CN116030072A

CN116030072A - Medical image processing method, medical image processing device, computer equipment and storage medium

Info

Publication number: CN116030072A
Application number: CN202111255694.XA
Authority: CN
Inventors: 陈俊强
Original assignee: Shanghai Weiwei Medical Technology Co ltd
Current assignee: Shanghai Weiwei Medical Technology Co ltd
Priority date: 2021-10-27
Filing date: 2021-10-27
Publication date: 2023-04-28

Abstract

The application relates to a medical image processing method, a medical image processing device, a computer device and a storage medium. The method comprises the following steps: acquiring a medical image to be processed; carrying out probability regression processing or probability threshold binarization processing on the medical image to be processed to obtain at least one initial region; classifying all the obtained initial areas to obtain a target area comprising a target object; and dividing the target area to obtain a target object. The method comprises the steps of firstly processing the medical image to be processed to obtain an initial region, then classifying the initial region to obtain a target region comprising a target object, and finally dividing the target object only in the target region, so that on one hand, the accuracy is improved through multi-step processing, and on the other hand, a part of data is reduced in each step of processing, so that the efficiency is also improved.

Description

Medical image processing method, medical image processing device, computer equipment and storage medium

Technical Field

The present disclosure relates to the field of industrial intelligence, and in particular, to a medical image processing method, apparatus, computer device, and storage medium.

Background

Cerebral microhemorrhages are old blood Huang Tiezhi (hemosiorin) deposits left by asymptomatic microhemorrhages of the brain. It can only be found by high intensity MRI or Gradient-echo T2 weighted MRI (Gradient-echo T2 weighted MRI) that it is manifested as multiple tiny MRI signal loss, accumulating in subcortical grey matter, and also seen in the cortex and brain bridge. Micro-bleeding is an early warning signal for amyloid angiopathy and hypertensive angiopathy ICH.

Because the micro-bleeding of the brain occurs in brain parenchyma, and the method has the characteristics of small volume, large quantity and the like, the manual marking method has the problems of time consumption, difficult identification and the like, so the automatic detection of the micro-bleeding point of the brain is particularly important. In the conventional technology, the treatment of cerebral micro-bleeding is usually performed by a semi-automatic detection method.

However, in the existing semi-automatic detection method, the detection function of the cerebral micro-bleeding position is realized by preprocessing the cerebral image, extracting the image characteristics, finally training a classifier to distinguish whether the current position has a focus or not, and the like.

Disclosure of Invention

In view of the foregoing, it is desirable to provide a medical image processing method, apparatus, computer device, and storage medium capable of improving accuracy and efficiency.

A medical image processing method, the method comprising:

acquiring a medical image to be processed;

carrying out probability regression processing or probability threshold binarization processing on the medical image to be processed to obtain at least one initial region;

classifying all the obtained initial areas to obtain a target area comprising a target object;

and dividing the target area to obtain a target object.

In one embodiment, the performing probability regression or probability threshold binarization on the medical image to be processed to obtain at least one initial region includes:

carrying out probability regression or probability threshold binarization processing on the medical image to be processed to obtain probability distribution corresponding to the medical image to be processed;

performing binarization processing on the medical image to be processed according to the probability distribution to obtain at least one area to be processed;

carrying out connected domain calculation according to the region to be treated to obtain at least one initial region;

or alternatively

Processing the medical image to be processed to obtain the probability corresponding to each point in the medical image to be processed;

and acquiring a probability threshold, and processing the probability according to the probability threshold to obtain at least one initial region in the medical image to be processed.

In one embodiment, the performing probability regression processing or probability threshold binarization processing on the medical image to be processed to obtain at least one initial region includes:

extracting features of the medical image to be processed through an encoding network of a thermodynamic diagram regression network to obtain first form information corresponding to a target object, wherein the first form information comprises one or more of texture features, geometric structure features and position features of the target object;

and carrying out probability regression or probability threshold binarization processing according to the first form information corresponding to the target object through a decoding network of the thermodynamic diagram regression network to obtain at least one initial region.

In one embodiment, the classifying the obtained initial region to obtain the target region including the target object includes:

sequentially inputting the initial areas into a two-class network to judge whether a target object exists in the initial areas;

and when the target object exists in the initial area, taking the initial area as a target area.

In one embodiment, the sequentially inputting the initial areas into the two classification networks to determine whether the target object exists in the initial areas includes:

Extracting the characteristics of the initial region through a coding network of a classification network to obtain second form information corresponding to a target object, wherein the second form information comprises one or more of texture characteristics, geometric structure characteristics and position characteristics of the target object;

and inputting the extracted second form information into a classification network of a classification network to judge whether a target object exists in the initial area.

In one embodiment, the dividing the target area to obtain the target object includes:

extracting features of the target area through a coding network of a full convolution segmentation network to obtain third form information corresponding to a target object, wherein the third form information comprises one or more of texture features, geometric structure features and position features of the target object;

and performing target object segmentation according to the third form information corresponding to the target object through a decoding network of the pre-trained full convolution segmentation network to obtain the target object.

In one embodiment, the medical image to be processed is a brain image, and the target object is a brain micro-bleeding target object; after the target area is segmented to obtain a target object, the method comprises the following steps:

Geometric and/or hemodynamic properties are calculated for the segmented brain microhemorrhage target object.

The network training method applied to the medical image processing method comprises a cascade thermodynamic diagram regression network, a two-class network and a full convolution segmentation network; the training method of the thermodynamic diagram regression network, the two-class network and the full convolution segmentation network comprises the following steps:

acquiring target object sample data;

acquiring a basic network and setting super parameters of the basic network;

defining a loss function of the base network;

training the basic network by utilizing target object sample data through a random gradient descent method to adjust characteristic parameters of the basic network, so that the value of the loss function meets the requirement, and obtaining the trained network.

A medical image processing apparatus, the apparatus comprising:

the medical image acquisition module to be processed is used for acquiring medical images to be processed;

the initial region acquisition module is used for carrying out probability regression or probability threshold binarization processing on the medical image to be processed to obtain at least one initial region;

the classification module is used for classifying the obtained initial area to obtain a target area comprising a target object;

And the segmentation module is used for segmenting the target area to obtain a target object.

A computer device comprising a memory storing a computer program and a processor implementing the steps of the method described above when the processor executes the computer program.

A computer readable storage medium having stored thereon a computer program which, when executed by a processor, performs the steps of the method described above.

