Colposcope cervical lesion area automatic labeling method and system thereof

Abstract

This invention belongs to the field of medical image processing technology and discloses an automatic annotation method and system for cervical lesion areas in colposcopy. The method includes: converting the cervical RGB image to the CIELAB color space to construct an acetic acid sensitivity index map to enhance color difference contrast; inputting the sensitivity index map and the original image into a dual-stream encoding and decoding segmentation network to accurately segment the conversion area; extracting three complementary features within the conversion area: acetic acid reaction features, vascular morphology features, and iodine staining spatial features; outputting four probabilities (normal, CIN1, CIN2, and CIN3) by a multi-branch hierarchical classifier and adjusting the segmentation threshold accordingly; and marking each level of area with different colored contours and indicating the biopsy location with arrows.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, specifically relating to a method and system for automatically annotating cervical lesion areas in colposcopy. Background Technology

[0002] Cervical cancer is a major public health problem threatening women's health worldwide. Statistics from the International Agency for Research on Cancer show that more than 600,000 new cases of cervical cancer are diagnosed annually, with over 340,000 deaths. Persistent high-risk human papillomavirus (HPV) infection is a core driving factor in the development of cervical cancer, and early detection and treatment of cervical intraepithelial neoplasia (CIN) are crucial in halting its progression. In clinical screening systems, colposcopy is recommended as a standard diagnostic tool when cytology or HPV testing results are abnormal. Colposcopy involves applying a 3% to 5% acetic acid solution to the cervical surface. Utilizing the principle of reversible coagulation between acetic acid and nuclear proteins in epithelial cells, abnormally proliferating areas exhibit a white discoloration, known as the acetowhite epithelial reaction. Experienced colposcopy physicians comprehensively assess the severity of the lesion and guide the biopsy location based on multiple characteristics, including the intensity, speed, and boundary morphology of the acetowhite reaction, as well as iodine staining and vascular abnormalities.

[0003] However, colposcopy is highly dependent on the operator's experience, and there are significant differences in diagnostic consistency among different physicians. Studies have shown that even for systematically trained colposcopy physicians, the sensitivity for detecting cervical intraepithelial neoplasia fluctuates between 81% and 96%, while the specificity is only 34% to 69%. Inter-observer consistency is considered the main limitation of this examination method. For primary healthcare institutions, the lack of experienced colposcopy specialists is particularly prominent. When faced with a smooth transition between the acetic acid test and the normal mucosa color of the cervical bleaching area, primary care gynecologists often struggle to accurately delineate the boundaries of the lesion area, easily leading to misplacement of biopsy samples and increasing the risk of missing high-grade lesions.

[0004] To address the aforementioned clinical challenges, researchers both domestically and internationally have conducted studies on colposcopy image-assisted diagnosis based on deep learning. For example, patent application CN108090906A discloses a cervical image processing method and device based on region nomination. This scheme uses a deep residual network ResNet50 combined with a feature pyramid network as the feature extraction backbone, locates the classification and location information of the target region through a region detection network, and designs a region screening classification network to reduce the false positive rate. This scheme has achieved a certain degree of differentiation between normal acetic acid white and lesion acetic acid white in cervical images. However, the above-mentioned technical solution has the following shortcomings: First, this scheme operates directly on the RGB color space without performing special color space enhancement processing for the color features of the acetic acid white reaction. When the color difference contrast between the lesion area and the normal mucosa is low, the feature extraction network has difficulty effectively capturing subtle differences in acetic acid white. Second, this scheme does not involve the precise segmentation of the transformation zone, which is a common site for cervical cancer, and its accurate definition is fundamental for the subsequent localization of lesion areas. Third, the scheme only outputs classification and location information through a single network, without comprehensively utilizing the dynamic characteristics of the acetic acid whitening reaction, vascular morphology characteristics, and iodine staining characteristics for multi-dimensional joint analysis, thus the grading accuracy needs improvement. Fourth, the scheme lacks precise guidance for biopsy sampling locations and fails to directly instruct physicians on the highest-level recommended biopsy area.

[0005] In summary, there is an urgent need for an automatic annotation method for colposcopy images that can enhance the color contrast between the acetic acid-white area and normal mucosa, accurately segment the transformation zone, comprehensively extract multi-dimensional lesion features, achieve four-level fine grading, and provide intelligent guidance for biopsy location. Summary of the Invention

[0006] To address the technical problems of insufficient color contrast enhancement in the acetic acid white region, missing transformation region segmentation, inadequate utilization of multi-dimensional features, and lack of precise biopsy location guidance in existing colposcopy image processing solutions, this invention provides an automatic annotation method and system for cervical lesion areas in colposcopy.

[0007] The first aspect of this invention provides a method for automatically marking cervical lesion areas using colposcopy, comprising the following steps:

[0008] Step S1, Construction of Acetic Acid Sensitivity Index Map: Obtain the cervical RGB image acquired by colposcopy after the acetic acid test, convert the cervical RGB image from the RGB color space to the CIELAB color space to obtain the L channel image and the A channel image; calculate the acetic acid sensitivity index pixel by pixel based on the brightness value of the L channel image and the red-green hue value of the A channel image to generate the acetic acid sensitivity index map, which enhances the color contrast between the lesion area and the normal mucosa.

[0009] Step S2, Dual-stream segmentation of the conversion region: The acetic acid sensitivity index map and the cervical RGB image are respectively used as two inputs and fed into the dual-stream encoder-decoder segmentation network. The multi-scale features of the two encoders are interactively fused by the cross-stream attention fusion module and then fed into the shared decoder to output the conversion region segmentation mask.

[0010] Step S3, multi-feature joint extraction: Within the transformation region defined by the transformation region segmentation mask, three types of lesion features are extracted, including acetic acid whitening reaction features, vascular morphology features, and iodine staining spatial features.

[0011] Step S4, lesion grading and classification: After fusing the features of the three types of lesions, input them into the multi-branch grading classifier, output the probability values ​​of each sub-region belonging to normal, CIN1, CIN2 and CIN3 respectively, and feed the probability values ​​back to adjust the segmentation threshold of the dual-stream encoding and decoding segmentation network in step S2.

[0012] Step S5, Visualization and Biopsy Guidance: The boundaries of each level of region are marked on the cervical RGB image with different colored outlines, and the center of the highest level region recommended for biopsy is indicated by an arrow.

[0013] The second aspect of the present invention provides an automatic labeling system for cervical lesions in colposcopy, including an acetic acid sensitivity index map construction module, a dual-flow segmentation module for the transformation zone, a multi-feature joint extraction module, a lesion grading and classification module, and a visualization labeling and biopsy guidance module, each module corresponding to each step in the above method.

