044] The processing module is configured to convert the probability map into an initial voxel-level segmentation result by a convolution network, and perform morphological processing on the initial voxel-level segmentation result to obtain a target voxel-level segmentation result in which isolated noise points are removed.

[0045] The calculation module is configured to calculate a prostate MRI volume based on the target voxel-level segmentation result.

[0046] In a third aspect, the present application provides an electronic device, comprising:

[0047] The memory is configured to store a computer program.

[0048] The processor is configured to implement the steps of the prostate MRI volume automatic calculation method based on the generative adversarial network according to the first aspect when executing the computer program.

[0049] In a fourth aspect, the present application provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program can implement the steps of the prostate MRI volume automatic calculation method based on the generative adversarial network according to the first aspect when executed by a processor.

[0050] In the present application, a prostate MRI volume automatic calculation method based on a generative adversarial network is provided, comprising: acquiring a first type of image reflecting an anatomical structure and a second type of image reflecting a blood vessel distribution of a target object, and aligning and integrating voxels of the first type of image and the second type of image to form a multi-modal image set; extracting boundary features of a peripheral zone and a central zone of the prostate in the multi-modal image set by a generation unit of an attention generative adversarial network, generating an initial image containing a prostate region, and evaluating the initial image by a discrimination unit of the attention generative adversarial network to obtain a probability map representing whether each voxel belongs to the prostate region; converting the probability map into an initial voxel-level segmentation result by a convolution network, and performing morphological processing on the initial voxel-level segmentation result to obtain a target voxel-level segmentation result in which isolated noise points are removed; and calculating a prostate MRI volume based on the target voxel-level segmentation result.

[0051] The prostate MRI volume automatic calculation method based on the generative adversarial network provided in the application can fuse anatomical structure and blood vessel distribution information by obtaining the first type of image and the second type of image and integrating them into a multi-modal image set, thereby providing a comprehensive data basis for subsequent feature extraction; the boundary feature can be extracted by the attention generative adversarial network, the initial image can be generated, and the probability map can be obtained, the network attention mechanism can be used to focus on the key area, and the accuracy of prostate region recognition can be improved; the target voxel-level segmentation result can be obtained by converting the probability map through the convolution network and morphological processing, noise interference can be removed, and the integrity of the segmentation result can be optimized; the prostate MRI volume can be calculated based on the target segmentation result, and the volume can be accurately measured based on the accurate segmentation.

[0052] Further, the generation unit and the discrimination unit of the attention generative adversarial network are iterated for multiple times, the multi-modal image set is preprocessed, the boundary feature and the structure and blood vessel feature are extracted and fused, the image and the gland feature are compared to obtain the difference information, the boundary feature focusing degree is enhanced according to the difference information, the feature fusion ratio is adjusted, the feature proportion is optimized for different associated regions, and the probability map is generated until the difference information change amplitude meets the standard; these steps can dynamically adjust the feature fusion ratio, enhance the recognition ability of the boundary region and the small gland, improve the accuracy of the probability map, and provide a more reliable basis for subsequent segmentation and volume calculation.

[0053] These aspects or other aspects of the application will be more apparent in the following description of the embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions of the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0055] Figure 1 A flowchart of a prostate MRI volume automatic calculation method based on a generative adversarial network provided by an embodiment of the application;

[0056] Figure 2 A specific implementation diagram of a prostate MRI volume automatic calculation method based on a generative adversarial network provided by an embodiment of the application;

[0057] Figure 3 A structure diagram of a prostate MRI volume automatic calculation system based on a generative adversarial network provided by an embodiment of the application. DETAILED DESCRIPTION

[0058] In order to solve the problems of insufficient single-modal fusion, weak micro-adenoid feature capture and inaccurate voxel labeling in the prior art, an embodiment of the present application provides a prostate MRI volume automatic calculation method based on a generative adversarial network. The method adopts the following design concept: two kinds of images reflecting anatomical structure and blood vessel distribution are acquired, corresponding small units in the two kinds of images are aligned in spatial position and integrated to form an image set containing two kinds of information; then a network capable of focusing on key regions is used to extract boundary features of different regions of the prostate from the image set to generate an initial image containing the prostate region, and a probability distribution image of each small unit belonging to the prostate region is obtained after evaluation; then a specific network is used to convert the probability distribution image into an initial region division result, and isolated noise is removed through processing to obtain an accurate target region division result; finally, the MRI volume of the prostate is calculated based on the target result. In this way, the fusion of two kinds of image information can more comprehensively reflect the features of the prostate, solving the boundary recognition problem caused by insufficient single image information; the network processing focusing on key regions and the noise removal step can more accurately identify micro-adenoids and reduce misjudgment, thereby improving the accuracy of volume calculation.

[0059] In order to enable personnel in the technical field to better understand the present application scheme, the present application will be further described in detail below in combination with the drawings and specific embodiments. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without making creative efforts fall within the scope of protection of the present application.

[0060] The core of the present application is to provide a prostate MRI volume automatic calculation method based on a generative adversarial network, and a flowchart of one specific embodiment thereof is shown in Figure 1 The method comprises:

[0061] S11, acquiring a first type of image for reflecting anatomical structure and a second type of image for reflecting blood vessel distribution of a target object, and aligning and integrating voxels of the first type of image and the second type of image in spatial coordinates to form a multi-modal image set.

[0062] The target object is an individual receiving prostate examination, the first type of image is an image capable of reflecting the anatomical structure of the prostate, such as an MRI image showing the size and shape of the adenoid, and the second type of image is an image capable of reflecting the blood vessel distribution of the prostate, such as an image showing the density of blood vessels, the voxel is the basic unit constituting the image, the spatial coordinate alignment is to adjust the positions of the voxels in the two kinds of images to make the voxels at the same position correspond to the same anatomical part, and the multi-modal image set is a set formed by integrating the voxel information of the two kinds of images after alignment.

[0063] In the embodiments of the present application, first, the first type of image and the second type of image of the target object are acquired, for example, the anatomical structure image and the blood vessel distribution image of patient A are acquired, second, the spatial position of each voxel in the second type of image is adjusted so that they match the voxel position of the corresponding anatomical part in the first type of image, for example, the voxel at the position of the “left lobe of prostate” in the second type of image is matched with the voxel at the same position in the first type of image, and finally, the anatomical structure information and the blood vessel distribution information contained in the position-matched voxels are associated, and all the associated voxels are combined to form a multi-modal image set.

[0064] S12, the boundary feature of the peripheral zone and the central zone of the prostate in the multi-modal image set is extracted by the generation unit of the attention generative adversarial network, an initial image containing the prostate region is generated, the initial image is evaluated by the discrimination unit of the attention generative adversarial network, and a probability map representing whether each voxel belongs to the prostate region is obtained.