According to the medical image processing method, the medical image processing device, the computer equipment and the storage medium, the probability regression processing or the probability threshold binarization processing is firstly carried out on the medical image to be processed to obtain the initial area, then the initial area is classified to obtain the target area comprising the target object, and finally the target object segmentation is only carried out on the target area, so that on one hand, the accuracy is improved through multi-step processing, and on the other hand, due to the fact that part of data is reduced in each step of processing, the efficiency is improved.

Drawings

FIG. 1 is a diagram of an application environment for a medical image processing method according to one embodiment;

FIG. 2 is a flow chart of a method of medical image processing according to one embodiment;

FIG. 3 is a schematic diagram of the structure of an initial region in one embodiment;

FIG. 4 is a schematic diagram of a thermodynamic diagram regression network in one embodiment;

FIG. 5 is a schematic diagram of a two-class network in one embodiment

FIG. 6 is a schematic diagram of a split network in one embodiment;

FIG. 7 is a flow chart of a network training method according to another embodiment;

FIG. 8 is a block diagram showing the structure of a medical image processing apparatus according to an embodiment;

fig. 9 is an internal structural diagram of a computer device in one embodiment.

Detailed Description

In order to make the objects, technical solutions and advantages of the present application more apparent, the present application will be further described in detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the present application.

The medical image processing method provided by the application can be applied to an application environment shown in fig. 1. Wherein the terminal 102 communicates with the medical imaging device 104 via a network. The terminal 102 may receive the image of the medical object to be processed scanned by the medical imaging device 104, and perform probability regression processing or probability threshold binarization processing on the image of the medical object to be processed to obtain at least one initial region; classifying the obtained initial region to obtain a target region comprising a target object; the target area is segmented to obtain a target object. The medical image to be processed is processed to obtain an initial area, the initial area is classified to obtain a target area comprising a target object, and finally, the target area is only segmented, so that the accuracy is improved through multi-step processing, and the efficiency is improved due to the fact that a part of data is reduced in each step of processing.

The terminal 102 may be, but is not limited to, various personal computers, notebook computers, smartphones, tablet computers, and portable wearable devices, and the medical imaging device 104 includes, but is not limited to, various imaging devices such as a CT imaging device (CT: computed Tomography, which is a cross-sectional scan around a certain part of a human body with a very sensitive detector using an X-ray beam that is precisely collimated, and by CT scanning, an accurate three-dimensional position image of a tumor or the like can be reconstructed), a magnetic resonance device (which is a type of tomographic imaging that obtains electromagnetic signals from a human body using a magnetic resonance phenomenon and reconstructs a human body information image), a positron emission type computed tomography (Positron Emission Computed Tomography) device, a positron emission type magnetic resonance imaging system (PET/MR), and the like.

In one embodiment, as shown in fig. 2, a medical image processing method is provided, and the method is applied to the terminal in fig. 1 for illustration, and includes the following steps:

s202: acquiring a medical image to be processed;

in particular, the medical object image to be processed may be a CTA (computed tomography angiography) volume data (e.g. three-dimensional data of a human body image) image, the size of which may be selected to be 512x512x130, although the size may be selected according to the specific image in practice.

Optionally, when the terminal acquires the medical object image to be processed, the terminal performs filtering processing on the medical object image to be processed to obtain a processed image. Specifically, the terminal may filter noise information in the image by using a three-dimensional gaussian filter, so as to obtain a preprocessed image.

S204: and carrying out probability regression processing or probability threshold binarization processing on the medical image to be processed to obtain at least one initial region.

Specifically, after probability regression processing or probability threshold binarization processing is performed on the medical image to be processed in the initial region, the initial region which possibly comprises the target object is selected according to the processed probability map. Wherein the probabilistic regression process or the probabilistic threshold binarization process may be a process of treating the medical image through a thermodynamic diagram regression network.

Taking a target object as brain micro-bleeding as an example for illustration, a terminal inputs a medical image to be processed into a thermodynamic diagram regression network to obtain an image conforming to a Gaussian regression curve, wherein the image comprises at least one initial region, the center of the initial region coincides with the center of the Gaussian regression curve, the initial region is expanded in a mode of reducing the probability of the brain micro-bleeding according to the trend of the Gaussian regression curve, and at least one initial region is finally obtained. Wherein, it can be combined with fig. 3, and fig. 3 is a to-be-treated area obtained after the initial area is treated.

S206: and classifying all the obtained initial areas to obtain a target area comprising the target object.

Specifically, classifying refers to classifying the initial region according to whether or not the target object exists, wherein the initial region where the target object exists is used as the target region, and the initial region where the target object does not exist is deleted to reduce the amount of data to be processed later. Optionally, the terminal normalizes the initial areas to a specified size and then inputs the normalized initial areas into the classification network.

The classification of the initial areas is mainly because a plurality of false positive initial areas possibly exist in the initial areas detected from the medical images to be processed of the whole brain through the thermodynamic diagram regression network, and therefore the terminal can remove the false positive initial areas through classifying the initial areas.

Specifically, the terminal may classify the initial region through the two classification models to obtain a target region including the target object. In other embodiments, the terminal may further classify the initial region by a preset image feature rule or the like to obtain a target region including the target object, which is not limited herein specifically. But to increase efficiency it is preferable to process in an end-to-end cascaded network.

Wherein, optionally, after the initial area and/or the target area are acquired, the acquisition of the maximum communication area may be included, that is, if the initial area and/or the target area with the distance smaller than the preset distance exists, the calculation of the maximum communication area may be performed, so as to reduce the number of the target areas that are subsequently segmented, and improve the efficiency.

S208: the target area is segmented to obtain a target object.

Specifically, the segmentation refers to a process of processing target areas to extract accurate target objects, wherein the target areas are determined areas including the target objects, and the terminal segments the target areas to obtain the target objects. Wherein optionally the terminal may segment the target region by means of a pre-trained segmentation network, such as a full convolution segmentation network, to obtain the target object, wherein optionally the terminal may segment the target region by means of parallel processing. For example, a plurality of division networks are connected in parallel after the classification network, and the terminal inputs the target area obtained by the classification network into the corresponding division network in parallel according to a preset rule, so as to improve the processing efficiency.

In order to fully understand the technical solution in this embodiment, a target object is taken as a brain micro-bleeding as an example, where the medical image to be processed is a complete brain image, and the terminal inputs the complete brain image into the thermodynamic diagram regression network to detect an initial area where micro-bleeding may exist from the whole brain image, where many false positive micro-bleeding initial areas may exist. The terminal thus inputs the initial regions into the classification network, wherein optionally the terminal normalizes the initial regions to a specified size and then inputs the normalized initial regions into the classification network. The initial region is further processed through a classification network to remove false positive micro-bleeding initial regions. And finally, dividing the target area to obtain the target object, namely inputting the target area into a dividing network to extract the target object, wherein the thermodynamic diagram regression network, the classifying network and the dividing network can be cascaded, namely end-to-end, so that the processing efficiency is improved.

