A deep learning-based automatic uterine fibroid typing method and device

Abstract

The application provides a kind of uterus myoma automatic typing method and device based on deep learning.The method comprises the following steps: constructing semantic segmentation model, input pelvic medical image is divided into mucosa, muscle layer, serosa outside 3 regions;Construct instance segmentation model, segment out uterine fibroids, get uterine fibroids volume V T ;Extract the outline of uterine fibroids, calculate the volume V O Of the uterine fibroids belonging to serosa outside area and the volume V N Belonging to mucosa inside area;Based on V O , V N And V T The uterine fibroids are typed based on FIGO.The whole process of typing uterine fibroids of the application does not need artificial participation, reduces the analysis pressure of imaging department doctors, helps to improve the accuracy and consistency of FIGO typing determination.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical imaging, and particularly relates to an automatic uterine fibroid typing method and device based on deep learning. BACKGROUND

[0002] Uterine fibroids are common benign tumors and one of the most common tumors in the human body, also known as fibromyoma and uterine fibroid. Since uterine fibroids are mainly formed by the hyperplasia of uterine smooth muscle cells, a small amount of fibrous connective tissue exists as a supporting tissue, so it is also called uterine smooth muscle tumor. Based on the relationship between uterine fibroids and uterine serosa and mucosa, the uterine fibroids can be typed according to FIGO (The International Federation of Gynecology and Obstetrics, International Federation of Gynecology and Obstetrics). The influence of uterine fibroids of different FIGO types on patients and the selection of treatment methods are different. The uterine fibroid FIGO typing method is as follows: type 0: submucosal myoma completely located in the uterine cavity; type 1: myoma mostly located in the uterine cavity, and the portion of myoma located in the myometrium is less than or equal to 50%; type 2: myoma protruding to the submucosa, and the portion of myoma located in the myometrium is greater than 50%; type 2-5: mixed type; type 3: myoma completely located in the myometrium, but close to the mucosa; type 4: myoma completely located in the myometrium, neither close to the serosa nor protruding to the mucosa; type 5: myoma protruding to the serosa, but the portion of myoma located in the myometrium is greater than or equal to 50%; type 6: myoma protruding to the serosa, but the portion of myoma located in the myometrium is less than 50%; type 7: pedunculated submucosal myoma; type 8: non-myometrial type (special sites such as cervical and broad ligament myoma).

[0003] At present, the FIGO typing mainly depends on the subjective judgment of doctors, and a large amount of time is consumed for analysis, and there is no automatic FIGO typing algorithm and device. The diagnosis of doctors depends on the seniority and diagnosis experience of doctors, and the subjectivity is strong, and the consistency is poor. Therefore, the application provides an automatic FIGO typing method based on a deep learning segmentation model, determines the regions of serosa and mucosa based on semantic segmentation, then obtains the region of uterine fibroids based on instance segmentation on medical images, and automatically obtains the FIGO typing after extracting relevant features. SUMMARY

[0004] In order to solve the above problems in the prior art, the application provides an automatic uterine fibroid typing method and device based on deep learning.

[0005] In order to achieve the above purpose, the application adopts the following technical solutions.

[0006] In a first aspect, the application provides an automatic uterine fibroid typing method based on deep learning, comprising the following steps:

[0007] A semantic segmentation model is constructed, and a pelvic medical image is input into the trained model to divide the image into three regions: intramucosal, muscular layer, i.e., the region between mucosa and serosa, and extraserosal;

[0008] An instance segmentation model is constructed, and a pelvic medical image is input into the trained model to segment uterine fibroids and obtain uterine fibroid volume V T ;

[0009] The contour of the uterine fibroid is extracted, the region adjacent to each pixel point on the contour is determined, and the volume V O of the uterine fibroid belonging to the extraserosal region and the volume V N belonging to the intramucosal region are calculated.

[0010] The FIGO classification of the uterine fibroid is performed based on V O , V N and V T .

[0011] Further, the output of the semantic segmentation model is two types of serosa and mucosa, and the activation function of the last layer is sigmoid; during model prediction, the region with a sigmoid output greater than 0.5 is the predicted type; if a pixel is predicted to be both serosa and mucosa, the final type of the pixel is mucosa.

[0012] Further, the method for determining the region adjacent to each pixel point on the contour includes:

[0013] A neighborhood centered on any pixel point A on the contour is selected.

