A medical image segmentation method and system

Abstract

Embodiments of the present specification provide a medical image segmentation method and system. The method comprises: acquiring a to-be-segmented medical image; segmenting the to-be-segmented medical image by using a first deep learning model to determine a first segmentation result; determining a second segmentation result by using a second deep learning model based on the to-be-segmented medical image and the first segmentation result, the second segmentation result comprising a segmentation region related to the first segmentation result, the second deep learning model being obtained by training based on a target sample medical image containing a segmentation difficult part; and determining a target segmentation result of the to-be-segmented medical image based on the first segmentation result and the second segmentation result.

Description

TECHNICAL FIELD

[0001] The present specification relates to the field of image processing, and in particular, to a medical image segmentation method and system. BACKGROUND

[0002] With the maturity of the application of deep learning in the field of medical image processing, more and more automatic organ segmentation methods begin to use pre-trained deep learning models to adaptively capture high-level features in medical images, which improves the organ segmentation speed and also obtains better generalization. However, due to the complex branching structure and topological morphology of the tree-like pipeline structure organ (for example, trachea), and the variation range of its pipe diameter size has more than ten levels, it has great challenge to the expression ability of the deep learning model. Especially for samples with image features such as distortion and blurred outline of tree-like pipeline structure organs caused by complex lesions, it further increases the fitting difficulty of the deep learning model.

[0003] Therefore, it is desirable to propose a medical image segmentation method and system which can quickly and accurately segment the tree structure organ, and at the same time, obtain the complete branching structure and topological morphology, so that the segmentation result has higher richness and integrity. SUMMARY

[0004] One aspect of the present specification provides a medical image segmentation method. The method comprises: obtaining a to-be-segmented medical image; segmenting the to-be-segmented medical image by using a first deep learning model to determine a first segmentation result; determining a second segmentation result by using a second deep learning model based on the to-be-segmented medical image and the first segmentation result, the second segmentation result including a segmentation region related to the first segmentation result, the second deep learning model being obtained by training based on a target sample medical image containing a segmentation difficult part; and determining a target segmentation result of the to-be-segmented medical image based on the first segmentation result and the second segmentation result.

[0005] In some embodiments, the determining the target segmentation result of the to-be-segmented medical image based on the first segmentation result and the second segmentation result comprises: determining an incremental region of the second segmentation result compared with the first segmentation result; updating the first segmentation result based on the incremental region to determine the target segmentation result of the to-be-segmented medical image.

[0006] In some embodiments, the determining the target segmentation result of the to-be-segmented medical image based on the first segmentation result and the second segmentation result comprises: taking the intersection of the first segmentation result and the second segmentation result as a seed region, performing seed growing on the union set of the first segmentation result and the second segmentation result, and determining the seed growing result as the target segmentation result of the to-be-segmented medical image.

[0007] In some embodiments, the seed growing on the union set of the first segmentation result and the second segmentation result comprises: 26-neighborhood seed growing on the union set of the first segmentation result and the second segmentation result.

[0008] In some embodiments, before determining the second segmentation result by using the second deep learning model, the method further comprises: determining whether a segmentation padding operation is needed based on the medical image to be segmented and the first segmentation result; when the segmentation padding operation is not needed, determining the first segmentation result as a target segmentation result of the medical image to be segmented; and when the segmentation padding operation is needed, inputting the medical image to be segmented and the first segmentation result into the second deep learning model as a two-channel input.

[0009] In some embodiments, the determining the target segmentation result of the medical image to be segmented based on the first segmentation result and the second segmentation result comprises: updating the first segmentation result based on the second segmentation result to determine an updated segmentation result; determining whether a current segmentation padding operation meets a termination condition; when the termination condition is met, determining the updated segmentation result as the target segmentation result of the medical image to be segmented; otherwise, performing at least one round of the segmentation padding operation by using the second deep learning model based on the medical image to be segmented and the updated segmentation result to determine the target segmentation result of the medical image to be segmented.

[0010] In some embodiments, the termination condition comprises: the segmentation padding operation reaches a preset iteration number, or an incremental area between the updated segmentation result and a segmentation result before updating is less than a preset threshold.

[0011] In some embodiments, the segmentation difficult part includes a part prone to breaking and a part prone to missing, and the training of the second deep learning model based on the target sample medical image containing the segmentation difficult part comprises: obtaining an initial sample medical image and determining a first segmentation prediction result of the initial sample medical image; identifying a segmentation difficult part in the initial sample medical image based on a difference between the first segmentation prediction result and a labeled gold standard; collecting a target sample medical image containing the segmentation difficult part, and taking the target sample medical image and the first segmentation prediction result as two-channel training samples; and taking the two-channel training samples as input to train an initialized second deep learning model to obtain the second deep learning model.

[0012] In some embodiments, the medical image comprises a CT image, and the segmentation comprises segmentation of a tree-like pipeline structure organ.

[0013] Another aspect of the present specification provides a medical image segmentation system. The system comprises: an acquisition module configured to acquire a medical image to be segmented; a first segmentation module configured to segment the medical image to be segmented by using a first deep learning model to determine a first segmentation result; a second segmentation module configured to determine a second segmentation result by using a second deep learning model based on the medical image to be segmented and the first segmentation result, the second segmentation result comprising a segmentation region related to the first segmentation result, the second deep learning model being obtained by training based on a target sample medical image comprising a segmentation difficult part; and a determination module configured to determine a target segmentation result of the medical image to be segmented based on the first segmentation result and the second segmentation result. BRIEF DESCRIPTION OF DRAWINGS

[0014] The present specification will be further described in the manner of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting, and in these embodiments, the same numbers represent the same structures, wherein:

[0015] Figure 1 is a schematic diagram of an application scenario of an exemplary medical image segmentation system according to some embodiments of the present specification;

[0016] Figure 2 is a schematic diagram of a module of an exemplary medical image segmentation system according to some embodiments of the present specification;

[0017] Figure 3 is a schematic diagram of a flow of an exemplary medical image segmentation method according to some embodiments of the present specification;

[0018] Figure 4 is a schematic diagram of a flow of an exemplary medical image segmentation method according to some other embodiments of the present specification;

[0019] Figure 5 is a schematic diagram of a flow of an exemplary CT image segmentation according to some embodiments of the present specification;

[0020] Figure 6 is a schematic diagram of a flow of an exemplary determination of a second deep learning model according to some embodiments of the present specification;

[0021] Figure 7 is a schematic diagram of an exemplary first segmentation result according to some embodiments of the present specification;

[0022] Figure 8 is a schematic diagram of an exemplary second segmentation result according to some embodiments of the present specification;

[0023] Figure 9is a schematic diagram of an exemplary target segmentation result according to some embodiments of the present specification. DETAILED DESCRIPTION

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present specification, the following will briefly introduce the drawings needed to be used in the embodiments description. Obviously, the drawings in the following description are only some examples or embodiments of the present specification, and for those skilled in the art, the present specification can also be applied to other similar scenarios without creative labor on the basis of these drawings. Unless it is clear from the language context or otherwise indicated, the same reference numbers in the drawings represent the same structure or operation.