According to the medical image processing method, the medical image to be processed is processed to obtain the initial area, the initial area is classified to obtain the target area comprising the target object, and finally the target object is segmented only in the target area, so that on one hand, the accuracy is improved through multi-step processing, and on the other hand, due to the fact that a part of data is reduced in each step of processing, the efficiency is improved.

In one embodiment, performing probability regression processing or probability threshold binarization processing on a medical image to be processed to obtain at least one initial region, including: carrying out probability regression processing or probability threshold binarization processing on the medical image to be processed to obtain probability distribution corresponding to the medical image to be processed; performing binarization processing on the medical image to be processed according to the probability distribution to obtain at least one region to be processed; carrying out connected domain calculation according to the region to be treated to obtain at least one initial region; or processing the medical image to be processed to obtain the probability corresponding to each point in the medical image to be processed; and acquiring a probability threshold value, and processing the probability according to the probability threshold value to obtain at least one initial region in the medical image to be processed.

Specifically, the probability regression processing refers to converting a medical image to be processed from one distribution to an image conforming to the probability of a gaussian distribution, wherein the probability of the gaussian distribution is used for representing the probability of the existence of a target object in the medical image to be processed, so that a plurality of regions conforming to the gaussian distribution are obtained after the medical image to be processed is subjected to probability distribution regression. The terminal performs binarization processing on the medical image to be processed according to the probability segmentation threshold to obtain at least one region to be processed, and specifically, the image after the binarization processing can be shown in fig. 3. After the to-be-processed area is obtained, the terminal performs connected domain calculation on the to-be-processed area to obtain at least one initial area, namely, the terminal performs connected domain calculation on the binarized foreground part, so that classification processing on small areas is avoided, the data volume of subsequent classification processing is reduced, and accuracy is improved. The algorithm for calculating the connected domain may be any existing algorithm, and is not limited herein.

The probability threshold binarization processing refers to processing a medical image to be processed to obtain probability corresponding to each point in the medical image to be processed, then obtaining a probability threshold, for example, 0.5, and the like, and performing binarization processing on the probability corresponding to each point based on the probability threshold, so as to obtain a foreground part and a background part, wherein the foreground part is an initial area.

In the above embodiment, by processing the medical image to be processed to detect the initial region of the brain micro-bleeding from the whole brain picture, rough processing of the image is realized, and a foundation is laid for accurate processing of the subsequent image.

In one embodiment, performing probability regression processing or probability threshold binarization processing on the medical image to be processed to obtain at least one initial region includes: extracting features of the medical image to be processed through an encoding network of a thermodynamic diagram regression network to obtain first form information corresponding to the target object, wherein the first form information comprises one or more of texture features, geometric structure features and position features of the target object; and carrying out probability regression processing or probability threshold binarization processing according to the first form information corresponding to the target object through a decoding network of the pre-trained thermodynamic diagram regression network to obtain at least one initial region.

Specifically, the first morphological information includes one or more of texture features, geometric features and position features of the target object. Taking brain micro-bleeding as an example, the method can comprise texture features, geometric structure features and position features of the brain micro-bleeding, and performing feature extraction on the medical image to be processed by using a thermodynamic diagram regression network to obtain first form information corresponding to a target object, so that the target object is positioned on the medical image to be processed according to the first form information, and an initial area can be obtained. The texture feature may refer to features such as surface gray values of the target object and other regions, the geometric structure feature may refer to shape features of the target object, and the position feature refers to a position of the target object in the medical image to be processed. Taking brain microhemorrhages as an example, brain microhemorrhages are slender, small in volume, large in number and in brain parenchymal areas.

Specifically, referring to fig. 4, fig. 4 is a schematic structural diagram of a thermodynamic diagram regression network in one embodiment. The method comprises the steps of constructing a deep full convolution neural network thermodynamic diagram regression model, and learning micro-bleeding characteristics from medical images to be processed to initially detect brain micro-bleeding image areas, wherein the learned micro-bleeding characteristics comprise micro-bleeding texture characteristics, micro-bleeding geometric structure characteristics and micro-bleeding position information characteristics.

The thermodynamic diagram regression network may be implemented by way of a deep full convolutional neural network that includes an encoding network and a decoding network. The encoding network is used for learning first morphological information of the target object from the medical image to be processed, in particular, the encoding network is used for learning the position characteristics of the target object; the decoding network is used for learning the obtained texture features and/or geometric structure features of the target object, and enabling the encoding network to find the region positions of the texture features and/or geometric structure features according to the texture features and/or geometric structure features. The encoding network includes a convolutional layer, a deconvolution layer. The decoding network includes a volume layer, a residual connection, and a max pooling layer. The deep convolutional neural network further includes a merge layer connecting the encoding network and the decoding network.

Specifically, extracting features of the medical image to be processed to obtain morphological information corresponding to the target object, including: performing feature extraction on the medical image to be processed by using a current convolutional layer of a coding network in the depth full convolutional neural network to obtain first form information; and performing image size conversion on the image after the first form information is extracted by utilizing a corresponding deconvolution layer of the coding network in the depth full convolution neural network so that the size of the converted image corresponds to the size of the image output by the convolution layer in the corresponding decoding network.

Specifically, according to first morphological information corresponding to a target object, positioning the target object position of a medical image to be processed by using a deep full convolution neural network to obtain an initial region, including: performing feature extraction on the medical image to be processed by using a convolution layer of a decoding network in the depth full convolution neural network to obtain first form information; the method comprises the steps of reserving the extracted form information by utilizing a maximum pooling layer of a decoding network in a depth full convolution neural network, and adjusting the size of an image output by a convolution layer of the decoding network; directly adding the input and the output of a convolution layer of a decoding network to be used as the input of a maximum pooling layer by utilizing residual connection of the decoding network in the depth full convolution neural network; and combining the image output by the convolution layer in the corresponding decoding network with the image output by the deconvolution layer of the coding network by utilizing the combining layer in the depth full convolution neural network, and taking the combined image as the input of the next convolution layer of the coding network.