[0014] Compared with existing technologies, the beneficial effects of this invention are as follows: First, by converting the cervical RGB image to the CIELAB color space and constructing an acetic acid sensitivity index map based on the ratio of the L channel to the A channel, the originally slight color difference between the lesion area and normal mucosa is significantly enhanced, effectively solving the problem of boundary recognition difficulties caused by the smooth color transition between the acetic acid region and normal mucosa. Second, the encoding and decoding segmentation network using a dual-stream input of the acetic acid sensitivity index map and the original image accurately segments the transformation area, taking into account both enhanced color difference information and original texture information, thus improving the accuracy of transformation area segmentation. Third, by comprehensively extracting three complementary features within the transformation area—acetic acid reaction features, green filter channel vascular morphology features, and iodine test spatial overlap features—multi-dimensional joint analysis is achieved, significantly improving the precision of lesion grading. Fourth, the classification probability value feedback of the multi-branch grading classifier adjusts the segmentation threshold of the dual-stream segmentation network, forming a closed-loop collaborative optimization mechanism, further improving the overall annotation accuracy of the system. Fifth, by overlaying different colored contours to annotate each level of region and indicating the biopsy sampling location of the highest-level region with arrows, intuitive and accurate biopsy guidance is provided for primary care physicians. Attached Figure Description

[0015] Figure 1This is a flowchart illustrating the automatic labeling method for cervical lesion areas via colposcopy provided in this embodiment of the invention.

[0016] Figure 2 This is a schematic diagram of the architecture of the colposcopic cervical lesion area automatic labeling system provided in an embodiment of the present invention. Detailed Implementation

[0017] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate understanding of the present invention and are not intended to limit it in any way.

[0018] like Figure 1 As shown in the figure, the automatic labeling method for cervical lesions via colposcopy provided in this embodiment of the invention includes five core steps, from S1 to S5. These steps form a deeply coupled forward data flow relationship, and the probability value from step S4 is fed back to step S2, constituting a closed-loop collaborative optimization mechanism. Each step is described in detail below.

[0019] Step S1: Construction of the Acetic Acid-White Sensitivity Index Map. The technical objective of this step is to convert the cervical RGB image acquired by colposcopy to a color space more suitable for describing the color changes of the acetic acid-white reaction, and on this basis, construct an acetic acid-white sensitivity index map that can significantly enhance the color contrast between the lesion area and the normal mucosa. This sensitivity index map will serve as a key input to the dual-stream encoding and decoding segmentation network in the subsequent step S2, and also as the basic data source for extracting acetic acid-white reaction features in step S3.

[0020] During clinical colposcopy, the operator first applies a 3% to 5% acetic acid solution to the cervical surface. Acetic acid undergoes a reversible coagulation reaction with nucleoproteins in the epithelial cells, causing abnormally proliferating areas with a high nucleocytoplasmic ratio to exhibit varying degrees of whitening. Preferably, the cervical RGB image obtained by this invention is acquired within 60 to 90 seconds after the application of acetic acid by the colposcope. Within this time window, the acetic whitening reaction is fully manifested and has not yet significantly subsided, thus reflecting the color characteristics of the lesion area well. In one embodiment of this invention, the colposcope's imaging resolution is not less than 1920 × 1080 pixels, and the ambient light source is kept constant during image acquisition to reduce interference from changes in external lighting on subsequent color space conversion.

[0021] After acquiring the cervical RGB image, this step first performs a conversion from the RGB color space to the CIELAB color space. It should be noted that the CIELAB color space is a device-independent color model defined by the International Commission on Illumination (ICI). Its L channel represents luminance information, with values ​​ranging from 0 to 100; the A channel represents chromaticity changes from green to red, with values ​​ranging from -128 to 127; and the B channel represents chromaticity changes from blue to yellow, with values ​​ranging from -128 to 127. The core reason for choosing the CIELAB color space in this invention is that, compared to the RGB color space, the CIELAB color space better matches the color perception characteristics of the human eye. In this space, two colors with the same numerical difference appear to have approximately equal color difference visually; this characteristic is called perceptual uniformity. More importantly, for the acetic acid reaction, the lesion area, after being treated with acetic acid, mainly exhibits a comprehensive color change characterized by increased luminance and a shift in red and green hues in a specific direction. This change is more concentrated and significant in the combined representation of the L and A channels than in the RGB three channels.

[0022] In the specific color space conversion process, this embodiment first normalizes the pixel values ​​of the RGB image from the integer range of 0 to 255 to the floating-point range of 0 to 1. Then, according to the standard conversion process recommended by the International Commission on Illumination (ICI), it sequentially goes through three stages: linearization, matrix transformation to the CIEXYZ color space, and conversion from the CIEXYZ color space to the CIELAB color space. Preferably, the reference white point used in the conversion process is the D65 standard light source, whose tristimulus values ​​are as follows: , , .

[0023] After completing the CIELAB color space conversion, the core innovation of this step lies in constructing an acetic acid sensitivity index map using the ratio of the L channel to the A channel. Specifically, for the coordinates in the CIELAB color space... The pixel at that location has an acetic acid sensitivity index. Calculate using the following formula:

[0024] ,

[0025] in: This is the brightness value of the pixel in the L channel, ranging from 0 to 100, and is dimensionless. This is the red-green hue value of the pixel in the A channel, ranging from -128 to 127, and is dimensionless. To prevent the smoothing constant from having a denominator of zero, it is set to 1.0 in this embodiment. Its function is to prevent numerical overflow when the value of channel A is close to zero. This is a brightness normalization factor used to eliminate differences in light source intensity between different colposcope devices. It is calculated by multiplying the reciprocal of the mean value of the L-channel within the local neighborhood of the pixel by the global mean value of the L-channel. ,in This represents the global mean of the L channel of the entire image. The size centered on this pixel is The mean value of the L channel within the local window, in this embodiment The value is 51 pixels, which was determined after a trade-off between ensuring local statistical stability and avoiding excessive smoothing. This is the chromaticity shift compensation amount, used to correct color response differences between different colposcope devices. In this embodiment... The value is 0.5, which was obtained by calibration on 500 cervical images acquired by different devices.

[0026] From a physical perspective, the design principle of the above formula is as follows: After the acetic acid treatment, the brightness (L) of the acetic acid-white region significantly increases, while the red-green hue (A) shifts positively with a relatively small amount of shift. Therefore, the ratio of the absolute value of the L channel to the absolute value of the A channel is significantly greater in the acetic acid-white region than in the normal mucosal region. This is achieved through a brightness normalization factor. Modulation can effectively eliminate the problem of uneven local brightness caused by different distances between the light source and the cervical surface. Preferably, after the calculation is completed, the acetic acid sensitivity index map is subjected to max-min normalization processing to map all pixel values ​​to the range of 0 to 1, so as to facilitate subsequent network input. In this embodiment, experimental verification shows that on a test set containing 3200 clinical colposcopy images, after using the acetic acid sensitivity index map constructed by this invention, the average color difference contrast between the lesion area and the normal mucosa area is improved by about 2.8 times, which is significantly better than the scheme of directly using RGB channel difference or HSV color space V channel for enhancement.