[0065] Among them, the attention generative adversarial network is a network containing two parts, the generation unit is the part of the network responsible for generating images, the discrimination unit is the part of evaluating whether the generated image is accurate, the boundary feature of the peripheral zone and the central zone of the prostate is the edge information that distinguishes the two, the initial image is the first image containing the prostate region generated by the generation unit, and the probability map is an image showing the possibility of each voxel belonging to the prostate region.

[0066] In the embodiments of the present application, first, the multi-modal image set is processed by the generation unit of the attention generative adversarial network, the boundary feature of the peripheral zone and the central zone of the prostate is extracted, and the structural feature of the first type of image and the blood vessel feature of the second type of image are fused to generate an initial image containing the prostate region, second, the discrimination unit compares the initial image with the gland feature in the multi-modal image set, identifies the small glands that are not accurately displayed, and outputs the difference information, then the generation unit focuses more on the boundary feature according to the difference information, adjusts the fusion ratio of the structural feature and the blood vessel feature to generate a new target image, and then the discrimination unit compares again and outputs the new difference, the above process is repeated until the difference changes small enough, finally, the possibility of each voxel belonging to the prostate region is determined based on the final target image, and a probability map is generated.

[0067] S13, the probability map is converted into an initial voxel-level segmentation result by a convolutional network, and morphological processing is performed on the initial voxel-level segmentation result to obtain a target voxel-level segmentation result with isolated noise points removed.

[0068] Wherein, the convolution network is a network capable of processing image data, the initial voxel-level segmentation result is a result of marking which voxels belong to the prostate region, the morphological processing is an arrangement operation on the initial segmentation result, the isolated noise points are a small number of mislabeled voxels discontinuous with surrounding prostate voxels, and the target voxel-level segmentation result is an accurate prostate region voxel labeling result after removing noise.

[0069] In the embodiment of the present application, first, the probability graph is processed by the convolution network, and the probability of each voxel is compared with the reference value. The voxels with a probability not lower than the reference value are marked as belonging to the prostate region to form an initial voxel-level segmentation result. Secondly, the continuous voxel groups (a set of voxels connected to each other) in the initial segmentation result are checked, and the number of voxels contained in each voxel group is counted. Then, the voxel groups with a number of voxels less than a preset threshold are determined as isolated noise points. Finally, these noise points are removed from the initial segmentation result, and the remaining voxel groups are integrated to form a target voxel-level segmentation result.

[0070] S14, calculating the prostate MRI volume based on the target voxel-level segmentation result.

[0071] Wherein, the target voxel-level segmentation result is an accurate labeling result of the prostate region voxels after noise removal, and the prostate MRI volume is the actual volume of the prostate in the MRI image calculated by the target segmentation result.

[0072] In the embodiment of the present application, first, the total number of voxels marked as belonging to the prostate region in the target voxel-level segmentation result is counted. Secondly, the size of a single voxel in the actual space is determined. Then, the volume of a single voxel is calculated. Finally, the total number of voxels in the prostate region is multiplied by the volume of a single voxel to obtain the prostate MRI volume.

[0073] Figure 2 A specific implementation diagram of a prostate MRI volume automatic calculation method based on a generative adversarial network provided in the embodiment of the present application is similar to the flow in Figure 1 , which will not be described here. Based on Figure 1 and Figure 2In the content of the foregoing, the present application provides the following specific examples: a prostate examination is performed on a patient A, first type image and second type image of the patient A are acquired, the second type image is adjusted in position to align with the first type image, structure information and blood vessel information of each aligned position are bound to form a multi-modal image set; then boundary features are extracted by a generation unit of an attention generation and adversarial network, structure and blood vessel features are fused to generate an initial image, a discrimination unit finds that a small amount of glands close to the peripheral zone are not clearly displayed after comparison, the generation unit accordingly increases the proportion of blood vessel features in the region to generate a target image, the adjustment is repeated for 3 times until the difference is small enough, and a probability map is generated based on the final target image; then voxels with a probability greater than or equal to 0.8 are marked as prostate voxels by a convolution network to obtain an initial segmentation result, and it is found that there is a 3-voxel isolated group (judged as noise), and the remaining voxel groups are integrated after the isolated group is removed to obtain a target segmentation result; finally, it is found that there are 8000 prostate voxels in the target result, the actual length, width and height of a single voxel are 0.5 mm, 0.5 mm and 1 mm respectively, the volume of a single voxel is 0.5*0.5*1=0.25 cubic millimeters, and the total volume is 8000*0.25=2000 cubic millimeters.

[0074] By performing S11-S14, the embodiment of the present application provides comprehensive data basis for feature extraction by integrating anatomical structure and blood vessel distribution information, enhances the recognition ability of the prostate boundary and the small amount of glands, and improves the accuracy of the probability map by iterative optimization of the attention generation and adversarial network, removes isolated noise by morphological processing, and optimizes the integrity of the segmentation result.

[0075] In a possible embodiment, S12, boundary features of the peripheral zone and the central zone of the prostate in the multi-modal image set are extracted by a generation unit of an attention generation and adversarial network, an initial image containing the prostate region is generated, an evaluation of the initial image is performed by a discrimination unit of the attention generation and adversarial network, and a probability map representing the probability that each voxel belongs to the prostate region is obtained, including:

[0076] Step 121, a generation unit of an attention generation and adversarial network is used for pre-processing the multi-modal image set, and boundary features corresponding to the peripheral zone and the central zone of the prostate are extracted from the pre-processing result, the boundary features are fused with structure features in the first type image and blood vessel features in the second type image, and an initial image containing the prostate region is generated.

[0077] Wherein, the generation unit of the attention generative adversarial network is responsible for generating images and focusing on key features; the multi-modal image set is an image set containing anatomical structure and blood vessel distribution information; preprocessing is a preliminary sorting operation on the multi-modal image set; the boundary feature of the peripheral zone and the central zone of the prostate is the edge information distinguishing the two; the structural feature of the first type of image is the information reflecting the size and shape of the gland; the vascular feature of the second type of image is the information reflecting the blood vessel distribution; and the initial image is the first generated image containing the prostate region.

[0078] In the embodiments of the present application, first, the generation unit preprocesses the multi-modal image set, such as removing irrelevant background to make the prostate region more prominent, for example, filtering the edge non-prostate tissue of the multi-modal image set of patient A; second, the boundary feature of the peripheral zone and the central zone is extracted from the preprocessing result, such as identifying the edge line between the two; then the boundary feature is fused with the structural feature and the vascular feature, such as combining the shape of the gland and the distribution of blood vessels; finally, the initial image is generated, such as forming an image containing the complete prostate region.