In conjunction with fig. 4, the convolution layers (including the convolution layers 5-2, 6-1, 6-2, 7-1, 7-2, 8-1, 8-2, 9-1, 9-2, and 9-3 in fig. 4) in the encoding network may learn to express the morphological information in the medical image to be processed, where the learning of the morphological information may be learned according to the annotated sample image, and the deconvolution layers (including the deconvolution layer 1, deconvolution layer 2, deconvolution layer 3, and deconvolution layer 4 in fig. 4) are used to increase the image output by the convolution layers to be combined with the size of the image output by the decoding network (including the convolution layers 1-1, 1-2, 2-1, 3-2, 4-1, 4-2, and 5-1 in fig. 4), so that the size of the image output by the deconvolution layer (including the deconvolution layer 1, deconvolution layer 2, deconvolution layer 3, and deconvolution layer 4 in fig. 4) in the decoding network may be combined with the size of the image output by the corresponding layers in the merging network, and the size of the image output by the deconvolution layer 4.

The convolution layers (including convolution layers 1-1, 1-2, 2-1, 2-2, 3-1, 3-2, 4-1, 4-2 and 5-1 in fig. 4) in the decoding network learn to express the morphological information in the medical image to be processed, wherein the learning of the morphological information can be obtained by learning according to the marked sample image, and the maximum pooling layer (including maximum pooling layer 1, maximum pooling layer 2, maximum pooling layer 3 and maximum pooling layer 4 in fig. 4) retains the extracted morphological information and adjusts the size of the image output by the convolution layers of the decoding network, in this embodiment, reduces the image size, and the residual connection directly intersects the input information and the output information so as to facilitate the subsequent optimization learning, i.e. directly adding the input and the output of the convolution layers of the decoding network as the input of the maximum pooling layer.

In one embodiment, classifying the obtained initial region to obtain a target region including a target object includes: sequentially inputting the initial areas into a two-class network to judge whether a target object exists in the initial areas; when the target object exists in the initial area, the initial area is taken as the target area, otherwise, the initial area is deleted.

Specifically, the pre-trained two-classification network may be a network for judging whether a target object exists in the initial area, that is, the terminal extracts second shape information of the initial area through the two-classification network, and then inputs the second shape information into the classification network to judge whether the target object exists in the initial area, wherein the classification network may include a full-connection layer and an output layer, and the full-connection layer is used for spatially distributing characteristics of the encoding network learning to the output tag.

In one embodiment, sequentially inputting the initial area into the two-class network to determine whether the target object exists in the initial area includes: extracting features of the initial region through a coding network of the two classification networks to obtain second form information corresponding to the target object, wherein the second form information comprises one or more of texture features, geometric structure features and position features of the target object; the extracted second shape information is input into a classification network of the classification network to judge whether a target object exists in the initial area.

Specifically, the second shape information includes one or more of texture features, geometric features, and position features of the target object. Taking brain micro-bleeding as an example, the method can include texture features, geometric structure features and position features of the brain micro-bleeding, and the classification network performs feature extraction on the initial region to obtain second form information corresponding to the target object, so that the initial region is classified according to the second form information to obtain the target region. The texture feature may refer to features such as surface gray values of the target object and other regions, the geometric structure feature may refer to shape features of the target object, and the position feature refers to a position of the target object in the medical image to be processed. Taking brain microhemorrhages as an example, brain microhemorrhages are slender, small in volume, large in number and in brain parenchymal areas.

Referring to fig. 5, fig. 5 is a schematic structural diagram of a classification network in an embodiment, in which by constructing a class two classification model, the characteristics of microhemorrhage are learned from data to distinguish whether there is microhemorrhage in the region, and the learned microhemorrhage characteristics include texture characteristics, geometric characteristics and position characteristics of microhemorrhage. The initial region is then classified to obtain a target region including the target object.

The second class classification model comprises a coding network and a classification network, wherein the coding network is used for learning useful characteristic information of micro bleeding from the image, namely the second state information; the classification network is used for judging whether the image has micro bleeding according to the learned characteristic information. The coding network is formed by a combination cascade of convolution layers (comprising convolution layers 1-1, 1-2, 2-1, 2-2, 3-1, 3-2, 4-1, 4-2 and 5-1 in fig. 5), residual connection and maximum pooling layers (comprising maximum pooling layer 1, maximum pooling layer 2, maximum pooling layer 3 and maximum pooling layer 4 in fig. 5), wherein the convolution layers are used for learning to express the second state information in the initial region; the maximum pooling layer is used for reducing the image size while retaining the second state information; the residual connection (e.g., the connection of the output of max-pooling layer 3 with the input of convolutional layer 4-2, the connection of the output of max-pooling layer with the input of convolutional layer 3-2, the connection of the output of max-pooling layer 1 with the input of convolutional layer 2-2, the connection of the output of input layer with the input of convolutional layer 1-2 in fig. 5) is to add the input information directly with the output information to facilitate the subsequent optimization learning. The classification network is composed of two fully connected layers and an output layer, wherein the fully connected layers are used for mapping learned characteristics to output label space distribution.

In one embodiment, segmenting the target region to obtain the target object includes: extracting features of the target area through a coding network of the full convolution segmentation network to obtain third form information corresponding to the target object, wherein the third form information comprises one or more of texture features, geometric structure features and position features of the target object; and performing target object segmentation according to third form information corresponding to the target object through a decoding network of the pre-trained full convolution segmentation network to obtain the target object.

Wherein the third morphological information comprises one or more of texture features, geometric features and position features of the target object. Taking brain micro-bleeding as an example, the full convolution segmentation network performs feature extraction on the target area to obtain third form information corresponding to the target object, so that the target object position of the target area is positioned according to the third form information, and a target object segmentation result can be obtained. The texture feature may refer to features such as surface gray values of the target object and other regions, the geometric structure feature may refer to shape features of the target object, and the position feature refers to a position of the target object in the target region.

Specifically, referring to fig. 6, fig. 6 is a network structure diagram of a full convolution split network in one embodiment. Wherein the full convolution partition network comprises an encoding network and a decoding network. The coding network term learns third morphological information of the target object from the target area, and specifically, the coding network is used for learning the position characteristics of the target object so as to realize image segmentation; the decoding network is used for learning the obtained texture features and/or geometric structure features of the target object, and the encoding network is enabled to find the region positions of the texture features and/or geometric structure features according to the texture features and/or geometric structure features so as to realize image segmentation. The coding network includes convolutional layers (including convolutional layer 1-1, convolutional layer 1-2, convolutional layer 2-1, convolutional layer 2-2, convolutional layer 3-1, convolutional layer 3-2, convolutional layer 4-1, convolutional layer 4-2, and convolutional layer 5-1 in FIG. 6), deconvolution layers (including deconvolution layer 1, deconvolution layer 2, deconvolution layer 3, and deconvolution layer 4 in FIG. 6). The decoding network includes convolutional layers (including convolutional layer 1-1, convolutional layer 1-2, convolutional layer 2-1, convolutional layer 2-2, convolutional layer 3-1, convolutional layer 3-2, convolutional layer 4-1, convolutional layer 4-2, and convolutional layer 5-1 in FIG. 6), residual connections, and max-pooling layers (including max-pooling layer 1, max-pooling layer 2, max-pooling layer 3, and max-pooling layer 4 in FIG. 6). The full convolution partition network further includes a merge layer connecting the encoding network and the decoding network.