[0014] If all the pixel points in the neighborhood except the uterine fibroid belong to one region of the three regions, the one region is the region adjacent to the pixel point A.

[0015] If the pixel points in the neighborhood except the uterine fibroid belong to multiple regions of the three regions, any one of the multiple regions is selected as the region adjacent to the pixel point A.

[0016] Further, the calculation method of V O and V N includes:

[0017] The pixel points on each layer of the uterine fibroid contour adjacent to the extraserosal region and the pixel points adjacent to the intramucosal region are connected, respectively, and the contours of the extraserosal region and the intramucosal region are combined to obtain two parts of each layer of the uterine fibroid belonging to the extraserosal region and the intramucosal region, respectively.

[0018] The volume VOi and the volume V belonging to the intramucosal region Ni , i = 1, 2, …, n, n is the number of layers of the pelvic medical image;

[0019] Calculate:

[0020] Further, the FIGO typing of the uterine fibroids based on V O , V N and V T includes:

[0021] S0, calculate: ratio_N = V N / V T , ratio_O = V O / V T ,

[0022] S1, if V N = V O = 0, the uterine fibroids are type 4; otherwise, go to step S2;

[0023] S2, if V T = V N , the uterine fibroids are type 0; otherwise, go to step S3;

[0024] S3, if V T = V O , the uterine fibroids are type 8; otherwise, go to step S4;

[0025] S4, if V N > 0 and V O > 0, the uterine fibroids are type 2-5; otherwise, go to step S5;

[0026] S5, if V O > 0, go to step S61; otherwise, go to step S71;

[0027] S61, if ratio_O > 0.98, the uterine fibroids are type 7; otherwise, go to step S62;

[0028] S62, if ratio_max_O > 0.5, the uterine fibroids are type 6; otherwise, go to step S63;

[0029] S63, if ratio_max_O > 0.1, the uterine fibroids are type 5; otherwise, the uterine fibroids are type 4;

[0030] S71, if ratio_max_N < 0.1, the uterine fibroids are type 3; otherwise, go to step S72;

[0031] S72. If ratio_max_N < 0.5, then the uterine fibroid is type 2; otherwise, the uterine fibroid is type 1.

[0032] Secondly, the present invention provides an automatic uterine fibroid classification device based on deep learning, comprising:

[0033] The first modeling module is used to construct a semantic segmentation model. The pelvic medical images are input into the trained model to divide the images into three regions: the intramucosa; the muscular layer, i.e., the region between the mucosa and the serosa; and the extramucosa.

[0034] The second modeling module is used to construct an instance segmentation model. Pelvic medical images are input into the trained model to segment uterine fibroids and obtain the uterine fibroid volume V. T ;

[0035] The volume calculation module is used to extract the outline of the uterine fibroid, determine the region adjacent to each pixel on the outline, and calculate the volume V of the uterine fibroid belonging to the extraserosa. O and the volume V belonging to the intramural region N ;

[0036] The fibroid classification module is used for V-based... O V N and V T The uterine fibroids were classified using FIGO classification.

[0037] Furthermore, the semantic segmentation model outputs two types: serous membrane and mucous membrane, and the activation function of the last layer is sigmoid. During model prediction, regions with a sigmoid output greater than 0.5 are the predicted types. If a pixel is predicted as both serous membrane and mucous membrane, then the final type of the pixel is mucous membrane.

[0038] Furthermore, methods for determining the regions adjacent to each pixel on the contour include:

[0039] Select a neighborhood centered on any pixel A on the contour;

[0040] If all pixels in the neighborhood other than the uterine fibroid belong to one of the three regions, then the one region is the region adjacent to pixel A.

[0041] If a pixel other than a uterine fibroid in the neighborhood belongs to multiple regions among the three regions, then any one of the multiple regions is selected as the region adjacent to pixel A.

