An AI vision recognition-based vocal cord targeting automatic drug administration judgment method and system

Abstract

This invention provides a method and system for automatically targeting and administering medication to the vocal cords based on AI visual recognition, belonging to the field of artificial intelligence technology. The method includes: real-time acquisition of laryngeal video image data; image enhancement and wavelet transform denoising preprocessing of the laryngeal video image data to obtain a preprocessed laryngeal image; coarse segmentation of the vocal cord candidate region using a trained convolutional neural network AI visual recognition model to obtain an initial vocal cord region mask; fine extraction of vocal cord edges using a preset active contour model based on the initial vocal cord region mask, construction of a geometric energy functional, and iterative optimization of the geometric energy functional to obtain a sub-pixel-level vocal cord contour curve; and geometric topological connectivity analysis based on the sub-pixel-level vocal cord contour curve to obtain the vocal cord region identification result. This invention meets the clinical needs for automated and intelligent targeted drug delivery for vocal cord diseases.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system for determining and automatically administering medication to vocal cords based on AI visual recognition. Background Technology

[0002] Currently, the diagnosis and treatment of vocal cord diseases in clinical practice largely rely on manual observation under laryngoscopy and manual nebulized drug administration. For common lesions such as vocal cord polyps, vocal cord nodules, and acute and chronic laryngitis, medical staff need to observe the morphology of the laryngeal tissue in real time using rigid or flexible laryngoscopes, rely on subjective experience to judge the location, size and severity of the lesion, and manually control the position of the nebulizer nozzle and the duration of drug administration.

[0003] This method has the following technical limitations: Manual visual recognition is susceptible to interference from image noise, uneven lighting, and dynamic laryngeal movements, making it difficult to accurately segment the vocal cord edges and lesion areas, and to achieve sub-pixel-level lesion quantification assessment, thus posing risks of missed or misdiagnosed cases. Manual nozzle positioning relies on experience-based operation, making it difficult to accurately map lesion information from two-dimensional images to the three-dimensional laryngeal cavity space. This can easily lead to drug spraying deviating from the lesion center, mismatch between drug dosage and lesion, reduced drug utilization and treatment efficacy, and may cause throat irritation or drug waste due to overdose, making it difficult to meet the clinical needs for accurate, automated, and targeted drug delivery for vocal cord lesions. Summary of the Invention

[0004] This invention provides a method and system for automatically determining targeted drug delivery for vocal cord diseases based on AI visual recognition, which meets the clinical needs for automated and intelligent targeted drug delivery for vocal cord diseases.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: In a first aspect, a method for automatically determining vocal cord-targeted drug delivery based on AI visual recognition is provided, the method comprising: Step 1: Real-time acquisition of laryngeal video image data; image enhancement processing and wavelet transform denoising preprocessing are performed on the laryngeal video image data to obtain the preprocessed laryngeal image; Step 2: The trained convolutional neural network AI visual recognition model is used to coarsely segment the vocal cord candidate region of the preprocessed laryngeal image to obtain the initial vocal cord region mask. Step 3: Based on the initial vocal cord region mask, the vocal cord edges of the preset active contour model are extracted in a refined manner, and a geometric energy functional is constructed. By iteratively optimizing the geometric energy functional, a sub-pixel level vocal cord contour curve is obtained. Based on the sub-pixel level vocal cord contour curve, geometric topological connectivity analysis is performed to obtain the vocal cord region recognition result. Step 4: Combine the lesion region coordinate range, lesion area ratio, and lesion type classification information from the vocal cord region identification results with the preset vocal cord lesion judgment criteria to quantitatively assess the vocal cord health status and obtain the assessment results. Step 5: Based on the evaluation results, when it is determined that there is a lesion area that needs to be administered, the pixel coordinate range, lesion area size and severity level of the lesion area in the laryngeal image are mapped to three-dimensional spatial coordinates to determine the final action point coordinates of the targeted drug delivery nozzle relative to the center of the lesion area. The spray flow rate and spray duration of the nebulized drug are determined according to the lesion area and severity level to generate a targeted drug delivery determination instruction.