[0079] Step 122, comparing the initial image with the gland feature in the multi-modal image set through the discriminant unit of the attention generative adversarial network, based on the comparison result, identifying the micro gland in the prostate region, to output the difference information of the initial image and the gland feature.

[0080] Wherein, the discriminant unit of the attention generative adversarial network is the part that evaluates the accuracy of the generated image; the initial image is the image containing the prostate region generated in step 121; the gland feature in the multi-modal image set refers to the size, shape, distribution, etc. of the normal gland; the micro gland is a smaller part of the gland; and the difference information is the mismatch information between the initial image and the gland feature.

[0081] In the embodiments of the present application, first, the discriminant unit compares the initial image with the gland feature in the multi-modal image set, such as comparing the initial image of patient A with the normal gland feature; second, based on the comparison result, the micro gland that is not accurately displayed is identified, for example, a certain micro gland is found to be missing; and finally, the difference information is output, such as specifying the location and shape of the missing gland.

[0082] Step 123, through the generation unit, enhancing the focusing degree on the boundary feature according to the difference information, based on the enhanced focusing degree, adjusting the fusion ratio of the structural feature and the vascular feature, to generate the target image.

[0083] Wherein, the generating unit is the part of generating the image in step 121; the difference information is the mismatch information output in step 122; the focusing degree of the boundary feature is the attention degree to the prostate boundary; the structure feature is the information of the gland shape in the first type of image; the blood vessel feature is the information of the blood vessel distribution in the second type of image; the fusion ratio is the proportion of different features in the fusion; and the target image is the image generated after optimization.

[0084] In the embodiment of the present application, first, the generating unit enhances the focusing on the boundary feature according to the difference information, such as focusing on the boundary where the small glands are located, for example, focusing on the corresponding area of the difference information of patient A; second, the fusion ratio of the structure feature and the blood vessel feature is adjusted based on the focusing degree, for example, increasing the proportion of the blood vessel feature in the small gland area; and finally, the target image is generated by fusing the adjusted features, such as the image that clearly shows the small glands.

[0085] Step 124, the target image is compared again by the judging unit, and new difference information is output, and the processing process of the generating unit and the judging unit is repeated until the change amplitude of the difference information is less than the preset difference threshold.

[0086] Wherein, the judging unit is the part of evaluating the image in step 122; the target image is the optimized image generated in step 123; the new difference information is the mismatch information output after the comparison of the judging unit; the repeated processing process is the loop of the generating and judging operations; the preset difference threshold is the standard for judging whether the difference is small enough; and the change amplitude of the difference information is the difference between the two times of difference.

[0087] In the embodiment of the present application, first, the judging unit compares the target image with the gland feature again, such as checking whether the target image of patient A is perfect; second, new difference information is output, for example, indicating that the edge of the small glands is not clear; then the process of adjusting the features by the generating unit and evaluating by the judging unit is repeated; and finally, the process is stopped when the difference change amplitude is less than the preset threshold, such as when the difference between the two times is small enough to be acceptable.

[0088] Step 125, after the repeated processing process is completed, the probability distribution of each voxel belonging to the prostate region is obtained based on the last generation of the target image, and a probability map is generated based on the probability distribution.

[0089] Wherein, the repeated processing process is the loop operation of step 124; the last generation of the target image is the image at the time when step 124 stops; the probability distribution of each voxel belonging to the prostate region is the possibility of each image basic unit belonging to the prostate; and the probability map is the image showing the probability distribution.

[0090] In the embodiments of the present application, first, the last generation of target images is obtained after the repeated processing ends, such as the image of patient A after 3 cycles; second, the probability distribution of each voxel is determined based on the image, for example, the center voxel of the gland has a high possibility and the edge is low; and finally, the probability map is generated based on the probability distribution, such as using different values to represent the possibility.

[0091] The present application provides the following specific examples: for a multi-modal image set of patient A, the generation unit first pre-processes to remove the background, extracts the boundary features of the peripheral zone and the central zone, and fuses the structure and blood vessel features to generate an initial image; the discrimination unit finds that the small glands near the peripheral zone are not completely displayed after comparison, and outputs difference information; the generation unit enhances the boundary focusing of this region accordingly, increases the fusion proportion of blood vessel features to generate a target image; the discrimination unit points out that the edge of the small glands is not clear again, and the generation unit generates a target image again after adjustment, and so on for 3 times, until the difference change amplitude is less than a preset threshold; and finally, based on the final target image, the probability of each voxel is determined to generate a probability map.

[0092] By performing steps 121-125, the embodiments of the present application generate an initial image by pre-processing and feature fusion, lay a foundation for subsequent optimization; with the help of the difference information output by the discrimination unit, the image optimization direction is clear; by adjusting the focusing degree and the fusion proportion, the accuracy of the target image is gradually improved; after multiple cycles of optimization, the image approaches the real prostate region; and finally, the probability map clearly reflects the possibility that the voxel belongs to the prostate, providing a reliable basis for subsequent segmentation.

[0093] In one possible embodiment, step 123, the focusing degree for the boundary feature is enhanced according to the difference information, the fusion proportion of the structure feature and the blood vessel feature is adjusted based on the enhanced focusing degree, and the target image is generated, including:

[0094] a1, analyzing the difference information to determine a first associated region corresponding to the difference information and the peripheral zone and the central zone boundary of the prostate, and a second associated region corresponding to the difference information and the small glands.

[0095] Wherein, the difference information is the content in the image that does not match the normal gland feature, the first associated region is the region in the difference information that involves the peripheral zone and the central zone boundary of the prostate at the same time, and the second associated region is the region in the difference information that involves the small glands, and the analysis is an analysis of the difference information to clarify the specific region it corresponds to.

[0096] In the embodiments of the present application, firstly, the difference information is analyzed, and the specific positions corresponding to the unmatched parts are carefully analyzed, for example, the difference information of patient A is analyzed, and it is determined which is related to the boundary of the peripheral zone and the central zone, and which is related to the micro gland; secondly, according to the analysis result, the region corresponding to the boundary of the peripheral zone and the central zone is determined as the first associated region, and the region corresponding to the micro gland is determined as the second associated region, for example, in the difference information of patient A, the fuzzy edge involving the double boundary is determined as the first associated region, and the micro missing part is determined as the second associated region.

[0097] a2, for the first associated region, the proportion of the boundary feature in feature fusion is increased to enhance the focusing degree.

[0098] Among them, the first associated region is the region related to the boundary of the peripheral zone and the central zone determined in the a1 step, the boundary feature is the edge information distinguishing the peripheral zone and the central zone, the proportion in the feature fusion is the proportion of a certain feature in the fusion, and the focusing degree is the attention and prominence of the boundary feature of the region.