[0042] Furthermore, V O V N The calculation methods include:

[0043] Connect the pixel points of the adjacent area of the serosa outer area and the adjacent area of the mucosa inner area on the contour of each layer of uterine fibroids respectively, and combine the contours of the serosa outer area and the mucosa inner area to obtain two parts of each layer of uterine fibroids belonging to the serosa outer area and the mucosa inner area respectively;

[0044] Calculate the volume V Oi of the serosa outer area and the volume V Ni of the mucosa inner area of each layer of uterine fibroids based on the two parts;

[0045] Calculate:

[0046] Further, the FIGO typing of the uterine fibroids based on V O , V N and V T includes:

[0047] S0, calculate: ratio_N=V N / V T , ratio_O=V O / V T ,

[0048] S1, if V N =V O =0, the uterine fibroids are type 4; otherwise, go to step S2;

[0049] S2, if V T =V N , the uterine fibroids are type 0; otherwise, go to step S3;

[0050] S3, if V T =V O , the uterine fibroids are type 8; otherwise, go to step S4;

[0051] S4, if V N >0 and V O >0, the uterine fibroids are type 2-5; otherwise, go to step S5;

[0052] S5, if V O >0, go to step S61; otherwise, go to step S71;

[0053] S61, if ratio_O>0.98, the uterine fibroids are type 7; otherwise, go to step S62;

[0054] S62, if ratio_max_O>0.5, the uterine fibroids are type 6; otherwise, go to step S63;

[0055] S63, if ratio_max_O>0.1, the uterine fibroids are type 5; otherwise, the uterine fibroids are type 4;

[0056] S71, if ratio_max_N<0.1, the uterine fibroids are type 3; otherwise, go to step S72;

[0057] S72, if ratio_max_N<0.5, the uterine fibroids are type 2; otherwise, the uterine fibroids are type 1.

[0058] Compared with the prior art, the present application has the following beneficial effects.

[0059] The present application divides the pelvic medical image into three regions by constructing a semantic segmentation model, constructs an instance segmentation model, segments the uterine fibroids, and obtains the uterine fibroid volume V T extracts the outline of the uterine fibroids, determines the region adjacent to each pixel point on the outline, and calculates the volume V O of the uterine fibroids belonging to the serosa outer region and the volume V N of the uterine fibroids belonging to the mucosa inner region, and performs FIGO typing on the uterine fibroids based on V O , V N and V T , thereby realizing automatic typing of uterine fibroids. The whole process of typing the uterine fibroids according to the present application does not require human intervention, greatly reducing the analysis pressure of the imaging department doctors, and helping to improve the accuracy and consistency of FIGO typing determination. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 is a flowchart of a uterine fibroid automatic typing method based on deep learning according to an embodiment of the present application.

[0061] Figure 2 is a uterine fibroid, mucosa and serosa annotation schematic diagram.

[0062] Figure 3 is a uterine fibroid outline point region determination method schematic diagram.

[0063] Figure 4 is a uterine fibroid three-part volume distribution schematic diagram.

[0064] Figure 5 is a flowchart of typing the uterine fibroids based on V O , V N and V T .

[0065] Figure 6A block diagram of an embodiment of an automatic uterine fibroid typing device based on deep learning. DETAILED DESCRIPTION

[0066] To make the objectives, technical solutions and advantages of the present application clearer and more apparent, the present application will be further described below with reference to the drawings and specific embodiments. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0067] Figure 1 A flowchart of an embodiment of an automatic uterine fibroid typing method based on deep learning, comprising the following steps:

[0068] Step 101, constructing a semantic segmentation model, inputting a pelvic medical image into the trained model to divide the image into three regions: endometrium, myometrium (the region between endometrium and serosa), and serosa;

[0069] Step 102, constructing an instance segmentation model, inputting a pelvic medical image into the trained model to segment uterine fibroids and obtain the volume V of uterine fibroids T ;

[0070] Step 103, extracting the contour of uterine fibroids, determining the region adjacent to each pixel point on the contour, and calculating the volume V O of the uterine fibroids belonging to the serosa region and the volume V N of the uterine fibroids belonging to the endometrium region.

[0071] Step 104, typing the uterine fibroids based on V O , V N , and V T .

