A bone age prediction method and device based on lightweight model

By fusing Shufflenet and Ghostnet models, combining channel segmentation and hollow space pyramid pooling layer, the problems of high time-consuming and high calculation consumption in the existing technology are solved, and a lightweight bone age prediction model is realized, which is suitable for embedded system applications. And suitable for diagnosis of various human organ diseases.

CN113989206BActive Publication Date: 2024-11-08HANGZHOU SHENRUI BOLIAN TECH CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111220778.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-10-20
Publication Date
2024-11-08
Estimated Expiration
2041-10-20

AI Technical Summary

Technical Problem

Existing deep learning methods are used in bone age assessment. The huge backbone model results in high calculation consumption, long time-consuming and unsuitable for embedded system applications. The bone age assessment is a subjective comparison process.

Method used

Using a bone age prediction method based on lightweight networks, the segmentation and prediction model is established by fusing Shufflenet and Ghostnet models, and the channel segmentation module, depth-direction convolution layer and hollow space pyramid pooling layer are used to reduce the computational amount and memory access cost. It is suitable for Embedded system.

Benefits of technology

It realizes automatic prediction of bone age, reduces calculation amount and time consumption, is suitable for embedded system calculation, and is suitable for manual image bone age prediction and other human organ disease diagnosis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113989206B_ABST
    Figure CN113989206B_ABST
Patent Text Reader

Abstract

The present invention provides a method and device for predicting bone age based on a lightweight model. The method comprises: obtaining a hand image of a person to be detected; establishing a segmentation model based on a modeling method integrating lightweight networks shufflenet and ghostnet, inputting the hand image into the trained segmentation model, and outputting a hand segmentation result; establishing a bone age prediction model based on a modeling method integrating lightweight networks shufflenet and ghostnet, inputting the hand image and the hand segmentation result into the trained bone age prediction model, and obtaining the bone age of the person to be detected. The present invention can reduce the amount of calculation and time consumption, reduce the memory access cost MAC and save manufacturing cost, and is suitable for a network design mode for embedded system calculation. It is worth noting that the method of the present invention is not only suitable for bone age prediction based on human hand images, but also suitable for disease type diagnosis and prediction based on any human organ image.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of medical imaging technology, and in particular relates to a bone age prediction method and device based on a lightweight model. Background Art

[0002] Bone age is the main method for evaluating the biological age of adolescents and children, and is widely used in clinical medicine, forensic medicine, sports medicine and other fields. Bone age assessment methods are divided into atlas method and scoring method. The atlas method is represented by GP atlas. This method compares the bone age film to be tested with a set of standard GP bone age films of the left hand, and selects the bone age value of the most similar "standard GP atlas bone age film" as the evaluation value of the bone age film to be tested. The scoring method is represented by TW2 / TW3 method and CH05 method. The development level of 20 bones in the wrist is evaluated respectively, and then the bone age assessment value is calculated. In summary, bone age assessment is a subjective comparison process, and bone age assessment is a time-consuming and labor-intensive task for doctors. Using deep learning methods to assist doctors in bone age assessment can greatly reduce the workload of doctors and effectively ensure the stability of prediction results. However, existing deep learning methods often use very large backbone models as networks, which not only waste time but also consume more energy when performing reasoning.

[0003] The invention patent with application number 202110613678.7 discloses a method for classifying diabetic retinopathy images based on an improved ResNeSt convolutional neural network model. The method is as follows: first, obtain lesion images from the hospital; preprocess the images, manually annotate them by ophthalmologists, and divide the data sets; then build the deep learning server platform required for the experiment, and then write the python code; introduce two lightweight and efficient convolution operations, OctConv and SPConv, into the ResNeSt convolutional neural network, and introduce the learning rate adjustment mechanism of Warm Restart and cosine annealing. The network design of this method is extremely fragmented, that is, there are multi-path branches in the network, and the fragmentation of the network structure will reduce the parallel efficiency; secondly, this network design contains a lot of element-level operations (element-level operations refer to the element-by-element operations on the output features of the neural network, such as element-by-element addition, subtraction, multiplication and division, etc.), which means that there will be a large MAC (memory access cost). The above two aspects jointly determine that the network cannot be applied to embedded systems. Summary of the invention

[0004] In order to solve the above problems existing in the prior art, the present invention provides a bone age prediction method and device based on a lightweight model.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions.