[0006] Furthermore, real-time acquisition of laryngeal video image data; image enhancement processing and wavelet transform denoising preprocessing are performed on the laryngeal video image data to obtain preprocessed laryngeal images, including: The micro-imaging module of the wireless laryngeal endoscope continuously captures dynamic video streams of the laryngeal cavity, extracts discrete video frames at preset frame intervals, and combines the discrete video frames into original laryngeal video image data. Multi-scale illumination compensation and local contrast stretching were performed on the original laryngeal video image data to obtain an enhanced laryngeal image with balanced illumination. Discrete wavelet transform decomposition is performed on the enhanced throat image to obtain a low-frequency approximate subband and a high-frequency detail subband. Based on the coefficient distribution characteristics of the high-frequency detail subband, an adaptive threshold shrinkage strategy is used to perform nonlinear filtering on the high-frequency coefficients to obtain the filtered wavelet coefficient set. By performing inverse wavelet transform on the wavelet coefficient set and the low-frequency approximate subband, a preprocessed throat image with noise suppression and clear texture is obtained.

[0007] Furthermore, the trained convolutional neural network AI visual recognition model is used to coarsely segment the vocal cord candidate region of the preprocessed laryngeal image to obtain an initial vocal cord region mask, including: The preprocessed laryngeal image is input into the encoder network of the convolutional neural network AI visual recognition model to extract multi-scale deep semantic feature maps of laryngeal anatomy, compress spatial dimensions and aggregate tissue contextual information to obtain high-level semantic feature representations. The high-level semantic feature representation is input into the decoder network to gradually restore the spatial resolution of the multi-scale deep semantic feature map, and cross-level feature fusion is performed to obtain the fused feature map. The fused feature map is input into a pixel-level classification convolutional layer, and the probability confidence of each image coordinate point belonging to the vocal cord tissue region is calculated to obtain the probability distribution map of the vocal cord region. Based on a preset confidence threshold, the probability distribution map of the vocal cord region is binarized and thresholded. Pixel regions with probability values ​​higher than the threshold are marked as candidate foreground regions of the vocal cord, and pixel regions with probability values ​​lower than the threshold are marked as background regions, thus obtaining the initial vocal cord region mask.

[0008] Furthermore, based on the initial vocal cord region mask, the preset active contour model is subjected to refined extraction of vocal cord edges, and a geometric energy functional is constructed. Through iterative optimization of the geometric energy functional, sub-pixel level vocal cord contour curves are obtained, including: Using the boundary pixel set of the initial vocal cord region mask as the initial closed curve, a geometric energy functional containing an internal geometric smoothing constraint term and an external image gradient driving term is constructed. Based on the variational evolution mechanism of the preset active contour model, the geometric energy functional is solved by minimizing energy, and the control equation for the spatial position evolution of the contour curve is derived. The control equation is solved by the discretized numerical iteration method. During the iteration process, the spatial coordinates of the contour nodes are obtained based on the vector superposition result of the gradient potential field traction force and the internal smoothing constraint force. The contour node positions are continuously updated iteratively, so that the closed curve migrates and fits to the actual anatomical edge of the vocal cord under the control of the energy functional convergence condition, until the contour deformation displacement of the adjacent iteration cycle is lower than the preset convergence threshold, thus obtaining the sub-pixel level vocal cord contour curve.

[0009] Furthermore, geometric topological connectivity analysis is performed based on sub-pixel level vocal cord contour curves to obtain vocal cord region identification results, including: Perform connected component traversal marking based on preset neighborhood rules on the internal pixel region enclosed by the sub-pixel level vocal cord contour curve to identify the set of independent foreground regions that are interconnected at the pixel level in the image space. By calculating the geometric center coordinates, outer envelope size and pixel filling area of ​​each independent foreground region, the target connected region located in the glottic region below the epiglottis and satisfying the condition of left-right symmetry distribution is obtained. By removing the irregularly clustered ventricular zone interference region above the target connected domain, as well as the rigid tissue interference region of the arytenoid cartilage with a small radius of curvature on the posterolateral side of the vocal cords, the set of the main connected domains of the vocal cords after geometric topological filtering is obtained. Based on the set of connected domains of the vocal cord body, contour boundary smoothing and adjacent pixel region fusion processing are performed to obtain vocal cord region recognition results that fully cover the vocal cord tissues on both sides.

[0010] Furthermore, by combining the lesion region coordinates, lesion area percentage, and lesion type classification information from the vocal cord region identification results with preset vocal cord lesion judgment criteria, the health status of the vocal cords is quantitatively assessed, yielding assessment results including: The results of vocal cord region identification are analyzed at the pixel level to extract the pixel coordinate boundary range of the lesion area, the area ratio of the lesion area to the total area of ​​the vocal cord, and the lesion type classification information based on color distribution and texture features. The coordinate range, area proportion, and lesion type classification information of the lesion area are compared and matched with the morphological size threshold, area proportion threshold, and pathological type grading rules in the preset vocal cord lesion judgment criteria to obtain the deviation of each lesion indicator. A multidimensional health status quantitative mapping model is constructed based on deviation, and the overall health status of the vocal cords is weighted and graded to obtain an assessment result that includes the severity level of the lesion and the health status score.

