Skin fungus microscopic examination image identification method and system

Abstract

The invention provides a skin fungus microscopic examination image identification method and system, and the method comprises the steps: determining a high-frequency component and a low-frequency component of a fungus microscopic examination image, and reconstructing the fungus microscopic examination image into a denoised clear image; synchronously extracting optical features of spore edges during optical diffraction and texture features of spore surface textures for multi-channel feature fusion to obtain an enhanced texture feature map during fungal microscopic examination; according to a spatial topological relation of different feature units in the enhanced textural feature map in a preset neighborhood, generating discrimination features of a spore configuration, and further determining a confidence level of a spore target in the different feature units in the enhanced textural feature map through the discrimination features; and performing adaptive threshold segmentation on the enhanced texture feature map based on all confidence levels, and generating a confidence detection report of the fungal spores on the skin of the patient through a segmentation result. By adopting the scheme of the invention, the key identification features of the spores can be completely reserved while the image noise is inhibited.

Description

Technical Field

[0001] This application relates to the field of image recognition technology, and more specifically, to a method and system for image recognition of skin fungi under a microscope. Background Technology

[0002] Image recognition refers to the process of transforming pixel-level visual information in digital images into symbolic representations with semantic categories and structural relationships through image acquisition and preprocessing, feature analysis and encoding. This enables a leap from low-level visual features to high-level semantic understanding of target objects, ultimately providing an interpretable cognitive reasoning basis for detection, retrieval and decision-making.

[0003] Skin fungal microscopic image recognition refers to the automated analysis and interpretation of microscopic images of patient skin scales, nail scales, or hair samples through optical microscope image acquisition, digital image preprocessing, fungal morphological feature extraction, and intelligent classification algorithms. This transforms complex image information from a microscopic perspective into fungal species identification results with clinical diagnostic value, assisting laboratory personnel in achieving rapid and objective pathogen screening and identification. Traditional skin fungal microscopic image recognition methods typically employ a uniform image processing workflow for global denoising and feature extraction, which struggles to effectively resolve the contradiction between image noise suppression and preservation of spore microscopic features. Furthermore, the lack of modeling capabilities for spore spatial configuration and topological relationships leads to insufficient recognition accuracy in complex backgrounds. For example, traditional methods may use a single filter for global denoising of the microscopic image, which, while suppressing background noise, simultaneously weakens the optical diffraction characteristics and surface microtexture of the spore edges, making it difficult to distinguish between real spores and morphologically similar image artifacts in subsequent feature extraction stages. Therefore, how to suppress image noise while fully preserving the key identification features of spores has become a challenge for the industry. Summary of the Invention

[0004] This application provides a method and system for identifying skin fungi under a microscope, which can completely preserve the key identification features of spores while suppressing image noise.

[0005] In a first aspect, this application provides a method for identifying skin fungi under a microscope, comprising the following steps: Obtain microscopic images of fungi on the patient's skin; The high-frequency and low-frequency components of the fungal microscopic image are determined, and then the fungal microscopic image is reconstructed into a clear image after denoising based on the noise difference between the high-frequency and low-frequency components. Optical features of spore edge optical diffraction and texture features of spore surface texture are extracted simultaneously from the denoised clear image. The optical features and texture features are fused using multi-channel features to obtain an enhanced texture feature map for fungal microscopy. Based on the spatial topological relationship of different feature units in the enhanced texture feature map in a preset neighborhood, a discrimination feature for spore configuration is generated, and then the confidence level of the presence of spore targets in different feature units in the enhanced texture feature map is determined by the discrimination feature; The enhanced texture feature map is subjected to adaptive threshold segmentation based on all confidence levels, and a confidence detection report of fungal spores on the patient's skin is generated based on the segmentation results.

[0006] In some embodiments, determining the high-frequency and low-frequency components of the fungal microscopic image specifically includes: The frequency domain transformation of the fungal microscopic image is performed to obtain the frequency domain representation of the fungal microscopic image; The high-frequency and low-frequency components of the fungal microscopic image are extracted from the frequency domain representation.

[0007] In some embodiments, reconstructing the fungal microscopic image into a denoised, clear image based on the noise difference between the high-frequency and low-frequency components specifically includes: The noise difference is determined based on the noise variance between the high-frequency component and the low-frequency component; Adaptive threshold denoising is performed on the high-frequency components based on the noise differences. The denoised high-frequency components and the low-frequency components are then subjected to inverse frequency domain transformation to reconstruct a clear image after denoising.

[0008] In some embodiments, simultaneously extracting the optical features of spore edge optical diffraction and the texture features of spore surface texture from the denoised clear image specifically includes: The denoised clear image is subjected to edge enhancement processing based on gradient operators, and then the initial gradient features characterizing the optical diffraction of spore edges are extracted. Local binary mode analysis is performed on the denoised clear image to extract the initial texture features characterizing the surface texture of the spores; The initial gradient features and the initial texture features are respectively quantized to obtain the optical features of optical diffraction at the spore edge and the texture features of the spore surface texture.

[0009] In some embodiments, multi-channel feature fusion of the optical features and the texture features to obtain an enhanced texture feature map for fungal microscopy specifically includes: The optical features and the texture features are each treated as independent feature channels to create a dual-channel feature map. Normalization is performed on each feature channel in the dual-channel feature map; Cross-channel feature aggregation is performed on each normalized feature channel to generate the enhanced texture feature map for fungal microscopy.

[0010] In some embodiments, generating discriminative features for spore configurations based on the spatial topological relationships of different feature units in a preset neighborhood in the enhanced texture feature map specifically includes: The enhanced texture feature map is divided into multiple feature units, and a preset neighborhood is established for each feature unit; Extract the spatial topological relationship between each feature unit and other feature units in its preset neighborhood; Based on the aforementioned spatial topological relationships, discriminative features characterizing spore configurations are constructed.