[0025] It should be understood that the "system", "device", "unit" and / or "module" used herein is a method for distinguishing different components, elements, parts, sections or assemblies at different levels. However, if other words can achieve the same purpose, the words can be replaced by other expressions.

[0026] As shown in the specification and claims, unless the context clearly indicates otherwise, "a", "one", "an" and / or "the" do not refer to the singular, but can also include the plural. Generally speaking, the terms "comprise" and "include" only indicate the inclusion of the steps and elements explicitly identified, and these steps and elements do not constitute an exclusive list, and the method or device can also include other steps or elements.

[0027] Flowcharts are used in the present specification to illustrate the operations performed by the system according to the embodiments of the present specification, and the relevant description is to help better understand the medical imaging method and / or system. It should be understood that the preceding or subsequent operations are not necessarily performed in sequence. On the contrary, each step can be processed in reverse order or simultaneously. At the same time, other operations can also be added to these processes, or one or more steps of operation can be removed from these processes.

[0028] Figure 1 is a schematic diagram of an exemplary medical image segmentation system according to some embodiments of the present specification.

[0029] As Figure 1 shown, the medical image segmentation system 100 can include an imaging device 110, a processing device 120, one or more terminals 130, a storage device 140, and a network 150. The components in the medical image segmentation system 100 can be connected in one or more of various ways. For example only, as Figure 1As shown in FIG. 1, the imaging device 110 can be connected to the processing device 120 through the network 150. As another example, the imaging device 110 can be directly connected to the processing device 120, as the imaging device 110 and the processing device 120 can be connected as indicated by the dashed bidirectional arrow in FIG. 1. As yet another example, the storage device 140 can be connected to the processing device 120 directly (not shown in FIG. 1) or through the network 150. As still another example, one or more terminals 130 can be connected to the processing device 120 directly (as shown by the dashed bidirectional arrow connecting the terminal 130 and the processing device 120) or through the network 150. Figure 1

[0030] The imaging device 110 can be configured to scan a target object within a detection region to obtain scan data (e.g., a to-be-segmented medical image, an initial sample medical image, a target sample medical image, etc.) of the target object. In some embodiments, the target object can include a biological object and / or a non-biological object. For example, the target object can include a particular portion of a body, such as a head, a chest, an abdomen, etc., or a combination thereof. As another example, the target object can be a human-made component of a living or non-living organic and / or inorganic matter. In some embodiments, the scan data related to the target object can include projection data, one or more scan images, etc., of the target object.

[0031] In some embodiments, the imaging device 110 can include a non-invasive bioimaging apparatus for disease diagnosis or research purposes. For example, the imaging device 110 can include a single modality scanner and / or a multi-modality scanner. The single modality scanner can include, for example, an ultrasound scanner, an X-ray scanner, a computed tomography (CT) scanner, a magnetic resonance imaging (MRI) scanner, an ultrasonograph, a positron emission tomography (PET) scanner, an optical coherence tomography (OCT) scanner, an ultrasound (US) scanner, an intravascular ultrasound (IVUS) scanner, a near-infrared spectroscopy (NIRS) scanner, a far-infrared (FIR) scanner, etc. The multi-modality scanner can include, for example, an X-ray imaging-magnetic resonance imaging (X-ray-MRI) scanner, a positron emission tomography-X-ray imaging (PET-X-ray) scanner, a single photon emission computed tomography-magnetic resonance imaging (SPECT-MRI) scanner, a positron emission tomography-computed tomography (PET-CT) scanner, a digital subtraction angiography-magnetic resonance imaging (DSA-MRI) scanner, etc. The scanners provided above are for illustrative purposes only and are not intended to limit the scope of the present specification. As used herein, the term “imaging modality” or “modality” broadly refers to an imaging method or technique that collects, generates, processes, and / or analyzes imaging information of a target object.

[0032] ​In some embodiments, data acquired by the imaging device 110 (e.g., a medical image to be segmented, etc.) can be transmitted to the processing device 120 for further analysis. Additionally or alternatively, data acquired by the imaging device 110 can be sent to a terminal device (e.g., the terminal 130) for display and / or a storage device (e.g., the storage device 140) for storage.

[0033] The processing device 120 can process data and / or information acquired and / or extracted from the imaging device 110, the terminal 130, the storage device 140, and / or other storage devices. For example, the processing device 120 can acquire an initial sample medical image of a target object from the terminal 130 or the storage device 140, and determine a first segmentation prediction result of the initial sample medical image. As another example, the processing device 120 can acquire a medical image to be segmented from the imaging device 110, segment the medical image to be segmented using a first deep learning model to determine a first segmentation result, determine a second segmentation result based on the medical image to be segmented and the first segmentation result using a second deep learning model, and determine a target segmentation result based on the first segmentation result and the second segmentation result.

[0034] In some embodiments, the processing device 120 can be a single server or a group of servers. The group of servers can be centralized or distributed. In some embodiments, the processing device 120 can be local or remote. In some embodiments, the processing device 120 can be implemented on a cloud platform. By way of example only, the cloud platform can include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an on-premises cloud, a multi-cloud, etc., or any combination thereof.

[0035] In some embodiments, the processing device 120 can be implemented on a computing device. In some embodiments, the processing device 120 can be implemented on a terminal (e.g., the terminal 130). In some embodiments, the processing device 120 can be implemented on an imaging device (e.g., the imaging device 110). For example, the processing device 120 can be integrated in the terminal 130 and / or the imaging device 110.

[0036] The terminal 130 can be connected with the imaging device 110 and / or the processing device 120 for inputting / outputting information and / or data. For example, a user can interact with the imaging device 110 through the terminal 130 to control one or more components of the imaging device 110 (e.g., input patient information, etc.). As another example, the imaging device 110 can output generated medical images (e.g., a medical image to be segmented and / or a target segmentation result thereof) to the terminal 130 for display to the user.

[0037] In some embodiments, the terminal 130 can include a mobile device 131, a tablet computer 132, a laptop computer 133, and / or the like, or any combination thereof. In some embodiments, the mobile device 131 can include a smart home device, a wearable device, a smart mobile device, a virtual reality device, an augmented reality device, and / or the like, or any combination thereof. In some embodiments, one or more terminals 130 can remotely operate the imaging device 110. In some embodiments, the terminal 130 can operate the imaging device 110 via a wireless connection. In some embodiments, one or more terminals 130 can be part of the processing device 120. In some embodiments, the terminal 130 can be omitted.

[0038] The storage device 140 can store data and / or instructions. In some embodiments, the storage device 140 can store data acquired from the terminal 130 and / or the processing device 120. For example, the storage device 140 can store a sample medical image, a first segmentation result, a second segmentation result, a target segmentation result, and / or the like. In some embodiments, the storage device 140 can store data and / or instructions that the processing device 120 can execute or use to perform the exemplary methods described in this specification.

