Vocal cord lesion feature extraction method and apparatus

Abstract

The application provides a vocal cord lesion feature extraction method and device. The method first acquires a laryngeal white light image and an NBI image, and calls a vocal cord segmentation model to segment the vocal cord in the two images to obtain a white light vocal cord image and an NBI vocal cord image. A lesion segmentation model is called to segment a target lesion in the white light vocal cord image to obtain a white light lesion image. Then, the lesion position, lesion convexity, lesion relative area and vocal cord symmetry quantification features are obtained from the white light vocal cord image and the white light lesion image. The lesion shape, lesion pedunculated / sessile learning feature and lesion color quantification feature are obtained from the white light lesion image. A blood vessel segmentation model is called to segment the blood vessels in the NBI vocal cord image, and the blood vessel morphology quantification feature is obtained from the segmented NBI vocal cord image. Finally, all the quantification features and learning features are integrated to obtain the vocal cord lesion feature, which can provide an accurate reference for the determination of the lesion type and improve the determination accuracy.

Description

Technical Field

[0001] This application relates to the field of medical auxiliary technology, and in particular to a method and apparatus for extracting vocal cord lesion features. Background Technology

[0002] Vocal cord polyps are benign proliferative lesions occurring in the superficial layer of the lamina propria of the vocal cords. They are also a special type of chronic laryngitis, with hoarseness as the most prominent clinical symptom. Clinical diagnosis can be made through laryngoscopy. Lesions on the vocal cords may be vocal cord nodules, vocal cord cysts, or vocal cord polyps. Currently, lesion characteristics are primarily identified manually to determine the lesion type. However, the characteristics of these three types are not easily distinguishable, and manual identification is not very accurate, easily leading to misdiagnosis and affecting subsequent treatment plans.

[0003] Therefore, there is a technical problem with the low accuracy of lesion feature identification when determining the type of vocal cord lesions, which needs to be improved. Summary of the Invention

[0004] This application provides a method and apparatus for extracting vocal cord lesion features, in order to alleviate the technical problem of low accuracy in lesion feature identification when determining the type of vocal cord lesions.

[0005] To address the aforementioned technical problems, this application provides the following technical solutions:

[0006] This application provides a method for extracting vocal cord lesion features, including:

[0007] Acquire white light image and NBI image of the larynx of the same case, call the trained vocal cord segmentation model to segment the vocal cords in the white light image and the NBI image respectively to obtain white light vocal cord image and NBI vocal cord image, call the trained lesion segmentation model to segment the target lesion in the white light vocal cord image to obtain white light lesion image;

[0008] First position information, first shape information, and first area information of the vocal cord reference region are obtained from the white light vocal cord image. Second position information and second area information of the target lesion are obtained from the white light lesion image. Based on the first position information, the first shape information, and the second position information, first relative position information of the target lesion and the vocal cord in a first direction and second relative position information in a second direction are obtained. A first quantization feature is obtained based on the first relative position information, and a second quantization feature is obtained based on the second relative position information. The protrusion information of the target lesion is obtained based on the first position information and the first shape information, and a third quantization feature is obtained based on the protrusion information. The relative area information of the target lesion and the vocal cord reference region is obtained based on the first position information, the first shape information, the first area information, and the second area information, and a fourth quantization feature is obtained based on the relative area information. The symmetry information of the vocal cord is obtained from the white light vocal cord image, and a fifth quantization feature is obtained based on the symmetry information.

[0009] The trained lesion shape recognition model is invoked to identify the shape of the target lesion in the white light lesion image, and a first recognition result is obtained. A first learning feature is obtained based on the first recognition result. The trained lesion stalk recognition model is invoked to identify whether the target lesion in the white light lesion image has a stalk, and a second recognition result is obtained. A second learning feature is obtained based on the second recognition result. The color information of the target lesion is obtained from the white light lesion image, and a sixth quantization feature is obtained based on the color information.

[0010] The trained blood vessel segmentation model is invoked to segment the blood vessels in the NBI vocal cord image, and the morphological information of the blood vessels is obtained from the segmented NBI vocal cord image. The seventh quantization feature is obtained based on the morphological information.

[0011] Based on all quantitative features and all learned features, vocal cord lesion features are obtained.

[0012] Meanwhile, this application also provides a vocal cord lesion feature extraction device, including:

[0013] The segmentation module is used to acquire white light images and NBI images of the larynx in the same case, call the trained vocal cord segmentation model to segment the vocal cords in the white light images and the NBI images respectively, and obtain white light vocal cord images and NBI vocal cord images. The trained lesion segmentation model is then called to segment the target lesion in the white light vocal cord image, and obtain white light lesion images.

[0014] A first obtaining module is configured to: obtain first position information, first shape information, and first area information of a vocal cord reference region from the white light vocal cord image; obtain second position information and second area information of a target lesion from the white light lesion image; obtain first relative position information and second relative position information of the target lesion and the vocal cord in a first direction and in a second direction based on the first position information, the first shape information, and the second position information; obtain a first quantization feature based on the first relative position information; obtain a second quantization feature based on the second relative position information; obtain protrusion information of the target lesion based on the first position information and the first shape information; obtain a third quantization feature based on the protrusion information; obtain relative area information of the target lesion and the vocal cord reference region based on the first position information, the first shape information, the first area information, and the second area information; obtain a fourth quantization feature based on the relative area information; and obtain symmetry information of the vocal cord from the white light vocal cord image, and obtain a fifth quantization feature based on the symmetry information.

[0015] The second module is used to call the trained lesion shape recognition model to identify the shape of the target lesion in the white light lesion image, obtain a first recognition result, and obtain a first learning feature based on the first recognition result; call the trained lesion stalk recognition model to identify whether the target lesion in the white light lesion image has a stalk, obtain a second recognition result, and obtain a second learning feature based on the second recognition result; and obtain the color information of the target lesion from the white light lesion image, and obtain a sixth quantization feature based on the color information.

[0016] The third module is used to call the trained blood vessel segmentation model to segment the blood vessels in the NBI vocal cord image, and to obtain the morphological information of the blood vessels from the segmented NBI vocal cord image, and to obtain the seventh quantization feature based on the morphological information.

[0017] The fourth module is used to obtain vocal cord lesion features based on all quantified features and all learned features.

[0018] Beneficial Effects: This application provides a method and apparatus for extracting vocal cord lesion features. The method first acquires a white light image and an NBI image of the larynx in the same case, and then calls a trained vocal cord segmentation model to segment the vocal cords in the white light image and the NBI image respectively, obtaining a white light vocal cord image and an NBI vocal cord image. Then, it calls a trained lesion segmentation model to segment the target lesion in the white light vocal cord image, obtaining a white light lesion image. Then, it obtains a first quantitative feature and a second quantitative feature of the lesion location, a third quantitative feature of the lesion protrusion, and a fourth quantitative feature of the relative area of ​​the lesion from the white light vocal cord image and the white light lesion image. It also obtains a fifth quantitative feature of the vocal cord symmetry from the white light vocal cord image, and a first learned feature of the lesion shape, a second learned feature of the lesion being pedunculated / unpedunculated, and a sixth quantitative feature of the lesion color from the white light lesion image. Finally, it calls a trained blood vessel segmentation model to segment the blood vessels in the NBI vocal cord image, obtaining a seventh quantitative feature of the blood vessel morphology from the segmented NBI vocal cord image. Finally, it synthesizes all quantitative features and all learned features to obtain the vocal cord lesion features. After acquiring white light and NBI images of the larynx, this application can automatically quantify and learn the lesion status of the target lesion in the vocal cords from multiple dimensions, and synthesize the features of the vocal cord lesion. These features can provide an accurate and comprehensive reference for determining the type of the target lesion, with high efficiency and high accuracy in determining the type of vocal cord lesion. Attached Figure Description

[0019] The technical solution and other beneficial effects of this application will become apparent from the following detailed description of specific embodiments in conjunction with the accompanying drawings.

[0020] Figure 1 This is a schematic diagram illustrating an application scenario of the vocal cord lesion feature extraction method provided in this application embodiment.

[0021] Figure 2 This is a flowchart illustrating the vocal cord lesion feature extraction method provided in an embodiment of this application.

[0022] Figure 3 This is a schematic diagram of vocal cord nodules, vocal cord polyps, and vocal cord cysts in the embodiments of this application.

[0023] Figure 4 This is a schematic diagram of the vocal cords in open and closed states in an embodiment of this application.

[0024] Figure 5 This is a schematic diagram of the vocal cord reference area in an embodiment of this application.

[0025] Figure 6 This is a schematic diagram showing the target lesion in different location intervals of the three partitions in the embodiments of this application.

[0026] Figure 7This is a schematic diagram showing different degrees of protrusion of the target lesion in the embodiments of this application.

[0027] Figure 8 This is a schematic diagram of the vocal cord segmentation region in the white light vocal cord image in the embodiments of this application.

[0028] Figure 9 This is a schematic diagram of the segmented blood vessel morphology in an embodiment of this application.

[0029] Figure 10 This is a schematic diagram of the vocal cord lesion feature extraction device provided in an embodiment of this application.

[0030] Figure 11 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0031] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0032] This application provides a method, apparatus, electronic device, and computer-readable storage medium for extracting vocal cord lesion features. The vocal cord lesion feature extraction apparatus can be integrated into an electronic device, which can be a server or a terminal, etc.

[0033] Please see Figure 1 , Figure 1 This is a schematic diagram of an application scenario for the vocal cord lesion feature extraction method provided in this application embodiment. The scenario may include a terminal and a server. Terminals, servers, and terminals and servers are connected and communicate through the Internet composed of various gateways. The application scenario includes an image receiving device 11 and a server 12. The image receiving device 11 may be a device with human-computer interaction function. The server 12 includes a local server and / or a remote server.