[0011] In some embodiments, adaptive threshold segmentation of the enhanced texture feature map based on all confidence levels, and then generating a confidence detection report of fungal spores on the patient's skin based on the segmentation results, specifically includes: The adaptive segmentation threshold is determined based on the confidence level distribution characteristics of all feature units; The enhanced texture feature map is binarized based on the adaptive segmentation threshold to obtain the spore target region; A confidence test report for fungal spores is generated based on the target spore region and its corresponding confidence level.

[0012] Secondly, this application provides a method for identifying skin fungal microscopic images, including: The acquisition module is used to acquire microscopic images of fungi on the patient's skin; The processing module is used to determine the high-frequency and low-frequency components of the fungal microscopic image, and then reconstruct the fungal microscopic image into a clear image after denoising based on the noise difference between the high-frequency and low-frequency components. The processing module is also used to simultaneously extract the optical features of optical diffraction at the spore edge and the texture features of the spore surface texture from the denoised clear image, and to perform multi-channel feature fusion of the optical features and the texture features to obtain an enhanced texture feature map for fungal microscopy. The processing module is further configured to generate a discrimination feature of spore configuration based on the spatial topological relationship of different feature units in the enhanced texture feature map in a preset neighborhood, and then determine the confidence level of the presence of spore targets in different feature units in the enhanced texture feature map through the discrimination feature; The execution module is used to perform adaptive threshold segmentation on the enhanced texture feature map based on all confidence levels, and then generate a confidence detection report of fungal spores on the patient's skin based on the segmentation results.

[0013] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method for identifying skin fungal microscopic images.

[0014] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method for identifying skin fungal microscopic images.

[0015] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: The skin fungal microscopic image recognition method and system provided in this application firstly reconstructs the fungal microscopic image into a denoised clear image based on the noise difference between the high-frequency and low-frequency components. Then, optical features of spore edge optical diffraction and texture features of spore surface texture are simultaneously extracted from this clear image. The optical and texture features are then fused using multi-channel features to obtain an enhanced texture feature map. This process can specifically suppress random noise in the image through frequency domain decomposition, while utilizing the sensitivity of optical features to edge diffraction effects and the characterization ability of texture features to surface microstructures, forming complementary feature enhancements on the basis of denoising. This provides a comprehensive feature representation that is both background-clean and detail-rich for subsequent recognition. This process provides a multi-dimensional information foundation for extracting key spore identification features. Subsequently, based on the spatial topological relationships of different feature units in a preset neighborhood in the enhanced texture feature map, a spore configuration discrimination feature is generated. This method uses discriminative features to determine the confidence level of spore targets in different feature units. It models the spatial configuration patterns of spores in actual samples, such as aggregation and directional arrangement. By analyzing the topological correlation between feature units and their neighboring features, it effectively distinguishes isolated noise points from regions conforming to real spore distribution patterns. This process provides a spore authenticity verification mechanism based on spatial distribution consistency. It dynamically quantifies the probability of each feature unit being a real spore through spatial topological relationships, providing a spatially context-verified confidence metric for subsequent segmentation. This ensures that the final identification result no longer relies solely on single-point feature responses but is a reliable decision that integrates local feature quality and spatial distribution rationality, effectively distinguishing real spores from morphologically similar image artifacts in complex backgrounds. In summary, this scheme can suppress image noise while fully preserving the key discriminative features of spores. Attached Figure Description

[0016] Figure 1 This is a schematic flowchart illustrating a method for identifying skin fungi under a microscope according to some embodiments of this application; Figure 2This is a schematic flowchart illustrating the process of determining an enhanced texture feature map according to some embodiments of this application; Figure 3 This is a schematic diagram illustrating the process of implementing binarized segmentation according to some embodiments of this application; Figure 4 This is a schematic diagram of the structure of a skin fungal microscopic image recognition system according to some embodiments of this application; Figure 5 This is an internal structural diagram of a computer device for implementing a method for identifying microscopic images of skin fungi, according to some embodiments of this application. Detailed Implementation

[0017] To better understand the technical solutions in this embodiment, the technical solutions in this embodiment will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0018] refer to Figure 1 The figure is a flowchart illustrating a method for identifying skin fungi under a microscope according to some embodiments of this application. This method mainly includes the following steps: In step 101, fungal microscopic images of the patient's skin are obtained.

[0019] In practice, raw image data of patient skin samples can be obtained through the standardized fungal microscopy procedures of the clinical laboratory department of a medical institution; preferably, an optical microscope equipped with a digital imaging system can be used to acquire images of skin scraping samples treated with potassium hydroxide solution or fluorescent staining, thereby obtaining fungal microscopic images of the patient's skin.

[0020] It should be noted that the fungal microscopic images of patient skin described in this application refer to digital images obtained through standard clinical microscopy techniques that can characterize the morphological features of fungi in the stratum corneum of the skin. Physically, they are two-dimensional visual information carriers of fungal spores, hyphae, and other target objects in patient skin samples recorded at a specific optical magnification. The image acquisition must simultaneously meet the following requirements: follow the sample preparation and image acquisition standards of the "Clinical Fungal Laboratory Testing Operating Procedures" to ensure the biological validity of the images; and associate complete laboratory metadata (including patient anonymization identifiers, sampling time, and staining methods) to ensure data traceability.

[0021] In step 102, the high-frequency and low-frequency components of the fungal microscopic image are determined, and then the fungal microscopic image is reconstructed into a clear image after denoising based on the noise difference between the high-frequency and low-frequency components.

[0022] In some embodiments, determining the high-frequency and low-frequency components of the fungal microscopic image can be achieved using the following steps: The frequency domain transformation of the fungal microscopic image is performed to obtain the frequency domain representation of the fungal microscopic image; The high-frequency and low-frequency components of the fungal microscopic image are extracted from the frequency domain representation.