[0027] It is worth noting that the acetic acid whitening sensitivity index map is not a simple single-channel image transformation, but a composite representation that integrates luminance and chromaticity information. Under this representation, areas with strong acetic acid whitening reaction exhibit high sensitivity index values, areas with weak acetic acid whitening reaction exhibit medium sensitivity index values, and normal mucosal areas exhibit low sensitivity index values. The numerical gradient distribution among the three is clear, laying a good data foundation for subsequent transformation region segmentation and feature extraction.

[0028] Step S2, Dual-stream segmentation of the transformation zone. The technical objective of this step is to accurately segment the extent of the cervical transformation zone. The transformation zone is the area where columnar epithelium transforms into squamous epithelium through metaplasia, and it is also a common site for cervical intraepithelial neoplasia and cervical cancer. Its accurate definition is fundamental for the subsequent localization and grading of lesions. This step receives the acetic acid sensitivity index map output from step S1 and the original cervical RGB image as dual-stream inputs, and achieves accurate segmentation of the transformation zone through a specially designed dual-stream encoding and decoding segmentation network.

[0029] The core consideration in designing the dual-input architecture of this invention lies in the fact that while the acetic acid sensitivity index map can significantly enhance the color contrast between the acetic acid-whitened region and the normal mucosa, it inevitably loses some texture details and microvascular morphology information contained in the original RGB image during the color space conversion process. Therefore, by simultaneously inputting two complementary information sources into the network, the network can take into account both the enhanced color difference information and the original texture information, achieving information complementarity. This is a key innovation of this invention, distinguishing it from existing technologies that use only a single input.

[0030] Specifically, the dual-stream codec segmentation network in this embodiment adopts an encoder-decoder architecture, which includes three core components: a sensitive exponential stream encoder, a raw image stream encoder, and a shared decoder, as well as a cross-stream attention fusion module that connects the two encoders.

[0031] Sensitive exponential stream encoder receive size is The acetic acid sensitivity index graph is used as input, where and The height and width of the image (in pixels) are represented by these dimensions, and the number of channels is 1. The encoder employs a four-level downsampling structure composed of residual convolutional blocks. Each level contains two residual convolutional blocks, each consisting of two convolutional layers and one residual skip connection. The kernel size is 3×3, the stride is 1, and the padding is 1. Downsampling is achieved by performing a 2×2 max-pooling operation with a stride of 2 after each residual convolutional block. The feature map scales after the four levels of downsampling are, respectively, 1 / 2, 1 / 8, and 1 / 16 of the original input scale, corresponding to 64, 128, 256, and 512 feature channels, respectively. Preferably, each convolutional layer is followed by a batch normalization layer and a ReLU activation function.

[0032] The raw image stream encoder receives a size of 100 ms. The original cervical RGB image is used as input, and its structure is consistent with that of the sensitivity index stream encoder, also a four-level downsampling structure, with corresponding feature channel numbers of 64, 128, 256, and 512 respectively. It should be noted that the parameters of the two encoders are independent and do not share weights. This is because the acetic acid sensitivity index image and the original RGB image have different data distribution characteristics and semantic representation methods, and independent encoder parameters can better adapt to their respective input characteristics.

[0033] The cross-stream attention fusion module is the core innovative component of the dual-stream segmentation network of this invention. It performs interactive fusion of the two-channel features at each level of downsampling features. Specifically, for the first... class( The fusion process of the sensitive exponential stream encoder is described by the following: Let the feature map output by the sensitive exponential stream encoder be... The feature map output by the original image stream encoder is Both have the same spatial dimensions and the same number of channels. Cross-stream attention fusion first performs a channel attention weighting operation: ... and By concatenating along the channel dimension, the number of channels is obtained as follows: The concatenated feature maps are then used to generate channel attention weight vectors through a global average pooling layer and two fully connected networks. The output dimension of the first fully connected layer is ( The channel compression ratio is 8 in this embodiment, and the activation function is ReLU; the output dimension of the second fully connected network is... The activation function is Sigmoid. The channel attention weight vector... After multiplying the concatenated feature map channel by channel, the number of channels is reduced to a smaller dimension using a 1×1 convolution. Channel attention fusion features are obtained. .

[0034] Based on channel attention weighting, a spatial attention selection operation is further performed: for The average and maximum values ​​are calculated along the channel dimension to obtain two spatial descriptor maps. These two spatial descriptor maps are then concatenated along the channel dimension and passed through a 7×7 convolutional layer and a sigmoid activation function to generate a spatial attention weight map. ,in and The first The height and width of the level feature map. The final cross-stream fusion feature. Obtained by multiplying the spatial attention weight map and the channel attention fusion feature pixel by pixel:

[0035] ,

[0036] in: This represents the element-wise Hadamard product operation. Through the sequential cascaded operation of channel attention and spatial attention described above, the cross-stream attention fusion module can adaptively select the feature information that is most discriminative for the transformation region segmentation from the two inputs, suppressing noise and interference from irrelevant regions.

[0037] Shared decoder receives four-level cross-stream fusion features to A bottom-up, stepwise upsampling structure is employed to restore spatial resolution. Each upsampling step includes 2x bilinear interpolation upsampling, skip connections for fusing features with the corresponding level, and two 3×3 convolutional layers. The decoder's final output is then processed by a 1×1 convolution and a sigmoid activation function to generate a transformation segmentation mask of the same size as the original input. Each pixel value represents the probability that the location belongs to the transformation region, and the value ranges from 0 to 1. In this embodiment, the default segmentation threshold is... The value is set to 0.5, meaning that pixels with a probability value greater than 0.5 are considered to be pixels within the conversion zone.

[0038] In one embodiment of the present invention, the training of the dual-stream codec segmentation network uses a weighted combination of binary cross-entropy loss and Dice loss as the loss function. Let the mask of the true transformation region of the training image be... The predicted mask is Then the comprehensive loss function Defined as:

[0039] ,

[0040] in: This is a binary cross-entropy loss used for pixel-by-pixel classification optimization; is the Dice coefficient, which ranges from 0 to 1 and is used to measure the overall overlap between the predicted mask and the real mask. and The weighting coefficients for the two losses are shown in this embodiment. , This combination achieves a balance between pixel-level classification accuracy and region-level overlap optimization. The Adam optimizer is used during training, with an initial learning rate set to [value missing]. The batch size is 8, and a total of 200 epochs are trained.

[0041] Step S3, Multi-feature Joint Extraction. The technical objective of this step is to extract three complementary lesion features within the transformation region defined by the transformation region segmentation mask output in step S2, providing multi-dimensional feature input for grading and classification in step S4. These three types of features characterize cervical lesions from three dimensions: color change in the acetic acid whitening reaction, vascular morphology under green filter light, and color difference in the iodine test. The complementarity of information among these features makes the grading accuracy after joint use far exceed that of any single feature type.