[0039] In some embodiments, the storage device 140 can include a mass storage device, a removable storage device, a volatile read / write memory, a read-only memory (ROM), and / or the like, or any combination thereof. An exemplary mass storage device can include a magnetic disk, an optical disk, a solid state drive, and / or the like. An exemplary removable storage device can include a flash drive, a floppy disk, an optical disk, a memory card, a compact disk, a magnetic tape, and / or the like. An exemplary volatile read / write memory can include a random access memory (RAM). In some embodiments, the storage device 140 can be implemented on a cloud platform. In some embodiments, the storage device 140 can be part of the processing device 120.

[0040] The network 150 can include any suitable network that can facilitate the exchange of information and / or data of the medical image segmentation system 100. In some embodiments, one or more components of the medical image segmentation system 100 (e.g., the imaging device 110, one or more terminals 130, the processing device 120, or the storage device 140) can communicate with one or more other components of the medical image segmentation system 100 to transmit information and / or data. In some embodiments, the network 150 can be any type of wired or wireless network or combination thereof. For example, the network 150 can be and / or include a public network (e.g., the Internet), a private network (e.g., a local area network (LAN), a wide area network (WAN), etc.), a wired network (e.g., an Ethernet network), a wireless network (e.g., an 802.11 network, a Wi-Fi network, etc.), a cellular network (e.g., a Long Term Evolution (LTE) network), a frame relay network, a virtual private network (VPN), a satellite network, a telephone network (e.g., a

[0041] It should be noted that the above description of the medical image segmentation system 100 is for illustrative purposes only and is not intended to limit the scope of the present specification. Various variations and modifications can be made to the present specification in light of the above teachings. However, such variations and modifications do not depart from the scope of the present specification. For example, the imaging device 110, the processing device 120, and the terminal 130 can share one storage device 140, or each can have a separate storage device.

[0042] With the maturity of the application of deep learning in the field of medical image processing, more and more automatic organ segmentation methods begin to use pre-trained deep learning models to adaptively capture high-level features in medical images, which improves the speed of organ segmentation and also obtains better generalization. However, due to the complex branching structure and topological morphology of the tree-like pipeline structure organ (e.g., trachea), and the variation range of its pipe diameter size has more than ten levels, it is a great challenge to the expression ability of a single deep learning model. The existing deep learning-based methods mostly treat tracheal segmentation as a single task and use a single deep learning model to process, ignoring the inherent bias brought by the training samples. For example, uniform sampling leads to a small proportion of difficult samples, and the trained model is not sensitive to the image features of difficult positions. Especially for samples with complex lesions leading to tracheal distortion, blurred outline and other image features, it further aggravates the fitting difficulty of the deep learning model.

[0043] This specification provides a medical image segmentation method and system. It utilizes a first deep learning model to segment the medical image to be segmented, determining a first segmentation result. Based on the medical image to be segmented and the first segmentation result, a second deep learning model is used to determine a second segmentation result, which includes segmented regions related to the first segmentation result. Then, based on the first and second segmentation results, a target segmentation result for the medical image to be segmented is determined. This approach can not only segment organs with tree-like tubular structures (e.g., tracheal trees) in a relatively short time, but also utilizes a second deep learning model highly sensitive to segmentation difficulties such as easily broken or missing parts for adaptive filling operations, improving the accuracy, richness, and completeness of the segmented tree-like tubular organ structures.

[0044] Figure 2 This is a schematic diagram of the modules of an exemplary medical image segmentation system according to some embodiments of this specification.

[0045] like Figure 2 As shown, in some embodiments, the medical image segmentation system 200 may include an acquisition module 210, a first segmentation module 220, a second segmentation module 230, and a determination module 240. In some embodiments, the medical image segmentation system 200 may be integrated into the imaging device 110 or the processing device 120.

[0046] The acquisition module 210 can be used to acquire medical images. In some embodiments, the acquisition module 210 can be used to acquire a medical image to be segmented. In some embodiments, the medical image may include a CT image.

[0047] In some embodiments, the acquisition module 210 can be used to acquire sample medical images. For example, the acquisition module 210 can be used to acquire initial sample medical images and / or target sample medical images.

[0048] The first segmentation module 220 can be used to determine a first segmentation result. In some embodiments, the first segmentation module 220 can be used to segment the medical image to be segmented using a first deep learning model to determine the first segmentation result.

[0049] In some embodiments, the first segmentation module 220 can be used to determine a first segmentation prediction result. For example, the first segmentation module 220 can be used to determine a first segmentation prediction result for an initial sample medical image.

[0050] The second segmentation module 230 can be used to determine a second segmentation result. In some embodiments, the second segmentation result may include a segmented region related to the first segmentation result. In some embodiments, the second segmentation module 230 can be used to determine the second segmentation result based on the medical image to be segmented and the first segmentation result using a second deep learning model.

[0051] In some embodiments, the second deep learning model can be trained based on target sample medical images containing segmentation difficult parts. In some embodiments, the segmentation difficult parts can include parts that are prone to breakage and missing. In some embodiments, the segmentation difficult parts in the initial sample medical images can be identified based on the difference between the first segmentation prediction result of the initial sample medical images and the annotation gold standard, the target sample medical images containing the segmentation difficult parts can be collected, and the target sample medical images and the first segmentation prediction result can be used as two-channel training samples. The initialized second deep learning model can be trained based on the two-channel training samples to obtain the second deep learning model.

[0052] In some embodiments, the second segmentation module 230 can determine whether segmentation filling operation is needed based on the medical image to be segmented and the first segmentation result. When the segmentation filling operation is needed, the second segmentation module 230 can input the medical image to be segmented and the first segmentation result into the second deep learning model as a two-channel input.

[0053] In some embodiments, the second segmentation module 230 can update the first segmentation result based on the second segmentation result to determine an updated segmentation result. In some embodiments, it can be determined whether the current segmentation filling operation meets a termination condition. When the termination condition is not met, the second segmentation module 230 can perform at least one round of segmentation filling operation based on the medical image to be segmented and the updated segmentation result using the second deep learning model to determine the target segmentation result of the medical image to be segmented. In some embodiments, the termination condition can include that the segmentation filling operation reaches a preset iteration number, or the incremental area between the updated segmentation result and the segmentation result before updating is less than a preset threshold.

[0054] The determination module 240 can be configured to determine the target segmentation result. In some embodiments, the determination module 240 can determine the target segmentation result of the medical image to be segmented based on the first segmentation result and the second segmentation result. In some embodiments, the determination module 240 can determine the incremental area of the second segmentation result compared to the first segmentation result, update the first segmentation result based on the incremental area, and determine the target segmentation result of the medical image to be segmented. In some embodiments, the determination module 240 can take the intersection of the first segmentation result and the second segmentation result as a seed region, perform seed growing on the union of the first segmentation result and the second segmentation result, and determine the seed growing result as the target segmentation result of the medical image to be segmented. In some embodiments, the determination module 240 can perform 26-neighborhood seed growing on the union of the first segmentation result and the second segmentation result.