[0072] In this embodiment, step 101 is mainly used to divide the input pelvic medical image into three regions. The three regions are endometrium, myometrium (endometrium is inside serosa, and myometrium is the region between endometrium and serosa), and serosa, as shown in Figure 4As shown. In this embodiment, a semantic segmentation model is constructed. Pelvic medical images are input into the trained model, dividing the image into three regions. The semantic segmentation model can classify each pixel of the input image. The nnU-Net framework can be used to train the semantic segmentation model, resulting in a 3D U-Net segmentation model. Model training requires the construction of training, validation, and test sets by collecting medical images containing pelvic areas. Commonly used image sequences for observing uterine fibroids and serosa / mucosa can be selected, such as sagittal T2WI and axial T2WI pelvic medical images. The entire data collection process must adhere to confidentiality principles, anonymizing the collected image data and removing all patient personal information. The collected images also need to be labeled. An experienced annotator, under the guidance of a radiologist with many years of experience in diagnosing female pelvic imaging, can delineate the uterine outline and endometrium on each patient's medical images, delineate all uterine fibroids, and label their FIGO classification according to guidelines. Figure 2 As shown, there are two uterine fibroids in the image, and each fibroid is labeled separately. To avoid adverse effects on model performance due to differences in image resolution and size, spatial resolution normalization is performed on the collected medical image sequences. First, the resolution and image size information of all medical image sequences are statistically analyzed, and the median is selected as the target resolution and target size. An interpolation algorithm is used to resample all images to the target resolution and crop them to the target size. At the same time, the doctor's drawing results are processed in the same way, i.e., resampled to the target resolution and cropped to the target size. Finally, all data are divided into training set, validation set, and test set according to the patient. The ratio of the three datasets can be selected according to the amount of data, such as 6:2:2. The pelvic medical images in this embodiment include, but are not limited to, MR images.

[0073] In this embodiment, step 102 is mainly used to segment uterine fibroids. This embodiment constructs an instance segmentation model, inputs pelvic medical images into the trained model, and segments the uterine fibroids. Image segmentation can be divided into instance segmentation and semantic segmentation. As mentioned earlier, semantic segmentation classifies each pixel of the input image, while instance segmentation is a combination of object detection and semantic segmentation. It detects the target in the image (object detection) and then labels each pixel (semantic segmentation). For instance segmentation, it is necessary not only to determine whether a pixel belongs to the target of interest but also to distinguish different objects of the same category. Any instance segmentation model capable of performing uterine fibroid segmentation can be used. Mask R-CNN can be used as the instance segmentation model. Since pelvic MR scans are generally thick-layer data, a pseudo-3D approach can be used for segmentation. Specifically, the target layer image and one image before and after it (a total of three images) are taken as input, and a 2D Mask R-CNN model is used to predict the segmentation result of the target image. For example, for the prediction of the k-th layer, the input image consists of three images from layers (k-1), k, and k+1). These three images are combined into three channels of the input tensor, and the output is the segmentation result of the k-th layer. This approach is similar to the RGB three-channel method for natural images, thus fully utilizing the pre-trained parameters of natural image datasets, such as ImageNet, and avoiding overfitting. It's worth noting that the above use of Mask R-CNN is just one example; other existing instance segmentation models can also be used. Furthermore, the segmentation method is not limited to pseudo-3D; 3D instance segmentation algorithms can also be used directly. After uterine fibroid segmentation, the volume of the uterine fibroid can be obtained.

[0074] In this embodiment, step 103 is mainly used to calculate the volume V of the uterine fibroid belonging to the extraserosa. O The volume V belonging to the intramural region N Calculate V O V N This is for step 104 based on V O V N and V T Uterine fibroids are classified. This embodiment first uses an image contour retrieval algorithm to extract the contour of the uterine fibroid. Since the lesion does not contain cavities, only the outermost contour is retrieved, obtaining all contour points of the uterine fibroid in the current layer. Then, based on the three regions divided from the input pelvic medical image in step 101, it is determined which region each pixel on the uterine fibroid contour belongs to. Based on this, the portions of the uterine fibroid belonging to the extraserous and intramucosal regions are determined, and the volume V is calculated for each. O and V N .like Figure 4 As shown in the figure, V Oi V NiThe volume of the i-th layer of the uterine fibroid belonging to the extra-serosal region and the volume belonging to the intra-mucosal region can be calculated by calculating V Oi Ni Summing i to obtain V O N It should be noted that V Oi Ni Here, V O N T are all volumes, not areas, because each layer image has a certain thickness, referred to as layer thickness (the layer thickness is a parameter of the medical image, representing the actual thickness of the human body corresponding to the layer image, which can be read from the stored Dicom medium). The volume can be obtained by multiplying the area and the layer thickness.

[0075] In this embodiment, step 104 is mainly used to classify the uterine fibroid based on V O N T The FIGO classification method of uterine fibroid is given in the foregoing (background section), which divides the uterine fibroid into 0-8 types, and there is a mixed type, i.e., 2-5 type, and the characteristics of each type are given. According to the characteristics, the classification of the uterine fibroid can be identified according to the values (or values after simple operation) of V O N T The following embodiment will give a specific classification method.

