A method for automatically measuring the thickness of subcutaneous soft tissue in the fetal abdomen in ultrasound images

Abstract

The present invention discloses a method for automatically measuring the thickness of the subcutaneous soft tissue of the fetal abdomen in an ultrasound image. The subcutaneous soft tissue area of ​​the abdomen that needs to be measured in the fetal abdominal circumference ultrasound image is extracted using a VGG‑SeUnet network, and each pixel in the image is classified by a pixel classifier to obtain the fetal subcutaneous soft tissue segmentation result; a DF_loss loss function is constructed to guide the training of the VGG‑SeUnet network; the thickness of the fetal subcutaneous soft tissue is then calculated by a distance transformation algorithm; finally, the thickness of the fetal subcutaneous soft tissue is converted into centimeters to obtain a fetal subcutaneous soft tissue thickness that meets clinical needs. The present invention is trained in an end-to-end manner through the VGG‑SeUnet network, without the need for manual feature extraction, and has superior segmentation performance compared to existing methods; the method of the present invention can automatically measure the thickness of the fetal abdominal subcutaneous soft tissue, thereby helping doctors to subsequently evaluate the development and nutritional status of the fetus.

Description

Technical Field

[0001] The present invention relates to the field of deep learning, and in particular to a method for automatically measuring the thickness of subcutaneous soft tissue in the fetal abdomen in ultrasound images. Background Art

[0002] Intrauterine fetal growth is a complex process. The literature [Xie Xing, Kong Beihua, Duan Tao. Obstetrics and Gynecology [M]. 9th Edition, People's Medical Publishing House, 2019: 53.] and [Pascal A., Govaert P., Oostra A., Naulaers G., Ortibus E., Van Den Broeck C.. Neurodevelopmental Outcome in Very Preterm and Very-Low-Birthweight Infants Born Over the Past Decade: A Meta-Analytic Review. Developmental Medicine & Child Neurology, 2018, 60(4): 342-355.] indicates that intrauterine nutritional imbalance (including intrauterine malnutrition and overnutrition) can lead to fetal maldevelopment. Traditional examinations make it difficult to directly and accurately determine the nutritional status of the intrauterine fetus. Therefore, in the literature [Song Qingyun, Luo Hong. A brief discussion on the importance of using ultrasound indicators to evaluate fetal growth and development [J]. Chinese Journal of Medical Ultrasound (Electronic Edition), 2021, 18 (08): 733-736.], it is pointed out that ultrasound imaging has been regarded as an important clinical method for evaluating fetal growth due to its advantages of being non-invasive, non-radioactive and easy to perform. In previous studies [Lu Qing, Sun Ying. Clinical observation on the prediction of fetal weight by ultrasound measurement of fetal abdominal subcutaneous tissue thickness [J]. Chinese Journal of Practical Gynecology and Obstetrics. 2007 (06): 450-451.], [ N., A..Can Fetal Abdominal Visceral Adipose Tissue and Subcutaneous Fat Thickness Be Used for Correct Estimation of Fetal Weight? A Preliminary Study. Journal of Obstetrics and Gynaecology, 2019, 39(5): 594-600.] and [Khalifa EA, Hassanein SA, Eid H.H..Ultrasound Measurement of Fetal Abdominal Subcutaneous Tissue Thickness As A Predictor of Large Versus Small Fetuses for Gestational Age. Egyptian Journal of Radiology and Nuclear Medicine, 2019, 50(1): 80.] found that fetal abdominal subcutaneous soft tissue thickness (FASSTT) was significantly correlated with actual birth weight (ABW) of newborns. Therefore, the measurement of FASSTT in ultrasound images is of great significance for assessing fetal nutritional status.