[0006] In a first aspect, the present invention provides a bone age prediction method based on a lightweight model, comprising the following steps:

[0007] Acquire a hand image of the detected person;

[0008] A segmentation model is established based on a modeling method that integrates lightweight networks shufflenet and ghostnet, the hand image is input into the trained segmentation model, and a hand segmentation result is output;

[0009] A bone age prediction model is established based on a modeling method that integrates lightweight networks shufflenet and ghostnet, and the hand image and hand segmentation results are input into the trained bone age prediction model to obtain the bone age of the person being tested.

[0010] Furthermore, the method also includes an image preprocessing step: denoising the acquired hand image; adjusting the contrast and brightness of the image; rotating the image to eliminate the tilt; cropping the image, and adjusting the image size to 512 pixels*512 pixels by adding padding.

[0011] Furthermore, the modeling method based on the fusion lightweight network shufflenet and ghostnet includes:

[0012] The input feature map is equally divided into a first feature map and a second feature map along the channel direction by using a channel segmentation module;

[0013] Input the first feature map into a combined module consisting of a convolutional layer, a batch normalization layer, and a rectified linear unit, and output a third feature map with half the number of channels;

[0014] Inputting the third feature map into a combination module consisting of a depthwise convolutional layer, a batch normalization layer, and a rectified linear unit, and connecting the output of the combination module with the third feature map to obtain a fourth feature map;

[0015] The fourth feature map is concatenated with the second feature map and input into a combined module consisting of a single-point convolution layer, a batch normalization layer, and a rectified linear unit to obtain a fifth feature map fused in the channel direction.

[0016] Furthermore, the segmentation model established based on the modeling method of fusing lightweight networks shufflenet and ghostnet is a DeASPP-Unet structure, which is obtained by replacing the cross-layer connection, downsampling layer, and upsampling layer in the original Unet with the atrous spatial pyramid pooling layer, the downsampling atrous spatial pyramid pooling layer, and the upsampling anti-atrous spatial pyramid pooling layer, respectively.

[0017] Furthermore, the atrous spatial pyramid pooling layer includes three parallel combination modules consisting of a convolution layer, a batch normalization layer and a rectified linear unit and one combination module consisting of a single point convolution layer, a batch normalization layer and a rectified linear unit. The convolution kernels of the three parallel combination modules are all 3*3, the step sizes are all 1, and the atrous rates are 2, 4, and 6, respectively.

[0018] Furthermore, the downsampling atrous spatial pyramid pooling layer includes three parallel combination modules consisting of a convolution layer, a batch normalization layer and a rectified linear unit and one combination module consisting of a single point convolution layer, a batch normalization layer and a rectified linear unit. The convolution kernels of the three parallel combination modules are all 3*3, the step sizes are all 2, and the atrous rates are 2, 4, and 6, respectively.

[0019] Furthermore, the upsampling anti-cavity spatial pyramid pooling layer includes three parallel combination modules consisting of a deconvolution layer, a batch normalization layer and a rectified linear unit and one combination module consisting of a single point convolution layer, a batch normalization layer and a rectified linear unit. The convolution kernels of the three parallel combination modules are all 3*3, the step size is all 1, the expansion rate is all 2, and the cavity rates are 2, 4, and 6 respectively.

[0020] In a second aspect, the present invention provides a bone age prediction device based on a lightweight model, comprising:

[0021] An image acquisition module, used to acquire a hand image of a detected person;

[0022] The first modeling module is used to establish a segmentation model based on a modeling method that integrates lightweight networks shufflenet and ghostnet, input the hand image into the trained segmentation model, and output a hand segmentation result;

[0023] The second modeling module is used to establish a bone age prediction model based on a modeling method that integrates lightweight networks shufflenet and ghostnet, and input the hand image and hand segmentation results into the trained bone age prediction model to obtain the bone age of the person being tested.

