Method and device for detecting tinea capitis based on multi-channel fluorescence sensing

Abstract

The application discloses a tinea capitis detection method and device based on multi-channel fluorescence sensing, comprising: emitting composite waveband excitation light to the scalp to be detected and collecting fluorescence spectrum signals and short-wave infrared diffuse reflection signals to construct multi-modal optical signals; performing adaptive filtering processing, blind source separation and correction on the multi-modal optical signals based on variational modal decomposition according to an environment baseline signal to generate net fluorescence signals; performing spectrum data cube construction and decomposition on the net fluorescence signals to extract RGB three-color fluorescence intensity values; generating a fluorescence distribution dot array and converting it into a recurrence graph matrix according to the RGB three-color fluorescence intensity values to identify a fungal infection area; performing superpixel segmentation and spectral clustering on the fungal infection area to obtain multi-scale texture features and heterogeneous sub-region labels; and identifying tinea capitis types and infection degree grades based on the recurrence graph matrix, the multi-scale texture features and the heterogeneous sub-region labels. The application can effectively improve the efficiency and accuracy of tinea capitis detection.

Description

Technical Field

[0001] This application relates to the field of tinea capitis detection technology, and in particular to a method and apparatus for detecting tinea capitis based on multi-channel fluorescence sensing. Background Technology

[0002] Tinea capitis is a common fungal skin infection in children. Based on the causative fungus and clinical characteristics, it is classified into four types: favus, white tinea, black dot tinea, and kerion. Current epidemiological data shows that the prevalence of tinea capitis in children under 3 years old is rising, with boys significantly more affected than girls. In adults, the prevalence of tinea capitis in women is nine times that of men, with most female patients presenting with the black dot type. Although early diagnosis and thorough treatment are crucial for preventing sequelae and blocking the spread of tinea capitis, early diagnosis faces multiple challenges in reality: the symptoms of early tinea capitis are very similar to those of common skin diseases such as seborrheic dermatitis, psoriasis, and scalp eczema, lacking specific characteristics, making accurate identification difficult for non-professionals and even some doctors in primary healthcare institutions, often leading to misdiagnosis and missed diagnosis. Meanwhile, traditional methods for detecting tinea capitis, such as fungal culture and direct microscopy, while having a certain degree of accuracy, suffer from problems such as long testing cycles, cumbersome procedures, and high professional requirements for testing personnel. Fungal culture usually takes several days or even weeks to obtain results, which will undoubtedly delay the condition of patients who need timely diagnosis and treatment. Summary of the Invention

[0003] The main objective of this application is to provide a method and device for detecting tinea capitis based on multi-channel fluorescence sensing, which aims to solve the technical problems of low efficiency and poor accuracy in traditional tinea capitis detection methods.

[0004] To achieve the above objectives, this application proposes a method for detecting tinea capitis based on multi-channel fluorescence sensing. This method is applied to a tinea capitis detection device based on multi-channel fluorescence sensing. The device includes a detection light generation module and a fluorescence recognition module. The method comprises:

[0005] A composite wavelength excitation light is emitted toward the area to be detected on the scalp, and the fluorescence spectrum signal and short-wave infrared diffuse reflection signal generated in the area to be detected are collected simultaneously to construct a multimodal optical signal;

[0006] The environmental baseline signal of the area to be detected is acquired, and the multimodal optical signal is subjected to adaptive filtering, blind source separation and correction based on variational mode decomposition according to the environmental baseline signal to generate a net fluorescence signal. The environmental baseline signal is composed of the dark field noise background and the ambient stray light spectrum superimposed.

[0007] Based on the net fluorescence signal, a spectral data cube is constructed and decomposed to extract the RGB three-color fluorescence intensity values;

[0008] A fluorescence distribution dot matrix is ​​generated based on the RGB three-color fluorescence intensity values. The fluorescence distribution dot matrix is ​​then converted into a recursive graph matrix using recursive graph analysis. Based on the recursive graph matrix, fungal infection areas in the area to be detected are identified.

[0009] Superpixel segmentation and spectral clustering were performed on the fungal infection area to obtain multi-scale texture features and heterogeneous sub-region labels;

[0010] Based on the recursive graph matrix, multi-scale texture features, and heterogeneous sub-region labels, the type and severity level of tinea capitis in the fungal infection area were identified.

[0011] In one embodiment, acquiring the environmental baseline signal of the area to be detected, and performing adaptive filtering, blind source separation, and correction on the multimodal optical signal based on variational mode decomposition according to the environmental baseline signal to generate a net fluorescence signal, includes:

[0012] Without emitting excitation light in the composite band, the background noise signal of the dark field and the ambient stray light spectrum signal around the area to be detected are collected.

[0013] The dark field noise background signal is superimposed with the environmental stray light spectrum signal to obtain the environmental baseline signal;

[0014] Variational mode decomposition is performed on the multimodal optical signal to obtain multiple intrinsic modes;

[0015] Based on the environmental baseline signal, the multiple intrinsic modes are adaptively filtered to obtain the filtered multimodal optical signal;

[0016] Blind source separation is performed on the filtered multimodal optical signal to obtain the target fluorescence signal;

[0017] An environmental interference weight is determined based on the environmental baseline signal, and the target fluorescence signal is differentially corrected according to the environmental interference weight and the environmental baseline signal to obtain a net fluorescence signal.

[0018] In one embodiment, the adaptive filtering of the plurality of intrinsic modes based on the environmental baseline signal to obtain the filtered multimodal optical signal includes:

[0019] Based on the environmental baseline signal, the power spectral density is calculated to determine the first power spectral density corresponding to the dark field noise background and the second power spectral density corresponding to the environmental stray light spectrum.

[0020] A priori noise fingerprint knowledge base is constructed based on the first power spectral density and the second power spectral density.

[0021] The third power spectral density of each intrinsic mode is determined, and the similarity between the third power spectral density and the power spectral density in the prior noise fingerprint knowledge base is calculated to obtain a similarity index.

[0022] The intrinsic modes that reach the preset threshold of the similarity index are identified as noise modes, and the noise modes are removed from multiple intrinsic modes to obtain a filtered intrinsic mode set.

[0023] The filtered intrinsic mode set is reconstructed to generate a filtered multimodal optical signal.

[0024] In one embodiment, the step of blind source separation of the filtered multimodal optical signal to obtain the target fluorescence signal includes:

[0025] Obtain the time delay signal corresponding to the filtered multimodal optical signal, and construct an observation signal based on the filtered multimodal optical signal and the corresponding time delay signal;

[0026] The observed signal is centered and whitened to obtain a whitened signal;

[0027] Independent component analysis is performed on the whitened signal, and the separation matrix is ​​solved iteratively by optimizing the algorithm;

[0028] The whitened signal is linearly transformed according to the separation matrix to obtain multiple independent source signals;

[0029] The signal that matches the fluorescence characteristics is selected from the multiple independent source signals as the target fluorescence signal.

[0030] In one embodiment, the step of constructing and decomposing a spectral data cube based on the net fluorescence signal and extracting RGB three-color fluorescence intensity values ​​includes:

[0031] The net fluorescence signal is subjected to spectral processing to obtain net fluorescence signals in different wavelength bands;

[0032] A three-dimensional spectral data cube is constructed based on the net fluorescence signal of the target band in the net fluorescence signals of the different bands, and the three-dimensional spectral data cube is decomposed into a non-negative matrix to obtain the fundamental matrix and the coefficient matrix.

[0033] The three-dimensional spectral data cube is reconstructed based on the fundamental matrix and coefficient matrix to obtain the reconstructed spectral data, and spectral feature vectors are extracted from the reconstructed spectral data.

[0034] The spectral feature vector is subjected to signal conditioning and amplification to obtain the processed spectral feature vector;

[0035] The RGB three-color fluorescence intensity values ​​are extracted from the processed spectral feature vector to obtain the RGB three-color fluorescence intensity values.

[0036] In one embodiment, the step of generating a fluorescence distribution dot matrix based on the RGB three-color fluorescence intensity values, converting the fluorescence distribution dot matrix into a recursive graph matrix using recursive graph analysis, and identifying fungal infection areas in the region to be detected based on the recursive graph matrix includes:

[0037] The spectral bias parameter is determined based on the RGB three-color fluorescence intensity values, and a fluorescence distribution dot matrix is ​​generated based on the spectral bias parameter and the RGB three-color fluorescence intensity values;

[0038] The fluorescence distribution dot matrix is ​​spatially filled using Hilbert curves to obtain a one-dimensional spatial sequence.

[0039] Based on the one-dimensional spatial sequence, the phase space is reconstructed to obtain multiple trajectory points in the phase space;

[0040] The Euclidean distance between multiple trajectory points in the phase space is calculated to obtain the distance matrix.

[0041] The distance matrix is ​​transformed into a recursive graph matrix by recursive quantization analysis, where the elements in the recursive graph matrix represent the recursive state of the corresponding trajectory point pair under a given threshold.

[0042] The recursive graph matrix is ​​input into the fungal infection identification model, and the recursive graph matrix is ​​used to extract features through the convolutional neural network in the fungal infection identification model to obtain a segmentation mask. The convolutional neural network includes an encoder-decoder structure and a skip connection structure.

[0043] The fluorescence distribution dot matrix is ​​segmented using the segmentation mask to identify fungal infection areas within the region to be detected.