[0003] Conventional FASSTT in ultrasound images involves segmenting the fetal abdominal subcutaneous soft tissue region and manually locating the thickest portion of this region, where the FASSTT can be measured. However, the fetal abdominal subcutaneous soft tissue region only occupies approximately 10% of the entire abdominal circumference in ultrasound images and is typically 4-7 mm thick. In an ultrasound image with a 96 dpi resolution, a manual annotation error of just 3-4 pixels will result in an error of 1 mm, or 14.3%-25% of the total thickness. Current methods rely heavily on the physician's experience; therefore, automated segmentation of fetal subcutaneous soft tissue in ultrasound images and accurate measurement of its thickness are of great importance.

[0004] In recent years, deep learning methods, particularly convolutional neural networks (CNNs), have been successfully applied to fetal biometry. The paper [Oghli MG, Shabanzadeh A., Moradi S., Sirjani N., Gerami R., Ghaderi P., Taheri MS, Shiri I., Arabi H., Zaidi H.. Automatic Fetal Biometry Prediction Using A Novel Deep Convolutional Network Architecture. Physica Medica, 2021, 88:127-137.] states that current progress has primarily focused on automated measurement of fetal head circumference, biparietal diameter, femur length, and abdominal circumference. However, there is still a lack of effective automated solutions for FASSTT measurement. There are two reasons for this: first, due to the complexity of fetal organs and tissues, no clear boundaries can be observed in the subcutaneous soft tissue area; second, FASSTT's automatic measurement relies on accurate tissue segmentation, where the pixels of the tissue area (positive sample) are much smaller than those of the background area (negative sample), and mainstream models give equal weight to pixels that are easy to classify (positive samples or negative samples) and pixels that are difficult to classify, and thus fail to pay high attention to difficult-to-classify samples, resulting in sample imbalance and reduced segmentation performance. Summary of the Invention

[0005] The purpose of the present invention is to address the deficiencies of the prior art and to provide a method for automatically measuring the thickness of subcutaneous soft tissue in the fetal abdomen in ultrasound images.

[0006] The object of the present invention is achieved through the following technical solution: a method for automatically measuring the thickness of subcutaneous soft tissue in the fetal abdomen in an ultrasound image, comprising the following steps:

[0007] (1) For the abdominal subcutaneous soft tissue area that needs to be measured in the fetal abdominal circumference ultrasound image, the VGG-SeUnet network is used to extract its features, and each pixel in the image is classified by the pixel classifier to obtain the fetal subcutaneous soft tissue segmentation result;

[0008] (2) The Dice_Loss loss function and the Focal_Loss loss function are combined to construct the DF_loss loss function to guide the training of the VGG-SeUnet network; the DF_loss loss function is used to evaluate the ratio between the intersection of the predicted tissue area and the real tissue area and the union of the two areas, and different weights are assigned to the easy-to-classify pixels and the difficult-to-classify pixels in the image to obtain the fetal subcutaneous soft tissue segmentation results;

[0009] (3) converting the fetal subcutaneous soft tissue segmentation results obtained in steps (1) and (2) into a binary image, wherein the tissue area of ​​the fetal subcutaneous soft tissue segmentation result is represented by 1, i.e., a non-zero pixel point, and the background area is represented by 0, i.e., a zero pixel point; then, using a distance transformation algorithm, a set is used to store the minimum distance between each non-zero pixel point and all zero pixel points, and the maximum distance is retrieved from the set; the pixel point corresponding to the retrieved maximum distance is regarded as the center of the maximum inscribed circle, and the maximum inscribed circle is drawn with the maximum distance as the radius, wherein the diameter of the maximum inscribed circle is the thickness of the fetal subcutaneous soft tissue;

[0010] (4) Convert the pixel unit of the fetal subcutaneous soft tissue thickness into centimeter unit to obtain the fetal subcutaneous soft tissue thickness that meets clinical needs.