[0024] Furthermore, the device also includes an image preprocessing module, which is used to perform denoising on the acquired hand image; adjust the contrast and brightness of the image; rotate the image to eliminate the tilt; crop the image, and adjust the image size to 512 pixels*512 pixels by adding padding.

[0025] Furthermore, the modeling method based on the fusion lightweight network shufflenet and ghostnet includes:

[0026] The input feature map is equally divided into a first feature map and a second feature map along the channel direction by using a channel segmentation module;

[0027] Input the first feature map into a combined module consisting of a convolutional layer, a batch normalization layer, and a rectified linear unit, and output a third feature map with half the number of channels;

[0028] Inputting the third feature map into a combination module consisting of a depthwise convolutional layer, a batch normalization layer, and a rectified linear unit, and connecting the output of the combination module with the third feature map to obtain a fourth feature map;

[0029] The fourth feature map is concatenated with the second feature map and input into a combined module consisting of a single-point convolution layer, a batch normalization layer, and a rectified linear unit to obtain a fifth feature map fused in the channel direction.

[0030] Furthermore, the segmentation model established based on the modeling method of fusing lightweight networks shufflenet and ghostnet is a DeASPP-Unet structure, which is obtained by replacing the cross-layer connection, downsampling layer, and upsampling layer in the original Unet with the atrous spatial pyramid pooling layer, the downsampling atrous spatial pyramid pooling layer, and the upsampling anti-atrous spatial pyramid pooling layer, respectively.

[0031] Furthermore, the atrous spatial pyramid pooling layer includes three parallel combination modules consisting of a convolution layer, a batch normalization layer and a rectified linear unit and one combination module consisting of a single point convolution layer, a batch normalization layer and a rectified linear unit. The convolution kernels of the three parallel combination modules are all 3*3, the step sizes are all 1, and the atrous rates are 2, 4, and 6, respectively.

[0032] Furthermore, the downsampling atrous spatial pyramid pooling layer includes three parallel combination modules consisting of a convolution layer, a batch normalization layer and a rectified linear unit and one combination module consisting of a single point convolution layer, a batch normalization layer and a rectified linear unit. The convolution kernels of the three parallel combination modules are all 3*3, the step sizes are all 2, and the atrous rates are 2, 4, and 6, respectively.

[0033] Furthermore, the upsampling anti-cavity spatial pyramid pooling layer includes three parallel combination modules consisting of a deconvolution layer, a batch normalization layer and a rectified linear unit and one combination module consisting of a single point convolution layer, a batch normalization layer and a rectified linear unit. The convolution kernels of the three parallel combination modules are all 3*3, the step size is all 1, the expansion rate is all 2, and the cavity rates are 2, 4, and 6 respectively.

[0034] Compared with the prior art, the present invention has the following beneficial effects.

[0035] The present invention obtains the hand image of the detected person, establishes a segmentation model and a bone age prediction model based on the modeling method of the fusion lightweight network shufflenet and ghostnet, and realizes an automatic prediction model of bone age. The present invention establishes a segmentation model and a bone age prediction model based on the modeling method of the fusion lightweight network shufflenet and ghostnet, which can not only reduce the amount of calculation and reduce time consumption, but also reduce the memory access cost MAC to save manufacturing costs, and is suitable for the network design mode of embedded system calculation. It is worth noting that the method described in the present invention is not only suitable for bone age prediction based on human hand images, but also suitable for disease type diagnosis and prediction based on any human organ image. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 The present invention is a flowchart of a method for predicting bone age based on a lightweight model according to an embodiment of the present invention.

[0037] Figure 2 This is a schematic diagram of the structure of the Shuffleghost module.

[0038] Figure 3 Schematic diagram of the DeASPP-Unet network structure.

[0039] Figure 4 Schematic diagram of the structure of the atrous spatial pyramid pooling layer.

[0040] Figure 5 Schematic diagram of the downsampling atrous spatial pyramid pooling layer structure.