[0042] Before starting feature extraction, this step first utilizes a transformation region segmentation mask. Define the region of interest in the image. Specifically, define the segmentation mask. After binarization (based on a threshold) After defining the boundary, connected component analysis is used to extract the largest connected region as the final transformation region. Preferably, morphological closing operations (structuring element is a 5×5 circle) are performed on the binarized mask to fill small holes, and morphological opening operations (structuring element is a 3×3 circle) are performed to remove isolated noise points. Within the determined transformation region, the transformation region is further divided into multiple sub-regions to support subsequent region-by-region hierarchical processing. In one embodiment of the present invention, the sub-region division adopts a superpixel-based method, specifically using a simple linear iterative clustering algorithm to cluster the pixels in the transformation region into several spatially compact and color-uniform superpixel blocks. The number of superpixels is adaptively determined according to the area of ​​the transformation region, preferably the average area of ​​each superpixel is 1 / 50 to 1 / 30 of the total area of ​​the transformation region.

[0043] The first type of feature is the acetic acid whitening reaction feature. This feature is used to quantitatively characterize the degree and rate of color change in each sub-region within the conversion zone under the action of acetic acid. Its extraction is based on the acetic acid whitening sensitivity index map output in step S1.

[0044] Mean strength of acetic acid whitening reaction It is obtained by averaging the pixel values ​​of each sub-region on the acetic acid sensitivity index map. Let the first... Each sub-region contains If there are 10 pixels, the average acetic acid whitening reaction intensity of the sub-region is calculated as follows:

[0045] ,

[0046] in: Indicates the first The set of coordinates of all pixels within a sub-region; This is the acetowhite sensitivity index value calculated in step S1 for this pixel. This index reflects the overall degree of acetowhite reaction in the sub-region; higher-grade lesions typically exhibit a higher average acetowhite reaction intensity.

[0047] The rate of whitening of acetic acid The acetic acid sensitivity index is obtained by calculating the time required for the index to rise from baseline to peak value within a preset time window after acetic acid application. In clinical practice, the acetic acid-whitening area corresponding to high-grade cervical intraepithelial neoplasia usually shows a rapid and obvious white change shortly after acetic acid application, while the acetic acid-whitening reaction of low-grade lesions or normalized epithelium is slower and weaker. To obtain information on the speed of acetic acid appearance, this invention preferably acquires multiple frames of colposcopy images continuously at 5-second intervals within a preset time window of 10 to 120 seconds after acetic acid application. The acetic acid sensitivity index map construction in step S1 is performed on each frame of image to obtain a time-series sensitivity index map. Let time be... The first collection The average acetic acid whitening reaction intensity of each sub-region is The baseline value is (in (The moment before acetic acid application), peak value is The rate of acetic acid whitening is defined as follows:

[0048] ,

[0049] in: For the first The moment when the acetic acid sensitivity index of each sub-region first reaches its peak, measured in seconds; The unit is The larger the value, the faster the acetic acid whitening occurs. In this embodiment, when multiple time-series images cannot be obtained (for example, when only a single frame of the image after acetic acid application can be obtained), the acetic acid whitening rate index can be estimated using an alternative method, that is, by converting the average acetic acid whitening reaction intensity in a single frame image with a preset empirical mapping function, but the accuracy will be reduced.

[0050] The second type of feature: vascular morphology. Vascular morphology is used to quantitatively characterize the density of vascular branches and the regularity of vascular course in each subregion within the transformation zone. In colposcopy, abnormal vascular morphology (such as coarse mosaic, coarse punctate vessels, and atypical vessels) is an important basis for assessing the lesion grade. The green filter channel, by filtering out red wavelengths and enhancing the contrast of hemoglobin absorption wavelengths, makes vascular morphology more clearly visible in the image.

[0051] This step first extracts the green channel image from the original cervical RGB image. Morphological vessel enhancement processing is performed within the transformation region. Specifically, this embodiment employs a multi-scale Frangi vessel filter to enhance the green channel image. This filter detects tubular structures by analyzing the Hessian matrix eigenvalues ​​of the image. The Frangi vessel filter enhances the green channel image at various scales. The response value below Calculate using the following formula:

[0052] ,

[0053] in: and The eigenvalues ​​of the Hessian matrix satisfy the following conditions: ; The eccentricity ratio is used to distinguish between tubular and patchy structures. It is a structural significance index used to suppress background noise; and These are the eccentricity sensitivity parameter and the structural salience threshold parameter, respectively, in this embodiment. The value is 0.5. A value of 15 indicates that this parameter combination has good sensitivity for detecting tubular structures in cervical vascular enhancement scenarios. For filter scale parameters, the multi-scale range used in this embodiment is: The final vascular enhancement map is calculated using the maximum response value across all scales.

[0054] After binarizing the enhanced vascular image (preferably using the Otsu adaptive thresholding method), skeletonization extraction is performed to obtain a vascular skeleton image with a width of one pixel. (Vascular branch density) Defined as the first The number of branch nodes of the vascular skeleton per unit area within a subregion is calculated using the following formula:

[0055] ,

[0056] in: For the first The number of branch nodes of the vascular skeleton in each sub-region. A branch node is a node with 3 or more adjacent pixels in the skeleton map. For the first The area of ​​each sub-region, in units of (The pixel area is converted to physical area using the scale calibration factor of the colposcopy system; in this embodiment, the calibration factor is approximately 0.01 mm per pixel.) Higher vascular branch density is generally associated with higher-grade lesions.

[0057] Regularity of blood vessel course It is obtained by calculating the standard deviation of the orientation angle of each segment in the vascular skeleton. Specifically, firstly, the first... The vascular skeleton within each sub-region is segmented into multiple independent vascular segments. For each vascular segment, the local orientation angle is calculated along its course at intervals of 5 pixels. (The angle relative to the horizontal direction, the range of values ​​is) to Then, the standard deviation of all local orientation angles within this sub-region is calculated as an indicator of the regularity of vascular course:

[0058] ,

[0059] in: For the first The total number of local orientation angles calculated across all vessel segments within each sub-region; This represents the standard deviation calculation. The smaller the standard deviation of the orientation angle, the more regular the course of the blood vessels and the more consistent the direction, usually corresponding to normal or low-grade lesions; the larger the standard deviation of the orientation angle, the more irregular the course of the blood vessels and the more chaotic the direction, usually corresponding to high-grade lesions or atypical vascular morphology.

[0060] The third type of feature: spatial characteristics of iodine staining. Spatial characteristics of iodine staining are used to quantitatively characterize the spatial correspondence between areas unstained by the iodine test and areas whitened by acetic acid. In colposcopy, an iodine test is usually performed after the acetic acid test, involving the application of compound iodine solution (Lugol iodine solution) to the cervical surface. Normal mature squamous epithelium is rich in glycogen and stains brownish-red after iodine application; however, cervical intraepithelial neoplasia areas, due to glycogen deficiency, appear unstained or lightly stained after iodine application, i.e., areas with a negative iodine test. The degree of spatial overlap between unstained areas and whitened areas has important auxiliary reference value for lesion grading.