[0023] In specific implementation, the frequency domain transformation of the fungal microscopic image to obtain its frequency domain representation can be achieved in the following way: First, the fungal microscopic image is converted from the RGB color model in the spatial domain to a luminance component model suitable for frequency domain analysis (e.g., converted to a grayscale image or its luminance Y channel is extracted); then, a two-dimensional fast Fourier transform algorithm or a discrete wavelet transform algorithm is used to perform global frequency domain mapping on the luminance component image; wherein, as a preferred embodiment, when using the two-dimensional fast Fourier transform, the image pixel matrix is ​​used as input, and its spectrum is calculated through complex number operations, and the output result is a frequency domain representation that simultaneously contains the amplitude spectrum and the phase spectrum. The amplitude spectrum characterizes the intensity distribution of different frequency components in the image, while the phase spectrum preserves the structural information of the image; in other embodiments, multi-resolution analysis tools such as discrete cosine transform and Gabor transform can also be used to extract local frequency domain features to better capture the multi-scale characteristics of spore edges and textures, which is not limited in this application.

[0024] It should be noted that the frequency domain representation mentioned in this application refers to the mathematical representation of the fungal microscopic image after conversion from the spatial domain to the frequency domain, which is used to reveal the distribution pattern of different spatial frequency components in the image and provide an analytical basis for subsequent separation of high-frequency details and low-frequency background.

[0025] In specific implementation, the high-frequency and low-frequency components of the fungal microscopic image are extracted from the frequency domain representation. Specifically, based on the frequency domain representation (such as the spectrum after Fourier transform or the subband coefficients after wavelet transform), the frequency domain components are divided into high-frequency and low-frequency parts according to a preset frequency cutoff threshold or scale parameter. First, for the frequency domain representation based on Fourier transform, the spectrum can be filtered by constructing an ideal low-pass filter and an ideal high-pass filter, or a Gaussian low-pass filter and a Gaussian high-pass filter. The frequency coefficients within the filter passband are extracted as low-frequency and high-frequency components, respectively. The low-frequency components mainly correspond to slowly changing smooth regions in the image (such as the uniform region inside the spore and the background), while the high-frequency components correspond to rapidly changing regions in the image. Rapidly changing edges and details (such as spore outlines, hyphal boundaries, and noise); then, for the frequency domain representation based on wavelet transform, the approximate subband coefficients at the lowest scale can be extracted as low-frequency components, and the coefficients of detail subbands (including horizontal, vertical, and diagonal directions) at each scale can be extracted as high-frequency components; wherein, the specific value of the frequency cutoff threshold or decomposition scale can be preset according to the typical spore size and noise characteristics of the fungal microscopic image. For example, for images containing a large number of fine hyphae, the frequency cutoff threshold can be appropriately increased to retain more high-frequency details. This application does not limit the specific value of the threshold; finally, through the division operation, the high-frequency and low-frequency components of the fungal microscopic image are separated from the integrated frequency domain representation.

[0026] It should be noted that the high-frequency components and low-frequency components mentioned in this application refer to the frequency components that are separated from the frequency domain representation, respectively characterizing rapidly changing details and slowly changing backgrounds of the image. The high-frequency components are mainly used to capture spore edges and noise features, while the low-frequency components are mainly used to describe the interior of the spores and smooth areas of the image background.

[0027] In some embodiments, reconstructing the fungal microscopic image into a denoised, clear image based on the noise difference between the high-frequency and low-frequency components can be achieved using the following steps: The noise difference is determined based on the noise variance between the high-frequency component and the low-frequency component; Adaptive threshold denoising is performed on the high-frequency components based on the noise differences. The denoised high-frequency components and the low-frequency components are then subjected to inverse frequency domain transformation to reconstruct a clear image after denoising.

[0028] In specific implementation, the noise difference based on the noise variance of the high-frequency component and the low-frequency component can be determined in the following way: First, the amplitude of the frequency domain coefficients contained in the high-frequency component is statistically analyzed, and its dispersion across the entire frequency distribution is calculated. The result is used as an estimate of the high-frequency noise variance. Simultaneously, the variance of the local coefficient blocks in the low-frequency component corresponding to the uniform background region of the image is calculated, and the average value is taken as an estimate of the low-frequency noise variance. Subsequently, the high-frequency noise variance estimate and the low-frequency noise variance estimate are proportionally calculated, and the resulting proportional value is defined as a noise difference index characterizing the degree of difference in noise levels between the two. The larger the proportional value, the more significant the noise pollution in the high-frequency region relative to the background. In other embodiments, the normalized variance difference or a difference measure based on logarithmic transformation can also be used to define the noise difference, which is not limited in this application.

[0029] It should be noted that the noise difference mentioned in this application refers to an index obtained by quantifying the difference in noise levels between high-frequency and low-frequency components, which is used to characterize the distribution characteristics of noise in the frequency domain of an image and to provide a basis for adaptive denoising.

[0030] In specific implementation, adaptive threshold denoising of the high-frequency components based on the noise difference can be achieved in the following way: First, establish a mapping relationship between the noise difference index and the denoising threshold. This mapping relationship is used to quantify the dimensionless noise difference into a specific threshold value. As a preferred embodiment, this mapping can be established through a preset linear function relationship. Specifically, a base threshold and a scaling factor are set. The noise difference index is multiplied by the scaling factor, and then added to the base threshold. The result is the final adaptive denoising threshold. The base threshold is used to ensure the suppression of normal level noise, and its value can be preset based on the average gray value and contrast of the fungal microscopic image. The scaling factor determines the noise difference index. The adjustment strength of the acoustic difference on the final threshold can be adjusted by setting a larger proportional coefficient when the desired noise difference has a significant impact on the denoising effect, and a smaller proportional coefficient when the noise difference has a small impact. This application does not limit the specific values ​​of the basic threshold and the proportional coefficient. In other embodiments, the mapping relationship can also be implemented using a nonlinear function (such as an exponential function or a piecewise function) or a predefined lookup table to adapt to the denoising requirements under different noise distribution characteristics. Subsequently, a soft threshold function is used to process the frequency domain coefficients in the high-frequency component using the adaptively calculated threshold: coefficients with amplitudes lower than the adaptive threshold are set to zero, and coefficients with amplitudes higher than the adaptive threshold are contracted towards zero, with the contraction amplitude equal to the adaptive threshold. Finally, the denoised high-frequency component is output.