[0076] As an optional embodiment, the output of the semantic segmentation model is two types of serosa and mucosa, and the activation function of the last layer is sigmoid; when the model is predicted, the region with a sigmoid output greater than 0.5 is the predicted type; if a pixel is predicted to be both serosa and mucosa, the final type of the pixel is mucosa.

[0077] ​​​This embodiment presents a specific technical solution for a semantic segmentation model. This semantic segmentation model is an improvement over existing segmentation models. Common multi-class segmentation models output the number of categories as the number of predicted targets + 1 (background). For example, the original model in this embodiment has a 3-channel output layer, representing the serous membrane, mucous membrane, and background. It uses a softmax activation function to obtain the prediction probabilities of the three channels, and takes the highest prediction probability as the pixel's prediction result; that is, multiple prediction results are mutually exclusive. In this embodiment, since the serous membrane and mucous membrane regions have an inclusion relationship (the mucous membrane is located within the serous membrane), some pixels can be both within the serous membrane and the mucous membrane. Therefore, this embodiment changes the output layer to a 2-channel convolutional layer to predict the serous membrane and mucous membrane separately, and changes the last nonlinear transformation from softmax to sigmoid. During model prediction, regions with a sigmoid transformation value greater than 0.5 are taken as the prediction target. Because the serous membrane and mucous membrane have an inclusion relationship, the mucous membrane has a higher priority than the serous membrane. In other words, if a pixel is predicted as both the serous membrane and the mucous membrane, then the final result for that pixel is the mucous membrane.

[0078] As an optional embodiment, the method for determining the region adjacent to each pixel on the contour includes:

[0079] Select a neighborhood centered on any pixel A on the contour;

[0080] If all pixels in the neighborhood other than the uterine fibroid belong to one of the three regions, then the one region is the region adjacent to pixel A.

[0081] If a pixel other than a uterine fibroid in the neighborhood belongs to multiple regions among the three regions, then any one of the multiple regions is selected as the region adjacent to pixel A.

[0082] This embodiment provides a technical solution for determining the region adjacent to any pixel on the outline of a uterine fibroid. The above technical solution uses pixel A as an example to describe in detail the method for determining the region adjacent to any pixel. For ease of understanding, Figure 3 The method of determining the 4-neighborhood is shown. It involves taking any pixel on the contour as the center, creating a 3*3 grid, and examining the regions to which the pixels other than uterine fibroids belong among the four pixels adjacent to the center point (upper, lower, left, and right). Figure 3 In the middle, the two pixels on the left and bottom belong to uterine fibroids, while the two pixels on the top and right both belong to the extraserosa region. Therefore, the adjacent region of the center point is the extraserosa region, marked as O. The regions adjacent to other pixels can be obtained using the same method, such as... Figure 3 As shown, regions marked J represent adjacent muscle layers, and regions marked N represent adjacent intramucosa regions.

[0083] As an optional embodiment, V O V N The calculation methods include:

[0084] Connect the pixels on the outline of each uterine fibroid layer whose adjacent regions are in the extraserous region and the pixels whose adjacent regions are in the intramucosal region. Combine the outlines of the extraserous region and the intramucosal region to obtain the two parts of each uterine fibroid layer that belong to the extraserous region and the intramucosal region respectively.

[0085] Based on the above two parts, the volume V of each layer of uterine fibroid belonging to the extraserosa region is calculated. Oi and the volume V belonging to the mucosal region Ni , i = 1, 2, ..., n, where n is the number of slices in the pelvic medical imaging;

[0086] calculate:

[0087] This embodiment provides a method for calculating V. O V N One technical solution. This embodiment calculates V layer by layer. Oi V Ni Then summate the results for each layer to obtain V. O V N This embodiment connects pixels on the outline of each uterine fibroid layer whose adjacent regions are in the extraserous region and whose adjacent regions are in the intramucosal region. For example, connecting... Figure 3 By identifying pixels marked as O and N, and based on the contours of the extraserous and intramucosal regions, the regions of the uterine fibroid protruding outward from the serosa and inward from the mucosa can be determined, thus yielding V. Oi V Ni ,like Figure 4 As shown.