[0034] The image receiving device 11 and the server 12 are located in a wireless network or a wired network to enable data interaction between them, wherein:

[0035] The user inputs a white light image and an NBI image of the larynx of the same case into the image receiving device 11, and the image receiving device 11 sends the two images and a vocal cord lesion feature extraction request to the server 12. Server 12 first calls the trained vocal cord segmentation model to segment the vocal cords in the white light image and the NBI image respectively, obtaining the white light vocal cord image and the NBI vocal cord image. Then, it calls the trained lesion segmentation model to segment the target lesion in the white light vocal cord image, obtaining the white light lesion image. Then, it calls the preset algorithm and preset model to obtain the first and second quantitative features of the lesion location, the third quantitative feature of the lesion protrusion, and the fourth quantitative feature of the relative area of ​​the lesion from the white light vocal cord image and the white light lesion image. It also obtains the fifth quantitative feature of the vocal cord symmetry from the white light vocal cord image, the first learned feature of the lesion shape, the second learned feature of the lesion being pedunculated / unpedunculated, and the sixth quantitative feature of the lesion color from the white light lesion image. Finally, it calls the trained blood vessel segmentation model to segment the blood vessels in the NBI vocal cord image, obtaining the seventh quantitative feature of the blood vessel morphology from the segmented NBI vocal cord image. Finally, based on all the quantitative features and all the learned features, it obtains the vocal cord lesion features and returns them to the image receiving device 11, which then displays the results to the user. This method can automatically quantify, learn, and synthesize the lesion status of the target lesion in the vocal cords from multiple dimensions, and return the final vocal cord lesion features to the user. The user can use these features as a reference to determine the type of the target lesion, such as vocal cord nodules, vocal cord cysts, or vocal cord polyps. This method is highly efficient, and because the extracted vocal cord lesion features are highly accurate and comprehensive, the accuracy in determining the type of vocal cord lesion is also high.

[0036] It should be noted that, Figure 1 The system scenario diagram shown is merely an example. The servers and scenarios described in this application embodiment are for the purpose of more clearly illustrating the technical solutions of this application embodiment and do not constitute a limitation on the technical solutions provided in this application embodiment. As those skilled in the art will know, with the evolution of systems and the emergence of new business scenarios, the technical solutions provided in this application embodiment are also applicable to similar technical problems. Detailed descriptions are provided below. It should be noted that the order of description of the following embodiments is not intended to limit the preferred order of embodiments.

[0037] Please see Figure 2 , Figure 2 This is a flowchart illustrating the vocal cord lesion feature extraction method provided in this application embodiment. The method specifically includes:

[0038] S1: Obtain the white light image and NBI image of the larynx of the same case, call the trained vocal cord segmentation model to segment the vocal cords in the white light image and NBI image respectively, and obtain the white light vocal cord image and NBI vocal cord image. Call the trained lesion segmentation model to segment the target lesion in the white light vocal cord image, and obtain the white light lesion image.

[0039] In this embodiment, a case refers to a patient with at least one lesion on their vocal cords. Images of the vocal cords from the same case are acquired using medical-aided instruments, resulting in both white light images and NBI images. Each image includes the vocal cords and all lesions within them. The vocal cord segmentation model can be a Unet++ model. Before use, a professional endoscopist delineates the vocal cord boundaries in both types of images from a certain number of historical cases, using this delineated data as training data. This training data is then used to train the vocal cord segmentation model, enabling it to segment the vocal cords. Similarly, the lesion segmentation model can be a Unet++ model. Before use, a professional endoscopist delineates the lesion boundaries in white light images from a certain number of historical cases, using this delineated data as training data. This training data is then used to train the lesion segmentation model, enabling it to segment lesions.

[0040] After obtaining two types of images, the trained vocal cord segmentation model is used to segment the vocal cords in both types of images, resulting in a white-light vocal cord image and an NBI vocal cord image containing the vocal cord segmentation boundaries. Then, the trained lesion segmentation model is used to segment the target lesion in the white-light vocal cord image, resulting in a white-light lesion image containing the target lesion segmentation boundaries. When multiple lesions exist, each lesion is sequentially used as the target lesion, and the corresponding white-light lesion image is obtained. Figure 3 As shown, in the vocal cords, the type of target lesion may be a vocal cord nodule, vocal cord polyp, or vocal cord cyst.

[0041] S2: Obtain first position information, first shape information, and first area information of the vocal cord reference region from the white light vocal cord image; obtain second position information and second area information of the target lesion from the white light lesion image; based on the first position information, first shape information, and second position information, obtain first relative position information of the target lesion and the vocal cord in the first direction and second relative position information in the second direction; obtain a first quantization feature based on the first relative position information; obtain a second quantization feature based on the second relative position information; obtain the protrusion information of the target lesion based on the first position information and the first shape information; obtain a third quantization feature based on the protrusion information; obtain relative area information of the target lesion and the vocal cord reference region based on the first position information, first shape information, first area information, and second area information; obtain a fourth quantization feature based on the relative area information; obtain symmetry information of the vocal cord from the white light vocal cord image; obtain a fifth quantization feature based on the symmetry information.

[0042] The white light vocal cord image contains the segmentation boundaries of the vocal cords. After obtaining the segmentation boundaries of the vocal cords, the first positional information of the vocal cords can be determined. For example... Figure 4 As shown, the vocal cords include the left vocal cord and the right vocal cord. In the following embodiments, one of the left vocal cord and the right vocal cord is referred to as the first vocal cord and the other as the second vocal cord. Figure 4 The left side of the image shows the vocal cords in their open state, and the right side shows them in their closed state. Between the first and second vocal cords in their open states, a triangular region is formed, which constitutes the vocal cord reference region. For different patients, once the vocal cord position is determined, the area of ​​the vocal cord reference region is also determined; therefore, the first area information of the vocal cord reference region can be calculated.

[0043] The white light lesion image contains the segmentation boundary of the target lesion. After obtaining the segmentation boundary of the target lesion, the first location information of the target lesion can be determined, and the second area information of the target lesion can be calculated.

[0044] Different types of target lesions exhibit varying distribution locations within the vocal cords. These differences are primarily reflected in two directions: longitudinal and transverse. When the target lesion is a nodule, it is mainly distributed longitudinally near the junction of the upper and middle thirds of the vocal cord. When the target lesion is a cyst, it is mainly distributed longitudinally within the middle third. When the target lesion is a polyp, it is mainly distributed transversely along the edge of the vocal cord. Therefore, in this embodiment, the longitudinal direction is taken as the first direction, and the transverse direction as the second direction. Based on the first position information of the vocal cords and the second position information of the target lesion, the first relative position information in the first direction and the second relative position information in the second direction are calculated respectively. This method measures the type of target lesion from the perspective of the relative position of the lesion and the vocal cord, and by considering the relative position dimensions in two directions, it achieves high accuracy.

[0045] In one embodiment, S2 specifically includes: obtaining the segmentation boundary of the first vocal tract and the second vocal tract from the white light vocal tract image; determining a first minimum bounding rectangle of the first vocal tract and a second minimum bounding rectangle of the second vocal tract based on the segmentation boundary; determining a first side and a second side relative to each other from the first minimum bounding rectangle based on the reference intersection point of the first minimum bounding rectangle and the second minimum bounding rectangle, wherein the first side and the third side intersect at the reference intersection point; determining the midpoint of the second side as a first reference point, determining the midpoint of the fourth side as a second reference point, determining a first reference side based on the reference intersection point and the first reference point, and determining a third side based on the reference intersection point and the second reference point. A second reference edge is defined, and a third reference edge is determined based on the first and second reference points. Based on the first, second, and third reference edges, a vocal cord reference region between the first and second vocal cords is obtained. Based on the positional information of the reference center, the first reference point, and the second reference point, the region size information, region position information, and first area information of the vocal cord reference region are obtained. Based on the region size information and region position information, the first position information and first shape information of the vocal cords are obtained. The segmentation boundary of the target lesion is obtained from the white light lesion image. Based on the segmentation boundary, the centroid position information and second area information of the target lesion are obtained. Based on the centroid position information, the second position information of the target lesion is obtained.

[0046] like Figure 5As shown, taking the target lesion as a vocal cord nodule, the first vocal cord as the left vocal cord, and the second vocal cord as the right vocal cord as an example, after obtaining the segmentation boundaries of the first and second vocal cords, the first minimum bounding rectangle P1P2P3P4 of the first vocal cord and the second minimum bounding rectangle P1'P2'P3'P4' of the second vocal cord are calculated respectively. The minimum bounding rectangle is the rectangle that can cover the entire vocal cord and has the smallest area. The two minimum bounding rectangles intersect at point A, which is used as the reference intersection point of the vocal cord reference area. Based on the reference intersection point A, the first side P1P4 and the second side P2P3 are determined from the first minimum bounding rectangle, and the third side P1'P4' and the fourth side P2'P3' are determined from the second minimum bounding rectangle. The first side P1P4 and the third side P1'P4' intersect at the reference intersection point A. Then, the midpoint of the second side P2P3 is determined as the first reference point B, and the midpoint of the fourth side P2'P3' is determined as the second reference point C. Based on the reference intersection point A and the first reference point B, the first reference side AB is determined. Based on the reference intersection point A and the second reference point C, the second reference side AC is determined. Based on the first reference point B and the second reference point C, the third reference side BC is determined. Finally, based on AB, AC and BC, the vocal cord reference region is obtained.