[0041] Figure 6 Schematic diagram of the upsampling anti-cavity spatial pyramid pooling layer structure.

[0042] Figure 7 The present invention is a block diagram of a bone age prediction device based on a lightweight model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solution and advantages of the present invention clearer and more understandable, the present invention is further described below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0044] Figure 1 The present invention is a flowchart of a method for predicting bone age based on a lightweight model, comprising the following steps:

[0045] Step 101, obtaining a hand image of a detected person;

[0046] Step 102, establishing a segmentation model based on a modeling method that integrates lightweight networks shufflenet and ghostnet, inputting the hand image into the trained segmentation model, and outputting a hand segmentation result;

[0047] Step 103, establish a bone age prediction model based on the modeling method of integrating lightweight networks shufflenet and ghostnet, input the hand image and hand segmentation result into the trained bone age prediction model to obtain the bone age of the person being tested.

[0048] In this embodiment, step 101 is mainly used to obtain a hand image of the person being tested. In this embodiment, bone age prediction is performed based on the hand image of the person being tested, so it is necessary to obtain a hand image of the person being tested from a hospital. Generally, it is an image of the left hand.

[0049] In this embodiment, step 102 is mainly used to obtain the hand segmentation result by establishing a segmentation model. In order to obtain a lightweight segmentation model, this embodiment proposes a modeling method that integrates lightweight networks shufflenet and ghostnet. Shufflenet and ghostnet are two commonly used lightweight CNN network models that both achieve model size reduction and speed by designing more efficient network structures. The core of Shufflenet is the use of two operations: pointwise groupconvolution and channel shuffle, which greatly reduces the computational complexity of the model while maintaining accuracy. Ghostnet decomposes the ordinary convolution layer into two parts. The first part contains normal convolution, but the number of convolutions will be strictly controlled. After the inherent feature map of the first part is given, a series of simple linear operations are then applied to generate more feature maps. Compared with ordinary convolutional neural networks, the total number of parameters and computational complexity required in Ghostnet have been reduced without changing the size of the output feature map. The modeling method proposed in this embodiment combines the advantages of the two lightweight models of Shufflenet and Ghostnet. For simplicity, we represent this modeling method that integrates Shufflenet and Ghostnet with the Shuffleghost module structure. The following embodiments will provide a modeling method using the Shuffleghost module structure. The segmentation model is trained by constructing a training data set containing manual annotations, and the hand image of the detected person is input into the trained segmentation model to obtain a hand segmentation result. Strictly speaking, the hand segmentation result is not a hand segmentation image, but a mask. The mask is generally the same size as the image. 0 is used at the corresponding position to indicate that the pixel at that position in the original image is the background, and 1 is used to indicate the foreground.

[0050] In the present embodiment, step 103 is mainly used to predict the bone age of the person being tested by establishing a bone age prediction model. In order to obtain a lightweight bone age prediction model, the present embodiment also establishes a bone age prediction model based on a modeling method that integrates lightweight networks shufflenet and ghostnet, and the bone age prediction model adopts a network structure similar to the MobilenetV2 network. The MobilenetV2 network draws on the idea of ​​Resnet, and uses the bottleneck module as the basic module for building a network. The bottleneck module consists of a single point convolution layer pointwise that amplifies the number of channels, a convolution layer along the depth direction, and a pointwise convolution layer that reduces the number of channels as the backbone, and the input feature map is mapped identically, and the two feature maps are finally added element by element. And the present embodiment uses the shuffleghost module to replace the bottleneck module in Mobilenetv2 to build the network. The structural parameters of the bone age prediction model are shown in Table 1.

[0051] Table 1 Structural parameters of the bone age prediction model

[0052]

[0053] The feature map size in the table is expressed as W*H*C, where W, H, and C represent width, height, and number of channels, respectively.

[0054] As an optional embodiment, the method also includes an image preprocessing step: denoising the acquired hand image; adjusting the contrast and brightness of the image; rotating the image so that the axis of symmetry of the hand image is vertically upward; cropping the image and adjusting the image size to 512 pixels * 512 pixels by adding padding.

