A method for processing nodule calcification medical images

Abstract

The present invention provides a method for processing nodule calcification medical images, which relates to the field of medical image processing technology, including: labeling the nodule calcification areas in each collected nodule calcification medical image to obtain corresponding labeled images; training an image processing model based on each labeled image, the image processing model including a residual network module, an anti-stacking module, a group hole convolutional network module, a stacking module, and a decision network module connected in sequence; inputting the medical image to be processed into the image processing model to obtain a semantic probability heat map of the nodule calcification area in the medical image to be processed, and using the semantic probability heat map as the processing result of the medical image to be processed. The beneficial effect is to overcome the problems of poor semantic resolution of calcified areas in nodule images and susceptibility of semantic feature extraction of calcified areas to interference from similar backgrounds in the prior art, and solve the problems of poor semantic probability heat map extraction effect of nodule calcification medical images due to the small size of the nodule calcification area.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a method for processing nodule calcification medical images. Background Art

[0002] Nodule calcification refers to the process in which certain tissues in the human body undergo necrosis under the influence of certain factors, and then calcium salts in the body are deposited in the necrotic foci, causing the lesions to be localized and stabilized. Accurate identification of nodule calcification points can help doctors better perform clinical diagnosis.

[0003] Traditional calcification point identification algorithms are mostly based on the brightness characteristics of medical images, using single or multiple threshold comparisons to determine whether a calcification point is present. Since calcification points are not necessarily the brightest areas in organs and nearby tissues, there are many other brightly lit tissues near nodules. In addition, the imaging quality of ultrasound machines varies, and traditional methods are prone to missed or misjudgment. There are also methods that use deep learning methods to extract calcification features in medical images. Compared to traditional methods, deep learning methods have better semantic probability heat map extraction effects when the contours of the nodule calcification area are relatively clear and complete. However, when the calcification area is small and similar to the background area, the traditional deep convolutional neural network has a weak ability to expand the receptive field in the forward propagation step, which limits the network's recognition ability. The network output results are not effective for semantic extraction of nodule calcification medical images.

[0004] Currently, existing methods for extracting calcification features from medical images using deep learning techniques are typically based on classic networks such as the AlexNet network and the FCN network, and only perform calcification region extraction on ultrasound images of the thyroid gland. In practical applications, in addition to the thyroid gland, the extraction of calcification regions in other areas, such as the breast, is also very important. In addition to ultrasound, CT, MRI, ultrasound angiography, and mammography are also commonly used imaging techniques. Therefore, existing calcification feature extraction methods limit their scope of application. Furthermore, classic networks such as AlexNet and FCN, as general-purpose feature extraction networks, are not designed and optimized for the unique features of nodule calcification medical images. Summary of the Invention

[0005] In view of the problems existing in the prior art, the present invention provides a method for processing nodule calcification medical images, comprising:

[0006] Step S1, collecting a plurality of nodule calcification medical images, and annotating the nodule calcification area in each of the nodule calcification medical images to obtain a corresponding annotated image;

[0007] Step S2, obtaining an image processing model through training based on each of the labeled images, wherein the image processing model includes a residual network module, a destacking module, a group hole convolutional network module, a stacking module, and a decision network module connected in sequence;

[0008] Step S3: Input a medical image to be processed into the image processing model, and process it in sequence through the residual network module, the destacking module, the group void convolutional network module, the stacking module and the decision network module to obtain a semantic probability heat map of the nodule calcification area in the medical image to be processed, and use the semantic probability heat map as the processing result of the medical image to be processed.

[0009] Preferably, step S3 includes:

[0010] Step S31, the residual network module performs feature extraction on the medical image to be processed to obtain a first feature map;

[0011] Step S32: the destacking module divides the first feature map into groups according to the channel dimension to obtain a plurality of destacking feature map groups;

[0012] Step S33, the group dilated convolutional network module performs group dilated convolution processing on each of the de-stacked feature map groups to obtain a corresponding dilated convolutional feature map group;

[0013] Step S34: the stacking module concatenates the dilated convolution feature map groups in the channel dimension to obtain a second feature map;

[0014] Step S35: The decision network module processes the second feature map to obtain the semantic probability heat map.

[0015] Preferably, the residual network module includes a convolutional layer, a maximum pooling layer and a residual layer arranged in series in sequence, the residual layer includes 6 sub-residual layers arranged in series, each of the sub-residual layers includes a parallel first branch and a second branch, the first branch includes two convolutional layers connected in series, the second branch is a convolutional layer, and the outputs of the first branch and the second branch are added as the output of the corresponding sub-residual layer.