[0099] In the embodiments of the present application, firstly, for the first associated region determined in the a1 step, the proportion of the current boundary feature in the fusion is analyzed, for example, the boundary feature proportion in the first associated region of patient A is 30%; secondly, the focusing degree is enhanced by increasing the proportion, for example, the proportion is increased from 30% to 50%, so that the boundary feature is more prominent.

[0100] a3, for the second associated region, according to the enhanced focusing degree, the proportion of the blood vessel feature in the feature fusion is increased.

[0101] Among them, the second associated region is the region related to the micro gland determined in the a1 step, the blood vessel feature is the information reflecting the distribution of blood vessels, the enhanced focusing degree is the improvement of the attention degree of the region, and the proportion in the feature fusion is the proportion of the blood vessel feature in the fusion.

[0102] In the embodiments of the present application, firstly, for the second associated region determined in the a1 step, the proportion of the current blood vessel feature in the fusion is analyzed in combination with the enhanced focusing degree of the a2 step, for example, the blood vessel feature proportion in the second associated region of patient A is 20%; secondly, the proportion is increased according to the focusing degree, for example, from 20% to 40%, and the blood vessel distribution information is highlighted.

[0103] a4, for the irrelevant region in the initial image, the basic proportion of the structure feature in the feature fusion is maintained.

[0104] Among them, the irrelevant region is the region in the initial image which is not involved in the difference information (matched with the normal gland feature), the structure feature is the information reflecting the size and shape of the gland, and the basic proportion is the proportion of the structure feature in the fusion.

[0105] In the embodiment of the present application, firstly, the irrelevant area is determined, that is, the area remaining after excluding the first and second associated areas of step a1, for example, most of the areas in the prostate image of patient A match the normal features, which are irrelevant areas; secondly, for these areas, the basic proportion of the structure features is maintained, for example, the original proportion of 50% is continued to be maintained without adjustment.

[0106] a5, the boundary features and structure features after adjusting the proportion, and the blood vessel features maintaining the basic proportion are fused to generate a target image.

[0107] Among them, the boundary features after adjusting the proportion are the boundary features with increased proportion in step a2, the structure features after adjusting the proportion are the structure information adjusted according to the area requirements, the blood vessel features maintaining the basic proportion are the blood vessel information not adjusted in step a4, the fusion is to combine different features, and the target image is the optimized image generated after fusion.

[0108] In the embodiment of the present application, firstly, the boundary features after adjusting the proportion in step a2, the blood vessel features after adjusting the proportion in step a3, and the structure and blood vessel features maintaining the basic proportion in step a4 are collected, for example, the boundary feature proportion of the first associated area in patient A is 50%, the blood vessel feature proportion of the second associated area is 40%, and the structure feature proportion of the irrelevant area is 50%; secondly, the features are fused according to the adjusted proportion, for example, the features of each area are mixed according to the corresponding proportion; finally, a target image is generated, for example, an image with clear boundaries and complete small glands is obtained.

[0109] The present application provides the following specific examples: analyzing the difference information of patient A, determining that the first associated area is "fuzzy edge at the junction of the peripheral zone and the central zone", and the second associated area is "small gland area near the peripheral zone not clearly displayed"; for the first associated area, the proportion of the boundary features in the fusion is increased from 30% to 50% to enhance the focus; for the second associated area, the proportion of the blood vessel features is increased from 20% to 40% combined with the degree of focus improvement; for other irrelevant areas, the basic proportion of the structure features is maintained at 50%; finally, the adjusted features are fused according to the corresponding proportion to generate a target image of patient A.

[0110] By performing a1-a5, the embodiment of the present application determines the optimization area by analyzing the difference information, increases the proportion of boundary features for the first associated area to enhance the boundary clarity, increases the proportion of blood vessel features for the second associated area to improve the display effect of small glands, maintains the proportion of structure features for irrelevant areas to ensure the stability of the image, and finally generates a target image that is more accurate in key areas and improves the overall quality.

[0111] In a possible embodiment, S13 converts the probability map into an initial voxel-level segmentation result by a convolution network, performs morphological processing on the initial voxel-level segmentation result, and obtains a target voxel-level segmentation result in which isolated noise points are removed, including:

[0112] Step 131: Extracting probability values of voxels in the probability map, and dividing voxels belonging to the prostate region according to the probability values of the voxels by a convolution network to form an initial voxel-level segmentation result.

[0113] The probability map is an image showing the possibility size of each voxel belonging to the prostate region, the probability value of the voxel is a possibility value of each image cell belonging to the prostate region, the convolution network is a tool for processing image data, and the initial voxel-level segmentation result is a preliminary result of marking voxels belonging to the prostate region.

[0114] In the embodiment of the application, firstly, the probability values of voxels in the probability map are extracted, that is, the possibility values of each cell belonging to the prostate region are obtained, for example, the probability values of voxels in the probability map of patient A are extracted, some of which are 0.8 and some of which are 0.3; secondly, the probability values are processed by a convolution network, and voxels belonging to the prostate region are divided according to the probability value size, for example, a reference value is set, and voxels with a probability value not lower than the reference value are marked as voxels belonging to the prostate region; and finally, the initial voxel-level segmentation result is formed, for example, in the initial segmentation result of patient A, voxels with a probability value greater than or equal to 0.6 are marked as voxels belonging to the prostate region.

[0115] Step 132: Checking voxel groups with continuous positions in the initial voxel-level segmentation result, and determining the number of voxels contained in each voxel group.

[0116] The initial voxel-level segmentation result is the preliminary segmentation result obtained in step 131, the voxel group with continuous positions is a set of voxels connected to each other, and the number of voxels contained in the voxel group is the number of voxels in each continuous set.

[0117] In the embodiment of the application, firstly, the voxel groups with continuous positions in the initial voxel-level segmentation result are checked, that is, it is found out which voxels are connected to each other to form multiple groups, for example, the initial segmentation result of patient A is viewed, and it is found that there are multiple voxel groups connected to each other; secondly, the number of voxels contained in each voxel group is determined, that is, the number of voxels in each group is counted, for example, the initial segmentation result of patient A has three continuous voxel groups, which contain 200, 50, and 3 voxels, respectively.

[0118] Step 133: Marking a voxel group with a number of voxels less than a preset number threshold as an isolated redundant voxel group.

[0119] Wherein, the number of voxels is the number of voxels in each voxel group determined in step 132, the preset number threshold is a standard value for judging whether the voxel group is redundant, and the isolated redundant voxel group is a voxel set with too few numbers and not connected to other voxel groups.