[0011] Furthermore, the step (1) includes the following sub-steps:

[0012] (1.1) collecting fetal abdominal circumference ultrasound imaging data, i.e., fetal abdominal circumference ultrasound images, and using the collected fetal abdominal circumference ultrasound imaging data as an image dataset;

[0013] (1.2) normalizing the image dataset and converting the resolution of the images in the image dataset to 512×512; then randomly selecting 20% ​​of the images in the image dataset as a test set and the remaining 80% as a training set;

[0014] (1.3) Inputting the training set into the VGG-SeUnet network for training, and extracting multi-scale feature fusion results reflecting tissue segmentation through the encoding-decoding U-shaped architecture of the jump connection in the VGG-SeUnet network;

[0015] (1.4) Introducing channel attention, the feature map obtained during each jump connection of the VGG-SeUnet network is reweighted by the channel attention;

[0016] (1.5) Apply 1×1 convolution to the weighted feature map through the pixel classifier to reduce its channel dimension from 64 to 32, and use the LeakyReLU function to activate the feature, and then classify each pixel in the image to obtain the segmentation result of the fetal subcutaneous soft tissue.

[0017] Furthermore, the channel attention in the step (1.4) is implemented by the following operations: the channel attention compresses each feature map by global average pooling to obtain a 1×1×C vector, and then normalizes the vector by two fully connected layers activated by Sigmund function to obtain the weight of each channel; finally, the channel is recalibrated by multiplying the weight with the original input channel.

[0018] Furthermore, in step (2), the expression of the DF_loss loss function is:

[0019]

[0020] Where DF_Loss represents the DF_loss loss function, X p represents the predicted tissue area, Y g represents the actual soft tissue area marked, N represents the total number of samples, and x i represents a single pixel in the image, c represents the annotation label, P(x i ) represents pixel x i Predict the probability of belonging to the soft tissue area, γ and α are hyperparameters that control positive and negative samples.

[0021] Compared with the prior art, the beneficial effect of the present invention is that the present invention combines the VGG-SeUnet network, the DF_Loss loss function and the morphology-based measurement algorithm to propose a method for automatically measuring the thickness of the fetal abdominal subcutaneous soft tissue. The present invention uses the segmentation network VGG-SeUnet for training in an end-to-end manner, without the need for manual feature extraction and without consuming a large amount of computing resources; compared with the most advanced existing networks, it has superior segmentation performance; and the morphology-based measurement algorithm of the present invention has measurement results comparable to manual ones. The method of the present invention can automatically measure the thickness of the fetal abdominal subcutaneous soft tissue, thereby helping doctors to subsequently evaluate the fetal development and nutritional status. Through experimental verification, the measurement results of the FASSTT of the present invention have an average error of 0.0615cm and a root mean square error of 0.081cm compared with the results provided by radiologists; the overall performance exceeds the existing methods, proving the effectiveness of the method of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a schematic flow diagram of the method of the present invention;

[0023] Figure 2 This is a structural diagram of the model for segmenting the fetal subcutaneous soft tissue area proposed by the present invention;

[0024] Figure 3 It is a structural diagram of the channel attention SE module in the present invention;

[0025] Figure 4 is a schematic diagram of the morphological measurement distance transformation algorithm of the present invention;

[0026] Figure 5 It is a schematic diagram of the fetal subcutaneous soft tissue prediction and segmentation results and measurement results of the present invention and other networks. DETAILED DESCRIPTION

[0027] The present invention mainly adopts U-shaped structure codec network, attention mechanism and morphological measurement methods to assist ultrasound imaging doctors to automatically measure the thickness of fetal abdominal subcutaneous soft tissue in clinical practice. The core technology of the present invention is to automatically measure the thickness of fetal abdominal subcutaneous soft tissue. A method for automatically measuring the thickness of fetal abdominal subcutaneous soft tissue in ultrasound images based on U-shaped attention network and maximum tangent circle detection is proposed to assist doctors in subsequent assessment of fetal development and nutritional status. Its main implementation process is as follows: Figure 1 The input image is a fetal abdominal circumference ultrasound image. First, a convolutional neural network is used to extract features of the fetal abdominal subcutaneous soft tissue. A pixel classifier is then used to classify the pixels in the image, distinguishing the fetal abdominal subcutaneous soft tissue area from the background. Next, the segmentation result is converted into a binary image, and a morphological method is used to draw the maximum inscribed circle to obtain the fetal abdominal subcutaneous soft tissue thickness (FASSTT).