[0061] This step involves acquiring a cervical image after the iodine test and extracting a mask of the unstained areas from the iodine test using a color thresholding segmentation method. Specifically, the cervical RGB image after the iodine test is converted to the HSV color space, and threshold ranges are set for the hue (H) channel, saturation (S) channel, and lightness (V) channel to segment colored and uncolored areas. Preferably, the segmentation threshold for uncolored areas is defined as follows: an H channel value between 20 and 60, an S channel value greater than 30, and a V channel value greater than 80 are considered colored areas; otherwise, they are considered uncolored areas. This threshold range is determined based on statistical analysis of the typical color distribution after iodine staining.

[0062] Spatial overlap between the unstained area and the acetic acid white area in the iodine test Calculate using the following formula:

[0063] ,

[0064] in: The acetic white region mask is obtained by binarizing the acetic white sensitivity index map in step S1 (using the global mean plus one standard deviation as the binarization threshold). For the first The set of pixels in each sub-region; Indicates the number of pixels in the set; This represents the intersection operation of sets; To prevent the smoothing constant from having a denominator of zero, a value of 1.0 is used. The overlap value ranges from 0 to 1. A larger value indicates a higher spatial consistency between the unstained area and the acetic acid-white area in the iodine test, which usually corresponds to a higher grade of lesion.

[0065] In summary, for each sub-region within the transformation region, step S3 ultimately outputs a five-dimensional feature vector. This feature vector will serve as the input to the grading classifier in step S4. The three types of features comprehensively characterize multiple aspects of the lesion from three complementary dimensions: color response, vascular morphology, and iodine staining spatial overlap. Their combined use can produce a grading discrimination effect that is significantly better than any single feature type, demonstrating the nonlinear gain advantage of multi-feature synergy.

[0066] Step S4, lesion grading and classification. The technical objective of this step is to fuse the multi-dimensional lesion features extracted in step S3 and feed them into the grading classifier, outputting probability values ​​for each sub-region within the transformation zone belonging to four levels: normal, CIN1, CIN2, and CIN3. Simultaneously, this step also implements a feedback adjustment mechanism for the probability values ​​to the dual-stream segmentation network in step S2, forming a closed-loop collaborative optimization between lesion grading and transformation zone segmentation.

[0067] In one embodiment of the present invention, the multi-branch hierarchical classifier includes two core components: a feature fusion layer and a four-branch probability output layer.

[0068] The purpose of the feature fusion layer is to combine the five-dimensional feature vector output from step S3. Effective fusion is achieved to generate a more discriminative fused feature representation. In this embodiment, the feature fusion layer employs a strategy combining channel concatenation and attention weighting. First, the five-dimensional feature vectors of each sub-region are concatenated along the channel dimension. Then, a fully connected network maps the concatenated features to a high-dimensional space. This fully connected network contains two hidden layers with 64 and 128 neurons respectively, and the activation function is ReLU. Preferably, a feature attention module is introduced after the fully connected network. This module generates a 128-dimensional attention weight vector through global average pooling and two fully connected network layers, which is used to adaptively assign different importance weights to features of different dimensions. The attention-weighted 128-dimensional feature vector is the fused feature vector. .

[0069] The four-branch probability output layer comprises four parallel fully connected subnetworks, corresponding to the four levels: normal, CIN1, CIN2, and CIN3. Each fully connected subnetwork contains one hidden layer (64 neurons, ReLU activation function) and one output layer (1 neuron, Sigmoid activation function). The output values ​​of the four subnetworks are normalized using Softmax to obtain the probability values ​​of each subregion belonging to the four levels. The sum of all probability values ​​is 1.

[0070] During training, the multi-branch hierarchical classifier uses weighted cross-entropy loss as the loss function. Let the first branch be the first branch. The true label of each sub-region is (corresponding to normal, CIN1, CIN2, and CIN3 respectively), then the loss function is:

[0071] ,

[0072] in: This represents the total number of sub-regions within the transformation zone; For the first The weight coefficients for each category are used to address the issue of uneven distribution of lesions at different levels in the training samples. In this embodiment... , , , The weighting increases with the lesion grade to increase attention to higher-grade lesions; This is the indicator function. The Adam optimizer is used during training, with an initial learning rate of... The cosine annealing learning rate scheduling strategy was adopted, and a total of 150 epochs were trained.

[0073] A key technical feature of this invention is that the probability value in step S4 can be used to adjust the segmentation threshold of the dual-stream codec segmentation network in step S2. This creates a closed-loop collaborative optimization between hierarchical classification and conversion zone segmentation. The core idea of ​​this closed-loop mechanism is: if the confidence level of the probability value output by the hierarchical classifier for the sub-region at the edge of the conversion zone is low, it indicates that the current conversion zone segmentation may have boundary offset or omissions, and in this case, it is necessary to appropriately expand the segmentation range of the conversion zone.

[0074] Specifically, the closed-loop feedback adjustment is performed according to the following process: First, after the initial classification is completed in step S4, the highest probability value output of each sub-region within the edge region of the conversion area (defined as a set of sub-regions no more than 20 pixels away from the boundary of the conversion area mask) is statistically analyzed. If there are sub-regions within the edge region. Below the preset confidence threshold (In this embodiment) If this occurs, a feedback adjustment mechanism is triggered. The adjustment method is as follows: the segmentation threshold in step S2 is adjusted. Decrease by a preset step size (In this embodiment) This process involves incorporating more edge pixels into the transformation region, and then re-executing steps S2 to S4. This iterative process continues until either of the following termination conditions is met: the highest probability value of all edge sub-regions is not less than... or segmentation threshold It has dropped to the preset minimum value. Or the preset maximum number of iterations has been reached. Through this closed-loop mechanism, the system can achieve a balance between segmentation accuracy and classification confidence, avoiding lesion omissions or mislabeling caused by inaccurate delineation of transformation zone boundaries.

[0075] Step S5, Visual Annotation and Biopsy Guidance. The technical purpose of this step is to overlay the grading results output from Step S4 onto the original colposcopy image in an intuitive visual form, and to guide the physician in indicating the optimal location for biopsy sampling. This step is the final output of the entire method, and its annotation results are directly intended for primary care gynecologists; therefore, the intuitiveness and accuracy of the annotation are crucial.

[0076] First, the lesion level of each sub-region is determined based on the probability values ​​of each sub-region output in step S4. In one embodiment of the present invention, the level determination rule is: the level with the highest probability value is taken as the determination level of that sub-region, i.e. ,in normal Preferably, when the highest probability value is lower than a preset effective judgment threshold... In such cases, the sub-region is marked as uncertain to alert physicians that extra attention is needed.

[0077] Then, the boundaries of each level of region are superimposed on the original cervical RGB image using different colored outlines. The color coding scheme in this embodiment is as follows: normal regions are marked with green (RGB values ​​0, 200, 0); CIN1 regions are marked with yellow (RGB values ​​255, 200, 0); CIN2 regions are marked with orange (RGB values ​​255, 128, 0); and CIN3 regions are marked with red (RGB values ​​255, 0, 0). The outlines are drawn by merging adjacent sub-regions of the same level, extracting the outer contour curve of the merged region, and drawing it on the original image with a line width of 2 pixels. Preferably, a semi-transparent fill of the same color is superimposed inside the outline, with a fill transparency of 0.2, so as to visually display the distribution range of each level of region while maintaining the visibility of the original image texture.