[0120] In the embodiments of the present application, firstly, the number of voxels in each voxel group determined in step 132 is compared with the preset number threshold, for example, the preset number threshold is 5, and the number of voxel groups of patient A is 200, 50 and 3 respectively; secondly, the voxel group with a number less than the preset number threshold is marked as an isolated redundant voxel group, for example, the voxel group with a number of 3 in patient A is less than the threshold 5, and is marked as a redundant voxel group.

[0121] Step 134, removing the redundant voxel group from the initial voxel-level segmentation result, retaining the remaining voxel group and integrating to form the target voxel-level segmentation result.

[0122] Wherein, the initial voxel-level segmentation result is the preliminary result obtained in step 131, the redundant voxel group is the voxel group with too few numbers marked in step 133, the remaining voxel group is the voxel group not marked as redundant, and the target voxel-level segmentation result is the final segmentation result obtained after removing the redundant voxel group.

[0123] In the embodiments of the present application, firstly, the redundant voxel group marked in step 133 is removed from the initial voxel-level segmentation result, for example, the redundant voxel group containing 3 voxels is removed in patient A; secondly, the remaining voxel group is retained, that is, the voxel group not removed is retained, for example, the voxel groups containing 200 and 50 voxels are retained in patient A; finally, the retained voxel groups are integrated to form the target voxel-level segmentation result, for example, the two voxel groups retained in patient A are combined to form a complete prostate region segmentation result.

[0124] The present application provides the following specific examples: extracting the probability values of each voxel from the probability map of patient A, wherein part of the probability values of the voxels are 0.7 and 0.9, and part are 0.2 and 0.4; setting a reference value 0.6 through a convolution network, marking the voxels with a probability value ≥0.6 as belonging to the prostate region to form an initial voxel-level segmentation result; checking the result to find three continuous voxel groups, counting the first voxel group containing 200 voxels, the second voxel group containing 50 voxels, and the third voxel group containing 3 voxels; the preset number threshold is 5, and the voxel group containing 3 voxels is compared with the threshold, and because 3 is less than 5, it is marked as a redundant voxel group; removing the redundant voxel group from the initial result, retaining the other two voxel groups and integrating to form the target voxel-level segmentation result of patient A.

[0125] By performing steps 131-134, the embodiment of the application obtains a preliminary segmentation result by extracting probability values and dividing voxels, provides a basis for subsequent processing; checks the continuous voxel group and determines the number, provides a basis for judging the redundant voxel group; marks and removes the redundant voxel group, reduces noise interference; the final integration of the target segmentation result is more accurate and complete, and lays a reliable foundation for subsequent prostate MRI volume calculation.

[0126] In one possible embodiment, step 131 divides voxels belonging to the prostate region according to the probability value of each voxel by using a convolutional network to form an initial voxel-level segmentation result, including:

[0127] b1, using a convolutional network to process the probability value of each voxel layer by layer, and setting a reference value for judging the attribution of the voxel.

[0128] The convolutional network is a tool for processing image data, the probability value of the voxel is a likelihood value of each image cell belonging to the prostate region, the layer-by-layer processing is processing data by levels to make the characteristics more obvious, and the reference value is a standard value for judging whether the voxel belongs to the prostate region.

[0129] In the embodiment of the application, the probability value of each voxel is processed layer by layer by using a convolutional network, that is, the likelihood value of each voxel is processed by levels to make the data characteristics clearer, for example, the voxel probability value of patient A is processed by layers to make the difference between different probability values more obvious; secondly, a reference value for judging the attribution of the voxel is set, that is, a standard value is determined to distinguish whether the voxel belongs to the prostate region, for example, the reference value is set to 0.6 after processing.

[0130] b2, comparing the probability value of each voxel with the reference value, assigning a first mark indicating that the voxel belongs to the prostate region to the voxel whose probability value is not lower than the reference value, and assigning a second mark indicating that the voxel does not belong to the prostate region to the voxel whose probability value is lower than the reference value.

[0131] The probability value of the voxel is a likelihood value of each image cell belonging to the prostate region, the reference value is the judgment standard set in step b1, the first mark is a mark indicating that the voxel belongs to the prostate region, and the second mark is a mark indicating that the voxel does not belong to the prostate region.

[0132] In the embodiments of the present application, firstly, the probability value of each voxel is compared with the reference value, that is, the probability value of each voxel is compared with the standard value, for example, the voxel probability value (such as 0.8, 0.3, etc.) of patient A is compared with the reference value 0.6; secondly, the first mark is given to the voxel whose probability value is not lower than the reference value, for example, the voxel whose probability value is 0.8 in patient A is marked as the first mark; the second mark is given to the voxel whose probability value is lower than the reference value, for example, the voxel whose probability value is 0.3 is marked as the second mark.

[0133] b3, fusing all the voxels with the first mark and the second mark to form an initial voxel-level segmentation result.

[0134] Wherein, the first mark is a mark indicating that the voxel belongs to the prostate region, the second mark is a mark indicating that the voxel does not belong to the prostate region, the fusion is to integrate all the marked voxels according to the original position, and the initial voxel-level segmentation result is a preliminary segmentation result containing all the marked voxels.

[0135] In the embodiments of the present application, firstly, all the voxels with the first mark and the second mark are collected, that is, all the marked image small units are summarized, for example, all the voxels with the first mark and the second mark in patient A are collected; secondly, these voxels are fused, that is, they are integrated together according to their original positions to form a complete marked image; finally, the initial voxel-level segmentation result is obtained, for example, a preliminary segmentation image of patient A containing all the marked voxels is formed.

[0136] The present application provides the following specific examples: using the convolution network to perform layer-by-layer processing on the probability value (such as 0.7, 0.9, 0.2, 0.4, etc.) of each voxel of patient A, so that the characteristics of different probability values are more clear, and then setting the reference value as 0.6; comparing the probability value of each voxel with 0.6, wherein 0.7 and 0.9 are not lower than 0.6, and the first mark indicating belonging to the prostate region is given to these voxels, 0.2 and 0.4 are lower than 0.6, and the second mark indicating not belonging to the prostate region is given to these voxels; finally, all the voxels with the first mark and the second mark are integrated and fused according to the original position to form the initial voxel-level segmentation result of patient A.

[0137] By performing b1~b3, the layer-by-layer processing of the convolution network in the embodiments of the present application makes the voxel probability value characteristics more distinguishable, and the set reference value provides a clear standard for voxel attribution judgment; by comparison and mark assignment, the voxels belonging to and not belonging to the prostate region are clearly divided; and finally, the initial voxel-level segmentation result formed by fusion contains complete marked information, which provides an accurate and comprehensive basis for subsequent processing, and improves the reliability of segmentation as a whole.