[0028] The present invention provides a method for automatically measuring the thickness of subcutaneous soft tissue in the fetal abdomen in an ultrasound image, comprising the following steps:

[0029] (1) Automatic segmentation of fetal subcutaneous soft tissue area: The VGG-SeUnet network is used to extract features of the abdominal subcutaneous soft tissue area that needs to be measured in the fetal abdominal circumference ultrasound image. Each pixel in the image is classified by the pixel classifier to obtain the fetal subcutaneous soft tissue segmentation result, thereby realizing automatic segmentation.

[0030] Step (1) is one of the cores of the present invention, and specifically includes the following sub-steps:

[0031] (1.1) Data preparation: Fetal abdominal circumference ultrasound imaging data, i.e., fetal abdominal circumference ultrasound images, are collected and used as an image dataset.

[0032] The data used in this paper is fetal abdominal circumference ultrasound imaging data, that is, fetal abdominal circumference ultrasound images. These data were collected from August 2020 to November 2021, and two senior ultrasound chief physicians annotated the segmentation boundaries and tissue thickness in each image. The measurement results are annotated in the form of file names, and the data are desensitized. The file name consists of a number and a length in centimeters. The image dataset contains 135 fetal abdominal circumference images and 135 annotated images in PNG format, with resolutions of 1280×872 and 96 pixels / inch (DPI), respectively.

[0033] Data inclusion criteria: 1) intrauterine singleton pregnancy; 2) Chinese nationality, residency in East China for more than two years; 3) the last ultrasound examination before delivery was less than or equal to 3 days; 4) clear gestational age, confirmed by ultrasound measurement of fetal crown-rump length before 13 + 6 weeks of gestation; if ultrasound CRL results were not obtained, gestational age could be confirmed by fetal BPD, HC, AC, and FL during the second trimester ultrasound; 5) pregnancy documentation completed in the hospital.

[0034] Exclusion criteria: 1) intrauterine multiple pregnancy; 2) unknown gestational age; 3) one or both spouses are non-Chinese; 4) fetal abnormalities indicated by ultrasound or chromosome examination.

[0035] (1.2) The image dataset is normalized and the resolution of the images in the image dataset is converted to 512×512; then 20% of the images in the image dataset are randomly selected as the test set, and the remaining 80% are used as the training set.

[0036] Specifically, data preprocessing is performed first: To reduce the computational complexity of the model and improve performance, the image dataset from step (1.1) is normalized and converted to a resolution of 512×512 before being sent to the network for training. Changing the image resolution does not change the DPI, as DPI represents the number of pixels that can be rendered per inch of output on a display device. This resolution meets the model's requirements, preserves the image's structural details, and significantly reduces the model's computational cost.

[0037] Dataset Split: The image dataset was split, randomly selecting 20% ​​of the images as the test set and 80% as the training set. Furthermore, given the small size of the dataset, the training images were randomly augmented by flipping, scaling, orientation distortion, and color space transformation to avoid overfitting and enhance the model's generalization capabilities.

[0038] (1.3) Backbone network feature extraction: The training set in step (1.2) is input into the VGG-SeUnet network for training. The multi-scale feature fusion results reflecting the high-level and low-level semantics of tissue segmentation are extracted through the encoding-decoding U-shaped architecture of the jump connection in the VGG-SeUnet network.

[0039] Automatically segmenting fetal abdominal subcutaneous soft tissue from ultrasound images is difficult because organs and tissues are intertwined and it is difficult to determine tissue boundaries. In addition, the lack of rich training samples may lead to overfitting, so it is not suitable to use complex networks. To this end, the present invention proposes VGG-SeUnet, which adopts an encoding-decoding U-shaped architecture with jump connections to extract multi-scale feature representations that reflect high-level and low-level semantics of tissue segmentation. Jump connections enhance the fusion of information at different levels of detail while reducing the amount of parameters and the risk of overfitting. The detailed structure of VGG SeUnet is shown in Figure 2. Figure 2 shown.