[0088] As an optional embodiment, the V-based O V N and V T FIGO typing of the uterine fibroids includes:

[0089] S0, Calculate: ratio_N = V N / V T ratio_O = V O / V T ,

[0090] S1, if V N =V O =0, then the uterine fibroid is type 4; otherwise, proceed to step S2;

[0091] S2, if V T = V N , then the uterine fibroids are type 0; otherwise, go to step S3;

[0092] S3, if V T = V O , then the uterine fibroids are type 8; otherwise, go to step S4;

[0093] S4, if V N > 0, and V O > 0, then the uterine fibroids are type 2-5; otherwise, go to step S5;

[0094] S5, if V O > 0, go to step S61; otherwise, go to step S71;

[0095] S61, if ratio_O > 0.98, then the uterine fibroids are type 7; otherwise, go to step S62;

[0096] S62, if ratio_max_O > 0.5, then the uterine fibroids are type 6; otherwise, go to step S63;

[0097] S63, if ratio_max_O > 0.1, then the uterine fibroids are type 5; otherwise, the uterine fibroids are type 4;

[0098] S71, if ratio_max_N < 0.1, then the uterine fibroids are type 3; otherwise, go to step S72;

[0099] S72, if ratio_max_N < 0.5, then the uterine fibroids are type 2; otherwise, the uterine fibroids are type 1.

[0100] The embodiment provides a technical solution for FIGO typing of uterine fibroids. The embodiment is based on V O , V N and V T to perform FIGO typing of uterine fibroids. First, four ratios, ratio_N, ratio_O, ratio_max_O and ratio_max_N, are calculated according to V O , V N and V T , and then the FIGO typing of the uterine fibroids is determined based on the relationship and value range of the parameters. For specific typing methods, refer to the flowchart shown in Figure 5 , which will not be described in detail here.

[0101] Figure 6 FIG. 1 is a schematic diagram of a device for automatically typing uterine fibroids based on deep learning according to an embodiment of the present application. The device comprises:

[0102] The first modeling module 11 is configured to construct a semantic segmentation model, and input the pelvic cavity medical image into the trained model to divide the image into three regions: intramucosal, muscular layer, and extraperitoneal.

[0103] The second modeling module 12 is configured to construct an instance segmentation model, and input the pelvic cavity medical image into the trained model to segment the uterine fibroids and obtain the uterine fibroid volume V T ;

[0104] The volume calculation module 13 is configured to extract the contour of the uterine fibroids, determine the region adjacent to each pixel point on the contour, and calculate the volume V O of the uterine fibroids belonging to the extraperitoneal region and the volume V N of the uterine fibroids belonging to the intramucosal region.

[0105] The fibroid classification module 14 is configured to perform FIGO classification on the uterine fibroids based on V O , V N , and V T .

[0106] The device of the embodiment can be used to implement the technical solutions of the method embodiment shown in Figure 1 , and the implementation principles and technical effects are similar, which will not be described here. The same is true for the following embodiments, which will not be described here.

[0107] As an optional embodiment, the output of the semantic segmentation model is two types of serosa and mucosa, and the activation function of the last layer is sigmoid; during model prediction, the region with a sigmoid output greater than 0.5 is the predicted type; if a pixel is predicted to be both serosa and mucosa, the final type of the pixel is mucosa.

[0108] As an optional embodiment, the method for determining the region adjacent to each pixel point on the contour includes:

[0109] selecting a neighborhood with any pixel point A on the contour as the center;

[0110] if the pixel points other than the uterine fibroids in the neighborhood belong to one region of the three regions, the region is the region adjacent to the pixel point A;

[0111] if the pixel points other than the uterine fibroids in the neighborhood belong to multiple regions of the three regions, any one of the multiple regions is selected as the region adjacent to the pixel point A.

[0112] As an optional embodiment, the calculation method of V O , V N includes:

[0113] Connect the pixel points of the adjacent region of the serosa outer region and the adjacent region of the mucosa inner region on the contour of each layer of uterine fibroids, and combine the contours of the serosa outer region and the mucosa inner region to obtain two parts of each layer of uterine fibroids belonging to the serosa outer region and the mucosa inner region respectively;

[0114] Calculate the volume V Oi of the serosa outer region and the volume V Ni of the mucosa inner region of each layer of uterine fibroids based on the two parts;

[0115] Calculate:

[0116] As an optional embodiment, the FIGO typing of the uterine fibroids based on V O , V N and V T includes:

[0117] S0, calculate: ratio_N=V N / V T , ratio_O=V O / V T ,

[0118] S1, if V N =V O =0, the uterine fibroids are type 4; otherwise, go to step S2;

[0119] S2, if V T =V N , the uterine fibroids are type 0; otherwise, go to step S3;

[0120] S3, if V T =V O , the uterine fibroids are type 8; otherwise, go to step S4;

[0121] S4, if V N >0 and V O >0, the uterine fibroids are type 2-5; otherwise, go to step S5;

[0122] S5, if V O >0, go to step S61; otherwise, go to step S71;

[0123] S61, if ratio_O>0.98, the uterine fibroids are type 7; otherwise, go to step S62;

[0124] S62, if ratio_max_O>0.5, the uterine fibroids are type 6; otherwise, go to step S63;

[0125] S63, if ratio_max_0>0.1, then the uterine fibroids are type 5; otherwise, the uterine fibroids are type 4;

[0126] S71, if ratio_max_N<0.1, then the uterine fibroids are type 3; otherwise, go to step S72;

[0127] S72, if ratio_max_N<0.5, then the uterine fibroids are type 2; otherwise, the uterine fibroids are type 1.

[0128] The above merely provides a specific implementation of the present application, but the protection scope of the present application is not limited thereto, any changes or replacements that can be easily thought of by those skilled in the art within the technical range disclosed by the present application shall be encompassed within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A deep learning-based automatic classification method for uterine fibroids, characterized by, The method comprises the following steps: a semantic segmentation model is constructed, and a pelvic medical image is input into the trained model to divide the image into three regions: intramucosal, myometrial, and extraserosal; The output of the semantic segmentation model is two types of serosa and mucosa, and the activation function of the last layer is sigmoid; during model prediction, the region with a sigmoid output greater than 0.5 is the predicted type; if a pixel is predicted to be both serosa and mucosa, the final type of the pixel is mucosa. Constructing an instance segmentation model, inputting the pelvic medical image into the trained model, segmenting out the uterine fibroids, and obtaining the uterine fibroid volume V T ; extracting the contour of the uterine fibroid, determining the region adjacent to each pixel point on the contour, respectively determining the part of the uterine fibroid belonging to the serosa outer region and the mucosa inner region, and calculating the volume V of the uterine fibroid belonging to the serosa outer region O and the volume V belonging to the mucosa inner region N ; Based on V O , V N and V T FIGO typing of the uterine fibroids is performed by means of a volumetric proportion parameter in combination with a multi-order threshold rule. 2.The deep learning-based automatic uterine fibroid typing method of claim 1, wherein, The method for determining the adjacent region of each pixel point on the contour comprises the following steps: 3.The deep learning-based automatic uterine fibroid typing method of claim 1, wherein, selecting a neighborhood with any pixel point A on the contour as the center; if all the pixel points in the neighborhood except the myoma are in one of the three regions, the one region is the adjacent region of the pixel point A; if the pixel points in the neighborhood except the myoma are in multiple regions of the three regions, any one of the multiple regions is selected as the adjacent region of the pixel point A. connect the pixel points with the adjacent region being the extraserosal region and the pixel points with the adjacent region being the intramucosal region on the contour of each layer of myoma, and combine the contours of the extraserosal region and the intramucosal region to obtain two parts of each layer of myoma belonging to the extraserosal region and the intramucosal region respectively; 4.The deep learning-based automatic uterine fibroid typing method of claim 3, wherein, V O , V N The calculation method comprises: S61, if ratio_O>0.98, the myoma is type 7; otherwise, step S62 is performed; calculating a volume V of each layer of the uterine fibroids belonging to the extraperitoneal region Oi and a volume V belonging to the intramucosal region Ni i = 1, 2, …, n, n is the number of layers of the pelvic medical image; Calculations: 5.The deep learning-based automatic uterine fibroid typing method of claim 4, wherein, The V O , V N , and V T FIGO typing of the uterine fibroids is performed by a volume ratio parameter in combination with a multi-order threshold rule, comprising: S0, compute: ratio_N = V N / V T , ratio_O = V O / V T , S1, if V N = V O = 0, then the uterine fibroids are type 4; otherwise, go to step S2; S2, if V T = V N then the uterine fibroids are type 0; otherwise, go to step S3; S3, if V T = V O then the uterine fibroid is type 8; otherwise, go to step S4; S4, if V N > 0, and V O > 0, the uterine fibroids are type 2-5; otherwise, go to step S5; S5, if V O > 0, then go to step S61; otherwise, go to step S71; S62, if ratio_max_O>0.5, the myoma is type 6; otherwise, step S63 is performed; S63, if ratio_max_O>0.1, the myoma is type 5; otherwise, the myoma is type 4; S71, if ratio_max_N<0.1, the myoma is type 3; otherwise, step S72 is performed; S72, if ratio_max_N<0.5, the myoma is type 2; otherwise, the myoma is type 1. The device comprises:

6. A deep learning-based automatic uterine fibroid typing device, characterized by, a first modeling module configured to construct a semantic segmentation model, and input a pelvic medical image into the trained model to divide the image into three regions: intramucosal, myometrial, and extraserosal; The output of the semantic segmentation model is two types of serosa and mucosa, and the activation function of the last layer is sigmoid; during model prediction, the region with a sigmoid output greater than 0.5 is the predicted type; if a pixel is predicted to be both serosa and mucosa, the final type of the pixel is mucosa. A second modeling module is configured to construct an instance segmentation model, input the pelvic medical image into the trained model, segment out the uterine fibroids, and obtain a uterine fibroid volume V T ; a volume calculation module configured to extract a contour of the uterine fibroid, determine a region adjacent to each pixel point on the contour, respectively determine a part of the uterine fibroid belonging to an extra-serosa region and a part of the uterine fibroid belonging to an intra-mucosa region, and calculate a volume V of the uterine fibroid belonging to the extra-serosa region O and a volume V of the uterine fibroid belonging to the intra-mucosa region N ​ a myoma typing module for typing the uterine fibroids based on V O , V N and V T FIGO typing the uterine fibroids by volume proportionality parameters combined with multi-order threshold rules.

7. The deep learning based automatic uterine fibroid typing device according to claim 6, wherein, The method for determining the adjacent region of each pixel point on the contour comprises the following steps: 8.The deep learning based automatic uterine fibroid typing device of claim 6, wherein, selecting a neighborhood with any pixel point A on the contour as the center; if all the pixel points in the neighborhood except the myoma are in one of the three regions, the one region is the adjacent region of the pixel point A; if the pixel points in the neighborhood except the myoma are in multiple regions of the three regions, any one of the multiple regions is selected as the adjacent region of the pixel point A. ​ 9.The deep learning based automatic uterine fibroid typing device of claim 8, wherein, V O , V N The calculation method of the present application comprises: Connect the pixel points of the serosa outer region and the pixel points of the mucosa inner region which are adjacent to each other on the contour of each layer of the uterine fibroids, and combine the contours of the serosa outer region and the mucosa inner region to obtain two parts of each layer of the uterine fibroids belonging to the serosa outer region and the mucosa inner region respectively; calculating a volume V of each layer of the uterine fibroid belonging to the extraperitoneal region based on the two parts Oi and a volume V belonging to the intramucosal region Ni , i = 1, 2, …, n, n being the number of layers of the pelvic medical image; Calculations: 10.The deep learning based automatic uterine fibroid typing device of claim 9, wherein, The V O , V N , and V T FIGO typing of the uterine fibroids is performed by a volume ratio parameter in combination with a multi-order threshold rule, comprising: S0, compute: ratio_N = V N / V T , ratio_O = V O / V T , S1, if V N = V O = 0, then the uterine fibroids are type 4; otherwise, go to step S2; S2, if V T = V N then the uterine fibroids are type 0; otherwise, go to step S3; S3, if V T = V O then the uterine fibroid is type 8; otherwise, go to step S4; S4, if V N > 0, and V O > 0, then the uterine fibroids are type 2-5; otherwise, go to step S5; S5, if V O > 0, then go to step S61; otherwise, go to step S71; S61, if ratio_O>0.98, the uterine fibroids are type 7; otherwise, go to step S62; S62, if ratio_max_O>0.5, the uterine fibroids are type 6; otherwise, go to step S63; S63, if ratio_max_O>0.1, the uterine fibroids are type 5; otherwise, the uterine fibroids are type 4; S71, if ratio_max_N<0.1, the uterine fibroids are type 3; otherwise, go to step S72; S72, if ratio_max_N<0.5, the uterine fibroids are type 2; otherwise, the uterine fibroids are type 1.

Images