[0138] In one possible implementation, S11, the voxels of the first type of image and the second type of image are spatially aligned and integrated to form a multi-modal image set, comprising:

[0139] Step 111, determining the spatial coordinates of the first type of image as reference coordinates, and adjusting the spatial positions of the voxels in the second type of image according to the reference coordinates, so that the spatial positions match the corresponding positions in the reference coordinates.

[0140] Wherein, the first type of image is an image reflecting the anatomical structure of the prostate, the second type of image is an image reflecting the blood vessel distribution of the prostate, the spatial coordinates are the position information of the voxels in the image, the reference coordinates are the spatial coordinates of the first type of image as the reference, the voxel is the basic unit constituting the image, and the spatial position matching refers to the alignment of the voxels in the second type of image with the voxels in the corresponding position of the first type of image.

[0141] In the embodiments of the present application, first, the spatial coordinates of the first type of image are determined as the reference coordinates, i.e. the positions of each voxel in the first type of image are taken as the reference, for example, the spatial coordinates of the first type of image (anatomical structure image) of patient A are taken as the reference; second, the spatial positions of the voxels in the second type of image are adjusted according to the reference coordinates, by moving or rotating the second type of image, so that the positions of the voxels in the second type of image are consistent with the positions of the corresponding anatomical parts in the reference coordinates, for example, the voxels in the position of "right lobe of prostate" in the second type of image (blood vessel distribution image) of patient A are adjusted, so that they correspond to the voxels in the same position in the first type of image.

[0142] Step 112, determining the voxels of the first type of image and the voxels of the second type of image with spatial position matching as corresponding voxel pairs.

[0143] Wherein, the voxels with spatial position matching are the voxels of the first type of image and the second type of image with position matching after step 111, and the corresponding voxel pair is a paired unit composed of two voxels with position matching.

[0144] In the embodiments of the present application, first, the voxels in the second type of image with position matching to the reference coordinates of the first type of image after step 111 are identified, for example, the voxel in the position (x1, y1, z1) in the second type of image of patient A matches the voxel in the same position in the first type of image; second, the first type of image voxels and the second type of image voxels with position matching are determined as corresponding voxel pairs, for example, the two voxels in the position (x1, y1, z1) in patient A are composed into a pair.

[0145] Step 113, associating the image information of the first type of image in the corresponding voxel pair with the image information of the second type of image.

[0146] Wherein, the corresponding voxel pair is the position-matched voxel pair determined in step 112, the image information of the first type of image refers to the content reflecting the anatomical structure (such as the shape of the gland), the image information of the second type of image refers to the content reflecting the blood vessel distribution (such as the blood vessel density), and the association is to bind the two types of information together.

[0147] In the embodiment of the present application, first, the image information of the first type of image and the image information of the second type of image in the corresponding voxel pair are extracted, for example, in a certain corresponding voxel pair of patient A, the first type of image voxel contains "gland edge" information, and the second type of image voxel contains "blood vessel dense" information; second, the two types of information are associated, that is, the structure information and the blood vessel information in the same corresponding voxel pair are bound, for example, the "gland edge" and the "blood vessel dense" information are bound together.

[0148] Step 114, combine all the associated corresponding voxel pairs to form a multi-modal image set.

[0149] Wherein, the associated corresponding voxel pair is the voxel pair bound with two types of information after step 113, and the multi-modal image set is a set containing two types of image information formed by combining all these voxel pairs.

[0150] In the embodiment of the present application, first, all the corresponding voxel pairs associated after step 113 are collected, for example, all the voxel pairs of patient A bound with structure and blood vessel information are collected; second, these voxel pairs are combined together according to their spatial positions to form a complete set, for example, all the associated voxel pairs are arranged in order of the spatial coordinates of the voxels; and finally, the multi-modal image set is formed, for example, a set containing the comprehensive information of the structure and the blood vessels of all positions of the prostate of patient A is obtained.

[0151] The present application provides the following specific examples: taking the spatial coordinates of the first type of image (anatomical structure image) of patient A as the reference coordinates, adjusting the positions of the voxels in the second type of image (blood vessel distribution image) so that the voxels of the "left lobe of prostate" and the "central region" and other parts in the second type of image are respectively matched with the voxel positions of the same parts in the first type of image; then, the matching voxels at positions (1, 2, 3) and (2, 3, 4) are identified, and each two voxels with the same position are determined as a corresponding voxel pair; for each voxel pair, the anatomical information such as "left lobe structure" and "central region shape" of the first type of image and the blood vessel information such as "left lobe blood vessel distribution" and "central region blood vessel density" of the second type of image are extracted, and the two types of information of the same voxel pair are bound; finally, all the associated voxel pairs are combined in order of their spatial coordinates to form a multi-modal image set containing the anatomical structure and the blood vessel distribution information of the prostate of patient A.

[0152] By performing steps 111-113, the embodiments of the present application realize accurate alignment of the two types of images in space by determining the reference coordinates and adjusting the positions, ensure that the information of the same anatomical part can be matched; by determining the corresponding voxel pairs and associating the information, each voxel contains dual information of anatomical structure and blood vessel distribution, which enriches the data dimension; and finally the multi-modal image set integrates the two key information, which provides a comprehensive and matched data basis for accurately extracting the boundary of the prostate, the gland and other features subsequently, and improves the reliability of subsequent processing.

[0153] In a possible embodiment, S14, based on the target voxel-level segmentation result, calculates the prostate MRI volume, including:

[0154] Step 141, counts the total number of voxels belonging to the prostate region in the target voxel-level segmentation result.

[0155] Wherein, the target voxel-level segmentation result is the result of accurately marking the voxels of the prostate region after denoising processing, the voxel belonging to the prostate region is the image basic unit marked as the prostate part, and the total number of voxels is the total number of these voxels.

[0156] In the embodiments of the present application, first, it is determined which voxels in the target voxel-level segmentation result belong to the prostate region, that is, the corresponding voxels are identified according to the previous marking, for example, in the target segmentation result of patient A, all voxels with the label of belonging to the prostate; second, the total number of these voxels is counted, which is counted one by one or automatically counted using tools, for example, there are 8000 voxels belonging to the prostate region in the target result of patient A.

[0157] Step 142, determines the actual length, actual width and actual height of a single voxel in the actual scanning space, and calculates the volume of the single voxel by the actual length, actual width and actual height.

[0158] Wherein, the single voxel is the basic unit of the image, the actual length, actual width and actual height are the dimensions of the unit in the real space, and the volume of the single voxel is the space size calculated by the three dimensions.

[0159] In the embodiments of the present application, first, the actual length, actual width and actual height of a single voxel in the actual scanning space are determined, which are usually obtained from the parameter settings of the scanning device, for example, in the scanning image of patient A, the actual length of a single voxel is 0.5 millimeter, the width is 0.5 millimeter, and the height is 1 millimeter; second, the volume of the single voxel is calculated by multiplying the actual length by the actual width and then by the actual height, for example, 0.5 millimeter multiplied by 0.5 millimeter multiplied by 1 millimeter equals 0.25 cubic millimeters.