[0055] In some embodiments, when the segmentation filling operation is not needed, the first segmentation result can be determined as the target segmentation result of the to-be-segmented medical image. In some embodiments, when the segmentation filling operation satisfies the termination condition, the updated segmentation result can be determined as the target segmentation result of the to-be-segmented medical image.

[0056] In some embodiments, the medical image segmentation system 200 can further include a training module (not shown in the figure) for obtaining the first deep learning model and / or the second deep learning model. For example, the training module can take the target sample medical image and the first segmentation prediction result as a two-channel training sample, and train the initialized second deep learning model by taking the two-channel training sample as input, to obtain the second deep learning model.

[0057] In some embodiments, the medical image segmentation system 200 can further include a data processing module (not shown in the figure) for data processing on the to-be-segmented medical image and / or the target segmentation result. For example, the data processing module can perform data preprocessing on the to-be-segmented CT image to obtain 3D volume data. The 3D volume data can reflect the three-dimensional structure information of the human organ, such as the volume, shape, size, etc. of the chest.

[0058] It should be noted that the above description of the medical image segmentation system 200 and its modules is for the convenience of description, and cannot limit the scope of the embodiments. It can be understood that, for those skilled in the art, after understanding the principle of the system, any combination of the modules or connection of the sub-systems with other modules can be made without departing from the principle. However, these modifications and changes are still within the scope of the present description.

[0059] Figure 3 is a flowchart of an exemplary medical image segmentation method according to some embodiments of the present description.

[0060] In some embodiments, the medical image segmentation method 300 can be performed by the imaging device 110 or the processing device 120. For example, the method 300 can be stored in the form of a program or instructions in a storage device (such as the storage device 140), and when the imaging device 110 or the processing device 120 executes the program or instructions, the method 300 can be implemented. In some embodiments, the method 300 can be performed by the medical image segmentation system 200.

[0061] In some embodiments, the step 310 can be performed by the processing device 120 or the acquisition module 210.

[0062] The to-be-segmented medical image can refer to a medical image of a target object that needs to be segmented, for example, a CT image, an MR image, a PET image, etc. In some embodiments, the to-be-segmented medical image can include a CT image, for example, a chest CT image, an abdominal CT image, etc. In some embodiments, the to-be-segmented medical image can include a 2D image, a 3D image, or a 4D image, etc. In some embodiments, the segmentation can include segmentation of a tree-like pipeline structure organ, for example, tracheal tree segmentation, blood vessel segmentation, etc.

[0063] In some embodiments, the to-be-segmented medical image can be obtained by an imaging device. For example, the to-be-segmented medical image can be obtained by scanning the target object by the imaging device 110. In some embodiments, the to-be-segmented medical image can be obtained from a storage device or a terminal. For example, the to-be-segmented medical image of the target object can be obtained from the storage device 140 or the terminal 130.

[0064] In step 320, the first deep learning model is used to segment the to-be-segmented medical image to determine a first segmentation result. In some embodiments, step 320 can be performed by the processing device 120 or the first segmentation module 220.

[0065] In some embodiments, the first deep learning model can be used to perform basic segmentation of organs in the medical image. For example, the first deep learning model can be used to perform preliminary tracheal extraction on the to-be-segmented medical image.

[0066] In some embodiments, the first deep learning model can be obtained based on sample images. In some embodiments, the sample images can include specific medical images sampled for a target region, and randomly sampled medical images. The medical images sampled for the target region can refer to medical images obtained by scanning the target object for a region that needs to be segmented, for example, a chest CT image. The randomly sampled medical images can refer to medical images obtained by uniform sampling of the entire image, which can or can not contain the target region. In some embodiments, the specific medical images sampled for the target region and the randomly sampled medical images can be obtained in a preset ratio and used as training samples for the first deep learning model. For example, 50% of the specific medical images and 50% of the randomly sampled medical images can be obtained to form the training samples for the first deep learning model. For another example, 60% of the specific medical images and 40% of the randomly sampled medical images can be obtained to form the training samples for the first deep learning model. In some embodiments, the preset ratio can be any other reasonable ratio, for example, 7:3, 8:2, etc., which is not limited in the present specification. In some embodiments, the first deep learning model can include but is not limited to a convolutional neural network model, a VB-Net neural network model, etc., which is not limited in the present specification.

[0067] By way of example only, 50% of the specific medical images can be acquired around the target region of the target object guided by a labeling gold standard (e.g., tracheal tree), and the remaining 50% of the images can be acquired using full-image uniform random sampling, and the two parts of images are used to form training samples. Further, the training samples can be used as the input of an initialized VB-Net neural network model to train a first deep learning model. The labeling gold standard can reflect the organ tissue contained in the medical image that needs to be segmented. For example, for tracheal tree segmentation, the labeling gold standard can be the tracheal tree structure marked (e.g., delineated, circled, etc.) in the corresponding CT image.

[0068] In some embodiments, the input of the first deep learning model can be a medical image to be segmented, and the output can be an organ segmentation mask. For example, a chest CT image 3D volume data can be input into the first deep learning model as an input, and a corresponding first segmentation result, a tracheal segmentation mask, can be determined. As another example, a 2D planar image of a target object can be input into the first deep learning model as an input, and a corresponding first segmentation result can be determined. It can be understood that the foregoing description is by way of example only, and in some embodiments, the input of the first deep learning model can be 2D image and / or 3D image data, which is not limited in the present specification.

[0069] In some embodiments, the medical image to be segmented can be pre-processed, and the pre-processed image can be used as the input of the first deep learning model to determine the first segmentation result. For example, the pre-processing can include but is not limited to data conversion, image window level and window width normalization, etc. In some embodiments, the corresponding 3D volume data can be obtained by pre-processing the medical image to be segmented.

[0070] Due to the influence of noise and other interference in medical image acquisition, the characteristics of tree-like organ structure (e.g., the trachea) that the closer to the end, the thinner, organ distortion caused by lesions (e.g., tracheal stenosis) or blurred visualization (e.g., motion-induced artifacts, etc.), the first segmentation result predicted by the first deep learning model is prone to have more organ missing and breaking problems near the end or the lesion area. By way of example only, Figure 7 is a schematic diagram of a first tracheal segmentation mask of an exemplary CT chest image according to some embodiments of the present specification, as shown in Figure 7 As shown in the dashed box, the end of the first tracheal segmentation mask has more organ missing and breaking problems.

[0071] Step 330, based on the medical image to be segmented and the first segmentation result, a second segmentation result is determined using a second deep learning model. In some embodiments, step 330 can be performed by processing device 120 or second segmentation module 230.

[0072] In some embodiments, the second deep learning model can be used to segment the difficult-to-segment part in the medical image to be segmented. In some embodiments, the difficult-to-segment part can include a part that is prone to breaking and missing (for example, Figure 7 The part missing and broken in the first trachea segmentation mask in the dashed box). In some embodiments, the second deep learning model can be obtained based on the target sample medical image containing the difficult-to-segment part. In some embodiments, the second deep learning model can include, but is not limited to, a convolutional neural network model, a VB-Net neural network model, etc., and the present specification does not limit this. More about the training of the second deep learning model can be found in Figure 6 and the related description, which will not be repeated here.