Specifically, extracting features of the target area to obtain third form information corresponding to the target object, including: extracting characteristics of a target area by utilizing a current convolution layer of a coding network in a full convolution segmentation network to obtain morphological information; and performing image size transformation on the target area after extracting the third form information by utilizing a corresponding deconvolution layer of the coding network in the full convolution segmentation network so that the transformed image size corresponds to the size of the image output by the convolution layer in the corresponding decoding network.

Specifically, according to the morphological information corresponding to the target object, positioning the target object position of the target area by using a full convolution segmentation network to obtain a target object segmentation result, including: performing feature extraction on the target area by using a convolution layer of a decoding network in the full convolution segmentation network to obtain morphological information; the maximum pooling layer of the decoding network in the full convolution segmentation network is utilized, the extracted form information is reserved, and the size of an image output by the convolution layer of the decoding network is adjusted; directly adding the input and the output of a convolution layer of a decoding network to be used as the input of a maximum pooling layer by utilizing residual connection of the decoding network in the full convolution segmentation network; and combining the image output by the convolution layer in the corresponding decoding network with the image output by the deconvolution layer of the encoding network by utilizing the combining layer in the full convolution segmentation network, and taking the combined image as the input of the next convolution layer of the encoding network.

Referring to fig. 6, the convolution layer in the encoding network may learn the third form information in the expression target area, where the learning of the third form information may be obtained by learning from the noted sample image, and the deconvolution layer is used to increase the image output by the convolution layer to correspond to the size of the image output by the convolution layer in the decoding network, so that the merging layer may merge the image output by the convolution layer in the decoding network corresponding to the size of the image with the image output by the deconvolution layer of the encoding network, as the input of the next convolution layer of the encoding network.

The convolution layer in the decoding network learns the third form information in the expression target area, wherein the third form information is learned according to the marked sample image, the maximum pooling layer retains the extracted third form information and adjusts the size of the image output by the convolution layer of the decoding network, in this embodiment, the image size is reduced, and the residual connection directly intersects the input information and the output information so as to facilitate the subsequent optimization learning, namely, the input and the output of the convolution layer of the decoding network are directly added as the input of the maximum pooling layer.

For convenience, the training process of the thermodynamic diagram regression network, the two classification network and the segmentation network is generally described in this embodiment, where, referring to fig. 7, fig. 7 is a flowchart of the training process of the network in one embodiment, the training process may include:

s702: target object sample data is acquired.

Specifically, the target object sample data is the data used for training last, wherein the deep learning can only have certain robustness by learning on certain data, so that in order to increase the robustness, a data amplification operation is needed, and in turn, the generalization capability of the full convolution network is increased.

The terminal firstly acquires initial sample data; carrying out random rigid transformation on the initial sample data to obtain sample data after data expansion; and taking the initial sample data and the sample data after the data expansion as target object sample data. Specifically, the same random rigid transformation is performed on the original image and the corresponding image label (which may represent the artificially labeled gold standard heart blood pool and myocardial image), and specifically may include, but is not limited to: rotation, scaling, translation, flipping, and gray scale transformation. For example, the original 20 images are amplified to 2000 cases, wherein, for example, 1600 cases are used as training samples for training, and 400 cases are used as test samples for testing, wherein each image comprises an original image and an image label. Specifically, the terminal generally performs data enhancement on the marked limited data, and expands the marked limited data to a large number, and in deep learning, 80% of data is generally selected for training, and 20% of data is generally selected for testing. Specifically, during image amplification, the terminal performs image extraction in at least one of the following modes: rotation-30 degrees to 30 degrees, scaling 0.8 to 1.2 times, translation-10 to 10 pixels, flipping (horizontal and vertical) and gray scale transformation (image normalization).

S704: and obtaining the basic network and setting the super parameters of the basic network.

Specifically, the obtaining of the basic network refers to the construction of a deep full convolution network structure or the construction of a two-class network, and the deep full convolution network and the two-class network which are specifically constructed can be shown in fig. 4 to 6.

The parameter setting of the basic network comprises characteristic parameters and super parameters, wherein the characteristic parameters are learned by the neural network continuously and iteratively and are used for learning image characteristics; the super parameters are set manually, and the network is trained by setting proper super parameters. As an example, the learning rate may be set to 0.001, the number of hidden layers is 16, 32, 64, 128, 256, the convolution kernel size is 3x3x3, the training iteration number is 30000, and the batch size of each iteration is 1. The characteristic parameters may include a weight parameter W and a bias parameter b of a network, where, taking a cardiac image as an example, the W and b parameters in the network are used to represent texture features, geometric information features, position information features and the like of micro-bleeding of the brain, and simple features and complex features of micro-bleeding of the local brain are expressed by W and b in a shallow layer and a deep layer, and the simple features may be edge and corner features, and the complex features are texture features and geometric shape features composed of the simple features and the like.

S706: a loss function of the underlying network is defined.

Specifically, the loss function is an objective function used to optimize the network by minimizing the loss function to make the underlying network learn better. The basic network learning image features need to be learned under a certain situation, namely, suitable loss functions need to be defined to learn effective features, bad loss functions cannot learn good features, and the characteristics are mainly defined by the loss functions, wherein in the embodiment, thermodynamic diagram regression network loss functions L (W, b) can be expressed as:

the bisectional network loss function L (W, b) can be expressed as:

the split network loss function L (W, b) can be expressed as:

wherein W and b represent the weight and bias parameters of the network, x ⁱ Representing input ith target object sample data, f _W,b (x ⁱ ) Model predictive result representing ith target object sample data, y ⁱ And (3) representing the real label of the sample data of the ith target object, wherein K is a smoothing parameter, and preventing the denominator from being zero and being unable to be calculated.

S708: training the basic network by utilizing target object sample data through a random gradient descent method to adjust characteristic parameters of the basic network, so that the value of the loss function meets the requirement, and obtaining the trained network.