[0044] In one embodiment, the step of performing superpixel segmentation and spectral clustering on the fungal infection region to obtain multi-scale texture features and heterogeneous sub-region labels includes:

[0045] The fungal infection region was segmented into multiple superpixel blocks using a simple linear iterative clustering algorithm.

[0046] Multimodal texture feature extraction is performed on each superpixel block to obtain the gray-level co-occurrence matrix texture features, local binary mode variance features, and Gabor multi-directional texture response features of each superpixel block;

[0047] The gray-level co-occurrence matrix texture features, local binary mode variance features, and Gabor multi-directional texture response features of each superpixel block are concatenated to obtain the superpixel feature vector.

[0048] A similarity matrix is ​​constructed based on the superpixel feature vectors, and a spectral clustering algorithm is used to perform cluster analysis on the similarity matrix to obtain multiple clusters;

[0049] The fungal infection area is divided into multiple heterogeneous sub-regions based on the clusters, and a corresponding label is assigned to each heterogeneous sub-region to obtain heterogeneous sub-region labels.

[0050] In one embodiment, the identification of the type and severity level of tinea capitis in a fungal infection area based on the recursive graph matrix, multi-scale texture features, and heterogeneous sub-region labels includes...

[0051] A multimodal feature map is constructed based on the recursive graph matrix, the multi-scale texture features, and the heterogeneous sub-region labels;

[0052] The multimodal feature map is input into a multimodal discrimination model based on graph attention network to obtain classification confidence. The multimodal discrimination model is pre-trained by transfer learning and its parameters are optimized by backpropagation algorithm.

[0053] Based on the classification confidence level, the type of tinea capitis in the fungal infection area was identified;

[0054] Obtain fluorescence point cloud data of the fungal infection area, and determine the lesion area ratio and fluorescence heterogeneity index of the fungal infection area based on the fluorescence point cloud data of the fungal infection area;

[0055] A comprehensive infection severity assessment index is calculated based on the proportion of lesion area in the fungal infection region, the fluorescence heterogeneity index, and the classification confidence level.

[0056] The level of infection severity is determined based on the comprehensive infection severity assessment index.

[0057] In one embodiment, the step of inputting the multimodal feature map into a multimodal discrimination model based on a graph attention network to obtain classification confidence includes:

[0058] The multimodal feature map is decomposed into a node feature matrix and an adjacency matrix. The node feature matrix contains fusion information of recursive graph matrix features, multi-scale texture features, and heterogeneous sub-region labels. The adjacency matrix represents the spatial correlation between nodes.

[0059] The attention weights between node pairs are calculated using the attention mechanism of the graph attention network in the multimodal discriminant model.

[0060] The node feature matrix is ​​weighted and aggregated based on the attention weights to obtain a weighted feature vector.

[0061] The weighted feature vector is input into the fully connected layer of the multimodal discrimination model for nonlinear transformation, and the initial classification probability of each type of tinea capitis is output.

[0062] The initial classification probabilities are normalized using Softmax to obtain the classification confidence scores for each type of tinea capitis.

[0063] Furthermore, to achieve the above objectives, this application also proposes a tinea capitis detection device based on multi-channel fluorescence sensing, the tinea capitis detection device based on multi-channel fluorescence sensing comprising:

[0064] The emission module is used to emit composite band excitation light to the area to be detected on the scalp, and simultaneously acquire the fluorescence spectrum signal and short-wave infrared diffuse reflection signal generated in the area to be detected to construct a multimodal optical signal;

[0065] The correction module is used to acquire the environmental baseline signal of the area to be detected, and perform adaptive filtering, blind source separation and correction on the multimodal optical signal based on variational mode decomposition according to the environmental baseline signal to generate a net fluorescence signal. The environmental baseline signal is composed of the dark field noise background and the ambient stray light spectrum superimposed.

[0066] The decomposition module is used to construct and decompose the spectral data cube based on the net fluorescence signal, and extract the RGB three-color fluorescence intensity values;

[0067] The conversion module is used to generate a fluorescence distribution dot matrix based on the RGB three-color fluorescence intensity values, convert the fluorescence distribution dot matrix into a recursive graph matrix using recursive graph analysis, and identify the fungal infection area in the area to be detected based on the recursive graph matrix.

[0068] The clustering module is used to perform superpixel segmentation and spectral clustering on the fungal infection area to obtain multi-scale texture features and heterogeneous sub-region labels.

[0069] The identification module is used to identify the type and severity level of tinea capitis in the fungal infection area based on the recursive graph matrix, multi-scale texture features, and heterogeneous sub-region labels.

[0070] One or more technical solutions proposed in this application are applied to a scalp tinea detection device based on multi-channel fluorescence sensing, including a detection light generation module and a fluorescence recognition module. The device emits composite-band excitation light to the detection area of ​​the scalp and simultaneously acquires the fluorescence spectral signal and short-wave infrared diffuse reflectance signal generated in the detection area to construct a multimodal optical signal. It also acquires the environmental baseline signal of the detection area and performs adaptive filtering, blind source separation, and correction based on variational mode decomposition on the multimodal optical signal according to the environmental baseline signal to generate a net fluorescence signal. The environmental baseline signal is composed of dark-field noise background and environmental stray noise. The optical spectrum is superimposed; a spectral data cube is constructed and decomposed based on the net fluorescence signal to extract RGB three-color fluorescence intensity values; a fluorescence distribution dot matrix is ​​generated based on the RGB three-color fluorescence intensity values, and the fluorescence distribution dot matrix is ​​transformed into a recursive graph matrix using recursive graph analysis, and the fungal infection area in the detection area is identified based on the recursive graph matrix; superpixel segmentation and spectral clustering are performed on the fungal infection area to obtain multi-scale texture features and heterogeneous sub-region labels; the type and degree of tinea capitis in the fungal infection area are identified based on the recursive graph matrix, multi-scale texture features and heterogeneous sub-region labels. By using the above method to excite multimodal signals of the scalp with composite band excitation light and introducing variational mode decomposition and blind source separation techniques to denoise and correct the multimodal signals, the detection sensitivity of the fluorescence signal can be effectively improved, and the false negative rate can be greatly reduced. This method is suitable for early and atypical tinea capitis infections. Furthermore, by combining recursive graph analysis technology, the fungal infection area can be accurately identified, effectively avoiding misjudgment. Then, by using superpixel segmentation and spectral clustering techniques to obtain multi-scale texture features and heterogeneous sub-region labels, combined with the recursive graph matrix, the accurate classification of tinea capitis types and the quantitative assessment of infection degree can be achieved, effectively improving the detection efficiency and accuracy of tinea capitis. Attached Figure Description

[0071] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0072] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0073] Figure 1 This is a flowchart illustrating an embodiment of the method for detecting tinea capitis based on multi-channel fluorescence sensing in this application.

[0074] Figure 2This is a schematic diagram of the structure of a tinea capitis detection device based on a multi-channel fluorescence sensor, provided in an embodiment of the tinea capitis detection method based on multi-channel fluorescence sensing according to this application.

[0075] Figure 3 This is a schematic diagram of the module structure of the tinea capitis detection device based on multi-channel fluorescence sensing according to an embodiment of this application.

[0076] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0077] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0078] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0079] This application provides a solution that uses composite band excitation light to excite multimodal signals of the scalp and introduces variational mode decomposition and blind source separation techniques to denoise and correct the multimodal signals. This effectively improves the detection sensitivity of the fluorescence signal and greatly reduces the false negative rate, making it suitable for early and atypical tinea capitis infections. Furthermore, by combining recursive graph analysis technology, the fungal infection area can be accurately identified, effectively avoiding misjudgment. Then, by using superpixel segmentation and spectral clustering techniques to obtain multi-scale texture features and heterogeneous sub-region labels, combined with the recursive graph matrix, the accurate classification of tinea capitis types and the quantitative assessment of infection degree can be achieved, effectively improving the detection efficiency and accuracy of tinea capitis.

[0080] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of performing the above functions, such as a tinea capitis detection device based on multi-channel fluorescence sensing. The following description uses a tinea capitis detection device based on multi-channel fluorescence sensing as an example to illustrate this embodiment and the subsequent embodiments.

[0081] Based on this, embodiments of this application provide a method for detecting tinea capitis based on multi-channel fluorescence sensing, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the tinea capitis detection method based on multi-channel fluorescence sensing in this application.

[0082] In this embodiment, the method for detecting tinea capitis based on multi-channel fluorescence sensing includes steps S10~S60:

[0083] Step S10: Emit composite band excitation light to the area to be detected on the scalp, and simultaneously collect the fluorescence spectrum signal and short-wave infrared diffuse reflection signal generated in the area to be detected to construct a multimodal optical signal.

