Thyroid cancer image feature extraction method and device based on convolutional network, electronic equipment and storage medium

Abstract

The invention provides a feature extraction method and device of a thyroid cancer image based on a convolutional network, electronic equipment and a storage medium, and the method comprises the steps: obtaining a preset number of color Doppler ultrasound images of thyroid cancer, and carrying out the preprocessing of the color Doppler ultrasound images; extracting a thyroid cancer image from the preprocessed color Doppler ultrasound image, performing color space conversion on the thyroid cancer image through a preset color space, and segmenting a blood flow spectrum image in the thyroid cancer according to hue and saturation information in the color space; performing a mathematical morphology operation on the blood flow spectrum image to enhance the blood flow spectrum image; and inputting the blood flow spectrum image after the mathematical morphology operation into the trained feature extraction model for convolution processing so as to obtain the proportion and distribution condition of the blood flow spectrum signal in the thyroid cancer. Therefore, effective features in the thyroid cancer image can be accurately identified, and thyroid cancer and thyroid benign nodules can be effectively distinguished.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and particularly to a method, device, electronic device and storage medium for extracting features of thyroid cancer images based on a convolutional network. Background Art

[0002] Thyroid cancer is a relatively common malignant tumor of the endocrine system. The gray-scale features of thyroid benign nodules and thyroid malignant tumors often overlap. Studies have found that the blood flow features in thyroid malignant tumors are different from those in thyroid benign nodules in terms of morphology and distribution. However, due to its low sensitivity, blood flow features are usually not used in traditional ultrasound for evaluating thyroid cancer. Therefore, how to improve the sensitivity of blood flow features in thyroid cancer so that it can be used as a basis for distinguishing thyroid benign nodules and thyroid malignant tumors is a research hotspot in the industry. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to provide a method, device, electronic device and storage medium for extracting features of thyroid cancer images based on a convolutional network, which solves the problem that it is difficult to evaluate thyroid cancer through blood flow features in the prior art.

[0004] In a first aspect, the present application provides a method for extracting features of thyroid cancer images based on a convolutional network, including: Obtaining a preset number of color Doppler ultrasound images of thyroid cancer and preprocessing the color Doppler ultrasound images; Extracting thyroid cancer images from the preprocessed color Doppler ultrasound images, performing color space conversion on the thyroid cancer images through a preset color space, and segmenting the blood flow spectrum images in the thyroid cancer according to the hue and saturation information in the color space; Performing mathematical morphological operations on the blood flow spectrum images to enhance the blood flow spectrum images; Inputting the blood flow spectrum images after performing the mathematical morphological operations into a trained feature extraction model for convolutional processing to obtain the proportion and distribution of blood flow spectrum signals in thyroid cancer.

[0005] Optionally, the extracting thyroid cancer images from the preprocessed color Doppler ultrasound images includes: Annotating the thyroid cancer images in the color Doppler ultrasound images to obtain a plurality of annotated regions, expanding the annotated regions according to gray-scale similarity, and covering the thyroid cancer regions in the color Doppler ultrasound images; Adjust the contour line of the thyroid cancer area through the principle of energy minimization to make the contour line fit the boundary of the thyroid cancer area, and extract the thyroid cancer image based on the contour line.

[0006] Optionally, the mathematical morphology operation includes opening operation and closing operation; The performing mathematical morphology operation on the blood flow spectrum image to enhance the blood flow spectrum image includes: Perform binarization processing on the blood flow spectrum image based on a preset binarization threshold; Perform target annotation on the binarized blood flow spectrum image based on a preset structural element; Perform opening operation and / or closing operation on the blood flow spectrum image after the target annotation to enhance the blood flow spectrum image.

[0007] Optionally, the mathematical morphology operation further includes erosion operation and dilation operation; the performing opening operation on the blood flow spectrum image after the target annotation includes: ; Wherein, is the target image annotated based on the preset structural element, is the blood flow spectrum image, is the dilation operation symbol, is the erosion operation symbol, is the structural disk element with a width of 2 pixels in the opening operation; Smooth the contour line of the target image annotated based on the preset structural element and remove the connection between multiple target images annotated based on the preset structural element by first performing the erosion operation and then the dilation operation on the target image annotated based on the preset structural element.