[0055] This embodiment provides a technical solution for image preprocessing. Before the hand image of the person being tested is input into the model, or when the model training data set is constructed, some preprocessing of the acquired original hand image is required to improve the modeling or prediction accuracy. Image preprocessing generally includes the following: eliminating noise in the hand image; adjusting contrast and brightness; rotating the image to eliminate tilt (making the axis of symmetry of the hand image vertically upward) to make the hand position more standard; cropping the image to increase the proportion of the hand position in the image to ensure the information integrity of the effective area, and adjusting the image size to 512 pixels * 512 pixels by adding padding to prevent distortion. Generally, professional physicians are also required to annotate the image, including the C-bone and R-bone masks that are usually used, as well as bone age information. The data set is divided into 80%-10%-10% ratios, of which 80% is the training set, 10% is the validation set, and 10% is the test set.

[0056] As an optional embodiment, the modeling method based on the fusion of lightweight networks shufflenet and ghostnet includes:

[0057] The input feature map is equally divided into a first feature map and a second feature map along the channel direction by using a channel segmentation module;

[0058] Input the first feature map into a combined module consisting of a convolutional layer, a batch normalization layer, and a rectified linear unit, and output a third feature map with half the number of channels;

[0059] Inputting the third feature map into a combination module consisting of a depthwise convolutional layer, a batch normalization layer, and a rectified linear unit, and connecting the output of the combination module with the third feature map to obtain a fourth feature map;

[0060] The fourth feature map is concatenated with the second feature map and input into a combined module consisting of a single-point convolution layer, a batch normalization layer, and a rectified linear unit to obtain a fifth feature map fused in the channel direction.

[0061] This embodiment provides a technical solution based on a modeling method that integrates lightweight networks shufflenet and ghostnet. Figure 2 This is a schematic diagram of the structure of the module Shuffleghost that integrates shufflenet and ghostnet. Figure 2 First, the channel segmentation module evenly divides the input feature map along the channel direction to obtain the first feature map and the second feature map. Among them, the second feature map uses a shortcut connection to directly propagate downward without passing through any neural network; the first feature map is input into a combination module composed of a convolutional layer, a batch normalization layer (Batch Normalization, BN) and a rectified linear unit (RectifiedLinearUnit, ReLU), and outputs a third feature map with half the number of channels (for example, if the size of the first feature map is 128 (width) * 128 (height) * 32 (number of channels), the size of the third feature map output after the combination module becomes 128 * 128 * 32). The third feature map Figure 1 The first path is input into a combined module consisting of a depthwise convolution layer, a batch normalization layer, and a rectified linear unit, and its output is concatenated with the third feature map that is propagated downward unchanged in another path to obtain the fourth feature map. Then, the fourth feature map is connected with the second feature map that is directly propagated downward and input into a combined module consisting of a pointwise convolution layer, a batch normalization layer, and a rectified linear unit to obtain the fifth feature map. The last step is mainly used for channel fusion.

[0062] This embodiment first performs a channel segmentation operation on the input feature map, one branch of which directly propagates downward to the second connection operation layer, which can save memory usage; while the other branch needs to pass through a combination module consisting of a convolution layer, a batch normalization layer, and a rectified linear unit, and a combination module consisting of a convolution layer, a batch normalization layer, and a rectified linear unit along the depth direction. This operation can improve the model's ability to extract features, and at the same time, the use of a convolution layer along the depth direction ensures that the model has less computational complexity to ensure the lightweight of the model. The two connection operations avoid pixel-by-pixel operations, reducing computational complexity and memory access costs.

[0063] As an optional embodiment, a segmentation model established based on a modeling method that integrates lightweight networks shufflenet and ghostnet is a DeASPP-Unet structure, which is obtained by replacing the cross-layer connection, downsampling layer, and upsampling layer in the original Unet with a dilated spatial pyramid pooling layer, a downsampling dilated spatial pyramid pooling layer, and an upsampling anti-dilated spatial pyramid pooling layer, respectively.