[0084] It should be noted that the scalp tinea detection device based on multi-channel fluorescence sensing in this embodiment has a handheld comb-like structure, which facilitates sliding scanning detection on the scalp surface. The device includes an acrylic comb tooth section 1, a detection light generation module 2, a fluorescence recognition module 3, a main frame 4, and a light shield 5. The comb tooth section 1 is made of acrylic material with a light transmittance of not less than 92%, and multiple light-guiding optical fibers are embedded inside for contacting the scalp and collecting light signals. The detection light generation module 2 is located on the inner side of the main frame 4 near the comb tooth section, ensuring that the light emission direction is directly facing the scalp area below the comb teeth. It includes multiple ultraviolet LEDs with a center wavelength of 365nm and a center wavelength of... A 1350nm short-wave infrared LED serves as the excitation source, emitting excitation light in a composite wavelength band. The fluorescence recognition module 3, located outside the main frame 4, is coupled to the output ends of each optical fiber and is used to receive and analyze fluorescence signals. The main frame 4, made of resin, is used to fix and connect the modules. A light shield 5, positioned above the fluorescence recognition module, serves not only for mechanical stability but also for light shielding and interference prevention, effectively blocking external visible light and ultraviolet scattering. This ensures the sensor only receives effective reflected signals from the detection area. The device also includes a gasket located between the comb teeth and the main frame 4, made of elastic silicone. Figure 2 As shown, Figure 2 This is a schematic diagram of a tinea capitis detection device based on multi-channel fluorescence sensing.

[0085] Understandably, upon detecting contact between the comb teeth and the scalp, the detection light generation module immediately activates, emitting composite-band excitation light towards the area to be detected on the scalp. This composite-band excitation light is formed by encoding and modulating ultraviolet light with a center wavelength of 365nm and short-wave infrared light with a center wavelength of 1350nm. This composite-band excitation light simultaneously excites the scalp tissue to produce reflected fluorescence, which is collected by multiple optical fibers embedded within the comb teeth. The optical fibers in the comb teeth use SOG-100 multi-component glass fiber as the transmission medium for reflected fluorescence. Each fiber has a diameter of approximately 0.6mm and a length customized to 40–60mm according to the comb structure. The fiber tips are embedded in the comb tooth tips and polished to form a smooth incident surface, enabling efficient collection of diffuse reflected fluorescence from the scalp surface at different incident angles. The rear end of the optical fiber is fixed to the optical guide interface of the recognition module via a high-precision alignment structure, and is vertically coupled to the incident window of the filter and spectral sensor, thereby achieving low-loss transmission of light energy. The multi-component glass core of SOG-100 optical fiber has stable transmittance in the range of 400–1650 nm, which can effectively ensure the spectral integrity of fluorescence signals.

[0086] During the detection process, multiple optical fibers are distributed in a parallel array inside the comb teeth, forming a multi-point sampling network with a spacing of approximately 0.6 mm. The ends of each fiber bundle converge to the sensor receiving surface through an optical guide interface, so that the fluorescence signals collected at different locations spatially correspond to the comb tooth distribution, realizing multi-point synchronous detection of the scalp area. Each fiber is equivalent to an independent optical sampling channel, sampling sequentially at a period of 200 ms. Within the sampling period, the fluorescence spectrum signal and short-wave infrared diffuse reflectance signal collected by each channel are recorded synchronously, and the signals from each channel are integrated into a time-series data stream through multiplexing technology, thereby constructing a complete multimodal optical signal.

[0087] Step S20: Obtain the environmental baseline signal of the area to be detected, and perform adaptive filtering, blind source separation and correction on the multimodal optical signal based on variational mode decomposition according to the environmental baseline signal to generate a net fluorescence signal, wherein the environmental baseline signal is composed of the dark field noise background and the ambient stray light spectrum superimposed.

[0088] It should be noted that the environmental baseline signal is the background signal collected in the area to be detected when the light source is not turned on. It mainly consists of two parts: one part is the dark field noise background, which is the inherent noise generated by the electronic components and sensors of the device itself under no light conditions; the other part is the ambient stray light spectrum, which is the spectral components generated by the light emitted by various uncontrollable light sources in the surrounding environment in the area to be detected.

[0089] Understandably, after acquiring the environmental baseline signal, variational mode decomposition (VMD) is used to adaptively decompose the multimodal optical signal into a series of intrinsic mode functions (EMFs) with different center frequencies and bandwidths. These EMFs can better characterize the local features of the signal, which is helpful for subsequent filtering. Next, a blind source separation algorithm is used to separate the decomposed EMFs. Since the blind source separation algorithm does not require prior information about the signal and can achieve signal separation based solely on the statistical characteristics of the observed signal, it can effectively remove environmental baseline signal interference from the multimodal optical signal, obtaining the net fluorescence signal.

[0090] Step S30: Construct and decompose the spectral data cube based on the net fluorescence signal, and extract the RGB three-color fluorescence intensity values.

[0091] It should be noted that the net fluorescence signal is transmitted to the fluorescence recognition module for the construction and decomposition of a spectral data cube, thereby extracting the RGB fluorescence intensity values. The fluorescence recognition module scans the net fluorescence signal point by point, acquiring fluorescence intensity data at different wavelengths, and arranges them in three-dimensional space according to wavelength, scan position, and time order to form a spectral data cube. Subsequently, a specific algorithm decomposes the cube, extracting the fluorescence intensity values ​​corresponding to the red (R), green (G), and blue (B) bands. These three intensity values ​​not only reflect the fluorescence characteristics of the scalp region at specific wavelengths but also provide basic data for the subsequent generation of fluorescence distribution dot matrix maps.

[0092] In one feasible implementation, step S30 may include: performing spectral processing on the net fluorescence signal to obtain net fluorescence signals in different bands; constructing a three-dimensional spectral data cube based on the net fluorescence signal of the target band in the net fluorescence signals of the different bands, and performing non-negative matrix decomposition on the three-dimensional spectral data cube to obtain a fundamental matrix and a coefficient matrix; reconstructing the three-dimensional spectral data cube according to the fundamental matrix and the coefficient matrix to obtain reconstructed spectral data, and extracting spectral feature vectors from the reconstructed spectral data; performing signal conditioning and amplification processing on the spectral feature vectors to obtain processed spectral feature vectors; and extracting RGB three-color fluorescence intensity values ​​from the processed spectral feature vectors to obtain RGB three-color fluorescence intensity values.

[0093] It should be noted that the fluorescence recognition module includes a light guide interface unit, a filter unit, a spectral sensing unit, and a signal conditioning and amplification unit. The light guide interface is located on the lower surface of the fluorescence recognition module and is tightly coupled to the fiber optic end from the comb teeth, ensuring efficient transmission and minimal loss of the fluorescence signal in the optical path. The fiber optic light guide path extends from the tip of the comb teeth to the interface position of the fluorescence recognition module, and its end face is micro-polished to ensure that the coupling efficiency with the spectral sensing window is consistently above 90%. The filter unit uses a high-performance long-pass filter with a cutoff wavelength of 200–390nm, a starting transmission wavelength of 400nm, and a transmission region covering 408–1650nm. It can completely block the main peak band of ultraviolet LED light (365nm) and ambient ultraviolet scattered light, allowing only the target fluorescence band to enter the spectral sensing region. The filter is embedded between the light shield and the light guide interface and is fixed by an anti-reflective coating and mechanical fastening, ensuring a stable transmission curve and low reflection interference even after repeated use and angle changes. The spectral sensing unit can accurately capture the target fluorescence band signal after it has been processed by the filtering unit, perform spectral sensing, and extract the spectral feature vector. The signal conditioning and amplification unit conditions and amplifies the weak signal output by the spectral sensing unit to ensure the accuracy and stability of the signal in subsequent processing.

[0094] Understandably, the net fluorescence signal is efficiently transmitted to the filter unit through the optical guide interface unit. The filter unit performs spectral processing on the net fluorescence signal, separating the net fluorescence signals of different wavelengths. The high-performance long-pass filter accurately blocks ultraviolet light and ambient ultraviolet scattered light, thus enabling the spectral sensing unit to allow only fluorescence signals that meet the set wavelength range to enter.

[0095] In its implementation, the spectral sensing unit arranges the net fluorescence signal of the target band in three-dimensional space according to wavelength, scanning position, and time order, forming a spectral data cube. A non-negative matrix factorization (NMF) algorithm decomposes this spectral data cube, rearranging it into a two-dimensional matrix. This two-dimensional matrix is ​​then approximately decomposed into the product of two non-negative matrices: one is the base matrix, representing the basic components of the spectral data cube, and the other is the coefficient matrix, describing the linear combination of these basic components in the original data. This decomposition allows for the extraction of key feature information from complex spectral data, enabling the reconstruction of the three-dimensional spectral data cube to obtain reconstructed spectral data. The reconstructed spectral data removes noise and redundant information, more accurately reflecting the fluorescence characteristics of the scalp region. Subsequently, spectral feature vectors are extracted from the reconstructed spectral data. These vectors contain fluorescence intensity information at different wavelengths. The signal conditioning and amplification unit conditions and amplifies the extracted spectral feature vectors. Through internal precision circuitry, it performs gain adjustment, filtering, and noise reduction on weak signals to ensure that the signal strength and purity meet the requirements of subsequent processing, resulting in processed spectral feature vectors. The spectral feature vector is weighted and integrated in the R, G, and B bands according to the standard color matching function of the International Commission on Illumination to obtain the corresponding RGB three-color fluorescence intensity values.

[0096] Step S40: Generate a fluorescence distribution dot matrix based on the RGB three-color fluorescence intensity values, convert the fluorescence distribution dot matrix into a recursive graph matrix using recursive graph analysis, and identify the fungal infection area in the area to be detected based on the recursive graph matrix.