[0008] Optionally, the performing closing operation on the blood flow spectrum image after the target annotation includes: First perform the dilation operation and then the erosion operation on the target image annotated based on the preset structural element to fill the holes in the target image annotated based on the preset structural element and fill the broken parts of the contour line of the target image annotated based on the preset structural element.

[0009] Optionally, the inputting the blood flow spectrum image after performing the mathematical morphology operation into a trained feature extraction model for convolution processing to obtain the proportion and distribution of the blood flow spectrum signal in thyroid cancer includes: Performing convolution sampling on the blood flow spectral image after the execution of the mathematical morphology operation for a preset number of times through the trained feature extraction model to obtain a plurality of pixel intensity values in the blood flow spectral image and the texture features in the blood flow spectral image; Statistical analysis of the plurality of pixel intensity values and their discrete conditions; Using the principal component analysis method to perform dimensionality reduction processing on the texture features and the discrete conditions of the pixel intensity values to obtain the dimensionality-reduced blood flow spectral signal; Performing a visualization operation on the dimensionality-reduced blood flow spectral signal to obtain the proportion and distribution of the blood flow spectral signal in thyroid cancer.

[0010] Optionally, after performing a visualization operation on the dimensionality-reduced blood flow spectral signal to obtain the proportion and distribution of the blood flow spectral signal in thyroid cancer, it further includes: Outputting the result of the visualization operation and generating a preliminary diagnosis result of thyroid cancer according to the proportion and distribution of the blood flow spectral signal in the thyroid cancer.

[0011] In a second aspect, the present application provides a feature extraction device for thyroid cancer images based on a convolutional network, including: A preprocessing module, configured to obtain a preset number of color Doppler ultrasound images of thyroid cancer and perform preprocessing on the color Doppler ultrasound images; A target segmentation module, configured to extract a thyroid cancer image from the preprocessed color Doppler ultrasound image, perform color space conversion on the thyroid cancer image through a preset color space, and segment the blood flow spectral image in the thyroid cancer according to the hue and saturation information in the color space; A target enhancement module, configured to perform a mathematical morphology operation on the blood flow spectral image to enhance the blood flow spectral image; A feature extraction module, configured to input the blood flow spectral image after the execution of the mathematical morphology operation into a trained feature extraction model for convolutional processing to obtain the proportion and distribution of the blood flow spectral signal in thyroid cancer.

[0012] In a third aspect, the present application provides an electronic device, including: A memory for storing instructions executable by the processor; Wherein, the processor is configured to execute the instructions to implement the feature extraction method for thyroid cancer images based on a convolutional network according to any one of the first aspect.

[0013] Fourthly, the present application provides a computer-readable storage medium. When the instructions in the computer-readable storage medium are executed by a processor, a method for extracting features of a thyroid cancer image based on a convolutional network as described in any one of the first aspects is implemented.

[0014] For the method, device, electronic device and storage medium for extracting features of a thyroid cancer image based on a convolutional network provided by the present application, a thyroid cancer image is extracted from a color Doppler ultrasound image of thyroid cancer, and then a blood flow spectrum image in the thyroid cancer is obtained. The degree of canceration of the thyroid cancer is judged according to the blood flow distribution and proportion in the thyroid cancer. In this way, effective features in the thyroid cancer image can be accurately identified, and thyroid cancer and thyroid benign nodules can be effectively distinguished. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can also be obtained based on these drawings without creative efforts.

[0016] Figure 1 It is a schematic flowchart of a method for extracting features of a thyroid cancer image based on a convolutional network in an embodiment of the present application.

[0017] Figure 2 It is a curve graph of the pixel intensity distribution of each pixel point in a blood flow spectrum image in an embodiment of the present application.

[0018] Figure 3 It is a schematic structural diagram of a device for extracting features of a thyroid cancer image based on a convolutional network in an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] The specific embodiments of the present invention will be described in detail below with reference to the drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the description of the present invention without creative efforts belong to the scope of protection of the present invention.