[0078] Finally, an arrow indicates the center location of the highest-level region recommended for biopsy. Specifically, among all sub-regions determined to be lesion-level, the set of sub-regions with the highest CIN level is selected; if multiple sub-regions have the same highest level, the one with the highest probability value is selected. The arrow indicates the centroid coordinates of that sub-region. The centroid coordinates are calculated as follows: let the set of pixel coordinates contained in the sub-region be... Then the x-coordinate of the centroid centroid ordinate The arrow markers are red, 30 pixels long, and point towards the centroid. Preferably, text labels are added next to the arrows, displaying the grading result of the region and the corresponding probability value, such as CIN3 (95.2%).

[0079] In one embodiment of the present invention, when no lesion-level sub-regions are detected within the transformation zone (i.e., all sub-regions are judged to be normal), the system only displays the boundary outline of the transformation zone and green fill markings on the output image, does not generate arrow markings, and outputs a prompt message that no cervical intraepithelial neoplasia has been detected, and recommends routine follow-up.

[0080] like Figure 2 As shown, this embodiment of the invention also provides an automatic labeling system for cervical lesions using colposcopy. This system includes five functional modules, each of which corresponds one-to-one with steps S1 to S5 in the above method embodiment. The specific description of the corresponding steps in the method embodiment can be referred to in terms of the functional implementation.

[0081] The acetic acid sensitivity index map construction module is configured to communicate with a colposcopy image acquisition device to receive cervical RGB images acquired by the colposcope after an acetic acid test. This module internally includes a color space conversion unit and a sensitivity index calculation unit. The color space conversion unit converts the input RGB image to the CIELAB color space, outputting L-channel and A-channel images. The sensitivity index calculation unit calculates the acetic acid sensitivity index pixel-by-pixel according to the acetic acid sensitivity index calculation formula described in step S1 of the method embodiment and generates an acetic acid sensitivity index map. Preferably, this module also includes an image preprocessing unit, used to perform preprocessing operations such as highlight reflection point detection and repair, white balance correction, and resolution unification adjustment on the input image before color space conversion, to improve the robustness of subsequent processing.

[0082] The dual-stream segmentation module for the transformation zone is configured to receive the acetic acid sensitivity index map output by the acetic acid sensitivity index map construction module and the original cervical RGB image, and feed them as dual-stream inputs into the dual-stream encoding and decoding segmentation network. This module internally deploys a complete dual-stream encoding and decoding segmentation network including a sensitivity index stream encoder, an original image stream encoder, a cross-stream attention fusion module, and a shared decoder. Its network structure and parameter configuration are consistent with the description in step S2 of the method embodiment. This module outputs a transformation zone segmentation mask and receives a classification probability value feedback signal from the lesion grading and classification module, dynamically adjusting the segmentation threshold based on the feedback signal. In one embodiment of the present invention, the network model parameters of the module are obtained through offline training. After training, the parameters are stored in the system's memory in the form of a model weight file. During the inference phase, the pre-trained weights are directly loaded for forward computation.

[0083] The multi-feature joint extraction module is configured to extract three types of lesion features within the range defined by the transformation area segmentation mask output by the dual-stream segmentation module in the transformation area. This module internally includes a sub-region division unit, an acetic acid whitening feature extraction unit, a vascular feature extraction unit, and an iodine staining feature extraction unit. The sub-region division unit uses a simple linear iterative clustering algorithm to cluster pixels within the transformation area into several spatially compact superpixel sub-regions. The acetic acid whitening feature extraction unit calculates the mean acetic acid whitening reaction intensity and the acetic acid whitening appearance rate index of each sub-region based on the acetic acid whitening sensitivity index map. The vascular feature extraction unit performs multi-scale Frangi vascular enhancement and skeletonization processing on the green filter channel image, calculating the vascular branch density and vascular course regularity of each sub-region. The iodine staining feature extraction unit performs color threshold segmentation on the image after the iodine test, calculating the spatial overlap between the unstained iodine test area and the acetic acid whitening area. The algorithm details of each extraction unit are consistent with the description of step S3 in the method embodiment.

[0084] The lesion grading and classification module is configured to receive three types of lesion features output by the multi-feature joint extraction module. After feature fusion, these features are input into a multi-branch grading classifier, which outputs probability values ​​for each sub-region belonging to four levels: normal, CIN1, CIN2, and CIN3. This module internally deploys a complete multi-branch grading classifier containing a feature fusion layer and a four-branch probability output layer, and its network structure is consistent with the description in step S4 of the method embodiment. Furthermore, this module also includes a confidence assessment unit and a feedback signal generation unit. The confidence assessment unit statistically analyzes the classification probability values ​​of the sub-regions at the edge of the transformation zone, and the feedback signal generation unit generates a segmentation threshold adjustment signal based on the confidence assessment results and sends it to the dual-stream segmentation module of the transformation zone, achieving closed-loop collaborative optimization.

[0085] The Visual Annotation and Biopsy Guidance Module is configured to receive the probability values and level determination results of each sub-region output by the Lesion Grading and Classification Module, and perform visual annotation and biopsy position guidance on the original cervical RGB image. Inside this module, there are a level determination unit, a contour drawing unit, and a biopsy guidance unit. The level determination unit determines the lesion level of each sub-region based on the probability values. The contour drawing unit superimposes and annotates the boundaries of regions at each level on the original image with contours of different colors. The biopsy guidance unit determines the centroid position of the highest-level region for recommended biopsy sampling and marks it with an arrow. The output image of this module is presented to the operating physician through the display device of the colposcope system, providing an intuitive reference for biopsy sampling.

[0086] In an embodiment of the present invention, the above-mentioned modules can be integrally deployed on a computing terminal supporting the colposcope system. This computing terminal is preferably equipped with at least one graphics processor supporting CUDA computing to accelerate the inference calculation of the deep learning model. In the inference stage, from inputting a frame of cervical RGB image to outputting the annotated colposcope image, the end-to-end time consumption of the entire processing flow does not exceed 3s, meeting the timeliness requirements of clinical real-time auxiliary diagnosis. In addition, the modules communicate with each other through standardized data interfaces, supporting modular deployment and independent upgrade, facilitating subsequent function expansion and model iteration update. Preferably, this system can also be deployed on a cloud server. Primary medical institutions upload colposcope images to the cloud for processing through the network, and after processing, the annotation results are transmitted back to the local display terminal. This cloud deployment method can reduce the requirements for the hardware configuration of primary institutions and is conducive to popularization and application in areas with relatively scarce medical resources.