[0073] In some embodiments, the input of the second deep learning model can include the medical image to be segmented and the first segmentation result, and the output is the second segmentation result containing the difficult-to-segment part. For example, the medical image to be segmented and the first segmentation result can be input as two channels to the second deep learning model to obtain the corresponding second trachea segmentation mask.

[0074] In some embodiments, the second segmentation result includes a segmentation region related to the first segmentation result. The segmentation region related to the first segmentation result can refer to the part missing and / or broken in the first segmentation result. For example only, Figure 8 is a schematic diagram of the second segmentation result of an exemplary CT chest image according to some embodiments of the present specification, Figure 8 The second trachea segmentation mask in the dashed box contains Figure 7 The part missing and broken in the first trachea segmentation mask in the dashed box.

[0075] In some embodiments, based on the medical image to be segmented and the first segmentation result, it can be determined whether the segmentation filling operation is needed, and when needed, the medical image to be segmented and the first segmentation result are input as a two-channel input to the second deep learning model. In some embodiments, at least one round of segmentation filling operation can be performed iteratively using the second deep learning model. More related content can be found in Figure 4 and the related description, which will not be repeated here.

[0076] Step 340, determining the target segmentation result of the medical image to be segmented based on the first segmentation result and the second segmentation result. In some embodiments, step 340 can be performed by processing device 120 or determination module 240.

[0077] In some embodiments, the incremental region of the second segmentation result compared to the first segmentation result can be determined; the first segmentation result is updated based on the incremental region to determine the target segmentation result of the medical image to be segmented.

[0078] In some embodiments, the intersection of the first segmentation result and the second segmentation result can be taken as a seed region, seed growing is performed on the union of the first segmentation result and the second segmentation result, and the seed growing result is determined as the target segmentation result of the medical image to be segmented. In some embodiments, 26-neighborhood seed growing can be performed on the union of the first segmentation result and the second segmentation result to determine the target segmentation result. For example, in a 3-dimensional image, the processing device 120 can perform seed growing in all neighborhoods, i.e., 26 neighborhoods, of the top, bottom, front, back, left, and right of the union of the first segmentation result and the second segmentation result, respectively.

[0079] In some embodiments, the first segmentation result can be updated by any feasible region growing method such as connected component analysis, seed region growing, etc. to determine the target segmentation result, and the present specification does not limit this.

[0080] In some embodiments, when it is determined that no subsequent segmentation filling operation is needed, the current segmentation result of the medical image to be segmented can be determined as the target segmentation result. For example, when it is first determined that no segmentation filling operation is needed, the first segmentation result can be determined as the target segmentation result. For another example, during the segmentation filling operation, when the current segmentation filling operation satisfies the termination condition, the updated segmentation result obtained before the determination can be determined as the target segmentation result.

[0081] The first deep learning model trained by the sample images composed of the specific medical images sampled for the target region and / or the randomly sampled medical images performs preliminary organ extraction in the first stage; for the end branch missing and local breakage existing in the first-stage first segmentation result, the second-stage uses the target sample medical images collected for the difficult segmentation part and the first segmentation prediction result pair as input, trains to obtain a second deep learning model more sensitive to the image features of the difficult segmentation part, and realizes fine segmentation of the tree-structured organ at the missing and broken part. By using two separately trained deep learning models for two-stage segmentation, the entire framework has higher robustness and generalization for diversified medical images (for example, chest CT images), organ (for example, trachea) distortion caused by lesions, or poor quality data, improves the segmentation efficiency, and improves the richness and completeness of the segmentation result.

[0082] It should be noted that the above description of the method 300 is merely for example and illustration, and does not limit the scope of the present specification. Various modifications and changes can be made to the method 300 under the guidance of the present specification by those skilled in the art. However, these modifications and changes are still within the scope of the present specification.

[0083] Figure 4is a flowchart of an exemplary medical image segmentation method according to some embodiments of the present specification.

[0084] As shown in Figure 4 In some embodiments, as shown in the middle, a determination can be made as to whether a segmentation filling operation is needed. When needed, at least one iteration of segmentation filling operation can be performed using the second deep learning model until the current segmentation filling operation meets a termination condition.

[0085] In some embodiments, the medical image segmentation method 400 can be performed by the imaging device 110 or the processing device 120. For example, the method 400 can be stored in the form of a program or instructions in a storage device (such as the storage device 140), which can be implemented when the imaging device 110 or the processing device 120 executes the program or instructions. In some embodiments, the method 400 can be performed by the medical image segmentation system 200.

[0086] At step 410, a determination is made as to whether a segmentation filling operation is needed based on the medical image to be segmented and the first segmentation result. In some embodiments, step 410 can be performed by the processing device 120 or the medical image segmentation system 200 (e.g., the second segmentation module 230).

[0087] The segmentation filling operation is a process of performing a second segmentation using the second deep learning model and updating the first segmentation result or the updated segmentation result based on the obtained second segmentation result to obtain an organ segmentation result (i.e., a target segmentation result) that is complete or meets a preset condition (e.g., meets a diagnostic requirement).

[0088] In some embodiments, a determination can be made as to whether a segmentation filling operation is needed based on a difference between the first segmentation result 413 and a ground truth of the medical image to be segmented 415. For example, a region difference between the first segmentation result 413 and the ground truth can be calculated. When the region difference is less than a preset threshold, it is determined that a segmentation filling operation is not needed, otherwise, it is determined that a segmentation filling operation is needed.

[0089] In some embodiments, when a segmentation filling operation is not needed, the first segmentation result can be determined as the target segmentation result 430 of the medical image to be segmented; otherwise, step 420 is entered.

[0090] At step 420, a segmentation filling operation is performed on the first segmentation result using the second deep learning model. In some embodiments, step 420 can be performed by the processing device 120 or the medical image segmentation system 200 (e.g., the second segmentation module 230).

[0091] In some embodiments, the medical image to be segmented 415 and the first segmentation result 413 can be combined into a dual-channel input, which is input to the second deep learning model to determine the second segmentation result 425.

[0092] In some embodiments, the first segmentation result 413 can be updated based on the second segmentation result 425 to determine an updated segmentation result 429. In some embodiments, an incremental region of the second segmentation result 425 compared to the first segmentation result 413 can be determined, and the first segmentation result 413 can be updated based on the incremental region to determine the updated segmentation result 429. In some embodiments, an intersection of the first segmentation result 413 and the second segmentation result 425 can be taken as a seed region, and a seed growing can be performed on a union of the first segmentation result 413 and the second segmentation result 425 to determine the updated segmentation result 429. In some embodiments, a 26-neighborhood seed growing can be performed on the union of the first segmentation result 413 and the second segmentation result 425 to determine the updated segmentation result 429.

[0093] In some embodiments, it can be determined whether the current segmentation filling operation satisfies a termination condition. In some embodiments, the termination condition can include that the segmentation filling operation reaches a preset number of iterations, or that an incremental region between the updated segmentation result and the segmentation result before updating is less than a preset threshold.