[0020] As Figure 1 shown, the present invention provides a method for extracting features of a thyroid cancer image based on a convolutional network, including: Step S100: Obtain a preset number of color Doppler ultrasound images of thyroid cancer and preprocess the color Doppler ultrasound images; Step S200: Extract the thyroid cancer image from the preprocessed color Doppler ultrasound image, perform color space conversion on the thyroid cancer image through a preset color space, and segment the blood flow spectrum image in the thyroid cancer according to the hue and saturation information in the color space; Step S300: Perform mathematical morphological operations on the blood flow spectrum image to enhance the blood flow spectrum image; Step S400: Input the blood flow spectrum image after performing the mathematical morphological operations into a trained feature extraction model for convolution processing to obtain the proportion and distribution of the blood flow spectrum signal in thyroid cancer.

[0021] It can be understood that parameters such as the degree of vascularization, blood flow velocity, and resistance index in thyroid cancer can be used to evaluate the malignant risk of thyroid cancer. Color Doppler ultrasound images can provide information such as blood flow direction, velocity, and distribution in thyroid cancer, which helps to analyze the angiogenesis situation in thyroid cancer.

[0022] In step S100, the preprocessing of the color Doppler ultrasound image may include: using a filter to reduce random noise and artifacts brought by the device in the color Doppler ultrasound image. Enhancing the clarity of the blood flow spectrum signal by adjusting parameters such as contrast and brightness. Using a segmentation model to segment the blood flow from surrounding tissues, or segment the tumor nodule from surrounding tissues, etc. Adding annotations to information such as blood flow direction and velocity in the color Doppler ultrasound image. By preprocessing the color Doppler ultrasound image, it helps to improve the accuracy and efficiency of subsequent blood flow spectrum image extraction.

[0023] Step S200 further includes: annotating the thyroid cancer image in the color Doppler ultrasound image to obtain multiple annotated regions, expanding the annotated regions according to gray similarity, and covering the thyroid cancer region in the color Doppler ultrasound image; adjusting the contour line of the thyroid cancer region through the principle of energy minimization to make the contour line fit the boundary of the thyroid cancer region, and extracting the thyroid cancer image based on the contour line.

[0024] In the embodiment of the present application, the method of annotating the thyroid cancer image in the color Doppler ultrasound image is to use a convolutional neural network to learn and identify thyroid cancer from the original color Doppler ultrasound image, so as to annotate the thyroid cancer. The convolutional neural network includes CNN convolutional neural network, etc.

[0025] In addition, it can also be that professional doctors outline the thyroid cancer area on the image. A semi-automatic annotation method can also be adopted, combining user markings and image analysis algorithms, such as finding the thyroid cancer area based on the similarity around the existing contour line and performing manual fine-tuning. A fully automatic annotation method can also be adopted, that is, using image processing and pattern recognition algorithms to automatically identify the thyroid cancer area, including methods based on pixel classification, region growing, clustering analysis, etc.

[0026] Here, the principle of energy minimization is used to enhance the contour line of thyroid cancer, improving the recognition accuracy of the boundary of thyroid cancer.

[0027] In step S200, the original color Doppler ultrasound image of thyroid cancer contains multiple channels. By weighted averaging, the RGB color is converted to another color space, such as HSV (Hue-Saturation-Value), HIS (Hue-Intensity-Saturation), etc., which can make the high-brightness areas (such as blood flow areas) in the color Doppler ultrasound image of thyroid cancer more prominent for the recognition and separation of these areas. The blood flow spectral signal is distinguished and enhanced by the new hue and saturation values, and then the improved image processing algorithm is used to separate the blood flow spectral signal from the color Doppler ultrasound image, so as to obtain a clearer and more accurate blood flow spectral image.

[0028] It can be understood that in the color Doppler ultrasound image of thyroid cancer patients, it can be clearly observed that the blood flow in the cancerous area is richer, and the blood flow spectral signal is stronger compared with thyroid benign nodules. Therefore, by judging the deterioration of thyroid cancer through the intensity of the blood flow spectral signal in the blood flow spectrogram, the diagnostic accuracy can be improved.