[0016] Preferably, in step S32, group division is performed using a division method based on normal distribution weights, and the first feature map includes C-layer sub-feature maps. The calculation formula of the division method based on normal distribution weights is as follows:

[0017]

[0018] in,

[0019] Among them, F m represents the mth anti-stacked feature map group, k∈(0,C] represents the radius of the number of original sub-feature maps used to calculate the group feature map, f nRepresents the sub-feature map of the nth layer, ∪ represents the splicing of the sub-feature maps in the channel dimension, μ, σ represent the mean and standard deviation of the normal distribution respectively, a represents the length of the horizontal axis half interval taken under the normal distribution curve, and any two adjacent n values ​​n1, n2 satisfy |n1-n2|<2k-2.

[0020] Preferably, the group hole convolutional network module includes a plurality of sub-network modules arranged in parallel, and the number of the sub-network modules is the same as the number of the de-stacked feature map groups.

[0021] Preferably, the sub-network module includes a first convolutional layer, a first dilated convolutional layer, a second dilated convolutional layer, a second convolutional layer, a third dilated convolutional layer and a fourth dilated convolutional layer arranged in series.

[0022] Preferably, in step S33, the following formula is used to perform group dilation convolution processing on each of the de-stacked feature map groups:

[0023]

[0024] Among them, z represents the calculation result of the group void convolution, n1, m1, n2, and m2 represent the two-dimensional horizontal and vertical coordinates respectively, c represents the sub-feature map of the cth channel, l represents the lth layer of the sub-network module, k represents the convolution kernel, r represents the void expansion rate, and b represents the offset.

[0025] Preferably, the decision network module includes a bilinear interpolation upsampling module and a network output module connected in sequence, wherein the network output module includes a third convolution layer, a fourth convolution layer and a normalized mapping layer connected in sequence.

[0026] Preferably, in step S2, the image processing model is trained using a piecewise smooth piecewise exponential cumulative distribution loss function, and the calculation formula of the piecewise exponential cumulative distribution loss function is as follows:

[0027]

[0028] in,

[0029] Where E(x) represents the piecewise exponential cumulative distribution loss function, x∈Z 2 Represents the two-dimensional coordinates, k represents the channel number of the feature map after normalization mapping, and y represents the one-hot label form of the real annotation. is the output of channel number k after normalized mapping, d k (x) is the value of the feature map with channel number k at coordinate x, and the total number of channels is recorded as K.

[0030] Preferably, before executing step S2, a pre-processing process is further included, including:

[0031] Step A1, performing median filtering on each of the labeled images to obtain a corresponding first pre-processed image;

[0032] Step A2, performing high-pass filtering on each of the first pre-processed images to obtain a corresponding second pre-processed image;

[0033] Then in step S2, the image processing model is obtained by training based on each of the second pre-processed images.

[0034] The above technical solution has the following advantages or beneficial effects:

[0035] 1) It can process multimodal and multi-site nodule calcification medical images to obtain semantic probability heat maps, which can provide a high-quality semantic probability heat map foundation for applications such as nodule calcification area positioning, recognition, and segmentation;

[0036] 2) It overcomes the problems of poor semantic resolution of calcified areas in nodule images and susceptibility of semantic feature extraction of calcified areas to interference from similar backgrounds in existing technologies, and solves the problems of poor semantic probability heat map extraction of nodule calcification medical images due to the small size of nodule calcification areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 FIG1 is a flow chart of a method for processing nodule calcification medical images in a preferred embodiment of the present invention;

[0038] Figure 2 Schematic diagram of the network structure of the image processing model in a preferred embodiment of the present invention;

[0039] Figure 3 FIG1 is a flow chart of a process for processing a medical image to be processed in a preferred embodiment of the present invention;

[0040] Figure 4 Schematic diagram of the structure of the residual network module in a preferred embodiment of the present invention;

[0041] Figure 5 Schematic diagram of the distribution of network layers of the image processing model in a preferred embodiment of the present invention;

[0042] Figure 6 Schematic diagram of the structure of the sub-network module in a preferred embodiment of the present invention;

[0043] Figure 7 FIG. 1 is a structural diagram of a decision network module in a preferred embodiment of the present invention;

[0044] Figure 8Schematic diagram of the flow of the pretreatment process in a preferred embodiment of the present invention;

[0045] Figure 9 FIG1 is a schematic diagram of an original ultrasound image of a thyroid nodule calcification in a preferred embodiment of the present invention;

[0046] Figure 10 This is a schematic diagram of a semantic probability heat map obtained by processing an original ultrasound image of thyroid nodule calcification by an image processing model in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0047] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. The present invention is not limited to this embodiment, and other embodiments may also fall within the scope of the present invention as long as they conform to the gist of the present invention.