[0094] The preset number of iterations can reflect the number of rounds of segmentation filling operations that need to be performed. Each time the process of performing the second-stage segmentation based on the second segmentation result to determine the updated segmentation result is performed, it can be defined as one round of segmentation filling operation. For example, based on the first segmentation result, the medical image to be segmented, using the second deep learning model to obtain the second segmentation result, and based on the first segmentation result and the second segmentation result to determine the updated segmentation result, it can be defined as one round of segmentation filling operation; based on the updated segmentation result, the medical image to be segmented, using the second deep learning model to obtain the second segmentation result, based on the updated segmentation result and the second segmentation result to determine a new updated segmentation result, it can be defined as another round of segmentation filling operation.

[0095] In some embodiments, one or more rounds of the segmentation and padding operation can be performed. For example, the current segmentation and padding operation can be determined to satisfy the termination condition when the current segmentation and padding operation reaches a preset number of iterations, and subsequent segmentation and padding operations can be stopped. In some embodiments, the preset number of iterations can be automatically determined by the processing device based on historical data. For example, the processing device 120 can determine the preset number of iterations based on the number of segmentation and padding operations corresponding to the target segmentation result in historical segmentation operations. In some embodiments, the preset number of iterations can be determined based on clinical requirements. For example, the completeness and richness of the tracheal tree segmentation result required can be determined according to user settings or custom diagnostic requirements to determine the preset number of iterations. In some embodiments, the preset number of iterations can be determined based on user habits. For example, the corresponding preset number of iterations can be determined for each user based on the user's reading habits, historical diagnosis habits, and the like. In some embodiments, the preset number of iterations can be manually set by the user. For example, the doctor can set the number of iterations required through the terminal 130.

[0096] In some embodiments, the segmentation result before updating can include the first segmentation result, or the updated segmentation result obtained in the last round of segmentation and padding operation. In some embodiments, the preset threshold can be manually set by the user, or automatically determined by the processing device based on historical data and / or clinical requirements, and the like.

[0097] In some embodiments, when the termination condition is satisfied, the updated segmentation result 429 corresponding to the current segmentation and padding operation can be determined as the target segmentation result 430 of the to-be-segmented medical image 415. In some embodiments, when the termination condition is not satisfied, a new round of segmentation and padding operation can be performed based on the to-be-segmented medical image 415 and the updated segmentation result 429 using the second deep learning model 423 until the termination condition is satisfied to determine the target segmentation result 430 of the to-be-segmented medical image 415. For example, the to-be-segmented medical image 415 and the updated segmentation result 429 can be input as a dual-channel input into the second deep learning model 423 to determine a new second segmentation result 425, and the updated segmentation result 429 can be updated based on the determined second segmentation result 425 until the current segmentation and padding operation satisfies the termination condition.

[0098] Figure 9 is a schematic diagram of an example target segmentation result according to some embodiments of the present specification, as shown in the dashed box, the endotracheal mask 910 is filled compared to the endotracheal mask 710 at the same position in Figure 7 the missing and broken parts of the tracheal end are filled. Figure 9The lower part is an enlarged view of the trachea mask 910, and the trachea mask 920 is one of the tracheas in 910, as shown in the ellipse. The upper dashed line represents the first-stage segmentation result of the trachea, and the lower dashed line represents the missing part obtained through the three-round segmentation filling operation. The two parts together constitute the complete information of the trachea.

[0099] Although the second deep learning model is more sensitive to difficult segmentation sites of missing and broken parts, the predicted second segmentation result has fewer broken and missing parts, but at the same time, there are missegmentations outside the tree structure organ. For example Figure 8 As shown in the figure, there are some small fragments (shown in solid ellipses 830 and 840) around the trachea in the second trachea segmentation mask obtained by the second deep learning model. By determining the updated segmentation result based on the second segmentation result and the first segmentation result, the existence of missegmentation can be avoided while ensuring the completeness and richness of the segmentation result. In addition, the local region growth is used in the segmentation filling operation, which avoids introducing new missegmentations (regions not adjacent to the previous segmentation) in the filling and improves the reliability of the filling.

[0100] It should be noted that the above description of the method 400 is only for example and illustration, and does not limit the scope of the present specification. Those skilled in the art can make various modifications and changes to the method 400 under the guidance of the present specification. However, these modifications and changes are still within the scope of the present specification.

[0101] Figure 5 is a schematic diagram of an example CT image segmentation process according to some embodiments of the present specification.

[0102] As Figure 5 shown, taking a CT image as an example, in some embodiments, the processing device can perform tracheal tree segmentation on the CT image based on a two-stage deep learning model to obtain a complete segmentation result of the CT image.

[0103] In some embodiments, the CT image can be obtained from an imaging device (for example, the imaging device 110). In some embodiments, the CT image can be a chest CT image. In some embodiments, the CT image can include DICOM image data. For more information about image acquisition, please refer to Figure 3 and related descriptions, which will not be repeated here.

[0104] At step 520, data preprocessing is performed on the CT image to obtain 3D volume data. In some embodiments, the chest CT image data obtained at step 510 can be preprocessed to obtain chest 3D volume data. For example, the chest 3D volume data can be obtained by converting DICOM data to volume data, normalizing the window level and / or window width of image CT values, and the like. For example, [-625, 425] can be used as the mean and standard deviation to achieve normalization. In some embodiments, preprocessing can also include noise reduction, enhancement, and the like, which are not limited in the present specification.

[0105] At step 530, the 3D volume data is input into the first deep learning model to obtain a first segmentation result. In some embodiments, the 3D volume data (e.g., chest 3D volume data) can be input into the first deep learning model as input to obtain a first segmentation result (e.g., a first trachea segmentation mask as shown in FIG. 6B). Figure 7

[0106] At step 540, a segmentation filling operation is performed based on the first segmentation result.

[0107] In some embodiments, whether to perform the segmentation filling operation can be determined based on the CT image 3D volume data and the first segmentation result output by the first deep learning model. When the segmentation filling operation is not needed, the current segmentation result can be determined as the complete segmentation result of the CT image, i.e., the target segmentation result. For example, when it is determined for the first time based on the first segmentation result and the 3D volume data that the segmentation filling operation is not needed, the trachea segmentation mask corresponding to the first segmentation result can be determined as the complete segmentation result of the CT image. In some embodiments, when the segmentation filling operation is needed, the 3D volume data and the corresponding first segmentation result can be jointly input into the second deep learning model as a dual-channel input to determine a second segmentation result (e.g., a second trachea segmentation mask as shown in FIG. 6C), and the first segmentation result is updated based on the second segmentation result. Figure 8