[0047] After obtaining the vocal cord reference region, the size information of each reference side can be calculated based on the position information of the reference center A, the first reference point B, and the second reference point C. Then, the region size information and region position information of the triangular vocal cord reference region can be obtained. When the region size and region position are fixed, the position and shape of the vocal cord are also fixed. Therefore, the region size information and region position information of the vocal cord reference region can be used to characterize the first position information and first shape information of the entire vocal cord.

[0048] After obtaining the segmentation boundary of the target lesion from the white light lesion image, the location information of the centroid O point of the target lesion can be calculated, such as... Figure 5 The area within the circle is shown, and this location information is used to represent the second location information of the target lesion.

[0049] In one embodiment, S2 specifically includes: determining a reference centerline along a first direction based on the reference intersection point and the midpoint of the third reference edge; dividing the reference centerline into three equal parts; and sequentially generating a first bisector and a second bisector parallel to the third reference edge in the first direction away from the reference intersection point, so that the vocal cord reference region sequentially forms a first partition, a second partition, and a third partition in the first direction; obtaining a first distance between the lesion centroid and the first bisector based on region size information, region location information, and lesion centroid location information, and obtaining the relative positional relationship between the lesion centroid and the three partitions; comparing the first distance with a first threshold; and based on the first comparison result and... The relative positional relationship is used to obtain the first relative positional information of the target lesion and vocal cord in the first direction. Based on the first relative positional information, the first quantitative feature is obtained. Based on the centroidal position information of the lesion and the segmentation boundary of the first and second vocal cords, the target vocal cord where the target lesion is located is determined. Based on the minimum bounding rectangle of the target vocal cord, the target reference edge is determined. The second distance between the centroid of the lesion and the target reference edge in the second direction is obtained. The second distance and the second threshold are compared. Based on the second comparison result, the second relative positional information of the target lesion and vocal cord in the second direction is obtained. Based on the second relative positional information, the second quantitative feature is obtained.

[0050] The center of the third reference edge BC is determined from the vocal cord reference region. The line connecting the reference intersection point A and the center is the centerline of the triangle. The direction from the reference intersection point A to the center is taken as the first direction. The vocal cord reference region is divided into three equal parts along the first direction. In the first direction, the first and second bisectors are obtained sequentially, i.e., the 1 / 3 and 2 / 3 bisectors in the figure, respectively. This results in three partitions: the uppermost first partition, the middle second partition, and the lowermost third partition. Then, the first distance d1 between point O and the first bisector is calculated. d1 is compared with the first threshold α to obtain the first comparison result. The relative positional relationship between point O and the three partitions is obtained. Combining the first comparison result and the relative positional relationship, the first relative positional information of the target lesion and the vocal cord in the first direction is obtained. This first relative positional information is used to characterize the longitudinal distribution of the target lesion on the vocal cord. Based on the first relative positional information, the first quantification feature label1 can be obtained. The above process satisfies the following formula:

[0051]

[0052] like Figure 6As shown, the value of α is less than 1 / 3 of the height of the first partition. When d1≤α, it means that point O is within a certain position range on both sides of the first bisector line, and the target lesion may be a nodule. When d1>α and point O is located in the second partition, it means that point O is in the second partition and is far away from the first bisector line, and the target lesion may be a cyst. When point O is located in other positions, label1 is recorded as 2.

[0053] When determining the initial relative position information between the target lesion and the vocal cords, the vocal cords themselves are not perfectly regular in shape. If the zoning criteria are derived solely from information on one side of the vocal cords, and the longitudinal distribution of the target lesion is determined using these criteria, the calculation error for each side will be substantial. Furthermore, due to differences in shape, size, and position between the left and right vocal cords, the zoning criteria for both sides will be inconsistent, leading to significant differences in the calculated distribution between the left and right sides. Both of these factors contribute to the low accuracy of the final first quantitative feature.

[0054] In this embodiment, instead of directly dividing the vocal cord into three zones, the two smallest bounding rectangles of the two vocal cords are first obtained to reduce the influence of irregular vocal cord shape. Then, the first and second reference edges of the vocal cord reference area are obtained by combining the two smallest bounding rectangles, and the partitioning standard is obtained based on the center line on the third reference edge. This partitioning standard is obtained by combining the information of both vocal cords, which reduces the error and makes both vocal cords adopt a unified partitioning standard. Combining these factors, the accuracy of the first quantitative feature can be significantly improved, and the accuracy of determining the target lesion type is also improved.

[0055] After determining the location of point O, the target vocal cord containing the target lesion is identified from the left and right vocal cords. Figure 5 Taking the left vocal cord as the target, the first minimum bounding rectangle P1P2P3P4 of the left vocal cord is the target minimum bounding rectangle, and P3P4 is the target reference edge of the vocal cord. The second distance d2 between the centroid O of the lesion and the target reference edge P3P4 is calculated. This second distance is the vertical distance, and the vertical direction between them is the second direction. Comparing d2 with the second threshold β yields the second comparison result. Based on this result, the second relative position information of the target lesion and the vocal cord in the second direction is obtained. This second relative position information characterizes whether the target lesion is located at the edge of the vocal cord or not. Based on this second relative position information, the second quantification feature label2 can be obtained, which satisfies the following formula:

[0056]

[0057] When d2 is not greater than β, it indicates that the lesion is located at the edge of the vocal cords and may be a polyp. When d2 is greater than β, it indicates that the lesion is not located at the edge of the vocal cords and may be one of the other two types. In this embodiment, the regular edge of the smallest outer rectangle is used as the target reference edge, and the second distance is calculated. Compared with directly taking a point on the incompletely regular left vocal cord edge to participate in the distance calculation, the calculation method in this embodiment has a smaller error, and the accuracy of the subsequent second quantitative feature is also higher, thus improving the accuracy of determining the type of target lesion.

[0058] Depend on Figure 3 As can be seen, the degree of protrusion relative to the vocal cord edge varies depending on the type of lesion. Polyps have a higher degree of protrusion, nodules have a lower degree, and cysts can have any degree of protrusion. That is, different lesion types exhibit different degrees of protrusion in white light vocal cord images. Therefore, the protrusion information of the target lesion can be obtained from the white light vocal cord image, and the protrusion of the vocal cord can be quantified to obtain a third quantitative feature. This method measures the type of the target lesion from the dimension of lesion protrusion.

[0059] In one embodiment, S2 specifically includes: obtaining a first reference angle with the reference intersection as the vertex based on the first outer point of the second side, the second outer point of the fourth side, and the reference intersection; obtaining a second reference angle with the reference intersection as the vertex based on the first reference point, the second reference point, and the reference intersection; obtaining a first angle of the first reference angle and a second angle of the second reference angle based on the region size information and the region position information; obtaining the protrusion information of the target lesion based on the difference between the first angle and the second angle and the protrusion coefficient; and obtaining a third quantification feature based on the protrusion information.

[0060] Figure 7 (a), (b), and (c) illustrate three scenarios: no lesions in the vocal cords, significant lesion protrusion in the vocal cords, and slight lesion protrusion in the vocal cords, respectively. In the first minimum bounding rectangle P1P2P3P4, the first outermost point of the second side is point P2. In the second minimum bounding rectangle P1'P2'P3'P4', the second outermost point of the fourth side is point P2'. The first outermost point P2, the reference center A, and the second outermost point P2' form the first reference angle ∠P2AP2' with point A as its vertex. The first reference point B, the reference center A, and the second outermost point P2' form the second reference angle ∠BAC with point A as its vertex. Based on the region size and location information obtained from the above embodiments, the first angle θ1 of the first reference angle ∠P2AP2' and the second angle θ2 of the second reference angle ∠BAC can be calculated.

[0061] contrast Figure 7As shown in (a) and (b), when the lesion protrudes outward from the edge of the vocal cords, if the degree of vocal cord opening remains consistent, other dimensions remain unchanged. Only the width of the second side P2P3 in the first minimum bounding rectangle P1P2P3P4 increases, causing the position of the midpoint B to shift to the right, θ2 to decrease. Furthermore, as the degree of protrusion increases, the degree of rightward shift of B also increases, the value of θ2 becomes smaller, and the difference between θ1 and θ2 increases. Conversely, as the degree of protrusion decreases, the degree of rightward shift of B also decreases, and the difference between θ1 and θ2 decreases until it decreases to a value similar to... Figure 7 As shown in (c), the protrusion is 0, meaning it does not protrude outwards at all. In this case, the position of point B is the same as in (a). Therefore, the difference between θ1 and θ2, along with the protrusion coefficient π, can be used to characterize the protrusion of the target lesion. Using label3 to represent the third quantification feature, it satisfies the following formula:

[0062]

[0063] Since the value of θ2 is also related to θ1, we use θ1 as a reference and the relative angle change between θ2 and θ1 to characterize the protrusion of the target lesion, which is more accurate than using absolute angles.

[0064] When the lesion type is a nodule, the lesion area in the vocal cords is small; when the lesion type is a cyst, the lesion area in the vocal cords is large; and when the lesion type is a polyp, the lesion area in the vocal cords can be any size. That is, different lesion types have different area sizes in the white light vocal cord image. Compared to using absolute size to measure the area of ​​the target lesion, relative size is more accurate. In this embodiment, the relative area size between the target lesion and the vocal cord reference area is used for measurement. Furthermore, when using relative area measurement, the vocal cords need to be in a specific open state to participate in the calculation. This condition can be restricted by the first position information of the vocal cords. Therefore, the open state of the vocal cords can be determined first based on the first position information of the vocal cords to see if the condition is met. Then, the relative area information between the target lesion and the vocal cord reference area is quantified to obtain a fourth quantification feature, which is used to represent the area size of different types of lesions. This method measures the type of target lesion from the dimension of relative area.