[0048] In a preferred embodiment of the present invention, based on the above problems existing in the prior art, a method for processing nodule calcification medical images is provided. Figure 1 and Figure 2 Shown, including:

[0049] Step S1, collecting multiple nodule calcification medical images, and annotating the nodule calcification area in each nodule calcification medical image to obtain a corresponding annotated image;

[0050] Step S2, obtaining an image processing model based on training of each labeled image, the image processing model including a residual network module, a destacking module, a group hole convolutional network module, a stacking module, and a decision network module connected in sequence;

[0051] In step S3, a medical image to be processed is input into the image processing model, and is processed in sequence by the residual network module, the destacking module, the group void convolutional network module, the stacking module, and the decision network module to obtain a semantic probability heat map of the nodule calcification area in the medical image to be processed, and the semantic probability heat map is used as the processing result of the medical image to be processed.

[0052] Specifically, in this embodiment, the above-mentioned nodule calcification medical images include but are not limited to thyroid nodule images and breast nodule images, wherein the image types of nodule calcification medical images include but are not limited to ultrasound images, CT images, MRI images, ultrasound contrast images and mammography images, so that the trained image processing model can process multi-modal and multi-site nodule calcification medical images to obtain semantic probability heat maps. Figure 2As shown in the figure, by constructing a network architecture in which the residual network module 1, the destacking module 2, the group hole convolutional network module 3, the stacking module 4 and the decision network module 5 are connected in sequence, the trained image processing model can overcome the problems of poor semantic resolution of the calcified area of ​​the nodule image in the existing technology and the susceptibility of the semantic feature extraction of the calcified area to interference from similar backgrounds. It also solves the problems of poor semantic probability heat map extraction of nodule calcification medical images due to the small size of the nodule calcification area.

[0053] In a preferred embodiment of the present invention, Figure 3 As shown, step S3 includes:

[0054] Step S31: The residual network module performs feature extraction on the medical image to be processed to obtain a first feature map;

[0055] Step S32: the destacking module divides the first feature map into groups according to the channel dimension to obtain a plurality of destacking feature map groups;

[0056] Step S33: the group dilated convolution network module performs group dilated convolution processing on each de-stacked feature map group to obtain a corresponding dilated convolution feature map group;

[0057] Step S34: The stacking module concatenates the dilated convolution feature map groups in the channel dimension to obtain a second feature map.

[0058] Step S35: The decision network module processes the second feature map to obtain a semantic probability heat map.

[0059] Specifically, in this embodiment, the image processing model takes the input of the residual network module as input and the output of the decision network module as output. In the process of processing the medical image to be processed, the medical image to be processed is first input into the residual network model for feature extraction to obtain the corresponding first feature map. As a preferred embodiment, Figure 4 As shown, the residual network module 1 includes a convolutional layer 11, a maximum pooling layer 12 and a residual layer 13 arranged in series in sequence. The residual layer 13 includes 6 sub-residual layers 131 arranged in series. Each sub-residual layer 131 includes a first branch and a second branch in parallel. The first branch includes two convolutional layers connected in series, and the second branch is a convolutional layer. The outputs of the first branch and the second branch are added as the output of the corresponding sub-residual layer.