[0108] In some embodiments, an incremental region of the second segmentation result compared to the first segmentation result can be determined; the first segmentation result is updated based on the incremental region to determine an updated segmentation result. In some embodiments, the segmentation result can be updated by connected component analysis, seed growing, and the like. For example, the intersection of the first trachea segmentation mask obtained by the first deep learning model as shown in FIG. 6B and the second trachea segmentation mask obtained by the second deep learning model as shown in FIG. 6C can be used as a seed region, and 26-neighborhood seed growing is performed on the union set of the first trachea segmentation mask and the second trachea segmentation mask, and the result of the seed growing is used as the filling updated trachea segmentation mask, i.e., the updated segmentation result. Figure 7 Figure 8 In some embodiments, the intersection of the first trachea segmentation mask obtained by the first deep learning model as shown in FIG. 6B and the second trachea segmentation mask obtained by the second deep learning model as shown in FIG. 6C can be used as a seed region, and 26-neighborhood seed growing is performed on the union set of the first trachea segmentation mask and the second trachea segmentation mask, and the result of the seed growing is used as the filling updated trachea segmentation mask, i.e., the updated segmentation result.​​​

[0109] In some embodiments, the updated segmentation result can be taken as an input of a new round of segmentation filling operation to determine whether the termination condition is met. In some embodiments, when the termination condition is not met, the updated segmentation result and the CT image 3D volume data can be taken as new dual-channel inputs to the second deep learning model to determine a new second segmentation result. In some embodiments, the updated segmentation result obtained in the last round of segmentation filling operation can be updated based on the new second segmentation result, and it is determined again whether the termination condition is met, and based on this iteration, the execution is performed until the termination condition is met. For example, the incremental region of the updated trachea segmentation mask and the previous trachea segmentation mask can be determined, and when the incremental region is less than a preset threshold, it is determined that the termination condition is met, and the segmentation filling operation is terminated; otherwise, the updated trachea segmentation mask is taken as an input of a new round, and the segmentation filling operation is performed again.

[0110] Step 550, determining the complete segmentation result of the CT image.

[0111] In some embodiments, when the current segmentation filling operation meets the termination condition, the current segmentation result can be determined as the complete segmentation result of the CT image. For example, the updated trachea segmentation mask that meets the termination condition can be determined as the complete trachea segmentation result of the chest CT image.

[0112] It should be noted that the above description of the method 500 is only for example and illustration, and does not limit the scope of the present specification. Those skilled in the art can make various modifications and changes to the method 500 under the guidance of the present specification. However, these modifications and changes are still within the scope of the present specification.

[0113] Figure 6 is a schematic flow diagram of an example of determining a second deep learning model according to some embodiments of the present specification.

[0114] In some embodiments, the method 600 can be performed by the imaging device 110 or the processing device 120. For example, the method 600 can be stored in the form of a program or instructions in a storage device (such as the storage device 140), and when the imaging device 110 or the processing device 120 executes the program or instructions, the method 600 can be implemented. In some embodiments, the method 600 can be performed by the medical image segmentation system 200 (for example, the training module).

[0115] Step 610, obtaining an initial sample medical image and determining a first segmentation prediction result of the initial sample medical image.

[0116] In some embodiments, the initial sample medical image can include one or more medical images. In some embodiments, the initial sample medical image can be acquired from a storage device (e.g., the storage device 140), an imaging device (e.g., the imaging device 110), or a terminal device (e.g., the terminal 130).

[0117] In some embodiments, the first segmentation prediction result of the initial sample medical image can be determined by using the first deep learning model. For example, the initial sample medical image 613 can be input into the first deep learning model 615 to obtain the corresponding first segmentation prediction result 617.

[0118] At step 620, a segmentation difficult part in the initial sample medical image is identified.

[0119] In some embodiments, the segmentation difficult part in the initial sample medical image can be identified based on the difference between the first segmentation prediction result 617 and the annotation gold standard. In some embodiments, the difference between the first segmentation prediction result 617 and the annotation gold standard can be determined as the corresponding segmentation difficult part.

[0120] In some embodiments, the target sample medical image 630 containing the segmentation difficult part can be acquired in a targeted manner, and the target sample medical image 630 and the first segmentation prediction result 617 can be used as two-channel training samples. For example, the target sample medical image 630 can be scanned and acquired for the broken and missing parts in the first segmentation prediction result 617, and the target sample medical image 630 and the first segmentation prediction result 617 can be used as a training sample pair.

[0121] In some embodiments, the determined two-channel training samples (i.e., the target sample medical image 630 and the first segmentation prediction result 617) can be used as input to train the initialized second deep learning model 640 to obtain the trained second deep learning model 650.

[0122] Since the second deep learning model is trained in a targeted manner using the target sample medical image containing the segmentation difficult part, and the first segmentation result (with broken and missing parts) is used as a guide, the second deep learning model is more sensitive to the areas with missing and broken results, so that the predicted segmentation mask for filling (i.e., the second segmentation result) has fewer broken and missing parts.

[0123] It should be noted that the above description of the method 600 is only for example and illustration, and does not limit the scope of the present specification. Various modifications and changes can be made to the method 600 under the guidance of the present specification. However, these modifications and changes are still within the scope of the present specification.

[0124] Some embodiments of the present specification also provide a computer readable storage medium storing computer instructions, when the computer instructions are read by a computer, the computer executes the medical image segmentation method (for example, the method 300-method 600) described above.

[0125] Some embodiments of the present specification also provide a medical image segmentation device, which comprises a processor and a memory, the memory is used to store instructions, when the processor executes the instructions, the medical image segmentation method (for example, the method 300-method 600) described above is realized.

[0126] The beneficial effects brought by the embodiments of the present specification may include but are not limited to: (1) a second deep learning model is trained by using a sample medical image containing a difficult-to-segment part, and the obtained model is more sensitive to the characteristics of the part prone to breaking and missing, which helps to improve the segmentation accuracy of the difficult-to-segment part; (2) the second segmentation result obtained by using the second deep learning model which is highly sensitive to the difficult-to-segment part is used to adaptively fill the first segmentation result obtained by using the first deep learning model, which improves the accuracy, richness and integrity of the segmented tree-shaped trachea, and at the same time, complete segmentation can be realized in a relatively short time; (3) using a two-stage deep learning model for organ segmentation can make the whole framework have higher robustness and generalization for diversified medical images (for example, chest CT images) and data with poor scanning quality caused by organ distortion due to lesions; (4) by setting the iteration termination condition of the segmentation filling operation, different degrees of organ segmentation demand scenarios can be more flexibly adapted; (5) by taking the intersection of the first segmentation result and the second segmentation result as a seed region and performing seed growing on the union set of the first segmentation result and the second segmentation result, the special local region growing is used to determine the fine segmentation region that needs to be filled into the previous segmentation, which ensures that only the fine segmentation region adjacent to the previous segmentation is filled into the previous segmentation, so that new mis-segmentation can be avoided.

[0127] It should be noted that different embodiments may produce different beneficial effects, and in different embodiments, the beneficial effects that may be produced can be any one or a combination of the above, or any other beneficial effects that can be obtained.