[0040] The VGG-SeUnet network consists of an encoder and a decoder, with the encoder being the VGG16 feature extraction network. The VGG16 feature extraction network is divided into five modules (Blocks 1-5) as the encoder, each corresponding to the decoder and then skip-connected. This not only enhances encoder performance but also allows transfer learning using pre-trained weights, further strengthening the model's feature extraction capabilities. The VGG16 network consists of convolutional and pooling layers. The network input data is a (512, 512, 3) image. In Block 1, the input image undergoes two 3x3 convolutions and a ReLu activation function, resulting in a feature map of (512, 512, 64). This is then downsampled through a 2x2 maxpool layer, which compresses the image's height and width without changing the number of channels, resulting in an input of (256, 256, 128) for Block 2. Block 2 is identical to Block 1, and downsampling results in an input of (128, 128, 256) for Block 3. Block3, Block4, and Block5 all perform three 3*3 convolutions and ReLu activation functions on the input image, and finally obtain (32, 32, 512) which is connected to the decoder part. The extracted deep features are upsampled by transposed convolution, and each layer halves the number of feature channels of the previous layer. Finally, the features obtained by each module of the encoder are fused with the upsampling results through jump connections to achieve enhanced feature extraction.

[0041] (1.4) Introducing channel attention: The feature map obtained during each jump connection of the VGG-SeUnet network is reweighted by the channel attention.

[0042] Considering that the problem of feature information loss is inevitable during the convolution operation, and the channel selection is extremely important for the segmentation task, in the network model of step (1.3), during each jump connection, a channel attention module (Squeeze-Excitation Blcok, Se-Blcok) is introduced to reweight the new feature map. By modeling the correlation between different channels and key information, the original feature layer is weighted, highlighting the channels with key information and suppressing the remaining channels, so as to better perform feature fusion and enhance segmentation performance. Its structure is as follows: Figure 3 First, each feature map is compressed by global average pooling to obtain a 1×1×C vector. Then, two fully connected layers and sigmoid activation functions are introduced to normalize it and obtain the weight of each channel. Finally, the weight is multiplied by the original input to achieve channel recalibration.

[0043] (1.5) Pixel classifier: The pixel classifier applies 1×1 convolution to the weighted feature map in step (1.4) to reduce the channel dimension from 64 to 32, and uses the LeakyReLU function to activate the feature before passing it to the subsequent loss function. This classifies each pixel in the image and obtains the segmentation result of the fetal subcutaneous soft tissue.

[0044] (2) The Dice_Loss loss function and the Focal_Loss loss function are combined to construct the DF_loss loss function to guide the training of the VGG-SeUnet network; the DF_loss loss function is used to evaluate the ratio between the intersection of the predicted tissue area and the real tissue area and the union of the two areas, and different weights are assigned to the easy-to-classify pixels and the difficult-to-classify pixels in the image to obtain the fetal subcutaneous soft tissue segmentation results.

[0045] Automatic measurement of fetal abdominal subcutaneous soft tissue thickness FASSTT relies on the accurate distinction between the abdominal subcutaneous soft tissue area and the background area. However, the number of pixels in the tissue area to be segmented (positive sample) is significantly less than the number of pixels in the background area (negative sample). Sample imbalance hinders the training of segmentation networks based on traditional cross-entropy loss functions. The Dice loss function avoids the sample imbalance problem by evaluating the ratio between the intersection of the predicted and true tissue areas and the union of the two areas, considering only positive samples. However, the Dice_Loss loss does not take into account the case of small targets. Once some pixels of a small target are predicted incorrectly, the loss value will change drastically, resulting in drastic changes in the gradient and unstable training. The expression of the Dice loss function is as follows:

[0046]

[0047] Where, represents the predicted tissue area, Represents a pixel, represents the actual soft tissue area marked, Indicates the labeled pixel point, X={x i} i＝1,…,N represents all pixels in the image, x i Represents a single pixel in an image.