Specifically, in this embodiment, the terminal trains the base network by using a random gradient descent method, and the main training process is to use the random gradient descent method to iteratively train and update the weight parameters and the bias parameters.

Specifically, the terminal trains the base network using a gradient descent method, and then updates and optimizes weight parameters and bias parameters in the base network using a back propagation algorithm. Specifically, the gradient descent method judges that the place with the maximum slope of the curve is the direction from the highest to the best value, and the back propagation method adopts a chain derivation method of probability to calculate the partial derivative so as to update the weight, and the parameter is updated through continuous iterative training so as to learn the image. The method for updating the weight parameters and the bias parameters by the back propagation algorithm is as follows:

first, forward propagation is performed, parameters are updated through continuous iterative training to learn the features of the image, and activation values of all layers (convolution layers, deconvolution layers) are calculated, i.e., the image is subjected to convolution operation to obtain an activated image.

And then to the output layer (nth _l Layer), calculate the sensitivity value

Wherein y is the true value of the sample,

for the predicted value of the output layer, +.>

Representing the partial derivative of the output layer parameters.

Second, for l=n _l -1,n _l -2..the layers of the..calculate the sensitivity value

/>

Wherein W is ^l Representing parameters of the first layer, delta ^l+1 Representing the sensitivity value of layer l+1, f' (z ^l ) Representing the partial derivative of the first layer;

finally, the weight parameter and the bias parameter of each layer are updated:

wherein W is ^l And b ^l Respectively representing the weight parameters and the bias parameters of the layer,

for learning rate a ^l Representing the output value of the first layer, delta ^l+1 Representing the sensitivity value of the l+1 layer.

The whole basic network is trained until the parameters of the basic network are converged to the error requirement, wherein the convergence to the error requirement can be that the loss function value is minimized or not changed greatly any more, and the trained network is obtained so as to be convenient for subsequent use.

In one embodiment, the medical image to be processed is a brain image and the target object is a brain microhemorrhage target object; after dividing the target area to obtain the target object, the method includes: geometric and/or hemodynamic properties are calculated for the segmented resulting brain microhemorrhage target object according to preset rules.

Specifically, in this embodiment, the medical image to be processed is a brain image, the target object is a brain micro-bleeding target object, the terminal firstly acquires the whole brain image, then sequentially inputs the brain image into three cascaded neural networks, firstly processes the brain image through a thermodynamic diagram regression network to obtain an initial region, namely, a result of the preliminary micro-bleeding region is obtained, then classifies the initial region through a second classification network, removes some initial regions without the micro-bleeding image, the remaining initial regions are target regions, and finally segments and extracts the brain micro-bleeding target object by using a full convolution segmentation network for the target region. The whole segmentation process comprises the following steps: firstly, loading network parameters, and then, carrying out forward propagation to finally obtain brain micro-bleeding segmentation results.

Finally, the geometrical properties and hemodynamic properties of the brain micro-bleeding target object can be further automatically calculated after the brain micro-bleeding target object is obtained, so as to further assist a doctor in performing surgical treatment.

The terminal can calculate and obtain the attribute parameters of the brain micro-bleeding target object according to the brain micro-bleeding target object obtained by segmentation, wherein the attribute parameters comprise one or more of the volume, the surface area and the long and short axes of the brain micro-bleeding target object; and/or; and simulating blood flow parameters in the brain micro-bleeding target object by the terminal to obtain a hemodynamic attribute result corresponding to the brain micro-bleeding target object. Thereby assisting the doctor in carrying out subsequent treatment according to the quantitative parameters of the brain micro-bleeding target object.

In the above embodiment, the brain micro-bleeding is segmented through the cascade thermodynamic diagram regression network, the two classification networks and the segmentation network, so that the accuracy of the whole segmentation algorithm is improved, the complex interaction operation is reduced, the end-to-end algorithm flow is realized, and the doctor can be better assisted in processing.

It should be understood that, although the steps in the flowcharts of fig. 2 and 7 are shown in order as indicated by the arrows, these steps are not necessarily performed in order as indicated by the arrows. The steps are not strictly limited to the order of execution unless explicitly recited herein, and the steps may be executed in other orders. Moreover, at least a portion of the steps in fig. 2 and 7 may include a plurality of steps or stages, which are not necessarily performed at the same time, but may be performed at different times, and the order of the execution of the steps or stages is not necessarily sequential, but may be performed in rotation or alternatively with at least a portion of the steps or stages in other steps or steps.

In one embodiment, as shown in fig. 8, there is provided a medical image processing apparatus including: a medical image acquisition module 801 to be processed, an initial region acquisition module 802, a classification module 803, and a segmentation module 804, wherein:

a medical image to be processed acquisition module 801, configured to acquire a medical image to be processed;

an initial region obtaining module 802, configured to perform probability regression processing or probability threshold binarization processing on a medical image to be processed, to obtain at least one initial region;

a classification module 803, configured to classify all the obtained initial areas to obtain a target area including a target object;

the segmentation module 804 is configured to segment the target area to obtain a target object.

In one embodiment, the initial region acquisition module 802 includes:

the probability regression or probability threshold binarization processing unit is used for carrying out probability regression processing or probability threshold binarization processing on the medical image to be processed to obtain probability distribution corresponding to the medical image to be processed;

the binarization processing unit is used for carrying out binarization processing on the medical image to be processed according to the probability distribution to obtain at least one area to be processed;

the connected domain calculation unit is used for carrying out connected domain calculation according to the region to be processed to obtain at least one initial region; or alternatively

The probability calculation unit is used for processing the medical image to be processed to obtain the probability corresponding to each point in the medical image to be processed;

and the binarization processing unit is used for acquiring a probability threshold value and processing the probability according to the probability threshold value to obtain at least one initial area in the medical image to be processed.

In one embodiment, the initial region acquisition module 802 includes:

the first feature extraction unit is used for extracting features of the medical image to be processed through an encoding network of the thermodynamic diagram regression network to obtain first form information corresponding to the target object, wherein the first form information comprises one or more of texture features, geometric structure features and position features of the target object;

the first decoding unit is used for carrying out probability regression processing or probability threshold binarization processing according to the first form information corresponding to the target object through a decoding network of a pre-trained thermodynamic diagram regression network to obtain at least one initial region.

In one embodiment, classification module 803 includes:

the judging unit is used for sequentially inputting the initial areas into the two classification networks so as to judge whether a target object exists in the initial areas; when the target object exists in the initial area, the initial area is taken as the target area, otherwise, the initial area is deleted.