[0128] The above has described the basic concepts, and it is obvious that the above detailed disclosure is only used as an example and does not limit the present specification. Although it is not explicitly stated here, those skilled in the art can make various modifications, improvements and corrections to the present specification. Such modifications, improvements and corrections are suggested in the present specification, so such modifications, improvements and corrections still belong to the spirit and scope of the exemplary embodiments of the present specification.

[0129] Furthermore, this specification uses specific terms to describe embodiments thereof. For example, "an embodiment," "one embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic associated with at least one embodiment of this specification. Therefore, it should be emphasized and noted that references to "an embodiment," "one embodiment," or "an alternative embodiment" in different locations throughout this specification do not necessarily refer to the same embodiment. Moreover, certain features, structures, or characteristics in one or more embodiments of this specification can be appropriately combined.

[0130] Furthermore, unless expressly stated in the claims, the order of processing elements and sequences, the use of numbers and letters, or other names described in this specification are not intended to limit the order of the processes and methods described herein. Although various examples have been discussed in the foregoing disclosure of some embodiments of the invention that are currently considered useful, it should be understood that such details are for illustrative purposes only, and the appended claims are not limited to the disclosed embodiments; rather, the claims are intended to cover all modifications and equivalent combinations that conform to the spirit and scope of the embodiments described herein. For example, while the system components described above can be implemented using hardware devices, they can also be implemented solely using software solutions, such as installing the described system on existing servers or mobile devices.

[0131] Similarly, it should be noted that, in order to simplify the description disclosed herein and thus aid in the understanding of one or more embodiments of the invention, the foregoing description of embodiments in this specification may sometimes combine multiple features into a single embodiment, drawing, or description thereof. However, this method of disclosure does not imply that the subject matter of this specification requires more features than those mentioned in the claims. In fact, the embodiments contain fewer features than all the features of a single embodiment disclosed above.

[0132] In some embodiments, numbers describing the quantity of components and attributes are used. It should be understood that such numbers used in the description of embodiments are modified in some examples with the terms "approximately," "approximately," or "generally." Unless otherwise stated, "approximately," "approximately," or "generally" indicates that the numbers are allowed to vary by ±20%. Accordingly, in some embodiments, the numerical parameters used in the specification and claims are approximate values, which may be changed depending on the characteristics required by individual embodiments. In some embodiments, numerical parameters should take into account specified significant digits and employ a general method of digit reservation. Although the numerical ranges and parameters used to confirm their breadth of range in some embodiments of this specification are approximate values, in specific embodiments, such values ​​are set as precisely as feasible.

[0133] Each patent, patent application, patent publication, and other material cited in this specification is hereby incorporated by reference in its entirety herein for the teachings relevant to the sentence and / or paragraph in which the reference is presented. Document histories, to the extent not inconsistent with the pertinent U.S. patent application file history, are also incorporated by reference herein. To the extent that material incorporated by reference contradicts or contradicts any portion of this specification, including definition, the portion of the material incorporated by reference prevails. Note, however, that in the event of inconsistencies between any such material and the present specification, including definitions, the present specification, including definitions, will control.

[0134] Finally, it should be understood that the embodiments described herein are merely exemplary of the principles of the present description. Other embodiments can be devised without departing from the scope of the present description. Accordingly, the embodiments described herein are not intended to limit the scope of the present description, but rather are intended to be exemplary thereof.

Claims

1. A medical image segmentation method characterized by, The method comprises: obtaining a medical image to be segmented; segmenting the medical image to be segmented by using a first deep learning model to determine a first segmentation result; based on the medical image to be segmented and the first segmentation result, determining a second segmentation result by using a second deep learning model, the second segmentation result comprising a segmentation region related to the first segmentation result, the second deep learning model being obtained by training based on a target sample medical image containing a segmentation difficult part; wherein the second segmentation result is obtained by inputting the medical image to be segmented and the first segmentation result as two channels into the second deep learning model; based on the first segmentation result and the second segmentation result, determining a target segmentation result of the medical image to be segmented.

2. The method of claim 1, wherein, The method comprises: determining an incremental region of the second segmentation result compared with the first segmentation result; updating the first segmentation result based on the incremental region to determine the target segmentation result of the medical image to be segmented.

3. The method of claim 1, wherein, The method comprises: taking the intersection of the first segmentation result and the second segmentation result as a seed region, performing seed growing on the union set of the first segmentation result and the second segmentation result, and determining the seed growing result as the target segmentation result of the medical image to be segmented.

4. The method of claim 3, wherein, The method comprises: performing 26-neighborhood seed growing on the union set of the first segmentation result and the second segmentation result.

5. The method of claim 1, wherein, Before determining the second segmentation result by using the second deep learning model, the method further comprises: based on the medical image to be segmented and the first segmentation result, determining whether a segmentation filling operation is needed; when the segmentation filling operation is not needed, determining the first segmentation result as the target segmentation result of the medical image to be segmented; when the segmentation filling operation is needed, inputting the medical image to be segmented and the first segmentation result into the second deep learning model as a two-channel input.

6. The method of claim 5, wherein, The method comprises: updating the first segmentation result based on the second segmentation result to determine an updated segmentation result; determining whether the current segmentation filling operation meets a termination condition; when the termination condition is met, determining the updated segmentation result as the target segmentation result of the medical image to be segmented; otherwise, based on the medical image to be segmented and the updated segmentation result, performing at least one round of the segmentation filling operation by using the second deep learning model to determine the target segmentation result of the medical image to be segmented.

7. The method of claim 6, wherein, The termination condition comprises: the segmentation filling operation reaches a preset iteration number, or an incremental region between the updated segmentation result and a segmentation result before updating is less than a preset threshold.

8. The method of claim 1, wherein, The difficult-to-segment positions include positions prone to breaking and positions prone to missing, and training the second deep learning model based on target sample medical images containing difficult-to-segment positions comprises: an initial sample medical image is acquired, and a first segmentation prediction result of the initial sample medical image is determined; difficult-to-segment positions in the initial sample medical image are identified based on differences between the first segmentation prediction result and a labeling gold standard; target sample medical images containing the difficult-to-segment positions are collected in a targeted manner, and the target sample medical images and the first segmentation prediction result are used as two-channel training samples; the initialized second deep learning model is trained by taking the two-channel training samples as input, and the second deep learning model is obtained.

9. The method of claim 1, wherein, The medical images include CT images, and the segmentation includes segmentation of an organ with a tree-like pipeline structure.

10. A medical image segmentation system, characterized by, comprises: an acquisition module configured to acquire a medical image to be segmented; a first segmentation module configured to segment the medical image to be segmented by using a first deep learning model to determine a first segmentation result; a second segmentation module configured to determine a second segmentation result by using a second deep learning model based on the medical image to be segmented and the first segmentation result, the second segmentation result including a segmentation region related to the first segmentation result, the second deep learning model being obtained by training based on target sample medical images containing difficult-to-segment positions; wherein the second segmentation result is obtained by inputting the medical image to be segmented and the first segmentation result as two-channel input to the second deep learning model; a determination module configured to determine a target segmentation result of the medical image to be segmented based on the first segmentation result and the second segmentation result.

Images