[0048] The Focal_Loss loss function solves this problem by giving different weights to samples of different difficulty levels. The expression of the Focal_Loss loss function is as follows:

[0049]

[0050] Where c∈{1,0} represents the annotation label, P(x i ) represents x i Predict the probability of belonging to the soft tissue area, γ and α are the hyperparameters that control positive and negative samples. If Y g Occupies a smaller area in the image, and α will become very large, where N represents the total number of samples, N g Represents the number of difficult-to-classify samples, which ensures that pixels potentially related to the tissue region of interest contribute more to the loss function. This paper combines the Dice_Loss loss function and the Focal_Loss loss function to construct the DF_loss loss function to guide the training of the VGG-SeUnet network and solve the problem of sample imbalance. The formula is as follows:

[0051]

[0052] (3) Morphological measurement method: The fetal subcutaneous soft tissue segmentation results obtained in steps (1) and (2) are converted into a binary image. In the binary image, the tissue area of ​​the fetal subcutaneous soft tissue segmentation result is represented by non-zero pixels and zero pixels. The minimum distance between each non-zero pixel and all zero pixels is stored in a set using a distance transformation algorithm. Then, the maximum distance is retrieved from the set and its corresponding pixel is regarded as the center of the maximum inscribed circle. The maximum inscribed circle is drawn with the maximum distance as the radius, where the diameter of the maximum inscribed circle is the fetal subcutaneous soft tissue thickness (FASSTT).

[0053] Step (3) is also one of the cores of the present invention, and specifically includes the following sub-steps.

[0054] (3.1) Binary image conversion.

[0055] In order to avoid background interference, the fetal subcutaneous soft tissue segmentation results are converted into binary images for further processing. In the binary map, the ROI area (tissue area) is represented by 1 (i.e., non-zero pixels) and the background area is represented by 0 (i.e., zero pixels).

[0056] (3.2) Retrieve the largest inscribed circle.

[0057] A distance transform algorithm is used to store the minimum distance between each non-zero pixel and all zero pixels, and then retrieve the maximum value from the record. The pixel with the maximum distance is regarded as the center of the largest inscribed circle, and a circle with a radius equal to the maximum distance is drawn. The diameter represents FASSTT. The distance transform algorithm is as follows Figure 4 shown.

[0058] Specifically, use set M to store each pixel in the binary image, then traverse the pixels in the image to find non-zero pixels, use set D to store the distance information calculated from each non-zero pixel to the zero pixel, find the distance from each non-zero point to the nearest zero pixel from set D, and use the pixel with the largest distance as the center [x, y] of the circle, and its distance as the radius. Twice the radius is FASSTT.

[0059] (4) Unit conversion: Convert the pixel unit of the fetal subcutaneous soft tissue thickness obtained in step (3) to centimeters to obtain the fetal subcutaneous soft tissue thickness that meets clinical needs. The conversion formula is as follows:

[0060] L c =L p ×2.54 / DPI

[0061] Among them L p Indicates pixel length, L c Indicates centimeter length, and 2.54 means one inch equals 2.54 centimeters. Since the image resolution is 96DPI, for ease of calculation, 38 pixels can be used to approximate one centimeter in the conversion to obtain the actual subcutaneous soft tissue thickness.

[0062] Experimental verification

[0063] The implementation example of the method of the present invention was implemented on a machine equipped with an AMD Ryzen 7 5800H CPU, an NVidia RTX3070 GPU and 32GB of memory. Using all the parameter values ​​listed in the specific implementation method, we obtained Figure 5 The experimental results in

[15] verified that the algorithm proposed in this paper can accurately segment the fetal subcutaneous soft tissue area and achieve accurate measurement of FASSTT.

[0064] like Figure 5As shown, the dotted circle represents the location of the maximum inscribed circle in the annotated image, and the solid circle represents the predicted location of the maximum inscribed circle in the current image. Compared with existing methods, the image segmented by our method (VGG-SeUnet-DF_loss) and the location of the drawn maximum inscribed circle are both closest to the annotated image. Compared with the results provided by radiologists, the average measurement error of our method is only 0.0615 cm, and the root mean square error is 0.081 cm. The overall performance of our method exceeds that of existing methods, demonstrating its effectiveness.