[0097] It should be noted that a fluorescence distribution dot matrix is ​​generated based on the extracted RGB three-color fluorescence intensity values. This dot matrix is ​​presented in the form of a two-dimensional matrix, where each element represents the fluorescence intensity value of a specific location on the scalp in the G channel.

[0098] Recursive graph analysis constructs a matrix reflecting the dynamic characteristics of the system—the recursive graph matrix—by calculating the temporal or spatial recursive relationships between points in a point matrix. This recursive graph matrix is ​​then input into a fungal infection identification model. This model, based on deep learning algorithms and trained with a large amount of labeled fungal infection sample data, can accurately identify fungal infection areas.

[0099] In one feasible implementation, step S40 may include: determining a spectral bias parameter based on the RGB three-color fluorescence intensity values, and generating a fluorescence distribution dot matrix based on the spectral bias parameter and the RGB three-color fluorescence intensity values; performing space-filling curve encoding on the fluorescence distribution dot matrix according to the Hilbert curve to obtain a one-dimensional spatial sequence; reconstructing the phase space based on the one-dimensional spatial sequence to obtain multiple trajectory points in the phase space; calculating the Euclidean distance between the trajectory points based on the multiple trajectory points in the phase space to obtain a distance matrix; converting the distance matrix into a recursive graph matrix using a recursive quantization analysis method, wherein the elements in the recursive graph matrix represent the recursive state of the corresponding trajectory point pair under a given threshold; inputting the recursive graph matrix into a fungal infection identification model, and extracting features from the recursive graph matrix using a convolutional neural network in the fungal infection identification model to obtain a segmentation mask, wherein the convolutional neural network includes an encoder-decoder structure and a skip connection structure; segmenting the fluorescence distribution dot matrix using the segmentation mask to identify the fungal infection region in the region to be detected.

[0100] It should be noted that the spectral bias parameter reflects the direction of the shift in the center of the spectral energy distribution and the dominant component of the fluorescence color. The formula for calculating the spectral bias parameter S is:

[0101]

[0102] in, , , The values ​​are the fluorescence intensity values ​​for the R, G, and B channels, respectively. It should be a very small positive number to prevent the denominator from being zero.

[0103] It is worth noting that, through extensive sample calibration, significant differences in S-values ​​and RGB channel proportions were observed among different fungal infection types. If the S-value is less than -0.3 and the green channel proportion exceeds 60%, the main fluorescence peak is located at approximately 520 nm, appearing bright green, indicating tinea capitis, with common pathogens including *Microsporum canis* and *Microsporum gypseum*. When the S-value is between -0.3 and 0.0, and the green channel proportion is between 40% and 60%, while the red channel proportion is significantly higher, the main peak shifts to the 540–570 nm range, appearing dark green, corresponding to favism or superficial hair infection. If the S-value is between 0.0 and 0.4, and the red and green channel intensities are similar, the main fluorescence peak is approximately 580 nm, appearing yellowish-white or copper-orange, indicating tinea versicolor caused by *Malassezia*. When the S-value is greater than 0.5 and the red channel intensity exceeds 70%, the main peak is located at 620–640 nm, with coral red fluorescence, indicating infection by *Corynebacterium micranthum*, characteristic of erythroplasia. However, relying solely on the S-value and RGB channel ratio for preliminary judgment may still result in a certain false positive rate. To improve the accuracy of fungal infection area identification, this implementation method introduces recursive graph analysis, which can deeply explore the dynamic characteristics of the fluorescence distribution dot matrix.

[0104] Understandably, after obtaining the spectral bias parameters, the fluorescence intensity G of the G channel is determined based on the RGB three-color fluorescence intensity values. _c (i,j), based on the spectral bias parameter S and the fluorescence intensity G of the G channel. _c (i,j) generates a fluorescence distribution dot matrix D(i,j), where i and j represent the row and column indices in the dot matrix, respectively. The Hilbert curve, as a space-filling curve, can map a two-dimensional fluorescence distribution dot matrix into a one-dimensional spatial sequence. This encoding method preserves spatial neighborhood information in the dot matrix, ensuring that adjacent points in the one-dimensional sequence are also adjacent in the original two-dimensional dot matrix, providing a foundation for subsequent phase space reconstruction.

[0105] Phase space reconstruction is the process of reconstructing a one-dimensional spatial sequence into a higher-dimensional phase space using certain methods to reveal the intrinsic dynamic characteristics of the system. Based on the obtained one-dimensional spatial sequence, the delayed embedding theorem is used for phase space reconstruction. By selecting an appropriate delay time and embedding dimension, each point in the one-dimensional spatial sequence is expanded into a trajectory point in the higher-dimensional phase space according to certain rules. These trajectory points exhibit the dynamic evolution characteristics of the system in the higher-dimensional space.

[0106] Based on multiple trajectory points in the reconstructed phase space, the Euclidean distance between any two trajectory points is calculated. The Euclidean distance reflects the degree of separation of trajectory points in the phase space, thus yielding a distance matrix. This matrix contains the dynamic correlation information between points in the fluorescence distribution lattice map. Using recursive quantization analysis, the distance matrix is ​​transformed into a recursive graph matrix R, as shown in the following equation:

[0107]

[0108] in, Let i be the element in the i-th row and j-th column of the recursive graph matrix R. For a given threshold, and Let i and j be the i-th and j-th trajectory points in phase space, respectively. ||·|| represents the Euclidean distance, and Θ is the Heaviside step function, which takes the value 1 when the value in parentheses is greater than or equal to 0 and 0 when it is less than 0.

[0109] In the recursive graph matrix, each element represents the recursive state of the corresponding trajectory point pair under a given threshold. Specifically, it indicates whether the distance between two trajectory points in phase space is less than the threshold. If it is less than the threshold, it means that the two trajectory points have similar states during the dynamic evolution of the system, and the corresponding position in the recursive graph matrix is ​​marked as 1. If it is greater than or equal to the threshold, it means that the states of the two trajectory points are significantly different, and the corresponding position is marked as 0. The recursive graph matrix constructed in this way can intuitively present the dynamic recursive relationships between points in the fluorescence distribution lattice map.

[0110] The obtained recurrence graph matrix is ​​input into a pre-trained fungal infection identification model. This model's convolutional neural network (CNN) possesses powerful feature extraction capabilities, and its encoder-decoder structure automatically learns multi-level features from the recurrence graph matrix. The encoder extracts abstract features from the recurrence graph matrix stepwise through multiple convolutional and pooling layers, converting the original high-dimensional recurrence graph matrix into a low-dimensional feature representation. The decoder then remaps the low-dimensional features back to the high-dimensional space through deconvolutional and upsampling layers, achieving segmentation of the fluorescence distribution dot matrix. The introduction of skip connections allows the shallow features extracted by the encoder to be directly passed to the decoder, preserving more detailed information and improving segmentation accuracy. After feature extraction from the recurrence graph matrix using the CNN, a segmentation mask is obtained, which accurately identifies the location and extent of fungal infection regions in the fluorescence distribution dot matrix. Finally, the segmentation mask is used to segment the fluorescence distribution dot matrix, accurately identifying the fungal infection regions within the detection area and providing a reliable basis for the detection of tinea capitis.

[0111] Step S50: Perform superpixel segmentation and spectral clustering on the fungal infection area to obtain multi-scale texture features and heterogeneous sub-region labels.

[0112] It should be noted that superpixel segmentation is a technique that divides an image into multiple uniform regions with similar colors, textures, and other features. This reduces the computational load of subsequent processing while preserving important structural information of the image. In this embodiment, a simple linear iterative clustering algorithm is used for superpixel segmentation. This algorithm iteratively optimizes the clustering of image pixels into superpixels, where pixels within each superpixel have similar fluorescence properties.

[0113] Understandably, spectral clustering is a graph-based clustering method that treats data points as vertices in a graph. It constructs a similarity matrix by calculating the similarity between vertices and then uses the matrix's eigenvectors for clustering. After obtaining the superpixel segmentation results, a similarity matrix is ​​constructed between superpixels, reflecting the degree of similarity in fluorescence distribution, texture, etc. By performing eigenvalue decomposition on the similarity matrix, eigenvectors are obtained. These eigenvectors are then used for clustering, grouping superpixels with similar features into the same category, thus obtaining multi-scale texture features and heterogeneous sub-region labels.

[0114] It is worth noting that multi-scale texture features can describe the texture information of fungal infection areas at different scales. Texture features at different scales can capture various texture variations from microscopic to macroscopic, providing rich information for accurately identifying fungal infection types. Heterogeneous sub-region labels identify sub-regions with different characteristics in the fluorescence distribution dot matrix. These sub-regions may correspond to different types of fungal infections or different stages of infection development. Analysis of multi-scale texture features and heterogeneous sub-region labels can further deepen our understanding of the characteristics of fungal infection areas, providing strong support for accurate subsequent determination of fungal infection types.

[0115] In one feasible implementation, step S50 may include: performing superpixel segmentation on the fungal infection region using a simple linear iterative clustering algorithm to obtain multiple superpixel blocks; extracting multimodal texture features from each superpixel block to obtain gray-level co-occurrence matrix texture features, local binary mode variance features, and Gabor multidirectional texture response features for each superpixel block; concatenating the gray-level co-occurrence matrix texture features, local binary mode variance features, and Gabor multidirectional texture response features of each superpixel block to obtain a superpixel feature vector; constructing a similarity matrix based on the superpixel feature vector, performing cluster analysis on the similarity matrix using a spectral clustering algorithm to obtain multiple clusters; dividing the fungal infection region into multiple heterogeneous sub-regions according to the clusters, and assigning a corresponding label to each heterogeneous sub-region to obtain heterogeneous sub-region labels.