[0064] This embodiment provides a technical solution for a segmentation model based on the Shuffleghost module structure. The segmentation model established is a DeASPP-Unet (De-Atrous Spatial Pyramid Pooling) structure, such as Figure 3 As shown. DeASPP-Unet is a variation of the original Unet: the original cross-layer connection is replaced by the Atrous Spatial Pyramid Pooling (ASPP), the original down-sampling layer is replaced by the Down Sample Atrous Spatial Pyramid Pooling (Down SampleAtrous Spatial Pyramid Pooling), and the original up-sampling layer is replaced by the De-Atrous Spatial Pyramid Pooling (De-Atrous Spatial Pyramid Pooling). This embodiment can increase the receptive field by using convolution kernels with different atrous rates for feature extraction; at the same time, multi-way parallel convolution increases the model's ability to extract features. The combined effect of these two aspects allows the network depth of the segmentation model to be appropriately reduced, which can avoid increasing the amount of calculation too much.

[0065] As an optional embodiment, the atrous spatial pyramid pooling layer includes three parallel combination modules consisting of a convolution layer, a batch normalization layer and a rectified linear unit and one combination module consisting of a single point convolution layer, a batch normalization layer and a rectified linear unit. The convolution kernels of the three parallel combination modules are all 3*3, the step sizes are all 1, and the atrous rates are 2, 4, and 6, respectively.

[0066] This embodiment provides a technical solution for the hollow space pyramid pooling layer. Its specific structural parameters are as follows: Figure 4 As shown, no detailed description is given here.

[0067] As an optional embodiment, the downsampling hole spatial pyramid pooling layer includes three parallel combination modules consisting of a convolution layer, a batch normalization layer and a rectified linear unit and one combination module consisting of a single point convolution layer, a batch normalization layer and a rectified linear unit. The convolution kernels of the three parallel combination modules are all 3*3, the step sizes are all 2, and the hole rates are 2, 4, and 6 respectively.

[0068] This embodiment provides a technical solution for downsampling the hole space pyramid pooling layer. Its specific structural parameters are as follows: Figure 5 As shown, no detailed description is given here.

[0069] As an optional embodiment, the upsampling anti-cavity spatial pyramid pooling layer includes three parallel combination modules consisting of a deconvolution layer, a batch normalization layer and a rectified linear unit, and one combination module consisting of a single point convolution layer, a batch normalization layer and a rectified linear unit. The convolution kernels of the three parallel combination modules are all 3*3, the step size is all 1, the expansion rate is all 2, and the cavity rates are 2, 4, and 6 respectively.

[0070] This embodiment provides a technical solution for upsampling the anti-cavity spatial pyramid pooling layer. Its specific structural parameters are as follows: Figure 6 As shown, no detailed description is given here.

[0071] Figure 7 The present invention is a schematic diagram of a bone age prediction device based on a lightweight model according to an embodiment of the present invention, wherein the device comprises:

[0072] An image acquisition module 11 is used to acquire a hand image of a detected person;

[0073] The first modeling module 12 is used to establish a segmentation model based on a modeling method that integrates lightweight networks shufflenet and ghostnet, input the hand image into the trained segmentation model, and output a hand segmentation result;

[0074] The second modeling module 13 is used to establish a bone age prediction model based on a modeling method that integrates lightweight networks shufflenet and ghostnet, and input the hand image and hand segmentation results into the trained bone age prediction model to obtain the bone age of the person being tested.

[0075] The device of this embodiment can be used to perform Figure 1The technical solution of the method embodiment shown has similar implementation principles and technical effects, which will not be described in detail here. The same is true for the following embodiments, which will not be described in detail.

[0076] As an optional embodiment, the device also includes an image preprocessing module, which is used to denoise the acquired hand image; adjust the contrast and brightness of the image; rotate the image to eliminate the tilt; crop the image, and adjust the image size to 512 pixels * 512 pixels by adding padding.