[0065] In one embodiment, S2 specifically includes: obtaining the opening degree coefficient of the vocal cord reference region based on the region size information; when the opening degree coefficient is not less than the degree coefficient threshold, obtaining the relative area of ​​the target lesion and the vocal cord reference region based on the second area information and the first area information, comparing the relative area with the third threshold, obtaining the relative area information of the target lesion and the vocal cord reference region based on the third comparison result, and obtaining the fourth quantitative feature based on the relative area information.

[0066] As can be seen from the above embodiments, the first position information of the vocal cords is characterized by the region size information of the vocal cord reference area, such as... Figure 5 As shown, based on the size information of this region, the length l of the third reference edge BC can be obtained. BC And the length l of the first reference edge AB AB Calculate the ratio l between the two. BC / l AB A larger ratio indicates a greater degree of opening in the vocal cord reference region, and vice versa. This ratio is compared with the degree coefficient threshold ε. If the ratio is not less than ε, it means that the current degree of opening can meet the needs of subsequent calculations. Let S1 represent the second area of ​​the target lesion, S0 represent the first area of ​​the vocal cord reference region, δ represent the third threshold, and label3 represent the fourth quantification feature. The value of label4 then satisfies the following formula:

[0067]

[0068] When the ratio is greater than δ, it indicates that the target lesion is large, possibly a cyst or polyp. When the ratio is not greater than δ, it indicates that the target lesion is small, possibly a nodule or polyp.

[0069] When determining the relative area information between the target lesion and the vocal cords, since the areas of the left and right vocal cords are somewhat different, if the size of the target lesion is measured only based on the relative area between the target lesion and one vocal cord, the relative area calculated for the same patient when the target lesion is located on the left vocal cord and when it is located on the right vocal cord will be different. This leads to inconsistent rules for measuring the size of the target lesion and may result in two different measurement results, making the accuracy of the fourth quantitative feature low.

[0070] In this embodiment, instead of directly calculating the relative area of ​​the target lesion with the left or right vocal cord, the two smallest bounding rectangles of the two vocal cords are first obtained, and then the two smallest bounding rectangles are combined to obtain the vocal cord reference area. Since the vocal cord reference area combines the area information of the two vocal cords and serves as the only relative area reference, the accuracy of the fourth quantitative feature can be significantly improved by combining these factors, and the accuracy of determining the type of target lesion is also improved.

[0071] When the lesion type is nodule, the lesion is symmetrically distributed in the vocal cords, meaning it is distributed simultaneously in both the left and right vocal cords. When the lesion type is cyst or polyp, the lesion is distributed only in one of the left and right vocal cords. In other words, different lesion types exhibit different distribution symmetries in white light vocal cord images. This asymmetry also results in asymmetry between the two vocal cords. Therefore, symmetry information of the vocal cords can be obtained from white light vocal cord images, and the symmetry of the vocal cords can be quantified to obtain the fifth quantitative feature. This method measures the type of target lesion from the dimension of symmetry in lesion distribution.

[0072] In one embodiment, S2 specifically includes: obtaining the first vocal tract area, the centroid position, and the first texture entropy of the first vocal tract from the white light vocal tract image, and obtaining the second vocal tract area, the centroid position, and the second texture entropy of the second vocal tract; obtaining the area difference between the first and second vocal tracts based on the first and second vocal tract areas, obtaining the positional difference between the first and second vocal tracts based on the centroid positions, and obtaining the texture difference between the first and second vocal tracts based on the first and second texture entropies; obtaining preset weights for each difference degree, obtaining symmetry coefficients based on each difference degree and the corresponding preset weights, comparing the symmetry coefficients with a fourth threshold, obtaining symmetry information of the vocal tracts based on the fourth comparison result, and obtaining a fifth quantization feature based on the symmetry information.

[0073] like Figure 8 As shown, taking the first vocal cord as the left vocal cord, the second vocal cord as the right vocal cord, and the target lesion located on the left vocal cord as an example, the white light vocal cord image is first binarized, and then the area S of the first vocal cord of the left vocal cord is calculated based on the connected components. Z and the area S of the second vocal cord of the right vocal cord Y Calculate the centroid position (x) of the first vocal tract segment of the left vocal tract based on the connected domain. Z y Z ) and the centroid position of the second vocal cord in the right vocal cord segmentation region (x Y y Y ), and calculate the first texture entropy Q of the left vocal cord. Z And the second texture entropy Q of the right vocal cord Y The texture entropy value satisfies the following formula:

[0074] Q=-∑∑p(g1,g2)logp(g1,g2) (Formula 5)

[0075] In the formula, (g1, g2) represents a pair of gray values ​​between two adjacent pixels in a certain vocal cord region, and p(g1, g2) represents the overall probability of each gray value pair, which satisfies the following formula:

[0076]

[0077] In the formula, R is the normalization coefficient, and N represents the width and height of the image. The value of R satisfies the following formula:

[0078]

[0079] In the formula, θ = 0 indicates that two pixels are adjacent horizontally, θ = 90 degrees indicates that two pixels are adjacent vertically, θ = 45 degrees indicates that two pixels are adjacent diagonally to the left, and θ = 135 degrees indicates that two pixels are adjacent diagonally to the right. Substituting the grayscale value pairs in the left vocal cord region and the grayscale value pairs in the right vocal cord region into formulas 5 to 7 above, the first texture entropy Q of the left vocal cord can be calculated. Z And the second texture entropy Q of the right vocal cord Y .

[0080] After obtaining S Z S Y 、(x Z y Z ), (x Y y Y ), Q Z And Q Y Then, the area difference S between the left and right vocal cords was calculated separately. Z / S Y Location difference |x Z -x Y | / |y Z -y Y | and texture difference Q Z / Q Y The preset weights of the three factors, such as 0.3, 0.3, and 0.4, are obtained. Then, by combining the differences and their corresponding preset weights, a symmetry coefficient can be obtained. This symmetry coefficient is then compared with a fourth threshold. Compare them, if they are not greater than This indicates symmetry, meaning the target lesion exists in both vocal cords. If it is larger than... This indicates asymmetry, meaning the target lesion exists only on one side of the vocal cord. Using label5 to represent the fifth quantification feature, its value satisfies the following formula:

[0081]

[0082] In this embodiment, the symmetry of the left and right vocal cords is measured by three aspects: area difference, position difference, and texture difference. This provides a more comprehensive measurement of distribution symmetry and makes the fifth quantitative feature more accurate.

[0083] S3: Call the trained lesion shape recognition model to identify the shape of the target lesion in the white light lesion image, obtain the first recognition result, and obtain the first learning feature based on the first recognition result; call the trained lesion stalk recognition model to identify whether the target lesion in the white light lesion image has a stalk, obtain the second recognition result, and obtain the second learning feature based on the second recognition result; obtain the color information of the target lesion from the white light lesion image, and obtain the sixth quantization feature based on the color information.

[0084] The lesion shape recognition model can be the Unet++ model. Before using this model, a professional endoscopist first delineates the lesion boundaries in white light lesion images of a certain number of historical cases, identifies the shape, and marks the corresponding label as training data. Then, the lesion shape recognition model is trained with this training data, so that the model has the ability to recognize the shape of lesions.

[0085] When the lesion type is a nodule, the lesion shape is nearly circular; when the lesion type is a polyp, the lesion is an irregular, arbitrary shape; and when the lesion type is a cyst, the lesion shape is nearly circular or elliptical. That is, in a white light lesion image, each lesion type exhibits a different shape. Based on this, a trained lesion shape recognition model can be used to identify the lesion in the white light lesion image, obtaining a first recognition result indicating that the target lesion is nearly circular or elliptical, or an irregular shape. Different values ​​are output for the first recognition results in different cases, thus obtaining the first learned feature of the target lesion. This method measures the type of the target lesion from the shape dimension.

[0086] Let label6 represent the first learned feature of the target lesion. The trained lesion shape recognition model is called to identify the lesion image under white light. When the first recognition result is nearly circular, elliptical, or irregular, the value of label6 satisfies the following formula:

[0087]

[0088] The lesion pedicle recognition model can be the VGG16 model. Before using this model, professional endoscopists first identify whether the lesions in the white light lesion images of a certain number of historical cases are pedunculated, and mark the corresponding pedunculated / stalkless labels as training data. Then, the lesion pedicle recognition model is trained with this training data, so that the model has the ability to identify whether the lesion is pedunculated or not.

[0089] When the lesion type is a nodule or cyst, the lesion is sessile; when the lesion type is a polyp, the lesion may be sessile. That is, in a white light image of a lesion, different lesion types may be sessile or pedunculated. Based on this, a trained lesion stalk recognition model can be used to identify the lesion in a white light image, obtaining a second recognition result of whether the target lesion is sessile or pedunculated. Different values ​​are output for the second recognition result in different cases, resulting in a second learned feature of the target lesion. This method measures the type of the target lesion from the dimension of whether it is pedunculated.

[0090] Let label7 represent the second learned feature of the target lesion. The trained lesion stalk recognition model is called to identify the lesion in the white light image. When the second recognition result is stalkless or pedunculated, the value of label7 satisfies the following formula:

[0091]

[0092] When the lesion type is a nodule, the lesion color is grayish-white; when the lesion type is a polyp, the lesion color is grayish-white or light red; and when the lesion type is a cyst, the lesion color is transparent or yellow. That is, in a white light lesion image, each lesion type exhibits a different color. Based on this, color information of the target lesion can be obtained from the white light lesion image, and different values ​​can be quantified for different color information to obtain the sixth quantification feature. This method measures the type of the target lesion from the color dimension.