[0031] It should be noted that the initial gradient features and initial texture features mentioned in this application refer to the original features extracted from the denoised clear image, which characterize the optical diffraction effect of the spore edge and the surface microstructure, respectively, and are used to jointly describe the morphological characteristics of the spore.

[0032] In specific implementation, the inverse frequency domain transformation of the denoised high-frequency components and the low-frequency components to reconstruct the denoised clear image can be achieved in the following way: First, the high-frequency components after adaptive threshold denoising and the low-frequency components that maintain their original state are recombined in the frequency domain to form a complete optimized frequency domain representation; then, according to the specific algorithm used in the initial frequency domain transformation, the corresponding inverse transformation algorithm is selected to transform the optimized frequency domain representation back to the spatial domain: for example, when the initial Fourier transform is used, the inverse Fourier transform algorithm is applied; when the initial wavelet transform is used, the inverse wavelet transform algorithm is applied; finally, the reconstructed spatial domain image data is output as the denoised clear image.

[0033] In step 103, optical features of spore edge optical diffraction and texture features of spore surface texture are extracted simultaneously from the denoised clear image. The optical features and texture features are then fused using multi-channel features to obtain an enhanced texture feature map for fungal microscopy.

[0034] In some embodiments, the simultaneous extraction of optical features of spore edge optical diffraction and texture features of spore surface texture from the denoised clear image can be achieved by the following steps: The denoised clear image is subjected to edge enhancement processing based on gradient operators, and then the initial gradient features characterizing the optical diffraction of spore edges are extracted. Local binary mode analysis is performed on the denoised clear image to extract the initial texture features characterizing the surface texture of the spores; The initial gradient features and the initial texture features are respectively quantized to obtain the optical features of optical diffraction at the spore edge and the texture features of the spore surface texture.

[0035] In specific implementation, the edge enhancement processing based on gradient operators on the denoised clear image, and the extraction of initial gradient features characterizing the optical diffraction of the spore edge, can be achieved by the following steps: First, a gradient operator is selected, and the gradient operator is convolved with the denoised clear image in two dimensions to calculate the approximate value of the first derivative of the image at each pixel. In a preferred embodiment, convolution operations can be performed simultaneously using gradient operator convolution kernels in the horizontal and vertical directions to obtain the horizontal gradient component map and the vertical gradient component map of the image. Then, the square root operation is performed on the horizontal gradient value and the vertical gradient value corresponding to each pixel, and the result is used as the comprehensive gradient magnitude of the pixel. The comprehensive gradient magnitude of all pixels constitutes the initial gradient feature map characterizing the optical diffraction intensity of the spore edge. In other embodiments, the sum of the absolute values ​​of the two directional gradient components can be directly taken as the gradient magnitude, or the Canny operator and its non-maximum suppression strategy can be used for edge refinement. This application does not limit this.

[0036] In specific implementation, the following steps can be used to perform local binary pattern analysis on the denoised clear image to extract the initial texture features characterizing the spore surface texture: First, select a local neighborhood of a preset size centered on each pixel in the denoised clear image; then, cyclically compare the gray values ​​of all surrounding pixels in the local neighborhood with the gray value of the center pixel. If the surrounding pixel value is greater than or equal to the center pixel value, mark it as 1 at that position; otherwise, mark it as 0, thereby generating a circularly distributed binary pattern sequence; subsequently, read the binary sequence clockwise and convert it into a corresponding decimal number. The value is assigned to the center pixel position to complete the local binary pattern encoding of that point. After traversing the entire image using a sliding window, a complete local binary pattern encoding map is generated. Finally, the local binary pattern encoding map is divided into grid blocks. The statistical histogram of all possible pattern values ​​in each image sub-block is calculated independently. The histograms of all sub-blocks are connected in a predetermined order to form a high-dimensional joint histogram feature vector. This feature vector is the initial texture feature characterizing the micro-texture structure of the spore surface. The size and shape of the local neighborhood, as well as the grid granularity of the image blocks, can be preset according to the roughness of the spore surface texture.

[0037] In specific implementation, feature quantization is performed on the initial gradient features and the initial texture features to obtain the optical features of optical diffraction at the spore edge and the texture features of the spore surface texture. This can be achieved through the following steps: First, for the initial gradient features, a statistical distribution-based quantization method is used to calculate the mean, standard deviation, skewness coefficient, and kurtosis coefficient of its gradient amplitude distribution. These four statistical quantities are then combined in a predetermined order to form a four-dimensional feature vector, which is the quantized optical feature. Simultaneously, for the initial texture features, principal component analysis is used to reduce their dimensionality. By calculating the covariance matrix and eigenvalues ​​and eigenvectors of the high-dimensional joint histogram feature vector, the eigenvectors corresponding to the k largest eigenvalues ​​are selected to construct a projection transformation matrix. The original high-dimensional texture features are projected into this k-dimensional subspace, and the projected k-dimensional feature vector is the quantized texture feature. The feature dimension k retained after dimensionality reduction can be preset according to the complexity of the spore texture pattern in the fungal microscopic image and its impact on the subsequent feature fusion effect.

[0038] In some embodiments, reference Figure 2 As shown in the figure, this is a schematic flowchart illustrating the process of determining an enhanced texture feature map according to some embodiments of this application. The enhanced texture feature map for fungal microscopy can be obtained by multi-channel feature fusion of the optical features and the texture features through the following steps: First, in step 1031, the optical features and the texture features are used as independent feature channels to create a dual-channel feature map; Then, in step 1032, each feature channel in the dual-channel feature map is normalized. Finally, in step 1033, cross-channel feature aggregation is performed for each normalized feature channel to generate the enhanced texture feature map for fungal microscopy.