[0160] Step 143, multiply the total number of voxels by the volume of a single voxel to obtain the volume of the prostate MRI.

[0161] Wherein, the total number of voxels is the total number of voxels belonging to the prostate region counted in step 141, the volume of a single voxel is the spatial size of each voxel calculated in step 142, and the volume of the prostate MRI is the actual volume of the prostate corresponding in the image obtained by multiplying the two.

[0162] In the embodiments of the present application, the total number of voxels in step 141 and the volume of a single voxel in step 142 are first obtained, for example, the total number of voxels of patient A is 8000, and the single volume is 0.25 cubic millimeters; second, the total number of voxels is multiplied by the volume of a single voxel to obtain the volume of the prostate MRI, for example, 8000 multiplied by 0.25 equals 2000 cubic millimeters.

[0163] The present application provides the following specific examples: for the target voxel segmentation result of patient A, all voxels labeled as belonging to the prostate region are identified, and the total number is 8000; from the scanning device parameters, it is determined that the actual length of a single voxel of the patient image is 0.5 millimeters, the actual width is 0.5 millimeters, and the actual height is 1 millimeter, and the volume of a single voxel is calculated as 0.5 multiplied by 0.5 multiplied by 1, which is equal to 0.25 cubic millimeters; the total number of voxels 8000 is multiplied by the volume of a single voxel 0.25 cubic millimeters, that is, 8000 multiplied by 0.25 equals 2000, and the volume of the prostate MRI of patient A is 2000 cubic millimeters.

[0164] By performing steps 141-143, the embodiments of the present application determine the total number of basic units constituting the prostate region by counting the total number of voxels, establish the correspondence between the image unit and the real space by calculating the volume of a single voxel combined with the scanning device parameters, and finally obtain the prostate MRI volume by multiplying the two, which accurately reflects the actual size of the prostate, and provides a scientific and reliable quantitative basis for clinical needs such as disease diagnosis and disease assessment.

[0165] Figure 3 A structure diagram of a prostate MRI volume automatic calculation system based on a generative adversarial network provided by the embodiments of the present application is shown in FIG. 1, which includes: Figure 3

[0166] The acquisition module 31 is configured to acquire a first type of image for reflecting an anatomical structure and a second type of image for reflecting a blood vessel distribution of a target object, and align and integrate the voxels of the first type of image and the second type of image in space coordinates to form a multi-modal image set.

[0167] ​The generating module 32 is configured to extract boundary features of the peripheral zone and the central zone of the prostate in the multi-modal image set by a generative unit of an attention generative adversarial network, generate an initial image containing the prostate region, and evaluate the initial image by a discriminative unit of the attention generative adversarial network to obtain a probability map representing whether each voxel belongs to the prostate region.

[0168] The processing module 33 is configured to convert the probability map into an initial voxel-level segmentation result by a convolutional network, and perform morphological processing on the initial voxel-level segmentation result to obtain a target voxel-level segmentation result in which isolated noise points are removed.

[0169] The calculating module 34 is configured to calculate the prostate MRI volume based on the target voxel-level segmentation result.

[0170] The prostate MRI volume automatic calculation system based on the generative adversarial network according to the embodiments of the present application is used to implement the prostate MRI volume automatic calculation method based on the generative adversarial network described above, and therefore the specific embodiments of the prostate MRI volume automatic calculation system based on the generative adversarial network can be seen from the foregoing embodiments of the prostate MRI volume automatic calculation method based on the generative adversarial network, and the specific embodiments can be referred to the descriptions of the corresponding embodiments, which will not be described herein again.

[0171] The present application also provides an electronic device, comprising: a memory configured to store a computer program; and a processor configured to implement the steps of the prostate MRI volume automatic calculation method based on the generative adversarial network described above when executing the computer program.

[0172] The present application also provides a computer readable storage medium having a computer program stored thereon, and the computer program is configured to implement the steps of the prostate MRI volume automatic calculation method based on the generative adversarial network described above when executed by a processor.

[0173] In an exemplary embodiment, the computer readable storage medium described above can include, but is not limited to, a U disk, a read-only memory, a random access memory, a mobile hard disk, a magnetic disk or an optical disk, and various media that can store computer programs.

[0174] The embodiments of the present application also provide a computer program product, and the computer program product includes a computer program, and the computer program is configured to implement the steps of any of the prostate MRI volume automatic calculation methods based on the generative adversarial network described above when executed by a processor.

[0175] Those skilled in the art will further realize that the mere concepts, teachings, and embodiments described herein are merely meant to provide an enabling description of embodiments of the present application and are not intended to limit the scope of the present application. Accordingly, embodiments as described herein contemplate all modifications that come within the scope of the present application as recited by the claims set forth below and any equivalents thereto, with the scope of the present application being measured by the broadest interpretation of those claims set forth below.

[0176] The above provides a kind of prostate MRI volume automatic calculation method based on generative adversarial network, system, electronic equipment and storage medium provided in the present application in detail.The principle and implementation mode of the present application are described in this paper by applying specific examples, the above example is only used to help understanding the method and its core idea of the present application.It should be pointed out that, for the ordinary skilled in the art, without departing from the principle of the present application, the present application can be improved and modified, these improvements and modifications also fall within the scope of the present application.

Claims

1. A method for automatic calculation of prostate MRI volume based on generative adversarial network, characterized in that, The method comprises the following steps: acquiring a first type of image reflecting an anatomical structure and a second type of image reflecting a blood vessel distribution of a target object, and aligning and integrating voxels of the first type of image and the second type of image to form a multi-modal image set; extracting boundary features of a peripheral zone and a central zone of the prostate in the multi-modal image set by a generation unit of an attention generative adversarial network, generating an initial image containing a prostate region, evaluating the initial image by a discrimination unit of the attention generative adversarial network, and obtaining a probability map representing whether each voxel belongs to the prostate region; converting the probability map into an initial voxel-level segmentation result by a convolutional network, and performing morphological processing on the initial voxel-level segmentation result to obtain a target voxel-level segmentation result in which isolated noise points are removed; calculating a prostate MRI volume based on the target voxel-level segmentation result; the step of extracting boundary features of a peripheral zone and a central zone of the prostate in the multi-modal image set by a generation unit of an attention generative adversarial network, generating an initial image containing a prostate region, and evaluating the initial image by a discrimination unit of the attention generative adversarial network to obtain a probability map representing whether each voxel belongs to the prostate region comprises the following steps: preprocessing the multi-modal image set by a generation unit of an attention generative adversarial network, extracting boundary features corresponding to the peripheral zone and the central zone of the prostate from the preprocessing result, fusing the boundary features with structural features in the first type of image and blood vessel features in the second type of image, and generating an initial image containing a prostate region; comparing the initial image with gland features in the multi-modal image set by a discrimination unit of the attention generative adversarial network, identifying micro glands in the prostate region based on a comparison result, and outputting difference information between the initial image and the gland features; enhancing a focusing degree on the boundary features according to the difference information by the generation unit, adjusting a fusion ratio of the structural features and the blood vessel features based on the enhanced focusing degree, and generating a target image; re-comparing the target image by the discrimination unit and outputting new difference information, repeating a processing process of the generation unit and the discrimination unit until a change amplitude of the difference information is less than a preset difference threshold; after the repeated processing process ends, obtaining a probability distribution of whether each voxel belongs to the prostate region based on the target image of the last generation, and generating a probability map based on the probability distribution.