[0093] In one embodiment, S3 specifically includes: obtaining the three-channel color features of each pixel of the target lesion from the white light lesion image; obtaining the three-channel color features of the target lesion based on the total number of pixels and the three-channel color features of each pixel; obtaining the mean value of the lesion color features based on the three-channel color features of the target lesion; comparing the mean value of the lesion color features with a fifth threshold; obtaining the color information of the target lesion based on the fifth comparison result; and obtaining a sixth quantization feature based on the color information.

[0094] The PIL built-in `getcolors()` function is used to obtain a list of lesion color features from a white light lesion image: `color = [(r1, g1, b1), (r2, g2, b2), ..., (r...`. n g n b n )], where (r i g i b iLet represent the three-channel color feature of the i-th pixel in the target lesion. Combining the pixel color features of n pixels in channel r, we obtain the pixel color feature of the target lesion in channel r. Combining the pixel color features of n pixels in channel g, we obtain the pixel color feature of the target lesion in channel g. Combining the pixel color features of n pixels in channel b, we obtain the pixel color feature of the target lesion in channel b. Calculating the average of the pixel color features of the target lesion in the three channels yields the mean color feature of the target lesion. The mean color features of the lesion calculated for different colors have a certain degree of difference; therefore, it can be used to represent the color distinction between different lesion types. Let label8 represent the sixth quantization feature, which satisfies the following formula:

[0095]

[0096] The above method integrates the pixel color features of each pixel in the target lesion under three channels for color quantization, which has high precision and makes the accuracy of the sixth quantization feature also high.

[0097] S4: Call the trained blood vessel segmentation model to segment blood vessels in the NBI vocal cord image, and obtain the morphological information of blood vessels from the segmented NBI vocal cord image. Based on the morphological information, obtain the seventh quantization feature.

[0098] The blood vessel segmentation model can be the Unet++ model. Before using this model, a professional endoscopist first delineates the blood vessel boundaries in a certain number of historical NBI vocal cord images and uses this as training data. Then, the blood vessel segmentation model is trained using this training data, so that the model has the ability to segment blood vessels.

[0099] When the lesion type is a nodule, no blood vessels are present. When the lesion type is a polyp, type I blood vessels (i.e., dendritic vessels) are present on the surface. When the lesion type is a cyst, the surface blood vessels are not type I, but the vascular texture is clear. In other words, in NBI lesion images, each lesion type exhibits differences in the presence / absence of blood vessels and the presence / absence of type I blood vessels. Based on this, a vascular segmentation model can be first used to segment blood vessels in the NBI vocal cord image. If no blood vessels are present, the absence of blood vessels is quantified numerically. If blood vessels are present, the morphological information of the blood vessels is extracted, and different values ​​are quantified for different morphologies, thus obtaining the seventh quantification feature. This method measures the type of the target lesion from the dimension of vascular morphology.

[0100] In one embodiment, S4 specifically includes: calling the trained vascular segmentation model to segment the NBI vocal cord image into blood vessels, and determining whether blood vessels exist based on the segmentation results; if not, generating morphological information of the absence of blood vessels, and obtaining the seventh quantitative feature of the lesion based on the morphological information; if yes, acquiring the centerline of M blood vessels and T blood vessel corner points, obtaining the blood vessel diameter of M blood vessel segments based on the centerline of M blood vessels, obtaining the slope of M blood vessel segments based on the T blood vessel corner points, and determining the blood vessel segment with the largest diameter among the M blood vessel segments as the target blood vessel segment; determining whether the slope of the target blood vessel segment is within a first threshold range, obtaining a first judgment result; determining whether the target blood vessel segment is the one with the smallest slope among the M blood vessel segments, obtaining a second judgment result; determining whether the slope of the remaining M-1 blood vessel segments is within a second threshold range, obtaining a third judgment result; determining whether M is not less than a sixth threshold, obtaining a fourth judgment result; obtaining the morphological information of the blood vessels based on the first judgment result, the second judgment result, the third judgment result, and the fourth judgment result, and obtaining the seventh quantitative feature based on the morphological information.

[0101] The seventh quantification feature is represented by label9. When no blood vessels are segmented, the morphological information indicates that no blood vessels exist, and label9 is set to 0. When blood vessels are segmented, it is calculated whether the blood vessel is a type I vessel. Specifically, for example... Figure 9 As shown, the Zhang-Suen thinning algorithm is used to find the centerline of the segmented blood vessels. Then, OpenCV's built-in corner detection algorithm is used to detect T corner points on the centerline, dividing the centerline into M segments based on these T corner points. For each segment, a tangent line is drawn at any point on the centerline, and a perpendicular line is drawn from this tangent line. The perpendicular line intersects the two sides of the blood vessel wall at two points; the distance between these two points is the diameter d of the blood vessel corresponding to the current pixel. ij Calculate the average diameter d of the blood vessel over K pixels in this segment. i The average value is then used as the diameter of the blood vessel segment, calculated using the following formula:

[0102]

[0103] For K vessel segments, K vessel diameters are obtained using the above formula. The segment with the largest diameter is then identified as the target vessel segment. For each vessel segment, the slope k can be calculated by connecting the two endpoints (corner points) of the segment. i Assuming the target vessel segment is the q-th segment, and the slope of the q-th segment is represented by k. q This indicates that the slope of each other blood vessel segment is represented by k. u This means that if the result of the first judgment is 1 / k q It is within the first threshold range, and the second judgment result is k. qIf the slope is the smallest among the M slopes, it indicates that the q-th segment of the vessel is a cross branch of a tree-like vessel, that is... Figure 9 The uppermost horizontal segment of the blood vessel. If the third judgment result is k u If all segments fall within the second threshold range, and the fourth judgment result is M ≥ the sixth threshold 4, then it indicates that the remaining M-1 segments of vessels are all vessels under the cross branches of the tree-like vessels, that is... Figure 9 The blood vessels in the lower middle section. When all four conditions are met, label9 is set to 1; when any one condition is not met, label9 is set to 2. The above process is represented by the following formula:

[0104]

[0105] S5: Based on all quantified features and all learned features, vocal cord lesion features are obtained.

[0106] In the above steps, multiple quantitative features and multiple learned features of the target lesion were obtained. These features together form the vocal cord lesion features, which can provide an accurate and comprehensive reference for the subsequent determination of the lesion type.

[0107] In one embodiment, after S5, the method further includes: obtaining the weights of each quantitative feature and each learned feature of the target lesion; obtaining the lesion coefficient of the target lesion based on each quantitative feature, each learned feature and the corresponding weight; and obtaining the lesion type information of the target lesion based on the lesion coefficient and the coefficient threshold.

[0108] In the above steps, the features label1 to label9 of the target lesion were obtained. The weights λ1 to λ9 of each quantitative feature were obtained using machine learning models such as decision trees and random forests. Then, the lesion coefficient of the target lesion was obtained by weighting these nine features. It satisfies the following formula:

[0109]

[0110] The coefficient thresholds include a first coefficient threshold τ1 and a second coefficient threshold τ2, which are used to determine the lesion coefficient. Compare with the first coefficient threshold τ1 and the second coefficient threshold τ2. If the comparison result is... If the value is not greater than τ1, then the type of the target lesion is determined to be a vocal cord nodule. If the comparison result is... If the value is greater than τ1 and not greater than τ2, then the type of the target lesion is determined to be a vocal cord polyp. If the comparison result is... If the value is greater than τ2, the target lesion is determined to be a vocal cord cyst, as shown in the following formula:

[0111]

[0112] Through the above process, the extracted vocal cord lesion features can be used to determine the type of target vocal cord lesion, providing an accurate reference for subsequent diagnosis and treatment.

[0113] Based on the above process, Table 1 illustrates the differences between the three lesion types in each dimension and the feature extraction methods in each dimension in the embodiments of this application. After acquiring the white light image and NBI image of the larynx, this application can automatically quantify and learn the lesion status of the target lesion in the vocal cords from multiple dimensions, and synthesize the vocal cord lesion features. Because the extracted vocal cord lesion features are highly accurate and comprehensive, the accuracy of determining the type of vocal cord lesion is also high.

[0114] Table 1

[0115]

[0116]

[0117] Based on the methods described in the above embodiments, this embodiment will further describe the process from the perspective of a vocal cord lesion feature extraction device. Please refer to [link to relevant documentation]. Figure 10 The vocal cord lesion feature extraction device may include:

[0118] The segmentation module 10 is used to acquire a white light image and an NBI image of the larynx of the same case, call the trained vocal cord segmentation model to segment the vocal cords in the white light image and the NBI image respectively to obtain a white light vocal cord image and an NBI vocal cord image, and call the trained lesion segmentation model to segment the target lesion in the white light vocal cord image to obtain a white light lesion image.

[0119] The first obtaining module 20 is configured to: obtain first position information, first shape information, and first area information of the vocal cord reference region from the white light vocal cord image; obtain second position information and second area information of the target lesion from the white light lesion image; obtain first relative position information and second relative position information of the target lesion and the vocal cord in a first direction and in a second direction based on the first position information, the first shape information, and the second position information; obtain a first quantization feature based on the first relative position information; obtain a second quantization feature based on the second relative position information; obtain protrusion information of the target lesion based on the first position information and the first shape information; obtain a third quantization feature based on the protrusion information; obtain relative area information of the target lesion and the vocal cord reference region based on the first position information, the first shape information, the first area information, and the second area information; obtain a fourth quantization feature based on the relative area information; and obtain symmetry information of the vocal cord from the white light vocal cord image, and obtain a fifth quantization feature based on the symmetry information.