[0039] In specific implementation, the optical features and texture features are treated as independent feature channels. The creation of a dual-channel feature map can be achieved through the following steps: First, the two one-dimensional feature vectors, optical features and texture features, are mapped back to two-dimensional feature matrices with the same spatial size as the original denoised clear image through a spatial grid rearrangement operation. The optical feature vector is reconstructed into four independent two-dimensional feature planes according to the physical meaning of its statistical moments (mean, variance, skewness, kurtosis), while the texture feature vector is reconstructed into k two-dimensional feature planes according to its reduced dimension k. Subsequently, all the reconstructed two-dimensional feature planes (4 from optical features and k from texture features) are stacked in the channel dimension to form an initial multi-channel feature map with (4+k) channels. In a preferred embodiment, if the two-dimensional reconstruction of the optical features and texture features results in a spatial resolution inconsistent with the original image, bilinear interpolation or nearest neighbor interpolation algorithms can be used to upsample it to a uniform size. This application does not limit the specific choice of the interpolation algorithm.

[0040] In specific implementation, the normalization processing of each feature channel in the dual-channel feature map can be achieved by the following steps: independently performing Z-Score standardization processing on the two-dimensional feature matrix of each channel in the dual-channel feature map, specifically: firstly, calculating the mean and standard deviation of all elements in each feature channel, then subtracting the mean from the value of each element in that channel and dividing by the standard deviation, so that the data distribution of each feature channel is adjusted to a standard normal distribution with a mean of 0 and a standard deviation of 1; in other embodiments, the Min-Max normalization method can also be used to linearly scale the value of each feature channel to the interval [0,1] or [-1,1], which is not limited in this application; the normalization processing aims to eliminate the dominance difference between channels caused by different feature dimensions and numerical ranges, and provide a consistent numerical basis for subsequent cross-channel feature aggregation.

[0041] In specific implementation, cross-channel feature aggregation is performed for each normalized feature channel to generate the enhanced texture feature map for fungal microscopy. This can be achieved through the following steps: First, an adjustable weighting coefficient is assigned to each feature channel in the normalized dual-channel feature map. The magnitude of the weighting coefficient can be preset according to the discrimination ability of the feature channel in distinguishing between spore and non-spore regions. For example, when the optical diffraction features of the spore edge are more critical for target identification, a relatively large weighting coefficient can be assigned to the channel originating from the optical feature. Then, the two-dimensional matrix of all feature channels is weighted and summed with their respective weighting coefficients. That is, the element values ​​at corresponding positions are multiplied by the coefficients and then added together to generate a single two-dimensional matrix that integrates optical and surface texture characteristics. Finally, the fused two-dimensional matrix is ​​the enhanced texture feature map for fungal microscopy, which retains the position and morphological information of the spore in space and comprehensively represents the composite features of edge diffraction and surface texture in numerical value.

[0042] In step 104, a discrimination feature for spore configuration is generated based on the spatial topological relationship of different feature units in the enhanced texture feature map in a preset neighborhood, and then the confidence level of the presence of spore targets in different feature units in the enhanced texture feature map is determined by the discrimination feature.

[0043] In some embodiments, generating discriminative features for spore configurations based on the spatial topological relationships of different feature units in a preset neighborhood in the enhanced texture feature map can be achieved through the following steps: The enhanced texture feature map is divided into multiple feature units, and a preset neighborhood is established for each feature unit; Extract the spatial topological relationship between each feature unit and other feature units in its preset neighborhood; Based on the aforementioned spatial topological relationships, discriminative features characterizing spore configurations are constructed.

[0044] In specific implementation, dividing the enhanced texture feature map into multiple feature units and establishing a preset neighborhood for each feature unit can be achieved in the following way: First, the enhanced texture feature map is divided into multiple non-overlapping or partially overlapping feature units using a regular grid partitioning method or an adaptive partitioning method based on superpixel segmentation. The regular grid partitioning method uniformly divides the feature map into rectangular units of fixed size, while the adaptive partitioning method generates irregularly shaped superpixel units using algorithms such as simple linear iterative clustering based on the continuity of texture features in the feature map. Then, a preset neighborhood is established for each feature unit. The preset neighborhood can be a fixed-size rectangular window centered on the current feature unit, or a variable-range neighborhood dynamically determined based on the feature similarity between feature units. In a preferred embodiment, when a fixed-size rectangular window is used, its window size can be preset according to the typical size and distribution density of spores in the fungal microscopic image. For example, a smaller neighborhood window can be set for densely distributed spore groups to capture local details. This application does not limit the specific size of the window.

[0045] In specific implementation, the extraction of spatial topological relationships between each feature unit and other feature units in its preset neighborhood can be achieved in the following way: First, a spatial relationship description model between feature units is established. This model is achieved by calculating the relative spatial positions of the current feature unit and all other feature units in its preset neighborhood. Specifically, this includes: calculating the Euclidean distance or Manhattan distance between the current feature unit and each neighboring unit as distance topological features, calculating the azimuth angle from the current feature unit to each neighboring unit as direction topological features, and statistically analyzing the similarity of the feature values ​​between the current feature unit and each neighboring unit as feature association topological features. Then, these multi-dimensional topological relationships, such as distance, direction, and feature association, are quantized and encoded to form a spatial topological relationship description vector for each feature unit. In other embodiments, a spatial relationship graph model of feature units can also be constructed, with feature units as nodes and spatial relationships between units as edges, and then the topological attributes of the graph can be extracted as spatial topological relationships. This application does not limit this approach.

[0046] It should be noted that the spatial topological relationship mentioned in this application refers to the relative position, connection relationship and distribution pattern of different feature units in the enhanced texture feature map within their preset neighborhood, which is used to characterize the spatial configuration features of the spore population.

[0047] In specific implementation, the discriminative features representing spore configurations based on the spatial topological relationships can be constructed in the following way: First, feature engineering is performed on the spatial topological relationship description vector of each feature unit, including but not limited to: calculating the statistical features of each dimension of the vector (such as mean, variance, extreme values, etc.), extracting the principal component analysis of the main components of the vector, or performing feature compression and reconstruction through neural networks such as autoencoders; then, the processed spatial topological features are combined with the original texture features of the corresponding feature units in the enhanced texture feature map through cross-modal feature splicing or weighted fusion to form a composite feature vector that comprehensively represents the spatial distribution pattern and local texture characteristics of spores; finally, the composite feature vector is used as the discriminative feature of the spore configuration. The discriminative feature includes both the texture characteristics of individual spores and the arrangement and aggregation patterns of spores in spatial distribution, providing a more discriminative feature representation for subsequent spore target identification.