[0060] More specifically, Figure 5As shown, the network structure of the residual network module 1 is 8 layers, and each layer is arranged in series. As a preference, the first layer is a convolution layer with a convolution kernel size of 7×7 and a step size of 1; the second layer is a maximum pooling layer with a pooling size of 2×2 and a step size of 2; the sub-residual layer of the third layer includes two parallel branches, wherein the first branch is a convolution layer with a convolution kernel size of 3×3 and a step size of 1, which is connected in series with a convolution kernel size of 3×3 and a step size of 1; the second branch is a convolution kernel size of 1×1 and a step size of 1. Convolutional layer; the outputs of each branch are element-wise added; the sub-residual layer of the fourth layer has the same network structure as the sub-residual layer of the third layer, and the outputs of each branch are element-wise added; the sub-residual layer of the fifth layer includes two parallel branches, where the first branch is a convolutional layer with a convolution kernel size of 3×3 and a stride of 2, which is connected in series with a convolution kernel size of 3×3 and a stride of 1; the second branch is a convolutional layer with a convolution kernel size of 1×1 and a stride of 2; the outputs of each branch are element-wise added; The sub-residual layer of the 6th layer consists of 2 parallel branches, where the first branch is a convolution layer with a convolution kernel size of 3×3 and a stride of 1, which is connected in series with a convolution kernel size of 3×3 and a stride of 1, and the second branch is a convolution layer with a convolution kernel size of 1×1 and a stride of 1; the outputs of each branch are element-wise added; the sub-residual layer of the 7th layer consists of 2 parallel branches, where the first branch is a convolution layer with a convolution kernel size of 3×3 and a stride of 2, which is connected in series with a convolution kernel size of 3×3 and a stride of 1. The second branch is a convolution layer with a convolution kernel size of 1×1 and a stride of 2. The outputs of each branch are element-wise added. The 8th sub-residual layer consists of two parallel branches. The first branch is a convolution layer with a convolution kernel size of 3×3 and a stride of 1, which is connected in series with a convolution kernel size of 3×3 and a stride of 1. The second branch is a convolution layer with a convolution kernel size of 1×1 and a stride of 1. The outputs of each branch are element-wise added. All convolution layer activation functions use the ReLU function.

[0061] After the first feature map is obtained based on the above-mentioned residual network module 1, the first feature map is then divided into groups based on the normal distribution weight according to the channel dimension by the destacking module 2 to obtain multiple destacking feature map groups. Specifically, assuming that the first feature map output by the residual network module 1 contains C layers of sub-feature maps, for the nth sub-feature map f n The k sub-feature maps before and after are calculated based on the normal distribution weight to obtain the mth anti-stacked feature map group F m , where the calculation formula based on the normal distribution weight is as follows:

[0062]

[0063] in,

[0064] Among them, F mrepresents the mth anti-stacked feature map group, k∈(0,C] represents the radius of the number of original sub-feature maps used to calculate the group feature map, f n represents the n-th layer sub-feature map, ∪ represents the concatenation of sub-feature maps in the channel dimension, μ and σ represent the mean and standard deviation of the normal distribution, respectively, a represents the length of the horizontal half interval under the normal distribution curve, n can be arbitrarily selected in the range of (0, C], and any two adjacent n values ​​n1 and n2 satisfy |n1-n2|<2k-2. The specific value of k can be pre-configured as long as k∈(0, C].

[0065] Furthermore, the group-void convolutional network module 3 includes multiple sub-network modules 31 arranged in parallel, and the number of sub-network modules 31 is the same as the number of de-stacked feature map groups. In other words, N sub-network modules 31 are constructed for the N de-stacked feature map groups output by the de-stacked module, and each sub-network module 31 processes a feature map in a group of de-stacked feature map groups.

[0066] Specifically, if Figure 6 As shown, the sub-network module 31 includes a first convolution layer 311, a first dilated convolution layer 312, a second dilated convolution layer 313, a second convolution layer 314, a third dilated convolution layer 315 and a fourth dilated convolution layer 316 arranged in series. Figure 5 As shown in the figure, the first convolution layer is a convolution layer with a convolution kernel of 3×3, the first dilated convolution layer is a dilated convolution layer with a convolution kernel of 3×3 and a dilation rate of 2, the second dilated convolution layer is a dilated convolution layer with a convolution kernel of 3×3 and a dilation rate of 5, the second convolution layer is a convolution layer with a convolution kernel of 3×3, the third dilated convolution layer is a dilated convolution layer with a convolution kernel of 3×3 and a dilation rate of 2, and the fourth dilated convolution layer is a dilated convolution layer with a convolution kernel of 3×3 and a dilation rate of 5; the stride of all convolution layers and dilated convolution layers is 1, and the activation function uses the ReLU function.

[0067] In a preferred embodiment of the present invention, in step S33, the following formula is used to perform group dilation convolution processing on each de-stacked feature map group:

[0068]

[0069] Among them, z represents the calculation result of the group void convolution, n1, m1, n2, and m2 represent the two-dimensional horizontal and vertical coordinates respectively, c represents the sub-feature map of the cth channel, l represents the lth layer of the sub-network module, k represents the convolution kernel, r represents the void expansion rate, and b represents the offset.