[0120] The second module 30 is used to call the trained lesion shape recognition model to identify the shape of the target lesion in the white light lesion image, obtain a first recognition result, and obtain a first learning feature based on the first recognition result; call the trained lesion stalk recognition model to identify whether the target lesion in the white light lesion image has a stalk, obtain a second recognition result, and obtain a second learning feature based on the second recognition result; obtain the color information of the target lesion from the white light lesion image, and obtain a sixth quantization feature based on the color information;

[0121] The third module 40 is used to call the trained blood vessel segmentation model to segment the blood vessels in the NBI vocal cord image, and to obtain the morphological information of the blood vessels from the segmented NBI vocal cord image, and to obtain the seventh quantization feature based on the morphological information.

[0122] The fourth module 50 is used to obtain vocal cord lesion features based on all quantified features and all learned features.

[0123] In one embodiment, the vocal cords include a first vocal cord and a second vocal cord, and the first obtaining module 20 includes:

[0124] The first determining submodule is used to obtain the segmentation boundary between the first vocal cord and the second vocal cord from the white light vocal cord image, and determine the first minimum bounding rectangle of the first vocal cord and the second minimum bounding rectangle of the second vocal cord based on the segmentation boundary.

[0125] The second determining submodule is used to determine a first side and a second side that are set opposite to each other from the first minimum bounding rectangle and a third side and a fourth side that are set opposite to each other from the second minimum bounding rectangle, based on the reference intersection point of the first minimum bounding rectangle and the second minimum bounding matrix, wherein the first side and the third side intersect at the reference intersection point;

[0126] The first submodule is used to determine the midpoint of the second side as the first reference point, the midpoint of the fourth side as the second reference point, determine the first reference side based on the reference intersection and the first reference point, determine the second reference side based on the reference intersection and the second reference point, determine the third reference side based on the first reference point and the second reference point, and obtain the vocal cord reference region between the first vocal cord and the second vocal cord based on the first reference side, the second reference side and the third reference side;

[0127] The second submodule is used to obtain the region size information, region position information and first area information of the vocal cord reference region based on the position information of the reference center, the first reference point and the second reference point, and to obtain the first position information and first shape information of the vocal cord based on the region size information and the region position information.

[0128] The third submodule is used to obtain the segmentation boundary of the target lesion from the white light lesion image, obtain the lesion centroid location information and second area information of the target lesion based on the segmentation boundary, and obtain the second location information of the target lesion based on the lesion centroid location information.

[0129] In one embodiment, the first obtaining module 20 includes:

[0130] A generation submodule is used to determine a reference centerline along a first direction based on the reference intersection point and the midpoint of the third reference edge, divide the reference centerline into three equal parts, and sequentially generate a first bisector and a second bisector parallel to the third reference edge in the first direction away from the reference intersection point, so that the vocal cord reference region sequentially forms a first partition, a second partition, and a third partition in the first direction.

[0131] The fourth submodule is used to obtain a first distance between the centroid of the lesion and the first bisector based on the region size information, the region location information, and the centroid location information of the lesion, and to obtain the relative positional relationship between the centroid of the lesion and the three partitions. It compares the first distance with a first threshold, and obtains the first relative positional information of the target lesion and the vocal cord in the first direction based on the first comparison result and the relative positional relationship. It also obtains a first quantitative feature based on the first relative positional information.

[0132] The fifth submodule is used to determine the target vocal cord where the target lesion is located based on the location information of the lesion centroid and the segmentation boundary of the first vocal cord and the second vocal cord; determine the target reference edge based on the target minimum bounding rectangle corresponding to the target vocal cord; obtain the second distance between the lesion centroid and the target reference edge in the second direction; compare the second distance with the second threshold; obtain the second relative position information of the target lesion and the vocal cord in the second direction based on the second comparison result; and obtain the second quantization feature based on the second relative position information.

[0133] In one embodiment, the first obtaining module 20 includes:

[0134] The sixth submodule is used to obtain a first reference angle with the reference intersection point as the vertex based on the first outer point of the second side, the second outer point of the fourth side, and the reference intersection point; and to obtain a second reference angle with the reference intersection point as the vertex based on the first reference point, the second reference point, and the reference intersection point.

[0135] The seventh submodule is used to obtain the first angle of the first reference angle and the second angle of the second reference angle based on the area size information and the area position information.

[0136] The eighth submodule is used to obtain the protrusion information of the target lesion based on the difference between the first angle and the second angle and the protrusion coefficient, and to obtain the third quantitative feature based on the protrusion information.

[0137] In one embodiment, the first obtaining module 20 includes:

[0138] The ninth submodule is used to obtain the opening degree coefficient of the vocal cord reference region based on the region size information;

[0139] The tenth submodule is used to obtain the relative area of ​​the target lesion and the vocal cord reference area based on the second area information and the first area information when the opening degree coefficient is not less than the degree coefficient threshold, compare the relative area with the third threshold, obtain the relative area information of the target lesion and the vocal cord reference area based on the third comparison result, and obtain the fourth quantization feature based on the relative area information.

[0140] In one embodiment, the first obtaining module 20 includes:

[0141] The first acquisition submodule is used to acquire the first vocal tract area, the centroid position of the first vocal tract, and the first texture entropy of the first vocal tract from the white light vocal tract image, and to acquire the second vocal tract area, the centroid position of the second vocal tract, and the second texture entropy of the second vocal tract.

[0142] The eleventh submodule is used to obtain the area difference between the first vocal cord and the second vocal cord based on the area of ​​the first vocal cord and the area of ​​the second vocal cord, to obtain the position difference between the first vocal cord and the second vocal cord based on the centroid position of the first vocal cord and the centroid position of the second vocal cord, and to obtain the texture difference between the first vocal cord and the second vocal cord based on the texture entropy of the first and the texture entropy of the second.

[0143] The twelfth submodule is used to obtain the preset weights of each degree of difference, obtain the symmetry coefficients according to each degree of difference and the corresponding preset weights, compare the symmetry coefficients with the fourth threshold, obtain the symmetry information of the vocal cords according to the fourth comparison result, and obtain the fifth quantization feature according to the symmetry information.

[0144] In one embodiment, the second obtaining module 30 includes:

[0145] The second acquisition submodule is used to acquire the three-channel color features of each pixel of the target lesion from the white light lesion image;

[0146] The thirteenth submodule is used to obtain the three-channel color features of the target lesion based on the total number of pixels and the three-channel color features of each pixel, and to obtain the average color feature value of the target lesion based on the three-channel color features of the target lesion.

[0147] The fourteenth submodule is used to compare the mean value of the lesion color features with the fifth threshold, obtain the color information of the target lesion based on the fifth comparison result, and obtain the sixth quantization feature based on the color information.

[0148] In one embodiment, the third obtaining module 40 includes:

[0149] The judgment submodule is used to call the trained blood vessel segmentation model to perform blood vessel segmentation on the NBI vocal cord image, and determine whether blood vessels exist based on the segmentation results;

[0150] The fifteenth submodule is used to generate morphological information of non-existent blood vessels if no, and to obtain the seventh quantitative feature of the lesion based on the morphological information.

[0151] The third determination submodule is used to obtain the center line of M blood vessels and T blood vessel corner points if the condition is met; to obtain the blood vessel diameter of M blood vessel segments based on the center line of the M blood vessels; to obtain the slope of the M blood vessel segments based on the T blood vessel corner points; and to determine the blood vessel segment with the largest blood vessel diameter among the M blood vessel segments as the target blood vessel segment.

[0152] The sixteenth submodule is used to determine whether the slope of the target blood vessel segment is within the first threshold range and obtain the first judgment result; determine whether the target blood vessel segment is the one with the smallest slope among the M blood vessel segments and obtain the second judgment result; determine whether the slope of the remaining M-1 blood vessel segments is within the second threshold range and obtain the third judgment result; determine whether M is not less than the sixth threshold and obtain the fourth judgment result.

[0153] The seventeenth submodule is used to obtain the morphological information of the blood vessel based on the first judgment result, the second judgment result, the third judgment result and the fourth judgment result, and to obtain the seventh quantitative feature based on the morphological information.

[0154] In one embodiment, vocal cord lesion feature extraction further includes:

[0155] The fifth module is used to obtain the weights of each quantitative feature and each learned feature of the target lesion, and to obtain the lesion coefficient of the target lesion based on each quantitative feature, each learned feature and the corresponding weight.

[0156] The sixth module is used to obtain lesion type information of the target lesion based on the lesion coefficient and coefficient threshold.

[0157] Unlike existing technologies, the vocal cord lesion feature extraction device provided in this application first acquires a white light image and an NBI image of the larynx of the same case, and then calls a trained vocal cord segmentation model to segment the vocal cords in the white light image and the NBI image respectively, obtaining a white light vocal cord image and an NBI vocal cord image. Then, it calls a trained lesion segmentation model to segment the target lesion in the white light vocal cord image, obtaining a white light lesion image. Then, it obtains the first and second quantitative features of lesion location, the third quantitative feature of lesion protrusion, the fourth quantitative feature of lesion relative area, and the fifth quantitative feature of vocal cord symmetry from the white light vocal cord image and the white light lesion image. It also obtains the first learned feature of lesion shape, the second learned feature of lesion with / without pedicle, and the sixth quantitative feature of lesion color from the white light lesion image. Finally, it calls a trained blood vessel segmentation model to segment the blood vessels in the NBI vocal cord image, obtaining the seventh quantitative feature of blood vessel morphology from the segmented NBI vocal cord image. Finally, it synthesizes all quantitative features and all learned features to obtain the vocal cord lesion features. After acquiring white light and NBI images of the larynx, this application can automatically quantify and learn the lesion status of the target lesion in the vocal cords from multiple dimensions, and synthesize the features of the vocal cord lesion. These features can provide an accurate and comprehensive reference for determining the type of the target lesion, with high efficiency and high accuracy in determining the type of vocal cord lesion.