[0116] It should be noted that a simple linear iterative clustering algorithm is used to perform detailed superpixel segmentation of the fungal infection region. This process divides the region into multiple superpixel blocks, each of which contains a set of pixels with similar fluorescence properties and texture features as much as possible. For each superpixel block S... k Three types of texture features are extracted from its internal pixels, including gray-level co-occurrence matrix texture features, local binary mode variance features, and Gabor multi-directional texture response features.

[0117] For gray-level co-occurrence matrix texture features, construct the gray-level co-occurrence matrix. Commonly used directions θ∈{0°,45°,90°,135°}, distance d=1, gray levels L=8. Four statistical features are extracted: energy, contrast, correlation, and entropy, which together constitute the gray-level co-occurrence matrix texture features, including energy features. It reflects the uniformity of the image's grayscale distribution and contrast characteristics. It reflects the image's sharpness, texture depth, and correlation characteristics. Entropy features measure the similarity of elements in a gray-level co-occurrence matrix along the row or column direction. This characterizes the complexity of the texture in the image.

[0118] The formula for calculating energy is:

[0119]

[0120] The formula for calculating contrast is:

[0121]

[0122] The formula for calculating correlation is:

[0123]

[0124] The formula for calculating entropy is:

[0125]

[0126] in, , Let represent the mean of the rows and columns of the gray-level co-occurrence matrix, respectively. , Let C(x) represent the standard deviations of the rows and columns of the gray-level co-occurrence matrix, respectively. , ) represents the position of the gray-level co-occurrence matrix located in ( , The element value at position ).

[0127] The local binary pattern variance feature is the variance of the local binary pattern histogram within a superpixel. Assuming P=8 neighborhood and R=1, the local binary pattern value for each pixel block is calculated as follows:

[0128]

[0129] Where s(x) is a sign function, s(x)=1 when x≥0, and s(x)=0 when x<0. This represents the grayscale value of the nth pixel in the 8-neighborhood centered at pixel p. This represents the grayscale value of the center pixel p.

[0130] After calculating the local binary mode value for each pixel, the histogram H of the local binary mode within the entire superpixel block is plotted, and the variance of this histogram is calculated. This variance value is the local binary mode variance feature, which can reflect the local changes in texture within the superpixel block, as shown in the following formula:

[0131]

[0132] in, Represents superpixel block S k The total number of pixels in the histogram H, where H(v) represents the number of pixels with gray level v in histogram H. This represents the mean of the histogram H.

[0133] Gabor multi-directional texture response features are obtained by filtering superpixel blocks using a set of Gabor filters with different directions and frequencies. The Gabor kernel function is defined as follows, and convolution is performed according to the direction of the Gabor kernel function to calculate the superpixel S. k The Gabor multi-directional texture response characteristics can be obtained by taking the mean value of the internal response amplitude.

[0134] The extracted gray-level co-occurrence matrix texture features, local binary mode variance features, and Gabor multi-directional texture response features are concatenated to form superpixel feature vectors. A similarity matrix is ​​constructed based on the Euclidean distance between the feature vectors. Each element in the similarity matrix represents the degree of similarity between two superpixel blocks in terms of texture features. The smaller the Euclidean distance, the more similar the texture features of the two superpixel blocks are, and the larger the corresponding element value in the similarity matrix.

[0135] After constructing the similarity matrix, spectral clustering is used for cluster analysis. The spectral clustering algorithm first performs eigenvalue decomposition on the similarity matrix, obtaining a series of eigenvectors that reflect the distribution structure of the data in low-dimensional space. Then, based on these eigenvectors, superpixel blocks are divided into different clusters, ensuring high similarity among superpixel blocks within the same cluster and significant differences between superpixel blocks in different clusters. Through this clustering operation, the fungal infection area can be divided into multiple heterogeneous sub-regions with different texture features. Each heterogeneous sub-region corresponds to a specific fungal infection characteristic or stage, such as the core region, active diffusion region, satellite microfoci region, and recovery region.

[0136] Specifically, spectral clustering algorithms map high-dimensional texture feature data to a low-dimensional space by calculating the eigenvalues ​​and eigenvectors of the similarity matrix. In the low-dimensional space, the similarity relationships between data points are clearer, facilitating clustering. Based on the data distribution structure reflected by the eigenvectors, the algorithm groups superpixel blocks with high similarity into the same cluster, while assigning superpixel blocks with low similarity to different clusters. After obtaining multiple clusters, the fungal infection area can be divided into multiple heterogeneous sub-regions based on these clusters.

[0137] Step S60: Based on the recursive graph matrix, multi-scale texture features, and heterogeneous sub-region labels, identify the type and severity level of tinea capitis in the fungal infection area.

[0138] It should be noted that when identifying the type and severity of tinea capitis in fungal infections, the recursive graph matrix can capture the dynamic changes in fluorescence signals within the infected area. Different types of tinea capitis exhibit different dynamic patterns in their fluorescence signals. For example, some types may cause periodic fluctuations in fluorescence signals, while others may show irregular changes. By analyzing specific patterns in the recursive graph matrix, the range of the tinea capitis type can be preliminarily determined. Multi-scale texture features describe the texture information of the fungal infection area at different scales. Different types of tinea capitis exhibit unique texture features at different scales. Heterogeneous sub-region labels identify sub-regions with different characteristics in the fluorescence distribution dot matrix. These sub-regions are closely related to the type and severity of tinea capitis. Different types of tinea capitis may produce different fluorescence characteristics in different sub-regions, and the higher the severity of infection, the more pronounced the characteristics of heterogeneous sub-regions may be. For example, the core area may correspond to severe infection, and its fluorescence characteristics differ significantly from other sub-regions; the active diffusion area may indicate that the infection is spreading, and its texture and fluorescence characteristics are also unique.

[0139] This study integrates information from three aspects: the recurrent graph matrix, multi-scale texture features, and heterogeneous sub-region labels, and employs a pre-trained classification model for identification. This model is trained on a large dataset of known tinea capitis types and infection severity levels, learning the mapping relationships between different types and severity levels and the aforementioned features. During the identification process, the recurrent graph matrix, multi-scale texture features, and heterogeneous sub-region labels of the area to be detected as fungal infection are input into the classification model. Based on the learned mapping relationships, the model outputs the corresponding tinea capitis type and infection severity level, thereby achieving accurate identification of the fungal infection area.

[0140] In one feasible implementation, step S60 may include: constructing a multimodal feature map based on the recursive graph matrix, the multi-scale texture features, and the heterogeneous sub-region labels; inputting the multimodal feature map into a multimodal discriminant model based on a graph attention network to obtain a classification confidence score, wherein the multimodal discriminant model is pre-trained through transfer learning and its parameters are optimized using a backpropagation algorithm; identifying the type of tinea capitis in the fungal infection area based on the classification confidence score; acquiring fluorescence point cloud data of the fungal infection area and determining the lesion area proportion and fluorescence heterogeneity index of the fungal infection area based on the fluorescence point cloud data of the fungal infection area; calculating a comprehensive infection severity assessment index based on the lesion area proportion, the fluorescence heterogeneity index, and the classification confidence score; and determining the infection severity level based on the comprehensive infection severity assessment index.

[0141] It should be noted that when constructing the multimodal feature map, the dynamic change information of fluorescence signal contained in the recursive graph matrix, the texture detail information at different scales depicted by the multi-scale texture features, and the information of different characteristic sub-regions identified by the heterogeneous sub-region labels are effectively fused to form a comprehensive and complete feature representation. This multimodal feature map can describe the fungal infection area from multiple dimensions, providing rich feature basis for subsequent accurate identification.

[0142] A multimodal discrimination model based on graph attention networks, through transfer learning pre-training, can quickly adapt to the current task of tinea capitis detection by leveraging knowledge learned from other related tasks or data. Based on pre-training, the model parameters are continuously optimized using a backpropagation algorithm, enabling the model to better fit the training data and improve its ability to distinguish different types and degrees of tinea capitis. After inputting the multimodal feature map into the model, it outputs a classification confidence score, which reflects the model's confidence in that the input feature map belongs to a certain type of tinea capitis.

[0143] Identifying the type of tinea capitis in the fungal infection area based on classification confidence can accurately determine the specific type of tinea capitis suffered by the patient. After obtaining the fluorescence point cloud data of the fungal infection area, the lesion area proportion (i.e., the proportion of the lesion area in the entire detection area) and the fluorescence heterogeneity index (which reflects the degree of uneven distribution of fluorescence signal within the lesion area) can be determined through analysis and processing of the fluorescence point cloud data.

[0144] A comprehensive infection severity assessment index is calculated based on the proportion of lesion area, fluorescence heterogeneity index, and classification confidence level. This index comprehensively considers factors such as lesion size, fluorescence signal distribution, and characteristics of tinea capitis, enabling a more comprehensive and accurate assessment of the severity of fungal infection. Finally, the infection severity level is determined based on the comprehensive infection severity assessment index, for example, it can be divided into different levels such as mild, moderate, and severe.