2. The method of claim 1, wherein, the step of enhancing a focusing degree on the boundary features according to the difference information, adjusting a fusion ratio of the structural features and the blood vessel features based on the enhanced focusing degree, and generating a target image comprises the following steps: analyzing the difference information, determining a first associated region corresponding to the difference information and the peripheral zone and the central zone boundary, and a second associated region corresponding to the difference information and the micro glands; for the first associated region, enhancing the focusing degree by increasing a proportion of the boundary features in feature fusion; For the second associated region, according to the enhanced focusing degree, the proportion of the blood vessel feature in feature fusion is increased; For the irrelevant region in the initial image, the basic proportion of the structural feature in feature fusion is maintained; The boundary feature and the structural feature after the proportion adjustment and the blood vessel feature maintaining the basic proportion are fused to generate a target image.

3. The method of claim 1, wherein, The method for converting the probability map into an initial voxel-level segmentation result through a convolution network, and performing morphological processing on the initial voxel-level segmentation result to obtain a target voxel-level segmentation result in which isolated noise points are removed, comprises: Extracting probability values of voxels in the probability map, and dividing voxels belonging to the prostate region according to the probability values of the voxels through a convolution network to form an initial voxel-level segmentation result; Checking voxel groups with continuous positions in the initial voxel-level segmentation result, and determining the number of voxels included in each voxel group; Marking a voxel group with a number of voxels less than a preset number threshold as an isolated redundant voxel group; Removing the redundant voxel group from the initial voxel-level segmentation result, retaining the remaining voxel groups and integrating them to form a target voxel-level segmentation result.

4. The method of claim 3, wherein, The method for dividing voxels belonging to the prostate region according to the probability values of the voxels through a convolution network to form an initial voxel-level segmentation result, comprises: Performing layer-by-layer processing on the probability values of the voxels by using a convolution network, and setting a reference value for judging the attribution of a voxel; Comparing the probability values of the voxels with the reference value, assigning a first identifier representing belonging to the prostate region to a voxel with a probability value not lower than the reference value, and assigning a second identifier representing not belonging to the prostate region to a voxel with a probability value lower than the reference value; Fusing all voxels with the first identifier and the second identifier to form an initial voxel-level segmentation result.

5. The method of claim 1, wherein, The method for aligning and integrating voxels of the first type of image and the second type of image in spatial coordinates to form a multi-modal image set, comprises: Determining the spatial coordinates of the first type of image as reference coordinates, and adjusting the spatial positions of voxels in the second type of image according to the reference coordinates, so that the spatial positions match corresponding positions in the reference coordinates; Determining voxels of the first type of image and voxels of the second type of image with matched spatial positions as corresponding voxel pairs; Associating image information of the first type of image and image information of the second type of image in the corresponding voxel pairs; Combining all associated corresponding voxel pairs to form a multi-modal image set.

6. The method of claim 1, wherein, The method for calculating the prostate MRI volume based on the target voxel-level segmentation result, comprises: Counting the total number of voxels belonging to the prostate region in the target voxel-level segmentation result; Determining the actual length, the actual width and the actual height of a single voxel in the actual scanning space, and calculating the volume of a single voxel through the actual length, the actual width and the actual height; Multiplying the total number of voxels and the volume of a single voxel to obtain the volume of the prostate MRI.

7. A system for automatic calculation of prostate MRI volume based on generative adversarial networks, characterized by, The method for calculating the prostate MRI volume based on the target voxel-level segmentation result, comprises: The acquisition module is configured to acquire a first type of image of a target object for reflecting an anatomical structure and a second type of image for reflecting a blood vessel distribution, align and integrate voxels of the first type of image and the second type of image in spatial coordinates to form a multi-modal image set; The generation module is configured to extract boundary features of a peripheral zone and a central zone of the prostate in the multi-modal image set by a generation unit of an attention generative adversarial network, generate an initial image containing a prostate region, and evaluate the initial image by a discrimination unit of the attention generative adversarial network to obtain a probability map representing whether each voxel belongs to the prostate region. The generation module is specifically configured to preprocess the multi-modal image set by a generation unit of an attention generative adversarial network, extract boundary features corresponding to the peripheral zone and the central zone of the prostate from the preprocessed result, fuse the boundary features with structural features in the first type of image and blood vessel features in the second type of image to generate an initial image containing a prostate region, compare the initial image with gland features in the multi-modal image set by a discrimination unit of the attention generative adversarial network, identify micro glands in the prostate region based on a comparison result to output difference information between the initial image and the gland features, enhance a focusing degree on the boundary features by the generation unit according to the difference information, adjust a fusion ratio of the structural features and the blood vessel features based on the enhanced focusing degree to generate a target image, compare the target image again by the discrimination unit and output new difference information, repeat a processing procedure of the generation unit and the discrimination unit until a change amplitude of the difference information is less than a preset difference threshold, and obtain a probability distribution of whether each voxel belongs to the prostate region based on the target image of the last generation to generate a probability map. The processing module is configured to convert the probability map into an initial voxel-level segmentation result by a convolution network, perform morphological processing on the initial voxel-level segmentation result to obtain a target voxel-level segmentation result in which isolated noise points are removed. The calculation module is configured to calculate a prostate MRI volume based on the target voxel-level segmentation result.

8. An electronic device, comprising: The memory is configured to store a computer program. The processor is configured to implement steps of the automatic prostate MRI volume calculation method based on the generative adversarial network according to any one of claims 1 to 6 when the computer program is executed. The computer readable storage medium stores a computer program, and the computer program can implement the automatic prostate MRI volume calculation method based on the generative adversarial network according to any one of claims 1 to 6 when executed by the processor.

9. A computer-readable storage medium, characterized in that, ​

Images