[0029] Here, the improved algorithms for separating the blood flow spectral signal can include the single-scale algorithm Retinex (SSR), the multi-scale algorithm Retinex (MSR), the multi-scale adaptive gain algorithm Retinex (MSRCR), the convolutional neural network algorithm, the wavelet transform image enhancement algorithm based on the improved MSR, etc.

[0030] In color Doppler ultrasound images, there are boundary lines for the boundaries of different tissues, organs or structures. One of the characteristics of the thyroid gland in patients with thyroid cancer is that there are nodules of irregular shapes or round shapes with different sizes on both sides of the thyroid gland, and the boundary with the surrounding tissues is unclear. There are relatively large swollen lymph nodes around the thyroid gland of patients with malignant tumors, and their state is similar to the edge of tumor tissue, and the boundary between these lymph nodes and tissues such as surrounding adipose tissue is unclear. Due to the above reasons, even if the color Doppler ultrasound image of thyroid cancer has been preprocessed, and then the blood flow spectrum image segmented after color space conversion may still have problems such as blurred image edges, which will affect subsequent feature extraction and ultimately lead to misdiagnosis of patients. Therefore, in step S300, a mathematical morphology operation is performed on the blood flow spectrum image to enhance the blood flow spectrum image, obtain a clearer blood flow spectrum signal, and improve the accuracy of diagnosis.

[0031] In one embodiment, the morphology operation includes opening operation and closing operation, and step S300 includes: Step S301, perform binaryzation processing on the blood flow spectrum image based on a preset binaryzation threshold; Step S302, perform target annotation on the binaryzation-processed blood flow spectrum image based on a preset structural element; Step S303, perform opening operation and / or closing operation on the blood flow spectrum image after the target annotation to enhance the blood flow spectrum image.

[0032] In this embodiment, in order to improve the accuracy of the binaryzation threshold, an adaptive threshold method is adopted in this embodiment to determine the preset binaryzation threshold based on the values of each pixel point in the blood flow spectrum image to be processed.

[0033] Before performing step S301, it further includes: determining a preset binaryzation threshold.

[0034] Optionally, there are various ways to determine the binaryzation threshold. In this embodiment, the pixel intensity values (or brightness values) of all pixel points in the blood flow spectrum image to be processed are arranged in sequence to obtain the number of pixel points corresponding to different gray values in the blood flow spectrum image to be processed. The number of pixel points is used as the horizontal axis, and the pixel intensity value of the pixel point is used as the vertical axis to form a Figure 2 curve graph as shown.

[0035] To further reduce the contingency of data, empirical parameters can also be determined according to the cut-off frequency of the wall filter, and some abnormally over-bright points can be filtered out based on the empirical parameters. For example, setting the empirical parameter as a, arranging in the order of the pixel intensity values of all pixel points, filtering out a% of the abnormally over-bright points. Exemplarily, the empirical parameter a can be 3. Optionally, a spectral image excluding the remaining part near the baseline can also be selected for the operation of arranging the pixel intensity values of pixel points.

[0036] After filtering out the abnormally over-bright points, based on Figure 2 the given curve graph and the brightness values of pixel points, determine two pixel points with the minimum and maximum brightness values of the pixel points, and draw a line connecting these two pixel points. As Figure 2 the dashed line in [Figure number not provided], find the point P on the sorted curve graph that is farthest from the perpendicular distance of this dashed line, and use P*b as the demarcation point between the signal and noise of the spectral image, that is, the binarization threshold. Optionally, b is an empirical parameter, and its specific value can be defined according to the signal-to-noise ratio of the ultrasonic diagnostic instrument, or can be set as a fixed value according to expert experience. Exemplarily, the value of b can be around 1, b can be less than 1 or greater than 1, and this embodiment does not limit this.

[0037] After determining the binarization threshold, perform binarization processing on the blood flow spectral image to be processed based on the binarization threshold, and obtain the blood flow spectral image to be processed after binarization processing.

[0038] Optionally, it includes the following two situations: One situation is that if the brightness value of the pixel point in the blood flow spectral image to be processed is less than the binarization threshold, then determine the brightness value of the pixel point as the first value.

[0039] In this embodiment, the binarization threshold is T h , compare the brightness values of all pixel points in the blood flow spectral image to be processed with the binarization threshold T h . If the brightness value of the pixel point is less than the binarization threshold T h , then set the brightness value of this pixel point as the first value. Optionally, the first value can be 0.