[0087] In an embodiment of the present invention, the data flow relationship between the modules is as follows: The acetic acid white sensitivity index map output by the acetic acid white sensitivity index map construction module and the original cervical RGB image are simultaneously transmitted to the transformation zone dual-stream segmentation module and the multi-feature joint extraction module; the transformation zone segmentation mask output by the transformation zone dual-stream segmentation module is transmitted to the multi-feature joint extraction module as a region of interest limitation condition; the three types of lesion feature vectors output by the multi-feature joint extraction module are transmitted to the lesion grading and classification module; the probability values and level determination results output by the lesion grading and classification module are transmitted to the visual annotation and biopsy guidance module, and at the same time, the classification probability value feedback signal is transmitted to the transformation zone dual-stream segmentation module. The above data flow relationship forms a closed-loop processing architecture with both forward transmission and feedback regulation, and the deep coupling between the modules ensures the optimization of the overall performance of the system.

[0088] To verify the technical effects of the automatic annotation method and system for colposcopic cervical lesion regions provided by the present invention, the inventors conducted a systematic performance evaluation experiment on a clinical dataset approved by the ethics review committee.

[0089] The dataset used in the experiment contained 3200 colposcopy images confirmed by histopathology, sourced from the gynecology outpatient departments of several tertiary-level hospitals. The dataset was distributed according to pathological diagnosis as follows: 1120 cases of normal or mild inflammation, 640 cases of CIN1 lesions, 720 cases of CIN2 lesions, and 720 cases of CIN3 lesions. Each image was accompanied by transformation zone boundaries and lesion region classification annotations jointly provided by three senior colposcopy specialists. The dataset was randomly divided into training, validation, and test sets in a 7:1:2 ratio.

[0090] On the test set, the dual-stream encoding / decoding segmentation network of this invention achieved a segmentation accuracy of 0.862 and an intersection-over-union (IoU) ratio of 0.791 for the transformed region segmentation task. In comparison, the standard U-Net network using only the original RGB image as a single-stream input has a Dice coefficient of 0.793, and the U-Net network using only the acetowhite sensitivity index map as a single-stream input has a Dice coefficient of 0.812. These comparisons show that the introduction of the acetowhite sensitivity index map and the design of the dual-stream fusion architecture improved the transformed region segmentation accuracy by approximately 2.4% and 6.3%, respectively, validating the effectiveness of multi-source complementary information fusion. Furthermore, after introducing a closed-loop feedback adjustment mechanism, the Dice coefficient for transformed region segmentation further increased to 0.878, indicating that dynamic threshold adjustment driven by classification probability values ​​can effectively improve the segmentation quality of edge regions.

[0091] On the test set, the multi-branch hierarchical classifier of this invention achieved an overall accuracy of 87.6% on a four-level classification task. The classification sensitivity and specificity for each level were as follows: normal level: sensitivity 91.2%, specificity 94.5%; CIN1 level: sensitivity 82.3%, specificity 90.8%; CIN2 level: sensitivity 85.7%, specificity 92.1%; CIN3 level: sensitivity 89.4%, specificity 93.6%. Compared with the scheme in CN108090906A that only distinguishes between normal and abnormal targets, this invention achieves a more refined four-level classification and has a significant advantage in the detection sensitivity of high-level lesions.

[0092] To verify the synergistic gain effect of using the three types of features in combination, the inventors conducted ablation experiments. Using only the acetic acid reaction feature, the overall accuracy of the four-level classification was 78.2%; using only the vascular morphology feature, it was 72.8%; and using only the iodine staining spatial feature, it was 68.5%. When used in pairs, the accuracy of acetic acid reaction plus vascular morphology features was 83.4%, the accuracy of acetic acid reaction plus iodine staining spatial features was 82.1%, and the accuracy of vascular morphology plus iodine staining spatial features was 79.6%. The accuracy of using all three types of features was 87.6%, significantly higher than any single feature or pairwise combination, demonstrating the nonlinear synergistic gain effect of multi-feature combination.

[0093] The biopsy location guidance function of this invention achieved a biopsy hit rate (i.e., the proportion of the arrow-indicated location within the highest-grade lesion area confirmed by histopathology) of 93.1% on the test set, which is superior to the average biopsy hit rate of 78.5% for primary care physicians without auxiliary annotation, representing an improvement of approximately 14.6 percentage points. This result indicates that the visual annotation and arrow guidance function provided by this invention can effectively compensate for the lack of colposcopy experience among primary care gynecologists, reducing the risk of missed diagnoses of high-grade lesions due to inaccurate delineation of lesion boundaries and deviation in biopsy sampling location.

[0094] To further verify the effect of the acetic acid sensitivity index map on enhancing color contrast, the inventors selected 500 difficult samples with smooth color transitions between lesion areas and normal mucosa for specific analysis. In the original RGB images, the average pixel intensity difference between the lesion areas and the surrounding normal mucosa in these difficult samples was only 12.3 gray values, making them almost visually indistinguishable. After conversion using the acetic acid sensitivity index map of this invention, the average sensitivity index difference between corresponding areas increased to 34.7, an increase of approximately 2.82 times. The lesion areas appeared as clearly high-value areas on the sensitivity index map, forming a clear gradient boundary with the surrounding normal mucosa. This result strongly demonstrates that the technical approach of constructing the acetic acid sensitivity index map based on the ratio of the L channel to the A channel in the CIELAB color space can fundamentally solve the core technical problem of smooth color transitions between acetic acid areas and normal mucosa.

[0095] System Real-Time Performance. On a standard computing terminal equipped with a single NVIDIA RTX 3060 graphics processor and an Intel Core i7 processor, the average end-to-end processing time for a single frame of cervical image with a resolution of 1920×1080 is 2.1 seconds. Specifically, step S1 takes approximately 0.08 seconds, step S2 approximately 0.92 seconds, step S3 approximately 0.35 seconds, step S4 approximately 0.15 seconds, step S5 approximately 0.06 seconds, and the closed-loop feedback iteration takes an additional approximately 0.54 seconds. This processing speed meets the real-time auxiliary diagnostic requirements of clinical colposcopy examinations and will not delay the physician's normal examination process. The experimental results show that the technical solution provided by this invention has achieved significant improvements in core performance indicators such as transformation region segmentation accuracy, lesion grading accuracy, and biopsy-guided hit rate.