[0048] It should be noted that the discriminative features described in this application refer to a composite feature vector constructed by fusing spatial topological relationships and original texture features, which is used to comprehensively characterize the spatial distribution pattern and local texture characteristics of spores, thereby improving the discriminative ability of spore identification.

[0049] In some embodiments, determining the confidence level of the presence of spore targets in different feature units of the enhanced texture feature map using the discriminative features can be achieved through the following steps: The discriminative features are input in parallel into multiple heterogeneous base classifiers of a pre-constructed spore target integrated classification model, thereby obtaining the preliminary classification probability of each feature unit corresponding to each base classifier; The preliminary classification results of all feature units corresponding to each feature unit in a preset neighborhood are aggregated, and then the neighborhood consistency confidence of each feature unit is calculated through a spatial context reasoning model. The initial classification probability of each feature unit in the base classifier is fused with the neighborhood consistency confidence at the decision level. Based on the fusion results, the confidence level of the presence of spore targets in different feature units of the enhanced texture feature map is determined.

[0050] In specific implementation, the discriminative features are input in parallel into multiple heterogeneous base classifiers of a pre-constructed spore target ensemble classification model to obtain the preliminary classification probability of each feature unit corresponding to each base classifier. This can be achieved in the following way: First, a heterogeneous base classifier set including support vector machines, random forests, and lightweight convolutional neural networks is constructed. The support vector machine uses radial basis function kernels to process nonlinear features, the random forest processes feature subsets in parallel through multiple decision trees, and the lightweight convolutional neural network extracts deep features through several convolutional layers and fully connected layers. Then, the discriminative features of each feature unit are simultaneously input into these three base classifiers to obtain the corresponding classification outputs. The support vector machine calculates the distance to the hyperplane and applies the sigmoid function to convert it into a probability. The random forest uses a voting mechanism to count the proportion of positive classes as a probability. The convolutional neural network outputs the probability through the final softmax layer. Finally, each feature unit obtains three independent preliminary classification probability values, that is, the preliminary classification probability of each feature unit corresponding to each base classifier. In other embodiments, other methods can also be used to achieve this, which are not limited here.

[0051] It should be noted that the preliminary classification probability mentioned in this application refers to the probability value output after the discriminative features are independently analyzed by multiple heterogeneous base classifiers, which is used to provide a preliminary judgment on the existence of spore targets from different feature perspectives.

[0052] In specific implementation, the preliminary classification results of all corresponding feature units within a preset neighborhood for each feature unit are aggregated, and then the neighborhood consistency confidence of each feature unit is calculated through a spatial context inference model. This can be achieved in the following way: For each feature unit, the preliminary classification probabilities of the three base classifiers of all feature units within the preset neighborhood for each feature unit are obtained to form a neighborhood classification probability set; then, a spatial context inference model is constructed based on a conditional random field or a Markov random field. This model establishes a joint probability distribution by defining the spatial correlation potential function between feature units and the observation potential function of the feature unit's own classification probability; next, inference calculation is performed through confidence propagation or mean field approximation algorithm to obtain the posterior probability of each feature unit after considering the spatial context; finally, the ratio of this posterior probability to the original independent classification probability is used as the neighborhood consistency confidence of the corresponding feature unit. In other embodiments, other methods can also be used, which are not limited here.

[0053] It should be noted that the neighborhood consistency confidence level mentioned in this application refers to the confidence metric obtained by analyzing the spatial consistency between the classification results of feature units and their neighboring units, which is used to evaluate the reliability of the preliminary classification results in the spatial context.

[0054] In specific implementation, the decision-level fusion of the preliminary classification probabilities of each feature unit in the base classifiers with the neighborhood consistency confidence can be achieved as follows: First, the preliminary classification probabilities of each feature unit in the three base classifiers are fused at the first level. The weighted probability value is calculated using a weighted average method. The weight coefficient of each base classifier is dynamically determined by its classification performance index on the validation set. Specifically, the harmonic mean of the accuracy and recall of each base classifier on the validation set is calculated as its performance score. The performance score of each base classifier is normalized and used as its corresponding weight coefficient. The preliminary classification probabilities of each base classifier are multiplied by their weight coefficients and summed to obtain the first-level fusion probability value. Then, the first-level fusion probability value is fused with the neighborhood consistency confidence at the second level. The first-level fusion is performed by designing an adaptive weighting function to calculate the final fusion confidence score. Specifically, a neighborhood consistency threshold is first set. When the neighborhood consistency confidence score is higher than this threshold, a weighted summation method is used to fuse the first-level fusion probability value with the neighborhood consistency confidence score. The fusion weight of the neighborhood consistency confidence score increases as its value increases. When the neighborhood consistency confidence score is lower than this threshold, the first-level fusion probability value is mainly relied upon as the final result. The neighborhood consistency threshold can be preset according to the spatial distribution density of feature units in the enhanced texture feature map. This application does not limit the specific value of the threshold. Finally, the fusion confidence score of each feature unit is output as the fusion result. Other methods can also be used in other embodiments, which are not limited here.

[0055] In specific implementation, the following method can be used to determine the confidence level of the presence of spore targets in different feature units of the enhanced texture feature map based on the fusion result. Specifically, a normalization method based on sample distribution is used, which involves calculating the mean and standard deviation of the fusion confidence scores of all samples in the training set, subtracting the mean from the current score and dividing by the standard deviation for Z-score standardization, and then converting it to a confidence level in the [0, 1] interval through a probability integral function. Thus, the confidence level of the presence of spore targets in different feature units of the enhanced texture feature map is obtained. In other embodiments, other methods can also be used, which are not limited here.

[0056] It should be noted that the confidence level mentioned in this application refers to the probability value obtained by fusing the judgment of the multi-classifier with spatial context information, which is used to quantify the credibility of the presence of spore targets in each feature unit and to provide a probabilistic basis for subsequent segmentation.