[0040] The other situation is that if the brightness value of the pixel point in the blood flow spectral image to be processed is greater than or equal to the binarization threshold, then determine the brightness value of the pixel point as the second value.

[0041] In this embodiment, compare the brightness values of all pixel points in the blood flow spectral image to be processed with the binarization threshold T h . If the brightness value of the pixel point is greater than or equal to the binarization threshold T h , then set the brightness value of this pixel point as the second value. Optionally, the second value can be 1.

[0042] In this embodiment, the main purpose of binarizing the blood flow spectral image to be processed is to completely separate the blood flow spectral signal from the background noise.

[0043] In this embodiment, in step S302, the preset structural element (the preset structural element can be circular, rectangular, triangular, etc.) is used to define the "window" in the mathematical morphology operation, that is, it can be used as a template for subsequent mathematical morphology operations, and can enhance specific details or rate regions in the blood flow spectral image. In this embodiment, the mathematical morphology operations include erosion operation and dilation operation. For the blood flow spectral image after target annotation, it includes: ; Among them, is the target image annotated based on the preset structural element, is the blood flow spectral image, is the dilation operation symbol, is the erosion operation symbol, is the structural disk element with a width of 2 pixels in the opening operation; By first performing an erosion operation on the target image annotated based on the preset structural element, and then performing a dilation operation, the contour line of the target image annotated based on the preset structural element is smoothed and the connection between multiple target images annotated based on the preset structural element is removed.

[0044] In this embodiment, performing a closing operation on the blood flow spectral image after target annotation includes: First performing a dilation operation on the target image annotated based on the preset structural element, and then performing an erosion operation, to fill the holes in the target image annotated based on the preset structural element and fill the broken parts of the contour line of the target image annotated based on the preset structural element.

[0045] It should be noted that for a target image, only an erosion operation or only a dilation operation can also be performed. Which type of mathematical morphology operation to perform on the target image depends on the specific state of the target image and the feature state to be extracted.

[0046] It can be understood that the erosion operation is also called morphological erosion. It aligns the center point of the element with the image. If the values of all pixels intersecting with the image are not less than the target pixel threshold, the center point is retained; otherwise, the center point is removed.

[0047] Similarly, the dilation operation is also called morphological dilation, that is, through the center point in the structural element, each pixel point in the image is scanned in turn. The value of the pixel point on the picture is the maximum value of all pixel points covered by the structural element. The image obtained after scanning is the dilation image of the original image.

[0048] Before performing the erosion operation on the image, a preset target pixel threshold should be determined first. Since there are relatively large swollen lymph nodes around the thyroid of patients with malignant tumors, the boundaries between the lymph nodes and surrounding adipose tissue and other tissues are unclear, which easily leads to the erosion of the image with unclear tissues during the erosion operation, resulting in the loss of important features. In this embodiment, important feature information in multiple image samples after the erosion operation is extracted in advance, and the model of the erosion operation is trained using this important feature information, so as to obtain the target pixel threshold, thereby avoiding the erosion of important features in the image with unclear tissues.

[0049] Specifically, the method for determining the target pixel threshold includes: Obtain multiple original image samples that have not undergone the erosion operation as the training sample data set. According to the distribution of the center points of the original image samples, a corresponding sample window is set outside the center point of each original image sample. The size and shape of the sample window are the same as the preset structural element, and each sample window contains at least one center point.

[0050] For the same original image sample, according to the random forest algorithm, randomly perform a masking operation on the sample image of the sample window where at least one center point is located (adjust the pixel intensity value in the sample window to zero), and then sequentially perform the erosion operation and dilation operation on the masked image sample to obtain a restored image sample. The dilation operation will perform a restoration operation on blocks such as cracks and dark regions according to the structural characteristics of the image sample, so as to fill the image sample into a complete image sample.