Claims

1. A method for automatically measuring the thickness of subcutaneous soft tissue in the fetal abdomen in ultrasound images, characterized in that: The following steps are involved: (1) For the abdominal subcutaneous soft tissue area that needs to be measured in the fetal abdominal circumference ultrasound image, the VGG-SeUnet network is used to extract its features, and each pixel in the image is classified by the pixel classifier to obtain the fetal subcutaneous soft tissue segmentation result; (2) The Dice_Loss loss function and the Focal_Loss loss function are combined to construct the DF_loss loss function to guide the training of the VGG-SeUnet network; the DF_loss loss function is used to evaluate the ratio between the intersection of the predicted tissue area and the real tissue area and the union of the two areas, and different weights are assigned to the easy-to-classify pixels and the difficult-to-classify pixels in the image to obtain the fetal subcutaneous soft tissue segmentation results; (3) converting the fetal subcutaneous soft tissue segmentation results obtained in steps (1) and (2) into a binary image, wherein the tissue area of ​​the fetal subcutaneous soft tissue segmentation result is represented by 1, i.e., a non-zero pixel point, and the background area is represented by 0, i.e., a zero pixel point; then, using a distance transformation algorithm, a set is used to store the minimum distance between each non-zero pixel point and all zero pixel points, and the maximum distance is retrieved from the set; the pixel point corresponding to the retrieved maximum distance is regarded as the center of the maximum inscribed circle, and the maximum inscribed circle is drawn with the maximum distance as the radius, wherein the diameter of the maximum inscribed circle is the thickness of the fetal subcutaneous soft tissue; (4) Convert the pixel unit of the fetal subcutaneous soft tissue thickness into centimeter unit to obtain the fetal subcutaneous soft tissue thickness that meets clinical needs.

2. The method for automatically measuring the thickness of the fetal abdominal subcutaneous soft tissue in an ultrasound image according to claim 1, characterized in that: The step (1) includes the following sub-steps: (1.1) collecting fetal abdominal circumference ultrasound imaging data, i.e., fetal abdominal circumference ultrasound images, and using the collected fetal abdominal circumference ultrasound imaging data as an image dataset; (1.2) normalizing the image dataset and converting the resolution of images in the image dataset to 512×512; Then randomly select 20% of the images in the image dataset as the test set and the remaining 80% as the training set; (1.3) Inputting the training set into the VGG-SeUnet network for training, and extracting multi-scale feature fusion results reflecting tissue segmentation through the encoding-decoding U-shaped architecture of the jump connection in the VGG-SeUnet network; (1.4) Introducing channel attention, the feature map obtained during each jump connection of the VGG-SeUnet network is reweighted by the channel attention; (1.5) Apply 1×1 convolution to the weighted feature map through the pixel classifier to reduce its channel dimension from 64 to 32, and use the LeakyReLU function to activate the feature, and then classify each pixel in the image to obtain the segmentation result of the fetal subcutaneous soft tissue.

3. The method for automatically measuring the thickness of the fetal abdominal subcutaneous soft tissue in ultrasound images according to claim 2, characterized in that: The channel attention in the step (1.4) is implemented by the following operations: the channel attention compresses each feature map by global average pooling to obtain a 1×1×C vector, and then normalizes the vector through two fully connected layers activated by Sigmund function to obtain the weight of each channel; finally, the channel is recalibrated by multiplying the weight with the original input channel.

4. The method for automatically measuring the thickness of the fetal abdominal subcutaneous soft tissue in an ultrasound image according to claim 1, wherein: In step (2), the expression of the DF_loss loss function is: Where DF_Loss represents the DF_loss loss function, X p represents the predicted tissue area, Y g represents the actual soft tissue area marked, N represents the total number of samples, and x i represents a single pixel in the image, c represents the annotation label, P(x i ) represents pixel x i Predict the probability of belonging to the soft tissue area, γ and α are hyperparameters that control positive and negative samples.

Images