[0096] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for automatic annotation of cervical lesion area in colposcopy, characterized in that, Includes the following steps: Step S1, Construction of Acetic Acid Sensitivity Index Map: Obtain the cervical RGB image acquired by colposcopy after the acetic acid test, convert the cervical RGB image from the RGB color space to the CIELAB color space to obtain the L channel image and the A channel image; calculate the acetic acid sensitivity index pixel by pixel based on the brightness value of the L channel image and the red-green hue value of the A channel image to generate the acetic acid sensitivity index map, which enhances the color contrast between the lesion area and the normal mucosa; Step S2, Dual-stream segmentation of the conversion zone: The acetic acid sensitivity index map and the cervical RGB image are respectively used as two inputs and sent to the dual-stream encoding and decoding segmentation network. The dual-stream encoding and decoding segmentation network includes a sensitivity index stream encoder and an original image stream encoder. The multi-scale features of the two encoders are processed by the cross-stream attention fusion module and then sent to the shared decoder to output the conversion zone segmentation mask. Step S3, multi-feature joint extraction: within the transformation area defined by the transformation area segmentation mask, three types of lesion features are extracted. The three types of lesion features include: acetic acid whitening reaction features extracted from the acetic acid whitening sensitivity index map, vascular morphology features extracted from the green filter channel image, and iodine staining spatial features extracted from the iodine test unstained area and the transformation area segmentation mask. Step S4, lesion grading and classification: After feature fusion of the three types of lesion features, the data is input into a multi-branch grading classifier. The multi-branch grading classifier outputs the probability values ​​of each sub-region in the transformation area belonging to normal, low-grade squamous intraepithelial lesion CIN1, high-grade squamous intraepithelial lesion CIN2, and high-grade squamous intraepithelial lesion CIN3, respectively. Step S5, Visualization and Biopsy Guidance: Determine the lesion level of each sub-region based on the probability value of each sub-region, mark the boundaries of each level region on the cervical RGB image with different colored outlines, and indicate the center position of the highest level region recommended for biopsy with an arrow mark, and output the marked colposcopy image.

2. The method for automatically marking cervical lesion areas using colposcopy according to claim 1, characterized in that, In step S1, the acetic acid sensitivity index is calculated as follows: in the CIELAB color space, based on the ratio of the L channel value to the A channel value, it is modulated by combining the luminance normalization factor and the chromaticity shift compensation amount. The luminance normalization factor is used to eliminate the influence of light source intensity fluctuations, and the chromaticity shift compensation amount is used to correct the color response differences between different colposcope devices.

3. The method for automatically marking cervical lesion areas using colposcopy according to claim 1, characterized in that, In step S2, the dual-stream encoding and decoding segmentation network adopts an encoder-decoder architecture. Both the sensitive exponential stream encoder and the original image stream encoder adopt a multi-level downsampling structure composed of residual convolutional blocks. The feature map scale after each downsampling level is 1 / 2, 1 / 4, 1 / 8 and 1 / 16 of the original input scale, respectively. The cross-stream attention fusion module performs channel attention weighting and spatial attention selection on each downsampled feature, and fuses the two features into a unified multi-scale feature representation.

4. The method for automatically marking cervical lesion areas using colposcopy according to claim 1, characterized in that, In step S3, the acetic whitening reaction characteristics include the average intensity of the acetic whitening reaction and the rate of acetic whitening appearance; the vascular morphology characteristics include the density of vascular branches and the regularity of vascular course; and the iodine staining spatial characteristics include the spatial overlap between the unstained area of ​​the iodine test and the acetic whitening area.

5. The method for automatically marking cervical lesion areas using colposcopy according to claim 4, characterized in that, In step S3, the average intensity of the acetic acid whitening reaction is obtained by averaging the pixel values ​​of each sub-region in the conversion zone on the acetic acid whitening sensitivity index map; the acetic acid whitening appearance rate index is obtained by calculating the time required for the acetic acid whitening sensitivity index to rise from the baseline value to the peak value within a preset time window after acetic acid application, wherein the preset time window is 10s to 120s after acetic acid application.

6. The method for automatically marking cervical lesion areas using colposcopy according to claim 4, characterized in that, In step S3, the vascular branch density is obtained by performing morphological vascular enhancement processing on the green filter channel image within the conversion area, extracting the skeleton of the enhanced vascular image, and counting the number of branch nodes of the vascular skeleton per unit area as the vascular branch density; the regularity of vascular course is obtained by calculating the standard deviation of the orientation angle of each segment in the vascular skeleton, and the smaller the standard deviation of the orientation angle, the more regular the vascular course.

7. The method for automatically marking cervical lesion areas using colposcopy according to claim 1, characterized in that, In step S4, the probability value feedback of the multi-branch hierarchical classifier adjusts the segmentation threshold of the dual-stream codec segmentation network in step S2. The multi-branch hierarchical classifier includes a feature fusion layer and a four-branch probability output layer. The feature fusion layer uses a combination of channel cascading and attention weighting to fuse the three types of lesion features to generate a fused feature vector. The four-branch probability output layer includes four parallel fully connected sub-networks that output the probability values ​​of normal, CIN1, CIN2, and CIN3, respectively. The sum of the probability values ​​is normalized to 1.

8. The method for automatically marking cervical lesion areas using colposcopy according to claim 7, characterized in that, In step S4, the probability value feedback adjusts the segmentation threshold of the dual-stream codec segmentation network in step S2 as follows: when the highest probability value output by the multi-branch hierarchical classifier for the edge sub-region of the conversion area is lower than the preset confidence threshold, the segmentation threshold of the dual-stream codec segmentation network is reduced to expand the segmentation range of the conversion area and steps S2 to S4 are re-executed until the highest probability value of the edge sub-region is not lower than the preset confidence threshold or the preset maximum number of iterations is reached.

9. The method for automatically marking cervical lesion areas using colposcopy according to claim 1, characterized in that, In step S5, the different colored outlines are set as follows: the normal area is marked with a green outline, the CIN1 area is marked with a yellow outline, the CIN2 area is marked with an orange outline, and the CIN3 area is marked with a red outline; the position indicated by the arrow mark is the centroid position of the sub-region with the highest probability value and the highest CIN level.

10. An automatic labeling system for cervical lesions using colposcopy, used to implement the automatic labeling method for cervical lesions using colposcopy as described in any one of claims 1-9, characterized in that, include: The acetic acid sensitivity index construction module is configured to acquire cervical RGB images collected by colposcopy after acetic acid test, convert the cervical RGB images from RGB color space to CIELAB color space to obtain L channel images and A channel images, calculate the acetic acid sensitivity index pixel by pixel based on the L channel images and the A channel images, and generate an acetic acid sensitivity index map that enhances the color difference contrast between the lesion area and the normal mucosa. The dual-stream segmentation module for the conversion zone is configured to send the acetic acid sensitivity index map and the cervical RGB image as two inputs into the dual-stream encoder-decoder segmentation network, and then send the multi-scale features of the two encoders to a shared decoder after interactive fusion of the multi-scale features of the two encoders through the cross-stream attention fusion module, and output the conversion zone segmentation mask. The multi-feature joint extraction module is configured to extract three types of lesion features, including acetic acid whitening reaction features, vascular morphology features, and iodine staining spatial features, within the transformation region defined by the transformation region segmentation mask. The lesion grading and classification module is configured to fuse the features of the three types of lesions and input them into a multi-branch grading classifier, output the probability values ​​of each sub-region belonging to normal, CIN1, CIN2 and CIN3 respectively, and feed the probability values ​​back to the dual-stream segmentation module of the transformation area to adjust the segmentation threshold. The visualization annotation and biopsy guidance module is configured to overlay the boundaries of each level of region on the cervical RGB image with different colored outlines, and to indicate the center position of the highest level region recommended for biopsy with an arrow mark.

Images