[0051] Since the image features of the masked sample window in the restored image sample are the image features restored according to the structures and contour lines of adjacent sample windows, its feature information may be similar to or quite different from that of the original image sample. By comparing the similarity between the restored image sample and the original image sample, it can be determined whether the important feature information in the masked sample window is likely to be lost during the erosion operation and dilation operation. When the similarity between the restored image sample and the original image sample is lower than the similarity threshold, it indicates that the important feature information of the masked sample window is likely to be lost during the erosion operation. At this time, the center point of the sample window should not be removed during the erosion operation; when the similarity between the restored image sample and the original image sample is higher than or equal to the similarity threshold, it indicates that the masked sample window is not likely to be lost during the erosion operation. At this time, the center point of the sample window can be removed during the erosion operation.

[0052] According to the organizational structure characteristics of the sample window, classify the features of the sample window corresponding to the restored image sample with a similarity lower than the similarity threshold, extract the pixel intensity values of the corresponding sample window, and perform error processing on the pixel intensity values (for example, remove the data with large pixel intensity value deviations), and the range data of each type of pixel intensity value obtained is the target pixel threshold.

[0053] By performing opening operations on the labeled targets in the blood flow spectrogram, it helps to eliminate the noise of the structural elements, clear the small island-like structures on the blood flow spectrogram, thereby enabling the blood vessel structure to be clearer and widening the gap of the blood flow spectral signals flowing away from the ultrasound probe. The closing operation can connect the relatively close structures in the blood vessels or blood flow spectral signals, making them form a whole and filling the ray penetration area and gaps. In addition, the closing operation can close the open blood vessel segments and make the entire blood flow path visible. In this way, it can provide clearer blood flow information in thyroid cancer and help make a more accurate diagnosis of the degree of deterioration of thyroid cancer.

[0054] In addition, mathematical morphological operations can also be used to separate thyroid cancer from surrounding tissues such as fat, making thyroid cancer present a clear independent state, which is convenient for analyzing the thyroid state.

[0055] Step S400 includes: Performing convolution sampling a preset number of times on the blood flow spectrogram image after the mathematical morphological operation through the trained feature extraction model to obtain multiple pixel intensity values in the blood flow spectrogram image and the texture features in the blood flow spectrogram image; Statistical analysis of the multiple pixel intensity values and their discrete conditions; Using the principal component analysis method to perform dimensionality reduction processing on the texture features and the discrete conditions of the pixel intensity values to obtain the dimensionality-reduced blood flow spectral signal; Performing a visualization operation on the dimensionality-reduced blood flow spectral signal to obtain the proportion and distribution of the blood flow spectral signal in thyroid cancer.

[0056] Here, the preset number of times can be any number. By performing multiple arbitrary convolution samplings on the blood flow spectrogram image until the key features in the blood flow spectrogram image are extracted, these key features can include the pixel intensity values and texture features of the blood flow spectrogram image, and these features can be used to represent information such as the blood flow state in the blood flow spectrogram image. By statistically analyzing the multiple pixel intensity values and their discrete conditions, the image quality of the blood flow spectrogram image can be improved and the blood flow characteristics such as the blood vessel structure and blood flow intensity included therein can be understood.

[0057] In this embodiment, the visualization of the dimensionality-reduced blood flow spectral signal can be displayed in various chart forms to obtain an intuitive representation of the blood flow spectral signal in thyroid cancer. By applying the dimensionality-reduced blood flow spectral signal to different types of charts, such as bar charts, line charts or pie charts, we can effectively visually represent the characteristics such as the distribution and proportion of the blood flow spectral signal. This visualization provides an easy understanding of the state of the blood flow spectral signal in thyroid cancer and helps analyze the blood flow dynamics and tumor characteristics.

[0058] In one embodiment, after performing a visualization operation on the dimension-reduced blood flow spectral signal to obtain the proportion and distribution of the blood flow spectral signal in thyroid cancer, the following steps are further included: Output the result of the visualization operation, and generate a preliminary diagnosis result of thyroid cancer based on the proportion and distribution of the blood flow spectral signal in the thyroid cancer.

[0059] Specifically, after obtaining the visualization result of the blood flow spectral signal, image analysis software can be used to score and quantify the blood flow visualization features, and determine the abnormal regions and features of the blood flow spectral signal. Then, a trained machine learning or deep learning model is used to analyze the quantified blood flow spectral signal. Based on the distribution and proportion of the blood flow spectral signal and the results of the machine learning model, a clinical doctor or an AI system comprehensively judges to generate a preliminary thyroid cancer diagnosis report, which at least includes the features of abnormal blood flow, the possible cancer risk level, and recommended treatment suggestions, etc.