In one embodiment, the determining unit includes:

the second feature extraction unit is used for extracting features of the initial region through a coding network of the two classification networks to obtain second form information corresponding to the target object, wherein the second form information comprises one or more of texture features, geometric structure features and position features of the target object;

and the classification unit is used for inputting the extracted second form information into a classification network of the classification network so as to judge whether a target object exists in the initial area.

In one embodiment, the segmentation module 804 includes:

the third feature extraction unit is used for extracting features of the target area through the coding network of the full convolution segmentation network to obtain third form information corresponding to the target object, wherein the third form information comprises one or more of texture features, geometric structure features and position features of the target object;

and the second decoding unit is used for dividing the target object according to the third form information corresponding to the target object through a decoding network of the pre-trained full convolution dividing network to obtain the target object.

In one embodiment, the medical image to be processed is a brain image and the target object is a brain microhemorrhage target object; the apparatus further comprises:

A calculation module for calculating geometrical and/or hemodynamic properties for the segmented resulting brain microhemorrhage target object.

A network training device in a medical image processing device according to any one of the above embodiments, the network comprising a cascade thermodynamic diagram regression network, a bi-classification network, and a full convolution segmentation network; the training device for thermodynamic regression network, two-class network and full convolution segmentation network comprises:

the sample acquisition module is used for acquiring sample data of a target object;

the parameter setting module is used for acquiring a basic network and setting super parameters of the basic network;

the loss function definition module is used for defining a loss function of the basic network;

and the training module is used for training the basic network by utilizing the target object sample data through a random gradient descent method so as to adjust the characteristic parameters of the basic network, so that the value of the loss function meets the requirement and a trained network is obtained.

For specific limitations of the medical image processing apparatus, reference may be made to the above limitations of the medical image processing method, and no further description is given here. The respective modules in the above-described medical image processing apparatus may be implemented in whole or in part by software, hardware, and a combination thereof. The above modules may be embedded in hardware or may be independent of a processor in the computer device, or may be stored in software in a memory in the computer device, so that the processor may call and execute operations corresponding to the above modules.

In one embodiment, a computer device is provided, which may be a terminal, and the internal structure thereof may be as shown in fig. 9. The computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected by a system bus. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The communication interface of the computer device is used for carrying out wired or wireless communication with an external terminal, and the wireless mode can be realized through WIFI, an operator network, NFC (near field communication) or other technologies. The computer program is executed by a processor to implement a medical image processing method. The display screen of the computer equipment can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer equipment can be a touch layer covered on the display screen, can also be keys, a track ball or a touch pad arranged on the shell of the computer equipment, and can also be an external keyboard, a touch pad or a mouse and the like.

It will be appreciated by those skilled in the art that the structure shown in fig. 9 is merely a block diagram of a portion of the structure associated with the present application and is not limiting of the computer device to which the present application applies, and that a particular computer device may include more or fewer components than shown, or may combine some of the components, or have a different arrangement of components.

In one embodiment, a computer device is provided comprising a memory and a processor, the memory having stored therein a computer program, the processor when executing the computer program performing the steps of: acquiring a medical image to be processed; carrying out probability regression processing or probability threshold binarization processing on the medical image to be processed to obtain at least one initial region; classifying all the obtained initial areas to obtain a target area comprising a target object; the target area is segmented to obtain a target object.

In one embodiment, the probability regression or probability threshold binarization of the medical image to be processed, implemented when the processor executes the computer program, results in at least one initial region, comprising: carrying out probability regression processing on the medical image to be processed to obtain probability distribution corresponding to the medical image to be processed; performing binarization processing on the medical image to be processed according to the probability distribution to obtain at least one region to be processed; carrying out connected domain calculation according to the region to be treated to obtain at least one initial region; or processing the medical image to be processed to obtain the probability corresponding to each point in the medical image to be processed; and acquiring a probability threshold value, and processing the probability according to the probability threshold value to obtain at least one initial region in the medical image to be processed.

In one embodiment, processing of a medical image to be processed, which is implemented when the processor executes the computer program, results in at least one initial region, comprising: extracting features of the medical image to be processed through an encoding network of a thermodynamic diagram regression network to obtain first form information corresponding to the target object, wherein the first form information comprises one or more of texture features, geometric structure features and position features of the target object; and carrying out probability regression processing or probability threshold binarization processing according to the first form information corresponding to the target object through a decoding network of the pre-trained thermodynamic diagram regression network to obtain at least one initial region.

In one embodiment, classifying the resulting initial region as a target region comprising a target object implemented when the processor executes the computer program comprises: sequentially inputting the initial areas into a two-class network to judge whether a target object exists in the initial areas; when the target object exists in the initial area, the initial area is taken as the target area, otherwise, the initial area is deleted.

In one embodiment, the method for sequentially inputting the initial regions into the two-class network when the processor executes the computer program to determine whether the target object exists in the initial regions includes: extracting features of the initial region through a coding network of the two classification networks to obtain second form information corresponding to the target object, wherein the second form information comprises one or more of texture features, geometric structure features and position features of the target object; the extracted second shape information is input into a classification network of the classification network to judge whether a target object exists in the initial area.

In one embodiment, a method for segmenting a target region to obtain a target object implemented when a processor executes a computer program includes: extracting features of the target area through a coding network of the full convolution segmentation network to obtain third form information corresponding to the target object, wherein the third form information comprises one or more of texture features, geometric structure features and position features of the target object; and performing target object segmentation according to third form information corresponding to the target object through a decoding network of the pre-trained full convolution segmentation network to obtain the target object.

In one embodiment, the medical image to be processed involved in execution of the computer program by the processor is a brain image, and the target object is a brain microhemorrhage target object; after segmenting the target region to obtain the target object, which is realized when the processor executes the computer program, the method comprises the following steps: geometric and/or hemodynamic properties are calculated for the segmented resulting brain microhemorrhage target object.

In one embodiment, a computer device is provided, comprising a memory and a processor, the memory having stored therein a computer program, the processor when executing the computer program involving a network comprising a cascaded thermodynamic diagram regression network, a two-class network, and a full convolution segmentation network; the training method of the thermodynamic diagram regression network, the two-class network and the full convolution segmentation network, which are realized when the processor executes the computer program, comprises the following steps: acquiring target object sample data; acquiring a basic network and setting super parameters of the basic network; defining a loss function of the base network; training the basic network by utilizing target object sample data through a random gradient descent method to adjust characteristic parameters of the basic network, so that the value of the loss function meets the requirement, and obtaining the trained network.