[0158] Accordingly, embodiments of this application also provide an electronic device, such as... Figure 11 As shown, the electronic device may include a radio frequency (RF) circuit 1001, a memory 1002 including one or more computer-readable storage media, an input unit 1003, a display unit 1004, a sensor 1005, an audio circuit 1006, a WiFi module 1007, a processor 1008 including one or more processing cores, and a power supply 1009, among other components. Those skilled in the art will understand that... Figure 11 The electronic device structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein:

[0159] The radio frequency circuit 1001 can be used for receiving and transmitting signals during information transmission or calls. Specifically, it receives downlink information from the base station and hands it over to one or more processors 1008 for processing; additionally, it transmits uplink data to the base station. The memory 1002 can be used to store software programs and modules. The processor 1008 executes various functional applications and vocal cord lesion type determination by running the software programs and modules stored in the memory 1002. The input unit 1003 can be used to receive input digital or character information and generate keyboard, mouse, joystick, optical, or trackball signal inputs related to customer settings and function control.

[0160] The display unit 1004 can be used to display information input by the customer or information provided to the customer, as well as various graphical client interfaces of the server. These graphical client interfaces can be composed of graphics, text, icons, videos, and any combination thereof.

[0161] The electronic device may also include at least one sensor 1005, such as a light sensor, a motion sensor, and other sensors. The audio circuitry 1006 includes a speaker that provides an audio interface between the client and the electronic device.

[0162] WiFi is a short-range wireless transmission technology. Electronic devices using the WiFi module 1007 can help customers send and receive emails, browse web pages, and access streaming media, providing customers with wireless broadband internet access. Although Figure 11 WiFi module 1007 is shown, but it is understood that it is not a necessary component of electronic devices and can be omitted as needed without changing the nature of the application.

[0163] The processor 1008 is the control center of the electronic device. It connects various parts of the phone through various interfaces and lines. By running or executing software programs and / or modules stored in the memory 1002, and calling data stored in the memory 1002, it performs various functions of the electronic device and processes data, thereby monitoring the phone as a whole.

[0164] The electronic device also includes a power supply 1009 (such as a battery) that supplies power to various components. Preferably, the power supply can be logically connected to the processor 1008 through a power management system, thereby enabling functions such as managing charging, discharging, and power consumption through the power management system.

[0165] Although not shown, electronic devices may also include cameras, Bluetooth modules, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 1008 in the server loads the executable files corresponding to the processes of one or more applications into the memory 1002 according to the following instructions, and the processor 1008 runs the applications stored in the memory 1002 to achieve the following functions:

[0166] Acquire white light image and NBI image of the larynx of the same case, call the trained vocal cord segmentation model to segment the vocal cords in the white light image and the NBI image respectively to obtain white light vocal cord image and NBI vocal cord image, call the trained lesion segmentation model to segment the target lesion in the white light vocal cord image to obtain white light lesion image;

[0167] First position information, first shape information, and first area information of the vocal cord reference region are obtained from the white light vocal cord image. Second position information and second area information of the target lesion are obtained from the white light lesion image. Based on the first position information, the first shape information, and the second position information, first relative position information of the target lesion and the vocal cord in a first direction and second relative position information in a second direction are obtained. A first quantization feature is obtained based on the first relative position information, and a second quantization feature is obtained based on the second relative position information. The protrusion information of the target lesion is obtained based on the first position information and the first shape information, and a third quantization feature is obtained based on the protrusion information. The relative area information of the target lesion and the vocal cord reference region is obtained based on the first position information, the first shape information, the first area information, and the second area information, and a fourth quantization feature is obtained based on the relative area information. The symmetry information of the vocal cord is obtained from the white light vocal cord image, and a fifth quantization feature is obtained based on the symmetry information.

[0168] The trained lesion shape recognition model is invoked to identify the shape of the target lesion in the white light lesion image, and a first recognition result is obtained. A first learning feature is obtained based on the first recognition result. The trained lesion stalk recognition model is invoked to identify whether the target lesion in the white light lesion image has a stalk, and a second recognition result is obtained. A second learning feature is obtained based on the second recognition result. The color information of the target lesion is obtained from the white light lesion image, and a sixth quantization feature is obtained based on the color information.

[0169] The trained blood vessel segmentation model is invoked to segment the blood vessels in the NBI vocal cord image, and the morphological information of the blood vessels is obtained from the segmented NBI vocal cord image. The seventh quantization feature is obtained based on the morphological information.

[0170] Based on all quantitative features and all learned features, vocal cord lesion features are obtained.

[0171] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the detailed description above, and they will not be repeated here.

[0172] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0173] Therefore, embodiments of this application provide a computer-readable storage medium storing a plurality of instructions that can be loaded by a processor to perform the following functions:

[0174] Acquire white light image and NBI image of the larynx of the same case, call the trained vocal cord segmentation model to segment the vocal cords in the white light image and the NBI image respectively to obtain white light vocal cord image and NBI vocal cord image, call the trained lesion segmentation model to segment the target lesion in the white light vocal cord image to obtain white light lesion image;

[0175] First position information, first shape information, and first area information of the vocal cord reference region are obtained from the white light vocal cord image. Second position information and second area information of the target lesion are obtained from the white light lesion image. Based on the first position information, the first shape information, and the second position information, first relative position information of the target lesion and the vocal cord in a first direction and second relative position information in a second direction are obtained. A first quantization feature is obtained based on the first relative position information, and a second quantization feature is obtained based on the second relative position information. The protrusion information of the target lesion is obtained based on the first position information and the first shape information, and a third quantization feature is obtained based on the protrusion information. The relative area information of the target lesion and the vocal cord reference region is obtained based on the first position information, the first shape information, the first area information, and the second area information, and a fourth quantization feature is obtained based on the relative area information. The symmetry information of the vocal cord is obtained from the white light vocal cord image, and a fifth quantization feature is obtained based on the symmetry information.

[0176] The trained lesion shape recognition model is invoked to identify the shape of the target lesion in the white light lesion image, and a first recognition result is obtained. A first learning feature is obtained based on the first recognition result. The trained lesion stalk recognition model is invoked to identify whether the target lesion in the white light lesion image has a stalk, and a second recognition result is obtained. A second learning feature is obtained based on the second recognition result. The color information of the target lesion is obtained from the white light lesion image, and a sixth quantization feature is obtained based on the color information.

[0177] The trained blood vessel segmentation model is invoked to segment the blood vessels in the NBI vocal cord image, and the morphological information of the blood vessels is obtained from the segmented NBI vocal cord image. The seventh quantization feature is obtained based on the morphological information.

[0178] Based on all quantitative features and all learned features, vocal cord lesion features are obtained.

[0179] The above provides a detailed description of a method, apparatus, electronic device, and computer-readable storage medium for extracting vocal cord lesion features according to the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the technical solutions and core ideas of this application. Those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for extracting a lesion feature of a vocal cord, characterized by, The method comprises the following steps: obtaining a white light image and an NBI image of a larynx of the same case, calling a trained vocal cord segmentation model to segment vocal cords in the white light image and the NBI image respectively to obtain a white light vocal cord image and an NBI vocal cord image, and calling a trained lesion segmentation model to segment a target lesion in the white light vocal cord image to obtain a white light lesion image; obtaining first position information, first shape information and first area information of a vocal cord reference region of the vocal cord from the white light vocal cord image, obtaining second position information and second area information of the target lesion from the white light lesion image, obtaining first relative position information of the target lesion and the vocal cord in a first direction and second relative position information in a second direction according to the first position information, the first shape information and the second position information, obtaining a first quantitative feature according to the first relative position information, obtaining a second quantitative feature according to the second relative position information, obtaining convexity information of the target lesion according to the first position information and the first shape information, and obtaining a third quantitative feature according to the convexity information; obtaining relative area information of the target lesion and the vocal cord reference region according to the first position information, the first shape information, the first area information and the second area information, and obtaining a fourth quantitative feature according to the relative area information; obtaining symmetry information of the vocal cord from the white light vocal cord image, and obtaining a fifth quantitative feature according to the symmetry information; calling a trained lesion shape recognition model to recognize a shape of the target lesion in the white light lesion image to obtain a first recognition result, and obtaining a first learning feature according to the first recognition result; calling a trained lesion peduncle recognition model to recognize whether the target lesion in the white light lesion image is pedunculated to obtain a second recognition result, obtaining a second learning feature according to the second recognition result, and obtaining color information of the target lesion from the white light lesion image, and obtaining a sixth quantitative feature according to the color information; calling a trained blood vessel segmentation model to segment blood vessels in the NBI vocal cord image, and obtaining morphological information of the blood vessels from the segmented NBI vocal cord image, and obtaining a seventh quantitative feature according to the morphological information; obtaining a vocal cord lesion feature according to all the quantitative features and all the learning features.