[0060] The method for extracting thyroid cancer image features based on a convolutional network provided by the embodiments of the present invention extracts thyroid cancer images from color Doppler ultrasound images of thyroid cancer, then obtains blood flow spectral images in thyroid cancer, and judges the degree of canceration of the thyroid cancer according to the blood flow distribution and proportion in the thyroid cancer. In this way, the effective features in the thyroid cancer images can be accurately identified, and thyroid cancer and thyroid benign nodules can be effectively distinguished, so as to more accurately judge whether thyroid cancer has deteriorated.

[0061] Based on the same inventive concept as the foregoing embodiments, this embodiment also provides a device for extracting thyroid cancer image features based on a convolutional network, as Figure 3 shown, the device includes: A preprocessing module, configured to obtain a preset number of color Doppler ultrasound images of thyroid cancer, and perform preprocessing on the color Doppler ultrasound images; A target segmentation module, configured to extract thyroid cancer images from the preprocessed color Doppler ultrasound images, perform color space conversion on the thyroid cancer images through a preset color space, and segment the blood flow spectral images in the thyroid cancer according to the hue and saturation information in the color space; A target enhancement module, configured to perform mathematical morphological operations on the blood flow spectral images to enhance the blood flow spectral images; A feature extraction module, configured to input the blood flow spectral images after performing the mathematical morphological operations into a trained feature extraction model for convolutional processing to obtain the proportion and distribution of the blood flow spectral signal in thyroid cancer.

[0062] Based on the same inventive concept as the foregoing embodiments, this embodiment also provides an electronic device, including: A memory for storing the processor-executable instructions; Wherein, the processor is configured to execute the instructions to implement the feature extraction method of thyroid cancer images based on a convolutional network as in the foregoing embodiments.

[0063] Based on the same inventive concept as the above embodiments, this embodiment also provides a computer-readable storage medium storing instructions for being loaded and executed by a processor to implement the feature extraction method of thyroid cancer images based on a convolutional network as described above.

[0064] In the embodiments of the mobile terminal and the computer-readable storage medium provided in this application, all the technical features of the above control method embodiments are included, and the description and explanation content in the specification are basically the same as those of the above method embodiments, and will not be repeated here.

[0065] This application embodiment also provides a computer program product, which includes computer program code. When the computer program code runs on a computer, it causes the computer to execute the methods in the above various possible implementation manners.

[0066] This application embodiment also provides a chip, including a memory and a processor. The memory is used to store a computer program, and the processor is used to call and run the computer program from the memory, so that the device installed with the chip executes the methods in the above various possible implementation manners.

[0067] The serial numbers of the above embodiments of this application are only for description and do not represent the advantages or disadvantages of the embodiments.

[0068] The above is only the specific implementation manner 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 those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the appended claims.

Claims

1. A method for extracting features of thyroid cancer images based on a convolutional network, characterized in that Including: Obtain a preset number of color Doppler ultrasound images of thyroid cancer, and preprocess the color Doppler ultrasound images; Extract thyroid cancer images from the preprocessed color Doppler ultrasound images, perform color space conversion on the thyroid cancer images through a preset color space, and segment the blood flow spectrum images in the thyroid cancer according to the hue and saturation information in the color space; Perform mathematical morphology operations on the blood flow spectrum images to enhance the blood flow spectrum images; Input the blood flow spectrum images after performing the mathematical morphology operations into a trained feature extraction model for convolution processing to obtain the proportion and distribution of blood flow spectrum signals in thyroid cancer.

2. The feature extraction method for thyroid cancer images based on a convolutional network according to claim 1, wherein The extracting thyroid cancer images from the preprocessed color Doppler ultrasound images includes: Annotate the thyroid cancer images in the color Doppler ultrasound images to obtain multiple annotated regions, expand the annotated regions according to gray similarity, and cover the thyroid cancer regions in the color Doppler ultrasound images; Adjust the contour line of the thyroid cancer region through the principle of energy minimization to make the contour line fit the boundary of the thyroid cancer region, and extract the thyroid cancer images based on the contour line.