[0145] In one feasible implementation, the step of inputting the multimodal feature map into a multimodal discriminant model based on a graph attention network to obtain classification confidence includes: decomposing the multimodal feature map into a node feature matrix and an adjacency matrix, wherein the node feature matrix contains fusion information of recursive graph matrix features, multi-scale texture features, and heterogeneous sub-region labels, and the adjacency matrix represents the spatial correlation between nodes; calculating the attention weights between node pairs through the attention mechanism of the graph attention network in the multimodal discriminant model; performing weighted aggregation on the node feature matrix based on the attention weights to obtain a weighted feature vector; inputting the weighted feature vector into the fully connected layer of the multimodal discriminant model for nonlinear transformation to output the initial classification probability of each tinea capitis type; and performing Softmax normalization on the initial classification probability to obtain the classification confidence of each tinea capitis type.

[0146] It should be noted that when the multimodal feature map is decomposed into a node feature matrix and an adjacency matrix, the node feature matrix integrates the dynamic characteristics of fluorescence signals reflected by the recursive graph matrix, the different scale texture details presented by the multi-scale texture features, and the information of different characteristic sub-regions identified by the heterogeneous sub-region labels. These information complement each other, providing a comprehensive data foundation for subsequent analysis. The adjacency matrix, by quantifying the spatial correlation between nodes, reveals the interrelationships between different sub-regions within the fungal infection area, which helps to more accurately understand the internal structure of the infection area.

[0147] Graph attention networks in multimodal discriminative models calculate attention weights between node pairs through an attention mechanism. This process can automatically identify and highlight node relationships that are more important to the classification task. For example, in some types of tinea capitis, the interactions between specific subregions may be more significant. The attention mechanism can capture this importance and assign higher weights to the corresponding node pairs.

[0148] The node feature matrix is ​​weighted and aggregated based on attention weights to obtain a weighted feature vector. This step further strengthens the role of key information in feature representation. The weighted feature vector integrates information from different nodes and highlights features that have a key impact on classification decisions.

[0149] The weighted feature vector is input into the fully connected layer of the multimodal discriminant model for nonlinear transformation. The fully connected layer learns complex nonlinear relationships to map the weighted feature vector to the classification space and outputs the initial classification probability of each type of tinea capitis. These initial probabilities reflect the model's preliminary judgment that the input features belong to different types of tinea capitis.

[0150] The initial classification probabilities are normalized using Softmax to obtain the classification confidence score for each type of tinea capitis. The Softmax function transforms the initial probabilities into a probability distribution, ensuring that the sum of the probabilities of all categories is 1, thus providing a more intuitive and easily interpretable classification result. The classification confidence score clearly indicates the model's predictive reliability for each type of tinea capitis.

[0151] This embodiment provides a method for detecting tinea capitis based on multi-channel fluorescence sensing, applied to a tinea capitis detection device based on multi-channel fluorescence sensing. It includes a detection light generation module and a fluorescence recognition module. The method emits a composite wavelength excitation light towards the detection area of ​​the scalp and simultaneously acquires the fluorescence spectral signal and short-wave infrared diffuse reflectance signal generated in the detection area to construct a multimodal optical signal. It then acquires the environmental baseline signal of the detection area and performs adaptive filtering, blind source separation, and correction based on variational mode decomposition on the multimodal optical signal to generate a net fluorescence signal. The environmental baseline signal is composed of a dark field noise background. The spectrum is superimposed with the ambient stray light spectrum; a spectral data cube is constructed and decomposed based on the net fluorescence signal to extract RGB three-color fluorescence intensity values; a fluorescence distribution dot matrix is ​​generated based on the RGB three-color fluorescence intensity values, and the fluorescence distribution dot matrix is ​​transformed into a recursive graph matrix using recursive graph analysis, and the fungal infection area in the area to be detected is identified based on the recursive graph matrix; superpixel segmentation and spectral clustering are performed on the fungal infection area to obtain multi-scale texture features and heterogeneous sub-region labels; the type and degree of tinea capitis in the fungal infection area are identified based on the recursive graph matrix, multi-scale texture features and heterogeneous sub-region labels. By using the above method to excite multimodal signals of the scalp with composite band excitation light and introducing variational mode decomposition and blind source separation techniques to denoise and correct the multimodal signals, the detection sensitivity of the fluorescence signal can be effectively improved, and the false negative rate can be greatly reduced. This method is suitable for early and atypical tinea capitis infections. Furthermore, by combining recursive graph analysis technology, the fungal infection area can be accurately identified, effectively avoiding misjudgment. Then, by using superpixel segmentation and spectral clustering techniques to obtain multi-scale texture features and heterogeneous sub-region labels, combined with the recursive graph matrix, the accurate classification of tinea capitis types and the quantitative assessment of infection degree can be achieved, effectively improving the detection efficiency and accuracy of tinea capitis.

[0152] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description and will not be repeated hereafter. Based on this, step S20 includes steps S201 to S207:

[0153] Step S201: Without emitting excitation light in the composite band, acquire the background noise signal of the dark field and the ambient stray light spectrum signal around the area to be detected.

[0154] It should be noted that the background noise signal in the dark field and the stray light spectrum signal around the area to be tested were collected under specific conditions where no excitation light of the composite band was emitted. The background noise signal in the dark field reflects the noise level of the detection equipment itself when there is no external excitation light interference. By collecting the background noise signal in the dark field, we can understand the noise characteristics of the equipment when there is no external light input, including electronic noise, thermal noise, etc. The stray light spectrum signal around the area to be tested reflects the interference of other light sources in the detection environment on the detection results, such as ambient light and natural light reflection.

[0155] Understandably, the background noise signal in the dark field can be acquired using the noise acquisition unit in the detection light generation module when the detection device is in a state where the excitation light emission function is turned off. This noise acquisition unit has high sensitivity and low noise characteristics, and can accurately capture the weak noise signals generated by the device itself. For the ambient stray light spectrum signal around the area to be detected, it is acquired using the ambient light acquisition submodule in the fluorescence recognition module. This submodule is equipped with a wide-spectrum response sensor that covers the visible light to short-wave infrared bands, and can comprehensively acquire various stray light signals in the environment.

[0156] Step S202: The dark field noise background signal and the environmental stray light spectrum signal are superimposed to obtain the environmental baseline signal.

[0157] It should be noted that the core purpose of superimposing the dark field noise background signal with the environmental stray light spectrum signal is to construct an environmental baseline signal that can comprehensively reflect the basic noise characteristics of the detection environment.

[0158] Understandably, signal processing algorithms can be used to precisely align and superimpose the two signals in the frequency or time domain, ensuring that the environmental baseline signal accurately covers the interference from the device's own noise and stray light. This environmental baseline signal will serve as an important reference for subsequent signal correction, effectively eliminating background noise components in multimodal optical signals, thereby improving the purity of the net fluorescence signal and detection sensitivity. This processing method maintains a high-precision capture capability for weak fluorescence signals even in complex detection environments.

[0159] Step S203: Perform variational mode decomposition on the multimodal optical signal to obtain multiple intrinsic modes.

[0160] It should be noted that variational mode decomposition (VMD) of multimodal optical signals is an advanced signal processing method. Its core idea is to decompose the complex multimodal signal X(t) into a series of eigenmode functions u with different frequency characteristics. k (t).

[0161] In its implementation, variational mode decomposition (VMD) automatically determines the center frequency and bandwidth parameters of each intrinsic mode by constructing and solving a constrained variational problem. This process essentially involves adaptively dividing the signal into frequency bands, ensuring that the resulting intrinsic modes do not overlap in the frequency domain while fully preserving all the characteristic information of the original signal, as shown in the following equation:

[0162]

[0163]

[0164] in, ω represents the k-th eigenmode function. k Represents the center frequency of the k-th eigenmode. For the Dirac function, This represents the convolution operation.

[0165] Understandably, the variational mode decomposition process is implemented through an iterative optimization algorithm. In each iteration, the center frequency and bandwidth parameters of the eigenmode functions are continuously adjusted, causing the decomposition result to gradually approach the optimal solution. Compared to the traditional fixed frequency band division method, this adaptive decomposition method can better adapt to the dynamic characteristics of the signal, especially for multimodal optical signals containing multiple frequency components and complex frequency distributions, such as those used for tinea pedis detection.

[0166] Step S204: Adaptive filtering is performed on the multiple intrinsic modes based on the environmental baseline signal to obtain the filtered multimodal optical signal.

[0167] It should be noted that the core of adaptive filtering of multiple intrinsic modes based on the environmental baseline signal lies in using prior knowledge, namely the dark field noise background Nd(t) and the environmental stray light spectrum Es(t), to identify and remove noise components from the modes decomposed by variational mode.

[0168] In one feasible implementation, step S204 may include: calculating the power spectral density based on the environmental baseline signal to determine a first power spectral density corresponding to the dark field noise background and a second power spectral density corresponding to the environmental stray light spectrum; constructing a priori noise fingerprint knowledge base based on the first and second power spectral densities; determining a third power spectral density for each intrinsic mode and performing a similarity calculation between the third power spectral density and the power spectral density in the priori noise fingerprint knowledge base to obtain a similarity index; labeling intrinsic modes whose similarity index reaches a preset threshold as noise modes and removing the noise modes from multiple intrinsic modes to obtain a filtered intrinsic mode set; and reconstructing the filtered intrinsic mode set to generate a filtered multimodal optical signal.