[0070] Furthermore, after the group atrous convolution module 3 processes multiple atrous convolution feature map groups, the constructed stacking module 4 is used to splice each atrous convolution feature map group in the channel dimension to obtain a second feature map; and then the second feature map is processed by the decision network module 5 to obtain a semantic probability heat map.

[0071] Specifically, if Figure 7 As shown, the decision network module 5 includes a bilinear interpolation upsampling module 51 and a network output module 52 connected in sequence, wherein the network output module 52 includes a third convolution layer 521, a fourth convolution layer 522 and a normalized mapping layer 523 connected in sequence. As a preferred embodiment, Figure 5 As shown in Figure 1, the convolution kernel size of the third and fourth convolution layers is 3×3, the stride is 1, and the activation function uses the ReLU function.

[0072] In a preferred embodiment of the present invention, in step S2, a piecewise smooth piecewise exponential cumulative distribution loss function is used to train the image processing model. The calculation formula of the piecewise exponential cumulative distribution loss function is as follows:

[0073]

[0074] in,

[0075] Where E(x) represents the piecewise exponential cumulative distribution loss function, x∈Z 2 Represents the two-dimensional coordinates, k represents the channel number of the feature map after normalization mapping, and y represents the one-hot label form of the real annotation. is the output of channel number k after normalized mapping, d k (x) is the value of the feature map with channel number k at coordinate x, and the total number of channels is recorded as K.

[0076] In a preferred embodiment of the present invention, before executing step S2, a pre-processing process is also included, such as Figure 8 Shown, including:

[0077] Step A1: performing median filtering on each labeled image to obtain a corresponding first pre-processed image;

[0078] Step A2, performing high-pass filtering on each first pre-processed image to obtain a corresponding second pre-processed image;

[0079] Then in step S2, an image processing model is obtained by training according to each second pre-processed image.

[0080] Specifically, in this embodiment, the median filter can effectively reduce the influence of speckle noise in the nodule calcification medical image, and the high-pass filter can make the contour line of the nodule calcification area in the second pre-processed image clearer.

[0081] As a preferred embodiment of the present invention, the method for processing nodule calcification medical images of the present invention can be used to process ultrasound images of nodule calcification in the thyroid gland to obtain corresponding semantic probability heat maps.

[0082] Specifically, at least 15,000 original ultrasound images of thyroid nodule calcifications can be collected, and the shapes of the thyroid nodule calcification areas in all the original ultrasound images of nodule calcifications can be outlined, wherein at least 10,000 images are randomly selected as a training set, at least 2,000 images are randomly selected from the remaining images as a validation set, and at least 3,000 images are randomly selected from the remaining images as a test set, such as Figure 9 Shown is an example of an original ultrasound image of nodule calcification.

[0083] After obtaining the above-mentioned original ultrasound images of the nodule calcification and outlining the thyroid nodule calcification area, the obtained image is preprocessed, including performing median filtering on the input original image to reduce the influence of speckle noise on the ultrasound image; performing high-pass filtering on the median filtered image to make the contour line of the calcification area in the image clearer. Subsequently, the image processing model is trained based on each processed image. The small batch gradient descent method can be used to train the weight parameters of the image processing model, and the model weight parameters are solidified after the training is completed. The trained model weight parameters are used to perform image processing on the new input thyroid nodule image after preprocessing, and the following can be obtained. Figure 10 Semantic probability heatmap shown.

[0084] The above description is only a preferred embodiment of the present invention and does not limit the implementation mode and protection scope of the present invention. For those skilled in the art, it should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the content of this specification and illustrations should be included in the protection scope of the present invention.