3. The feature extraction method for thyroid cancer images based on a convolutional network according to claim 1, wherein, The mathematical morphology operations include opening operation and closing operation; The performing mathematical morphology operations on the blood flow spectrum images to enhance the blood flow spectrum images includes: Perform binarization processing on the blood flow spectrum images based on a preset binarization threshold; Perform target annotation on the binarized blood flow spectrum images based on a preset structural element; Perform opening operation and / or closing operation on the blood flow spectrum images after performing the target annotation to enhance the blood flow spectrum images.

4. The feature extraction method for thyroid cancer images based on a convolutional network according to claim 3, wherein The mathematical morphology operations further include erosion operation and dilation operation; the performing opening operation on the blood flow spectrum images after performing the target annotation includes: ； Among them, is the target image labeled based on the preset structural elements, is the blood flow spectrum image, is the dilation operation symbol, is the erosion operation symbol, is a structural disc element with a width of 2 pixels in the opening operation; First perform erosion operation on the target image annotated based on the preset structural element, and then perform dilation operation to smooth the contour line of the target image annotated based on the preset structural element and remove the connection between multiple target images annotated based on the preset structural element.

5. The feature extraction method for thyroid cancer images based on a convolutional network according to claim 4, wherein The performing closing operation on the blood flow spectrum images after performing the target annotation includes: First perform dilation operation on the target image annotated based on the preset structural element, and then perform erosion operation to fill the holes in the target image annotated based on the preset structural element and fill the broken parts of the contour line of the target image annotated based on the preset structural element.

6. The feature extraction method for thyroid cancer images based on a convolutional network according to claim 1, wherein The inputting the blood flow spectrum images after performing the mathematical morphology operations into a trained feature extraction model for convolution processing to obtain the proportion and distribution of blood flow spectrum signals in thyroid cancer includes: Perform a preset number of convolution samplings on the blood flow spectrum images after performing the mathematical morphology operations through the trained feature extraction model to obtain multiple pixel intensity values in the blood flow spectrum images and the texture features in the blood flow spectrum images; Statistical analysis of the multiple pixel intensity values and their discrete conditions; The discrete situations of the texture features and the pixel intensity values are processed by principal component analysis for dimensionality reduction to obtain the blood flow spectral signal after dimensionality reduction; Visualization operations are performed on the blood flow spectral signal after dimensionality reduction to obtain the proportion and distribution of the blood flow spectral signal in thyroid cancer.

7. The feature extraction method of thyroid cancer images based on a convolutional network according to claim 6, characterized in that After performing visualization operations on the blood flow spectral signal after dimensionality reduction to obtain the proportion and distribution of the blood flow spectral signal in thyroid cancer, it further includes: Outputting the results of the visualization operations and generating a preliminary diagnosis result of thyroid cancer based on the proportion and distribution of the blood flow spectral signal in the thyroid cancer.

8. A feature extraction device for thyroid cancer images based on a convolutional network, characterized in that, It includes: A preprocessing module for acquiring a preset number of color Doppler ultrasound images of thyroid cancer and preprocessing the color Doppler ultrasound images; A target segmentation module for extracting thyroid cancer images from the preprocessed color Doppler ultrasound images, performing color space conversion on the thyroid cancer images through a preset color space, and segmenting the blood flow spectral images in the thyroid cancer according to the hue and saturation information in the color space; A target enhancement module for performing mathematical morphological operations on the blood flow spectral images to enhance the blood flow spectral images; A feature extraction module for inputting the blood flow spectral images after performing the mathematical morphological operations into a trained feature extraction model for convolution processing to obtain the proportion and distribution of the blood flow spectral signal in thyroid cancer.

9. An electronic device, characterized in that, It includes: A memory for storing instructions executable by the processor; Wherein, the processor is configured to execute the instructions to implement the method for feature extraction of thyroid cancer images based on a convolutional network as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are executed by the processor, the method for feature extraction of thyroid cancer images based on a convolutional network as described in any one of claims 1-7 is implemented.

Images