[0169] It should be noted that calculating the power spectral density based on the environmental baseline signal is to accurately quantify the energy distribution characteristics of the dark field noise background and the ambient stray light spectrum. The first power spectral density reflects the energy distribution of the device's own noise at different frequencies, while the second power spectral density reveals the interference characteristics of ambient stray light on the detection signal. By constructing a priori noise fingerprint knowledge base, these known noise characteristics can be stored and classified, providing a basis for subsequent noise identification.

[0170] After determining the third power spectral density of each intrinsic mode, similarity calculations can quantify the degree of matching between each intrinsic mode and the prior noise fingerprint. Intrinsic modes with a similarity index reaching a preset threshold are identified as noise modes. This process effectively distinguishes between real components and noise interference in the signal. The filtered intrinsic mode set obtained after removing noise modes eliminates most of the influence of environmental noise and inherent equipment noise.

[0171] Finally, the filtered intrinsic mode set is reconstructed. An inverse transform is used to recombine the decomposed modes into a complete multimodal optical signal, achieving filtering. This step ensures the integrity and continuity of the signal while significantly improving the signal-to-noise ratio. The generated filtered multimodal optical signal provides a cleaner data foundation for subsequent blind source separation and feature extraction, thereby improving the accuracy and reliability of the entire detection system.

[0172] Step S205: Perform blind source separation on the filtered multimodal optical signal to obtain the target fluorescence signal.

[0173] It should be noted that blind source separation of filtered multimodal optical signals is a technique for separating independent source signals from mixed signals in the absence of prior information. In the multimodal optical signal processing for tinea capitis detection, blind source separation can further remove interfering components from the signal and accurately extract the target fluorescence signal.

[0174] Understandably, blind source separation algorithms are based on the assumption of statistical independence of signals. They seek a linear transformation matrix that makes the transformed signal components as independent as possible. In this embodiment, the blind source separation algorithm can be Independent Component Analysis (ICA). By analyzing the statistical properties of the signal and utilizing features such as the non-Gaussianity of the signal, the separation matrix is ​​iteratively optimized, thereby decomposing the mixed signal into multiple independent source signals.

[0175] In one feasible implementation, step S204 may include: acquiring the time delay signal corresponding to the filtered multimodal optical signal; constructing an observation signal based on the filtered multimodal optical signal and the corresponding time delay signal; centering and whitening the observation signal to obtain a whitened signal; performing independent component analysis on the whitened signal and iteratively solving the separation matrix using an optimization algorithm; performing a linear transformation on the whitened signal according to the separation matrix to obtain multiple independent source signals; and selecting a signal that matches the fluorescence characteristics from the multiple independent source signals as the target fluorescence signal.

[0176] It should be noted that obtaining the time-delay signal corresponding to the filtered multimodal optical signal is to enrich the information dimension of the signal. Because the time-delay signal contains the characteristics of the original signal under different time delays, combining it with the original signal to construct the observation signal can provide more comprehensive signal information, which helps to more accurately separate independent source signals in the subsequent process.

[0177] The observed signal is centered and whitened. Centering is a mean-removal process that reduces the signal mean to zero and eliminates the DC component. Whitening, on the other hand, is a linear transformation that converts the signal into a unit variance, uncorrelated whitened signal. This process simplifies the complexity of subsequent independent component analysis and improves the separation effect.

[0178] When performing independent component analysis on the whitened signal, the optimization algorithm iteratively adjusts the parameters of the separation matrix to achieve optimal statistical independence between the transformed signal components. Finally, a linear transformation is performed on the whitened signal based on the solved separation matrix to obtain multiple independent source signals. Signals matching the fluorescence characteristics are selected from these independent source signals as target fluorescence signals, effectively extracting key feature information for tinea capitis detection.

[0179] Step S206: Determine the environmental interference weight based on the environmental baseline signal, and perform differential correction on the target fluorescence signal according to the environmental interference weight and the environmental baseline signal to obtain the net fluorescence signal.

[0180] It should be noted that the environmental interference weight is used to quantify the degree of interference of environmental factors on the target fluorescence signal. In the actual scenario of tinea capitis detection, environmental factors such as ambient light and equipment noise will inevitably affect the target fluorescence signal, causing signal distortion or deviation from the true value. By determining the environmental interference weight based on the environmental baseline signal, the specific impact of these environmental factors on the target fluorescence signal can be assessed more accurately.

[0181] Understandably, after obtaining the environmental interference weights, they are combined with the environmental baseline signal to perform differential correction on the target fluorescence signal. The differential correction process is essentially a signal adjustment process; it subtracts the signal deviation caused by environmental interference to obtain a purer, more accurate net fluorescence signal, as shown in the following equation:

[0182] R c =R0-W*R b

[0183] G c =G0-W*G b

[0184] B c =B0-W*B b

[0185] Among them, R c G c B c R0, G0, and B0 represent the net fluorescence signals of the R, G, and B channels, respectively; R0, G0, and B0 represent the original target fluorescence signals of the R, G, and B channels, respectively; W represents the environmental interference weight; and R0 represents the net fluorescence signals of the R, G, and B channels, respectively. b G b B b These are the environmental baseline signals for the R, G, and B channels, respectively.

[0186] This differential correction method can effectively eliminate the interference of environmental factors on the target fluorescence signal, thereby improving the accuracy and reliability of the signal.

[0187] In this embodiment, variational mode decomposition technology is introduced to preprocess multimodal optical signals, and noise is effectively eliminated by combining a priori noise fingerprint knowledge base. Then, the target fluorescence signal is accurately extracted by blind source separation algorithm, and finally the net fluorescence signal is obtained by differential correction. This effectively improves the accuracy and reliability of the signal and further enhances the accuracy of tinea capitis detection.

[0188] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the tinea capitis detection method based on multi-channel fluorescence sensing. Any simple modifications based on this technical concept are within the scope of protection of this application.

[0189] This application also provides a device for detecting tinea capitis based on multi-channel fluorescence sensing; please refer to [reference needed]. Figure 3 The tinea capitis detection device based on multi-channel fluorescence sensing includes:

[0190] The emission module 10 is used to emit composite band excitation light to the area to be detected on the scalp, and simultaneously collect the fluorescence spectrum signal and short-wave infrared diffuse reflection signal generated by the area to be detected to construct a multimodal optical signal.

[0191] The correction module 20 is used to acquire the environmental baseline signal of the area to be detected, and perform adaptive filtering, blind source separation and correction on the multimodal optical signal based on variational mode decomposition according to the environmental baseline signal to generate a net fluorescence signal. The environmental baseline signal is composed of the dark field noise background and the ambient stray light spectrum superimposed.

[0192] The decomposition module 30 is used to construct and decompose the spectral data cube based on the net fluorescence signal and extract the RGB three-color fluorescence intensity values.

[0193] The conversion module 40 is used to generate a fluorescence distribution dot matrix based on the RGB three-color fluorescence intensity values, convert the fluorescence distribution dot matrix into a recursive graph matrix using recursive graph analysis, and identify the fungal infection area in the area to be detected based on the recursive graph matrix.

[0194] Clustering module 50 is used to perform superpixel segmentation and spectral clustering on the fungal infection area to obtain multi-scale texture features and heterogeneous sub-region labels.

[0195] The identification module 60 is used to identify the type and degree of tinea capitis in the fungal infection area based on the recursive graph matrix, multi-scale texture features and heterogeneous sub-region labels.

[0196] The tinea capitis detection device based on multi-channel fluorescence sensing provided in this application, employing the tinea capitis detection method based on multi-channel fluorescence sensing in the above embodiments, can solve the technical problems of low efficiency and poor accuracy in traditional tinea capitis detection methods. Compared with the prior art, the beneficial effects of the tinea capitis detection device based on multi-channel fluorescence sensing provided in this application are the same as those of the tinea capitis detection method based on multi-channel fluorescence sensing provided in the above embodiments, and other technical features in the tinea capitis detection device based on multi-channel fluorescence sensing are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0197] The above are only some embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the content of this application specification and drawings, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for detecting tinea capitis based on multi-channel fluorescence sensing, characterized in that, The method includes: A composite wavelength excitation light is emitted toward the area to be detected on the scalp, and the fluorescence spectrum signal and short-wave infrared diffuse reflection signal generated in the area to be detected are collected simultaneously to construct a multimodal optical signal; The environmental baseline signal of the area to be detected is acquired, and the multimodal optical signal is subjected to adaptive filtering, blind source separation and correction based on variational mode decomposition according to the environmental baseline signal to generate a net fluorescence signal. The environmental baseline signal is composed of the dark field noise background and the ambient stray light spectrum superimposed. Based on the net fluorescence signal, a spectral data cube is constructed and decomposed to extract the RGB three-color fluorescence intensity values; A fluorescence distribution dot matrix is ​​generated based on the RGB three-color fluorescence intensity values. The fluorescence distribution dot matrix is ​​then converted into a recursive graph matrix using recursive graph analysis. Based on the recursive graph matrix, fungal infection areas in the area to be detected are identified. Superpixel segmentation and spectral clustering were performed on the fungal infection area to obtain multi-scale texture features and heterogeneous sub-region labels; Based on the recursive graph matrix, multi-scale texture features, and heterogeneous sub-region labels, the type and severity level of tinea capitis in the fungal infection area were identified.