[0057] In step 105, the enhanced texture feature map is subjected to adaptive threshold segmentation based on all confidence levels, and a confidence detection report of fungal spores on the patient's skin is generated based on the segmentation results.

[0058] In some embodiments, adaptive threshold segmentation of the enhanced texture feature map based on all confidence levels, followed by generating a confidence detection report of fungal spores on the patient's skin based on the segmentation results, can be achieved through the following steps: The adaptive segmentation threshold is determined based on the confidence level distribution characteristics of all feature units; The enhanced texture feature map is binarized based on the adaptive segmentation threshold to obtain the spore target region; A confidence test report for fungal spores is generated based on the target spore region and its corresponding confidence level.

[0059] In specific implementation, the adaptive segmentation threshold can be determined based on the confidence level distribution characteristics of all feature units in the enhanced texture feature map by the following method: First, construct a statistical histogram of the confidence levels of all feature units in the enhanced texture feature map, and determine the adaptive segmentation threshold by analyzing the distribution characteristics of the histogram. As a preferred embodiment, the optimal threshold can be automatically calculated using the maximum inter-class variance method. This method divides the confidence levels into foreground (spore target) and background categories by traversing all possible thresholds, calculates the variance between the two categories, and selects the threshold that maximizes the inter-class variance as the adaptive segmentation threshold. In other embodiments, the valley position between the two peaks can be selected as the adaptive segmentation threshold based on the bimodal characteristics of the confidence level histogram, or a method based on a probability distribution model can be used to determine the adaptive segmentation threshold by fitting a mixture Gaussian distribution of the confidence levels. This application does not limit the specific calculation method and value of the threshold.

[0060] For specific implementation, refer to Figure 3 As shown in the figure, this is a schematic diagram of the process of implementing binarization segmentation in some embodiments of this application. Binarization segmentation of the enhanced texture feature map based on the adaptive segmentation threshold to obtain the spore target region can be achieved in the following way: The confidence level of each feature unit in the enhanced texture feature map is compared with the adaptive segmentation threshold. When the confidence level is greater than or equal to the adaptive segmentation threshold, the feature unit is marked as foreground (spore target); when the confidence level is less than the adaptive segmentation threshold, it is marked as background, thereby generating an initial binary segmentation map. Subsequently, morphological post-processing is performed on the initial segmentation result, including using morphological opening operations to eliminate isolated noise points, using morphological closing operations to fill holes inside the spore region, and removing excessively small connected regions through area filtering. In the final optimized binary segmentation map, all connected foreground regions are the identified spore target regions.

[0061] In specific implementation, generating a confidence test report for fungal spores based on the spore target region and its corresponding confidence level can be achieved in the following way: First, a quantitative statistical analysis is performed on the spore target region, including counting the number of spore targets, calculating the area, perimeter, and other morphological parameters of each spore target, and calculating the average confidence level of the feature units within each spore target region as the comprehensive confidence level of the spore, thereby generating a structured test report; the test report includes at least: spore quantity statistics, spore morphological characteristics description, spore spatial distribution map, confidence level assessment of each spore target, and detection reliability index based on confidence level; wherein, as a preferred embodiment, the test results can also be divided into high-confidence test areas and low-confidence test areas according to the distribution of the comprehensive confidence level of the spores, providing differentiated diagnostic references for clinicians.

[0062] In another aspect, in some embodiments, this application provides a skin fungal microscopic image recognition system, with reference to... Figure 4 The figure is a schematic diagram of the structure of a skin fungal microscopic image recognition system 200 according to some embodiments of this application. The skin fungal microscopic image recognition system 200 includes: an acquisition module 201, a processing module 202, and an execution module 203, which are described below: The acquisition module 201 in this application is mainly used to acquire microscopic images of fungi on the patient's skin; Processing module 202, in this application, is mainly used to determine the high-frequency components and low-frequency components of the fungal microscopic image, and then reconstruct the fungal microscopic image into a clear image after denoising based on the noise difference between the high-frequency components and the low-frequency components. In addition, the processing module 202 in this application is also used to simultaneously extract the optical features of optical diffraction at the spore edge and the texture features of the spore surface texture from the denoised clear image, and to perform multi-channel feature fusion of the optical features and the texture features to obtain an enhanced texture feature map for fungal microscopy. In addition, the processing module 202 in this application is also used to generate a discrimination feature of spore configuration based on the spatial topological relationship of different feature units in the preset neighborhood in the enhanced texture feature map, and then determine the confidence level of the presence of spore targets in different feature units in the enhanced texture feature map through the discrimination feature; The execution module 203 in this application is mainly used to perform adaptive threshold segmentation on the enhanced texture feature map based on all confidence levels, and then generate a confidence detection report of fungal spores on the patient's skin based on the segmentation results.

[0063] In addition, this application also provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described method for identifying skin fungal microscopic images.

[0064] In some embodiments, reference Figure 5 This figure is an internal structural diagram of a computer device for implementing a method for identifying microscopic images of dermatophytes according to some embodiments of this application. The method for identifying microscopic images of dermatophytes in the above embodiments can be implemented through... Figure 5 The computer device shown is used to implement this, and the computer device 300 includes at least one processor 301, a communication bus 302, a memory 303, and at least one communication interface 304.

[0065] The processor 301 may be a general-purpose central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more devices used to control the execution of the skin fungus microscopic image recognition method in this application.

[0066] The communication bus 302 is used to transmit information between the aforementioned components.

[0067] Memory 303 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital versatile optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. Memory 303 may exist independently and be connected to processor 301 via communication bus 302. Memory 303 may also be integrated with processor 301.

[0068] The memory 303 stores program code for executing the scheme of this application, and its execution is controlled by the processor 301. The processor 301 executes the program code stored in the memory 303. The program code may include one or more software modules. In the above embodiments, the skin fungal microscopic image recognition method can be implemented by the processor 301 and one or more software modules in the program code in the memory 303.

[0069] Communication interface 304 uses any transceiver-like device to communicate with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0070] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single-CPU) processor or a multi-core (multi-CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores for processing data (e.g., computer program instructions).