[0077] As an optional embodiment, the modeling method based on the fusion of lightweight networks shufflenet and ghostnet includes:

[0078] The input feature map is equally divided into a first feature map and a second feature map along the channel direction by using a channel segmentation module;

[0079] Input the first feature map into a combined module consisting of a convolutional layer, a batch normalization layer, and a rectified linear unit, and output a third feature map with half the number of channels;

[0080] Inputting the third feature map into a combination module consisting of a depthwise convolutional layer, a batch normalization layer, and a rectified linear unit, and connecting the output of the combination module with the third feature map to obtain a fourth feature map;

[0081] The fourth feature map is concatenated with the second feature map and input into a combined module consisting of a single-point convolution layer, a batch normalization layer, and a rectified linear unit to obtain a fifth feature map fused in the channel direction.

[0082] As an optional embodiment, a segmentation model established based on a modeling method that integrates lightweight networks shufflenet and ghostnet is a DeASPP-Unet structure, which is obtained by replacing the cross-layer connection, downsampling layer, and upsampling layer in the original Unet with a dilated spatial pyramid pooling layer, a downsampling dilated spatial pyramid pooling layer, and an upsampling anti-dilated spatial pyramid pooling layer, respectively.

[0083] As an optional embodiment, the atrous spatial pyramid pooling layer includes three parallel combination modules consisting of a convolution layer, a batch normalization layer and a rectified linear unit and one combination module consisting of a single point convolution layer, a batch normalization layer and a rectified linear unit. The convolution kernels of the three parallel combination modules are all 3*3, the step sizes are all 1, and the atrous rates are 2, 4, and 6, respectively.

[0084] As an optional embodiment, the downsampling hole spatial pyramid pooling layer includes three parallel combination modules consisting of a convolution layer, a batch normalization layer and a rectified linear unit and one combination module consisting of a single point convolution layer, a batch normalization layer and a rectified linear unit. The convolution kernels of the three parallel combination modules are all 3*3, the step sizes are all 2, and the hole rates are 2, 4, and 6 respectively.

[0085] As an optional embodiment, the upsampling anti-cavity spatial pyramid pooling layer includes three parallel combination modules consisting of a deconvolution layer, a batch normalization layer and a rectified linear unit, and one combination module consisting of a single point convolution layer, a batch normalization layer and a rectified linear unit. The convolution kernels of the three parallel combination modules are all 3*3, the step size is all 1, the expansion rate is all 2, and the cavity rates are 2, 4, and 6 respectively.

[0086] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A bone age prediction method based on a lightweight model, characterized in that: The following steps are involved: Obtain a hand image of the person being tested; the hand image is obtained from a hospital; A segmentation model is established based on a modeling method that integrates lightweight networks shufflenet and ghostnet, the hand image is input into the trained segmentation model, and a hand segmentation result is output; A bone age prediction model is established based on a modeling method that integrates lightweight networks shufflenet and ghostnet, and the hand image and hand segmentation result are input into the trained bone age prediction model to obtain the bone age of the person being tested; The modeling method of the fused lightweight network shufflenet and ghostnet is implemented by the Shuffleghost module, including: The input feature map is equally divided into a first feature map and a second feature map along the channel direction by using a channel segmentation module; Input the first feature map into a combined module consisting of a convolutional layer, a batch normalization layer, and a rectified linear unit, and output a third feature map with half the number of channels; Inputting the third feature map into a combination module consisting of a depthwise convolutional layer, a batch normalization layer, and a rectified linear unit, and connecting the output of the combination module with the third feature map to obtain a fourth feature map; The fourth feature map is concatenated with the second feature map and then input into a combination module consisting of a single point convolution layer, a batch normalization layer, and a rectified linear unit to obtain a fifth feature map fused in the channel direction; The segmentation model established based on the modeling method of integrating the lightweight network shufflenet and ghostnet is the DeASPP-Unet structure, which is obtained by replacing the cross-layer connection, downsampling layer, and upsampling layer in the original Unet with the void space pyramid pooling layer, the downsampling void space pyramid pooling layer, and the upsampling anti-void space pyramid pooling layer respectively; The bone age prediction model was established based on the modeling method of integrating the lightweight networks shufflenet and ghostnet. It adopted a network structure similar to the Mobilenet V2 network and was obtained by replacing the bottleneck module in Mobilenetv2 with the Shuffleghost module.