2. The method as described in claim 1, characterized in that, The process of acquiring the environmental baseline signal of the area to be detected, and performing adaptive filtering, blind source separation, and correction based on variational mode decomposition on the multimodal optical signal according to the environmental baseline signal to generate a net fluorescence signal, includes: Without emitting excitation light in the composite band, the background noise signal of the dark field and the ambient stray light spectrum signal around the area to be detected are collected. The dark field noise background signal is superimposed with the environmental stray light spectrum signal to obtain the environmental baseline signal; Variational mode decomposition is performed on the multimodal optical signal to obtain multiple intrinsic modes; Based on the environmental baseline signal, the multiple intrinsic modes are adaptively filtered to obtain the filtered multimodal optical signal; Blind source separation is performed on the filtered multimodal optical signal to obtain the target fluorescence signal; An environmental interference weight is determined based on the environmental baseline signal, and the target fluorescence signal is differentially corrected according to the environmental interference weight and the environmental baseline signal to obtain a net fluorescence signal.

3. The method as described in claim 2, characterized in that, The adaptive filtering process performed on the multiple intrinsic modes based on the environmental baseline signal to obtain the filtered multimodal optical signal includes: Based on the environmental baseline signal, the power spectral density is calculated to determine the first power spectral density corresponding to the dark field noise background and the second power spectral density corresponding to the environmental stray light spectrum. A priori noise fingerprint knowledge base is constructed based on the first power spectral density and the second power spectral density. The third power spectral density of each intrinsic mode is determined, and the similarity between the third power spectral density and the power spectral density in the prior noise fingerprint knowledge base is calculated to obtain a similarity index. The intrinsic modes that reach the preset threshold of the similarity index are identified as noise modes, and the noise modes are removed from multiple intrinsic modes to obtain a filtered intrinsic mode set. The filtered intrinsic mode set is reconstructed to generate a filtered multimodal optical signal.

4. The method as described in claim 2, characterized in that, The step of performing blind source separation on the filtered multimodal optical signal to obtain the target fluorescence signal includes: Obtain the time delay signal corresponding to the filtered multimodal optical signal, and construct an observation signal based on the filtered multimodal optical signal and the corresponding time delay signal; The observed signal is centered and whitened to obtain a whitened signal; Independent component analysis is performed on the whitened signal, and the separation matrix is ​​solved iteratively by optimizing the algorithm; The whitened signal is linearly transformed according to the separation matrix to obtain multiple independent source signals; The signal that matches the fluorescence characteristics is selected from the multiple independent source signals as the target fluorescence signal.

5. The method as described in claim 1, characterized in that, The step of constructing and decomposing a spectral data cube based on the net fluorescence signal, and extracting RGB three-color fluorescence intensity values, includes: The net fluorescence signal is subjected to spectral processing to obtain net fluorescence signals in different wavelength bands; A three-dimensional spectral data cube is constructed based on the net fluorescence signal of the target band in the net fluorescence signals of the different bands, and the three-dimensional spectral data cube is decomposed into a non-negative matrix to obtain the fundamental matrix and the coefficient matrix. The three-dimensional spectral data cube is reconstructed based on the fundamental matrix and coefficient matrix to obtain the reconstructed spectral data, and spectral feature vectors are extracted from the reconstructed spectral data. The spectral feature vector is subjected to signal conditioning and amplification to obtain the processed spectral feature vector; The RGB three-color fluorescence intensity values ​​are extracted from the processed spectral feature vector to obtain the RGB three-color fluorescence intensity values.

6. The method as described in claim 1, characterized in that, The process of generating a fluorescence distribution dot matrix based on the RGB three-color fluorescence intensity values, converting the fluorescence distribution dot matrix into a recursive graph matrix using recursive graph analysis, and identifying fungal infection areas in the region to be detected based on the recursive graph matrix includes: The spectral bias parameter is determined based on the RGB three-color fluorescence intensity values, and a fluorescence distribution dot matrix is ​​generated based on the spectral bias parameter and the RGB three-color fluorescence intensity values; The fluorescence distribution dot matrix is ​​spatially filled using Hilbert curves to obtain a one-dimensional spatial sequence. Based on the one-dimensional spatial sequence, the phase space is reconstructed to obtain multiple trajectory points in the phase space; The Euclidean distance between multiple trajectory points in the phase space is calculated to obtain the distance matrix. The distance matrix is ​​transformed into a recursive graph matrix by recursive quantization analysis, where the elements in the recursive graph matrix represent the recursive state of the corresponding trajectory point pair under a given threshold. The recursive graph matrix is ​​input into the fungal infection identification model, and the recursive graph matrix is ​​used to extract features through the convolutional neural network in the fungal infection identification model to obtain a segmentation mask. The convolutional neural network includes an encoder-decoder structure and a skip connection structure. The fluorescence distribution dot matrix is ​​segmented using the segmentation mask to identify fungal infection areas within the region to be detected.

7. The method as described in claim 1, characterized in that, The process of performing superpixel segmentation and spectral clustering on the fungal infection region to obtain multi-scale texture features and heterogeneous sub-region labels includes: The fungal infection region was segmented into multiple superpixel blocks using a simple linear iterative clustering algorithm. Multimodal texture feature extraction is performed on each superpixel block to obtain the gray-level co-occurrence matrix texture features, local binary mode variance features, and Gabor multi-directional texture response features of each superpixel block; The gray-level co-occurrence matrix texture features, local binary mode variance features, and Gabor multi-directional texture response features of each superpixel block are concatenated to obtain the superpixel feature vector. A similarity matrix is ​​constructed based on the superpixel feature vectors, and a spectral clustering algorithm is used to perform cluster analysis on the similarity matrix to obtain multiple clusters; The fungal infection area is divided into multiple heterogeneous sub-regions based on the clusters, and a corresponding label is assigned to each heterogeneous sub-region to obtain heterogeneous sub-region labels.

8. The method as described in claim 1, characterized in that, The method identifies the type and severity of tinea capitis in the fungal infection area based on the recursive graph matrix, multi-scale texture features, and heterogeneous sub-region labels, including... A multimodal feature map is constructed based on the recursive graph matrix, the multi-scale texture features, and the heterogeneous sub-region labels; The multimodal feature map is input into a multimodal discrimination model based on graph attention network to obtain classification confidence. The multimodal discrimination model is pre-trained by transfer learning and its parameters are optimized by backpropagation algorithm. Based on the classification confidence level, the type of tinea capitis in the fungal infection area was identified; Obtain fluorescence point cloud data of the fungal infection area, and determine the lesion area ratio and fluorescence heterogeneity index of the fungal infection area based on the fluorescence point cloud data of the fungal infection area; A comprehensive infection severity assessment index is calculated based on the proportion of lesion area in the fungal infection region, the fluorescence heterogeneity index, and the classification confidence level. The level of infection severity is determined based on the comprehensive infection severity assessment index.

9. The method as described in claim 8, characterized in that, The step of inputting the multimodal feature map into a multimodal discrimination model based on a graph attention network to obtain classification confidence includes: The multimodal feature map is decomposed into a node feature matrix and an adjacency matrix. The node feature matrix contains fusion information of recursive graph matrix features, multi-scale texture features, and heterogeneous sub-region labels. The adjacency matrix represents the spatial correlation between nodes. The attention weights between node pairs are calculated using the attention mechanism of the graph attention network in the multimodal discriminant model. The node feature matrix is ​​weighted and aggregated based on the attention weights to obtain a weighted feature vector. The weighted feature vector is input into the fully connected layer of the multimodal discrimination model for nonlinear transformation, and the initial classification probability of each type of tinea capitis is output. The initial classification probabilities are normalized using Softmax to obtain the classification confidence scores for each type of tinea capitis.

10. A device for detecting tinea capitis based on multi-channel fluorescence sensing, characterized in that, The tinea capitis detection device based on multi-channel fluorescence sensing includes: The emission module is used to emit composite band excitation light to the area to be detected on the scalp, and simultaneously acquire the fluorescence spectrum signal and short-wave infrared diffuse reflection signal generated in the area to be detected to construct a multimodal optical signal; The correction module is used to acquire the environmental baseline signal of the area to be detected, and perform adaptive filtering, blind source separation and correction on the multimodal optical signal based on variational mode decomposition according to the environmental baseline signal to generate a net fluorescence signal. The environmental baseline signal is composed of the dark field noise background and the ambient stray light spectrum superimposed. The decomposition module is used to construct and decompose the spectral data cube based on the net fluorescence signal, and extract the RGB three-color fluorescence intensity values; The conversion module is used to generate a fluorescence distribution dot matrix based on the RGB three-color fluorescence intensity values, convert the fluorescence distribution dot matrix into a recursive graph matrix using recursive graph analysis, and identify the fungal infection area in the area to be detected based on the recursive graph matrix. The clustering module is used to perform superpixel segmentation and spectral clustering on the fungal infection area to obtain multi-scale texture features and heterogeneous sub-region labels. The identification module is used to identify the type and severity level of tinea capitis in the fungal infection area based on the recursive graph matrix, multi-scale texture features, and heterogeneous sub-region labels.

Images