Claims

1. A method for processing nodule calcification medical images, characterized in that: include: Step S1, collecting a plurality of nodule calcification medical images, and annotating the nodule calcification area in each of the nodule calcification medical images to obtain a corresponding annotated image; Step S2, obtaining an image processing model through training based on each of the labeled images, wherein the image processing model includes a residual network module, a destacking module, a group hole convolutional network module, a stacking module, and a decision network module connected in sequence; Step S3, inputting a medical image to be processed into the image processing model, and processing the image sequentially through the residual network module, the destacking module, the group void convolutional network module, the stacking module, and the decision network module to obtain a semantic probability heat map of the nodule calcification area in the medical image to be processed, and using the semantic probability heat map as the processing result of the medical image to be processed; The network structure of the residual network module is 8 layers, and each layer is arranged in series, where: The first layer is the convolutional layer; The second layer is the maximum pooling layer; The third sub-residual layer consists of two parallel branches, and the outputs of each branch are added element by element; The sub-residual layer of the fourth layer has the same network structure as the sub-residual layer of the third layer, and the outputs of each branch are superimposed element by element; The fifth to eighth sub-residual layers each consist of two parallel branches, and the outputs of each branch are element-wise added; The activation function of all convolutional layers uses the ReLU function; The step S3 comprises: Step S31, the residual network module performs feature extraction on the medical image to be processed to obtain a first feature map; Step S32: the destacking module divides the first feature map into groups according to the channel dimension to obtain a plurality of destacking feature map groups; Step S33, the group dilated convolutional network module performs group dilated convolution processing on each of the de-stacked feature map groups to obtain a corresponding dilated convolutional feature map group; Step S34: the stacking module concatenates the dilated convolution feature map groups in the channel dimension to obtain a second feature map; Step S35, the decision network module processes the second feature map to obtain the semantic probability heat map; In step S32, group division is performed using a division method based on normal distribution weights, and the first feature map includes Layer sub-characteristic map, the calculation formula of the division method based on normal distribution weight is as follows: ； in, ; in, Indicates the The anti-stacked feature map group, Indicates the radius of the number of original sub-feature maps used to calculate the group feature map. Indicates the The sub-feature map of the layer, Indicates that the sub-feature maps are spliced ​​in the channel dimension, denote the mean and standard deviation of the normal distribution, respectively. Indicates the length of the horizontal half interval taken under the normal distribution curve. Any two adjacent The value of satisfy .

2. The method for processing nodule calcification medical images according to claim 1, characterized in that: The residual network module includes a convolutional layer, a maximum pooling layer and a residual layer arranged in series in sequence, the residual layer includes 6 sub-residual layers arranged in series, each of the sub-residual layers includes a parallel first branch and a second branch, the first branch includes two convolutional layers in series, the second branch is a convolutional layer, and the outputs of the first branch and the second branch are added as the output of the corresponding sub-residual layer.

3. The method for processing nodule calcification medical images according to claim 1, characterized in that: The group hole convolutional network module includes multiple sub-network modules arranged in parallel, and the number of the sub-network modules is the same as the number of the de-stacked feature map groups.

4. The method for processing nodule calcification medical images according to claim 3, characterized in that: The sub-network module includes a first convolutional layer, a first dilated convolutional layer, a second dilated convolutional layer, a second convolutional layer, a third dilated convolutional layer and a fourth dilated convolutional layer arranged in series.

5. The method for processing nodule calcification medical images according to claim 4, characterized in that: In step S33, the following formula is used to perform group dilation convolution processing on each of the de-stacked feature map groups: ； in, represents the calculation result of group void convolution, Represent the two-dimensional horizontal and vertical coordinates, Indicates the The sub-feature map of each channel, Indicates the sub-network module layer, represents the convolution kernel, represents the void expansion rate, Indicates the offset.

6. The method for processing nodule calcification medical images according to claim 1, characterized in that: The decision network module includes a bilinear interpolation upsampling module and a network output module connected in sequence, wherein the network output module includes a third convolution layer, a fourth convolution layer and a normalization mapping layer connected in sequence.

7. The method for processing nodule calcification medical images according to claim 1, characterized in that: In step S2, the image processing model is trained using a piecewise smooth piecewise exponential cumulative distribution loss function, and the calculation formula of the piecewise exponential cumulative distribution loss function is as follows: ； in, ; in, represents the piecewise exponential cumulative distribution loss function, Represents the two-dimensional coordinates, j represents the channel number of the feature map after normalization mapping, Indicates the one-hot label form of the real annotation, is the output of channel number j after normalized mapping, The feature map with channel number j is at coordinate The total number of channels is recorded as .

8. The method for processing nodule calcification medical images according to claim 1, characterized in that: Before executing step S2, a pre-processing process is also included, including: Step A1, performing median filtering on each of the labeled images to obtain a corresponding first pre-processed image; Step A2, performing high-pass filtering on each of the first pre-processed images to obtain a corresponding second pre-processed image; Then in step S2, the image processing model is obtained by training based on each of the second pre-processed images.

Images