2. The bone age prediction method based on a lightweight model according to claim 1, characterized in that: The method also includes an image preprocessing step: performing denoising on the acquired hand image; adjusting the contrast and brightness of the image; rotating the image to eliminate tilt; and cropping the image, and adjusting the image size to 512 pixels*512 pixels by adding padding.

3. The bone age prediction method based on a lightweight model according to claim 1, characterized in that: The atrous spatial pyramid pooling layer includes three parallel combination modules consisting of a convolution layer, a batch normalization layer and a rectified linear unit and one combination module consisting of a single point convolution layer, a batch normalization layer and a rectified linear unit. The convolution kernels of the three parallel combination modules are all 3*3, the step size is all 1, and the atrous rates are 2, 4, and 6 respectively.

4. The bone age prediction method based on a lightweight model according to claim 1, characterized in that: The downsampling atrous spatial pyramid pooling layer includes three parallel combination modules consisting of a convolution layer, a batch normalization layer and a rectified linear unit and one combination module consisting of a single point convolution layer, a batch normalization layer and a rectified linear unit. The convolution kernels of the three parallel combination modules are all 3*3, the step size is all 2, and the atrous rates are 2, 4, and 6 respectively.

5. The bone age prediction method based on a lightweight model according to claim 1, characterized in that: The upsampling anti-cavity spatial pyramid pooling layer includes three parallel combination modules consisting of a deconvolution layer, a batch normalization layer and a rectified linear unit and one combination module consisting of a single point convolution layer, a batch normalization layer and a rectified linear unit. The convolution kernels of the three parallel combination modules are all 3*3, the step size is all 1, the expansion rate is all 2, and the cavity rates are 2, 4, and 6 respectively.

6. A bone age prediction device based on a lightweight model, characterized in that: include: An image acquisition module is used to acquire a hand image of the person being tested; the hand image is acquired from a hospital; The first modeling module is used to establish a segmentation model based on a modeling method that integrates lightweight networks shufflenet and ghostnet, input the hand image into the trained segmentation model, and output a hand segmentation result; The second modeling module is used to establish a bone age prediction model based on a modeling method that integrates lightweight networks shufflenet and ghostnet, and input the hand image and hand segmentation result into the trained bone age prediction model to obtain the bone age of the person being tested; The modeling method of the fused lightweight network shufflenet and ghostnet is implemented by the Shuffleghost module, including: The input feature map is equally divided into a first feature map and a second feature map along the channel direction by using a channel segmentation module; Input the first feature map into a combined module consisting of a convolutional layer, a batch normalization layer, and a rectified linear unit, and output a third feature map with half the number of channels; Inputting the third feature map into a combination module consisting of a depthwise convolutional layer, a batch normalization layer, and a rectified linear unit, and connecting the output of the combination module with the third feature map to obtain a fourth feature map; The fourth feature map is concatenated with the second feature map and then input into a combination module consisting of a single point convolution layer, a batch normalization layer, and a rectified linear unit to obtain a fifth feature map fused in the channel direction; The segmentation model established based on the modeling method of integrating the lightweight network shufflenet and ghostnet is the DeASPP-Unet structure, which is obtained by replacing the cross-layer connection, downsampling layer, and upsampling layer in the original Unet with the void space pyramid pooling layer, the downsampling void space pyramid pooling layer, and the upsampling anti-void space pyramid pooling layer respectively; The bone age prediction model was established based on the modeling method of integrating the lightweight networks shufflenet and ghostnet. It adopted a network structure similar to the Mobilenet V2 network and was obtained by replacing the bottleneck module in Mobilenetv2 with the Shuffleghost module.

7. The bone age prediction device based on a lightweight model according to claim 6, characterized in that: The device also includes an image preprocessing module, which is used to perform denoising on the acquired hand image; adjust the contrast and brightness of the image; rotate the image to eliminate tilt; crop the image, and adjust the image size to 512 pixels*512 pixels by adding padding.