[0071] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device may be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer device.

[0072] In addition, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for identifying skin fungal microscopic images.

[0073] In summary, the skin fungal microscopic image recognition method and system disclosed in this application acquires a fungal microscopic image of a patient's skin; determines the high-frequency and low-frequency components of the fungal microscopic image, and then reconstructs the fungal microscopic image into a denoised clear image based on the noise difference between the high-frequency and low-frequency components; simultaneously extracts optical features of spore edge optical diffraction and texture features of spore surface texture from the denoised clear image, and performs multi-channel feature fusion of the optical features and texture features to obtain an enhanced texture feature map for fungal microscopic examination; generates a spore configuration discrimination feature based on the spatial topological relationship of different feature units in a preset neighborhood in the enhanced texture feature map, and then determines the confidence level of the presence of spore targets in different feature units in the enhanced texture feature map through the discrimination feature; performs adaptive threshold segmentation on the enhanced texture feature map based on all confidence levels, and then generates a confidence detection report of fungal spores on the patient's skin based on the segmentation results; it can completely preserve the key identification features of spores while suppressing image noise.

[0074] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0075] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A skin fungal microscopy image recognition method, characterized by, The method comprises the following steps: obtaining a fungal microscopy image of the patient's skin; determining high-frequency components and low-frequency components of the fungal microscopy image, and reconstructing the fungal microscopy image into a clear image after denoising based on noise difference between the high-frequency components and the low-frequency components; synchronously extracting optical features of optical diffraction of spore edges and texture features of spore surface textures from the clear image after denoising, performing multi-channel feature fusion on the optical features and the texture features, and obtaining an enhanced texture feature map in fungal microscopy; generating a discriminative feature of spore configuration according to spatial topological relations of different feature units in the enhanced texture feature map in a preset neighborhood, and then determining a confidence level of the presence of a spore target in different feature units in the enhanced texture feature map through the discriminative feature; performing adaptive threshold segmentation on the enhanced texture feature map based on all confidence levels, and then generating a confidence detection report of fungal spores on the patient's skin through a segmentation result.

2. The method of claim 1, wherein, The method for determining the high-frequency components and the low-frequency components of the fungal microscopy image specifically comprises: performing frequency domain transformation on the fungal microscopy image to obtain a frequency domain representation of the fungal microscopy image; extracting the high-frequency components and the low-frequency components of the fungal microscopy image from the frequency domain representation.

3. The method of claim 1, wherein, The method for reconstructing the fungal microscopy image into a clear image after denoising based on noise difference between the high-frequency components and the low-frequency components specifically comprises: determining noise difference according to noise variance of the high-frequency components and the low-frequency components; performing adaptive threshold denoising processing on the high-frequency components based on the noise difference; performing inverse frequency domain transformation on the high-frequency components after denoising and the low-frequency components to reconstruct a clear image after denoising.

4. The method of claim 1, wherein, The method for synchronously extracting optical features of optical diffraction of spore edges and texture features of spore surface textures from the clear image after denoising specifically comprises: performing edge enhancement processing based on a gradient operator on the clear image after denoising, and then extracting initial gradient features representing optical diffraction of spore edges; performing local binary pattern analysis on the clear image after denoising, and then extracting initial texture features representing spore surface textures; respectively performing feature quantization on the initial gradient features and the initial texture features to obtain the optical features of optical diffraction of spore edges and the texture features of spore surface textures.

5. The method of claim 1, wherein, The method for performing multi-channel feature fusion on the optical features and the texture features to obtain an enhanced texture feature map in fungal microscopy specifically comprises: taking the optical features and the texture features as independent feature channels respectively, and creating a double-channel feature mapping graph; performing normalization processing on each feature channel in the double-channel feature mapping graph; performing cross-channel feature aggregation on each normalized feature channel to generate the enhanced texture feature map in fungal microscopy.

6. The method of claim 1, wherein, The method for generating a discriminative feature of spore configuration according to spatial topological relations of different feature units in the enhanced texture feature map in a preset neighborhood specifically comprises: dividing the enhanced texture feature map into multiple feature units, and establishing a preset neighborhood for each feature unit; extracting spatial topological relations between each feature unit and other feature units in the preset neighborhood of the feature unit; Construct a discriminant feature representing the spore configuration based on the spatial topological relationship.

7. The method of claim 1, wherein, The adaptive threshold segmentation is performed on the enhanced texture feature map based on all the confidence levels, and a confidence detection report of the fungal spores on the patient's skin is generated through the segmentation result, specifically including: An adaptive segmentation threshold is determined according to the confidence level distribution characteristics of all the feature units; The enhanced texture feature map is binarized and segmented based on the adaptive segmentation threshold, and a spore target region is obtained; A confidence detection report of the fungal spores is generated according to the spore target region and the corresponding confidence level.

8. A skin fungal microscopy image recognition system, characterized in that, It includes: An acquisition module is configured to acquire a fungal microscopic image of a patient's skin; A processing module is configured to determine high-frequency components and low-frequency components of the fungal microscopic image, and then reconstruct the fungal microscopic image into a clear image after denoising based on the noise difference between the high-frequency components and the low-frequency components; The processing module is further configured to simultaneously extract optical features of optical diffraction of spore edges and texture features of spore surface textures from the clear image after denoising, perform multi-channel feature fusion on the optical features and the texture features, and obtain an enhanced texture feature map during fungal microscopy; The processing module is further configured to generate discriminant features of spore configurations according to the spatial topological relationship of different feature units in a preset neighborhood in the enhanced texture feature map, and then determine the confidence levels of the presence of spore targets in different feature units in the enhanced texture feature map through the discriminant features; An execution module is configured to perform adaptive threshold segmentation on the enhanced texture feature map based on all the confidence levels, and then generate a confidence detection report of the fungal spores on the patient's skin through the segmentation result. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to realize the steps of the skin fungal microscopic image recognition method in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to realize the steps of the skin fungal microscopic image recognition method as claimed in any one of claims 1 to 7.

Images