2. The vocal cord lesion feature extraction method according to claim 1, characterized by, The vocal cord comprises a first vocal cord and a second vocal cord, and the steps of obtaining first position information, first shape information and first area information of a vocal cord reference region of the vocal cord from the white light vocal cord image, and obtaining second position information and second area information of the target lesion from the white light lesion image, comprise the following steps: obtaining segmentation boundaries of the first vocal cord and the second vocal cord from the white light vocal cord image, and determining a first minimum circumscribed rectangle of the first vocal cord and a second minimum circumscribed rectangle of the second vocal cord according to the segmentation boundaries. determining a first edge and a second edge arranged oppositely from the first minimum circumscribed rectangle according to a reference intersection point of the first minimum circumscribed rectangle and the second minimum circumscribed rectangle, and determining a third edge and a fourth edge arranged oppositely from the second minimum circumscribed rectangle, the first edge and the third edge intersecting at the reference intersection point; determining a first reference point as a midpoint of the second edge and a second reference point as a midpoint of the fourth edge, determining a first reference edge according to the reference intersection point and the first reference point, determining a second reference edge according to the reference intersection point and the second reference point, determining a third reference edge according to the first reference point and the second reference point, and obtaining a vocal cord reference area between the first vocal cord and the second vocal cord according to the first reference edge, the second reference edge and the third reference edge; obtaining area size information, area position information and first area information of the vocal cord reference area according to position information of the reference intersection point, the first reference point and the second reference point, and obtaining first position information and first shape information of the vocal cord according to the area size information and the area position information; obtaining a segmentation boundary of the target lesion from the white light lesion image, obtaining lesion centroid position information and second area information of the target lesion according to the segmentation boundary, and obtaining second position information of the target lesion according to the lesion centroid position information.

3. The vocal cord lesion feature extraction method according to claim 2, characterized by, obtaining first relative position information of the target lesion and the vocal cord in a first direction and second relative position information in a second direction according to the first position information, the first shape information and the second position information, and obtaining a first quantization feature according to the first relative position information and a second quantization feature according to the second relative position information, including: determining a reference center line along the first direction according to the reference intersection point and a midpoint of the third reference edge, tri-sectioning the reference center line, and generating a first bisector and a second bisector parallel to the third reference edge in a first direction away from the reference intersection point in sequence, so that the vocal cord reference area forms a first sub-area, a second sub-area and a third sub-area in the first direction in sequence; obtaining a first distance between the lesion centroid and the first bisector according to the area size information, the area position information and the lesion centroid position information, and obtaining a relative position relationship between the lesion centroid and the three sub-areas, comparing the first distance with a first threshold value, obtaining the first relative position information of the target lesion and the vocal cord in the first direction according to a first comparison result and the relative position relationship, and obtaining a first quantization feature according to the first relative position information; According to the lesion shape center position information and the segmentation boundaries of the first vocal cord and the second vocal cord, a target vocal cord where the target lesion is located is determined, a target reference edge is determined according to a target minimum circumscribed rectangle corresponding to the target vocal cord, a second distance between the lesion shape center and the target reference edge in a second direction is obtained, the second distance is compared with a second threshold value, and second relative position information of the target lesion and the vocal cord in the second direction is obtained according to a second comparison result, and second quantization features are obtained according to the second relative position information.

4. The vocal cord lesion feature extraction method according to claim 2, characterized by, According to the first position information and the first shape information, convexity information of the target lesion is obtained, and the step of obtaining third quantization features according to the convexity information includes: According to the first outer side point of the second edge, the second outer side point of the fourth edge and the reference intersection point, a first reference angle with the reference intersection point as a vertex is obtained, and according to the first reference point, the second reference point and the reference intersection point, a second reference angle with the reference intersection point as a vertex is obtained; According to the region size information and the region position information, a first angle of the first reference angle and a second angle of the second reference angle are obtained; According to a difference value between the first angle and the second angle and a convexity coefficient, the convexity information of the target lesion is obtained, and third quantization features are obtained according to the convexity information.

5. The method of claim 2, wherein the feature extraction of the lesion of the vocal cord is characterized by, According to the first position information, the first shape information, the first area information and the second area information, relative area information of the target lesion and the vocal cord reference region is obtained, and the step of obtaining fourth quantization features according to the relative area information includes: According to the region size information, an opening degree coefficient of the vocal cord reference region is obtained; When the opening degree coefficient is not less than a degree coefficient threshold value, relative areas of the target lesion and the vocal cord reference region are obtained according to the second area information and the first area information, the relative areas are compared with a third threshold value, relative area information of the target lesion and the vocal cord reference region is obtained according to a third comparison result, and fourth quantization features are obtained according to the relative area information.

6. The method of claim 1, wherein, According to the symmetry information, fifth quantization features are obtained. First vocal cord area, first vocal cord shape center position and first texture entropy of the first vocal cord are obtained from the white light vocal cord image, and second vocal cord area, second vocal cord shape center position and second texture entropy of the second vocal cord are obtained; According to the first vocal cord area and the second vocal cord area, area difference degrees of the first vocal cord and the second vocal cord are obtained, according to the first vocal cord shape center position and the second vocal cord shape center position, position difference degrees of the first vocal cord and the second vocal cord are obtained, and according to the first texture entropy and the second texture entropy, texture difference degrees of the first vocal cord and the second vocal cord are obtained; respectively, and a symmetry coefficient is obtained according to the differences and the corresponding preset weights; the symmetry coefficient is compared with a fourth threshold value, and symmetry information of the vocal cord is obtained according to a fourth comparison result; and a fifth quantitative feature is obtained according to the symmetry information.

7. The method of claim 1, wherein, The color information of the target lesion is obtained from the white light lesion image, and a sixth quantitative feature is obtained according to the color information, including: Three-channel color features of each pixel point of the target lesion are obtained from the white light lesion image; A three-channel color feature mean value of the target lesion is obtained according to the total number of pixel points and the three-channel color features of each pixel point, and a lesion color feature mean value of the target lesion is obtained according to the three-channel color feature mean value of the target lesion; The lesion color feature mean value is compared with a fifth threshold value, a color information of the target lesion is obtained according to a fifth comparison result, and a sixth quantitative feature is obtained according to the color information.

8. The method of claim 1, wherein, The trained blood vessel segmentation model is called to perform blood vessel segmentation on the NBI vocal cord image, and morphological information of a blood vessel is obtained from the segmented NBI vocal cord image, and a seventh quantitative feature is obtained according to the morphological information, including: The trained blood vessel segmentation model is called to perform blood vessel segmentation on the NBI vocal cord image, and whether there is a blood vessel is determined according to the segmentation result; If not, morphological information of a blood vessel that does not exist is generated, and a seventh quantitative feature of the lesion is obtained according to the morphological information; If yes, M blood vessel center lines and T blood vessel corner points are obtained, a blood vessel diameter of each of the M blood vessel segments is obtained according to the M blood vessel center lines, a slope of each of the M blood vessel segments is obtained according to the T blood vessel corner points, and a target blood vessel segment with the largest blood vessel diameter among the M blood vessel segments is determined; Whether the slope of the target blood vessel segment is within a first threshold range is determined, a first determination result is obtained, whether the target blood vessel segment is the blood vessel segment with the smallest slope among the M blood vessel segments is determined, a second determination result is obtained, whether the slopes of the remaining M-1 blood vessel segments are within a second threshold range is determined, a third determination result is obtained, and whether M is not less than a sixth threshold value is determined, a fourth determination result is obtained; The morphological information of the blood vessel is obtained according to the first determination result, the second determination result, the third determination result and the fourth determination result, and a seventh quantitative feature is obtained according to the morphological information.

9. The method of claim 1, wherein, After the vocal cord lesion feature is obtained according to all the quantitative features and all the learning features, the method further includes: Weights of the quantitative features and the learning features of the target lesion are obtained, and a lesion coefficient of the target lesion is obtained according to the quantitative features, the learning features and the corresponding weights; The lesion type information of the target lesion is obtained according to the lesion coefficient and a coefficient threshold value.

10. A vocal cord lesion feature extraction apparatus characterized by comprising: The method includes: A segmentation module is configured to obtain a white light image and an NBI image of a larynx of a same case, call a trained vocal cord segmentation model to segment a vocal cord in the white light image and the NBI image respectively, obtain a white light vocal cord image and an NBI vocal cord image, and call a trained lesion segmentation model to segment a target lesion in the white light vocal cord image, and obtain a white light lesion image; The first obtaining module is configured to: obtain first position information, first shape information and first area information of a vocal cord reference region of the vocal cord from the white-light vocal cord image, obtain second position information and second area information of the target lesion from the white-light lesion image, obtain first relative position information of the target lesion and the vocal cord in a first direction and second relative position information of the target lesion and the vocal cord in a second direction according to the first position information, the first shape information and the second position information, obtain a first quantization feature according to the first relative position information, obtain a second quantization feature according to the second relative position information, obtain convexity information of the target lesion according to the first position information and the first shape information, and obtain a third quantization feature according to the convexity information; obtain relative area information of the target lesion and the vocal cord reference region according to the first position information, the first shape information, the first area information and the second area information, and obtain a fourth quantization feature according to the relative area information; obtain symmetry information of the vocal cord from the white-light vocal cord image, and obtain a fifth quantization feature according to the symmetry information; The second obtaining module is configured to: call a trained lesion shape recognition model to recognize a shape of the target lesion in the white-light lesion image, and obtain a first recognition result, and obtain a first learning feature according to the first recognition result; call a trained peduncle recognition model to recognize whether the target lesion in the white-light lesion image has a peduncle, obtain a second recognition result, and obtain a second learning feature according to the second recognition result; obtain color information of the target lesion from the white-light lesion image, and obtain a sixth quantization feature according to the color information; The third obtaining module is configured to: call a trained blood vessel segmentation model to perform blood vessel segmentation on the NBI vocal cord image, obtain morphological information of the blood vessel from the segmented NBI vocal cord image, and obtain a seventh quantization feature according to the morphological information; The fourth obtaining module is configured to: obtain vocal cord lesion features according to all quantization features and all learning features.

Images