In one embodiment, a computer readable storage medium is provided having a computer program stored thereon, which when executed by a processor, performs the steps of: acquiring a medical image to be processed; carrying out probability regression processing or probability threshold binarization processing on the medical image to be processed to obtain at least one initial region; classifying all the obtained initial areas to obtain a target area comprising a target object; the target area is segmented to obtain a target object.

In one embodiment, the computer program, when executed by the processor, performs probability regression processing or probability threshold binarization processing on the medical image to be processed to obtain at least one initial region, including: carrying out probability regression processing on the medical image to be processed to obtain probability distribution corresponding to the medical image to be processed; performing binarization processing on the medical image to be processed according to the probability distribution to obtain at least one region to be processed; carrying out connected domain calculation according to the region to be treated to obtain at least one initial region; or processing the medical image to be processed to obtain the probability corresponding to each point in the medical image to be processed; and acquiring a probability threshold value, and processing the probability according to the probability threshold value to obtain at least one initial region in the medical image to be processed.

In one embodiment, processing of a medical image to be processed, which is implemented when a computer program is executed by a processor, results in at least one initial region, comprising: extracting features of the medical image to be processed through an encoding network of a thermodynamic diagram regression network to obtain first form information corresponding to the target object, wherein the first form information comprises one or more of texture features, geometric structure features and position features of the target object; and carrying out probability regression processing or probability threshold binarization processing according to the first form information corresponding to the target object through a decoding network of the pre-trained thermodynamic diagram regression network to obtain at least one initial region.

In one embodiment, classifying the resulting initial region as a target region comprising a target object, as implemented by a computer program when executed by a processor, comprises: sequentially inputting the initial areas into a two-class network to judge whether a target object exists in the initial areas; when the target object exists in the initial region, the initial region is taken as a target region.

In one embodiment, the method for sequentially inputting the initial regions into the two-class network when the computer program is executed by the processor to determine whether the target object exists in the initial regions comprises: extracting features of the initial region through a coding network of the two classification networks to obtain second form information corresponding to the target object, wherein the second form information comprises one or more of texture features, geometric structure features and position features of the target object; the extracted second shape information is input into a classification network of the classification network to judge whether a target object exists in the initial area.

In one embodiment, a method for partitioning a target region to obtain a target object implemented when a computer program is executed by a processor includes: extracting features of the target area through a coding network of the full convolution segmentation network to obtain third form information corresponding to the target object, wherein the third form information comprises one or more of texture features, geometric structure features and position features of the target object; and performing target object segmentation according to third form information corresponding to the target object through a decoding network of the pre-trained full convolution segmentation network to obtain the target object.

In one embodiment, the medical image to be processed involved when the computer program is executed by the processor is a brain image, and the target object is a brain micro-bleeding target object; after the computer program is executed by the processor to divide the target area to obtain the target object, the method includes: geometric and/or hemodynamic properties are calculated for the segmented resulting brain microhemorrhage target object.

In one embodiment, a computer readable storage medium is provided having a computer program stored thereon, the network involved when the computer program is executed by a processor comprising a cascaded thermodynamic diagram regression network, a binary classification network, and a full convolution segmentation network; the training method of the thermodynamic diagram regression network, the two-class network and the full convolution segmentation network, which is realized when the computer program is executed by the processor, comprises the following steps: acquiring target object sample data; acquiring a basic network and setting super parameters of the basic network; defining a loss function of the base network; training the basic network by utilizing target object sample data through a random gradient descent method to adjust characteristic parameters of the basic network, so that the value of the loss function meets the requirement, and obtaining the trained network.

Those skilled in the art will appreciate that implementing all or part of the above described methods may be accomplished by way of a computer program stored on a non-transitory computer readable storage medium, which when executed, may comprise the steps of the embodiments of the methods described above. Any reference to memory, storage, database, or other medium used in embodiments provided herein may include at least one of non-volatile and volatile memory. The nonvolatile Memory may include Read-Only Memory (ROM), magnetic tape, floppy disk, flash Memory, optical Memory, or the like. Volatile memory can include random access memory (Random Access Memory, RAM) or external cache memory. By way of illustration, and not limitation, RAM can be in the form of a variety of forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM), and the like.

The technical features of the above embodiments may be arbitrarily combined, and all possible combinations of the technical features in the above embodiments are not described for brevity of description, however, as long as there is no contradiction between the combinations of the technical features, they should be considered as the scope of the description.

The above examples merely represent a few embodiments of the present application, which are described in more detail and are not to be construed as limiting the scope of the invention. It should be noted that it would be apparent to those skilled in the art that various modifications and improvements could be made without departing from the spirit of the present application, which would be within the scope of the present application. Accordingly, the scope of protection of the present application is to be determined by the claims appended hereto.

Claims

1. A medical image processing method, the method comprising:

acquiring a medical image to be processed;

and dividing the target area to obtain a target object.

2. The method according to claim 1, wherein the performing a probability regression process or a probability threshold binarization process on the medical image to be processed to obtain at least one initial region comprises:

carrying out probability regression processing on the medical image to be processed to obtain probability distribution corresponding to the medical image to be processed;

or alternatively

3. The method according to claim 1, wherein the performing a probability regression process or a probability threshold binarization process on the medical image to be processed to obtain at least one initial region comprises:

and carrying out probability regression processing or probability threshold binarization processing according to the first form information corresponding to the target object through a decoding network of the thermodynamic diagram regression network to obtain at least one initial region.

4. The method of claim 1, wherein classifying the obtained initial region to obtain a target region including a target object, comprises:

5. The method of claim 4, wherein sequentially inputting the initial region into the two classification networks to determine whether the target object exists in the initial region comprises:

6. The method of claim 1, wherein the segmenting the target region to obtain a target object comprises:

7. The method according to any one of claims 1 to 6, wherein the medical image to be processed is a brain image and the target object is a brain microhemorrhage target object; after the target area is segmented to obtain a target object, the method comprises the following steps:

8. A network training method for use in the medical image processing method of any one of claims 1 to 7, wherein the network comprises a cascaded thermodynamic diagram regression network, a binary classification network and a full convolution segmentation network; the training method of the thermodynamic diagram regression network, the two-class network and the full convolution segmentation network comprises the following steps:

Acquiring target object sample data;

acquiring a basic network and setting super parameters of the basic network;

defining a loss function of the base network;

9. A medical image processing apparatus, the apparatus comprising:

the initial region acquisition module is used for carrying out probability regression processing or probability threshold binarization processing on the medical image to be processed to obtain at least one initial region;

10. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method of any of claims 1 to 7 or 8 when the computer program is executed.

11. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method of any of claims 1 to 7 or 8.