[0011] Furthermore, based on the evaluation results, when a lesion area requiring drug administration is determined, the pixel coordinate range, lesion size, and severity level of the lesion area in the laryngeal image are mapped to three-dimensional spatial coordinates to determine the final point of action of the targeted drug delivery nozzle relative to the center of the lesion area. The spray flow rate and duration of the nebulized drug are then determined based on the lesion area and severity level to generate a targeted drug delivery determination instruction, including: When the severity level of the lesion in the assessment results reaches the preset drug administration trigger threshold, the pixel coordinate range, lesion area size, and severity level data of the lesion area in the laryngeal image are extracted. Based on the pixel coordinate range, the transformation relationship between the two-dimensional image pixel coordinate system and the three-dimensional anatomical coordinate system of the laryngeal cavity is established by using the pre-calibrated intrinsic parameter matrix and depth mapping parameters of the wireless laryngeal endoscope. The pixel coordinate range is then inversely projected and calculated into spatial point cloud distribution data of the lesion area in three-dimensional space. The three-dimensional centroid of the spatial point cloud distribution data is calculated to obtain the coordinates of the three-dimensional geometric center. The spatial relative offset vector between the three-dimensional geometric center coordinates and the current pose data is calculated to determine the final action point coordinates of the targeted drug delivery nozzle. The size of the lesion area and the severity level are input into the preset drug delivery dose matching strategy to calculate and determine the parameters of the unit time spray flow rate and total spray duration of the nebulized drug. The final point of action coordinates, injection flow rate per unit time, and total injection duration parameters are encapsulated into control commands and formatted with a communication protocol to obtain the targeted drug delivery determination command.

[0012] Secondly, a vocal cord-targeted automatic drug delivery determination system based on AI visual recognition includes: The acquisition module is used to acquire laryngeal video image data in real time; it performs image enhancement processing and wavelet transform denoising preprocessing on the laryngeal video image data to obtain the preprocessed laryngeal image. The segmentation module is used to coarsely segment the vocal cord candidate region of the preprocessed laryngeal image by the trained convolutional neural network AI visual recognition model to obtain the initial vocal cord region mask. The module is used to refine the vocal cord edge extraction of the preset active contour model based on the initial vocal cord region mask, construct the geometric energy functional, and obtain the sub-pixel level vocal cord contour curve by iteratively optimizing the geometric energy functional; and to perform geometric topological connectivity analysis based on the sub-pixel level vocal cord contour curve to obtain the vocal cord region recognition result. The assessment module is used to combine the coordinate range of the lesion area, the proportion of the lesion area and the lesion type classification information in the vocal cord region identification results with the preset vocal cord lesion judgment criteria to quantitatively assess the health status of the vocal cords and obtain the assessment results. The mapping module is used to map the pixel coordinate range, lesion area size, and severity level of the lesion area in the laryngeal image to three-dimensional spatial coordinates based on the evaluation results when it is determined that there is a lesion area that needs to be drugged. This is to determine the final action point coordinates of the targeted drug delivery nozzle relative to the center of the lesion area, and to determine the spray flow rate and spray duration of the nebulized drug according to the lesion area and severity level, so as to generate a targeted drug delivery determination instruction.

[0013] Thirdly, a computing device, comprising: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0014] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0015] The above-described solution of the present invention has at least the following beneficial effects: By employing techniques such as laryngeal image enhancement and wavelet transform denoising preprocessing, convolutional neural network vocal cord region coarse segmentation, active contour model combined with geometric energy functional iterative optimization to obtain subpixel-level vocal cord contours, geometric topological connectivity analysis to eliminate interference regions, pixel-level lesion analysis, multi-dimensional quantitative assessment, mapping of two-dimensional pixel coordinates to three-dimensional laryngeal space, automatic matching of drug flow rate and duration, and generation of targeted drug delivery instructions, this approach overcomes the technical problems of low accuracy in manual laryngoscopy observation and identification, blurred lesion boundaries, easy omissions and misjudgments, inability to achieve subpixel-level quantization, inability of two-dimensional images to accurately map three-dimensional space leading to drug delivery positioning deviations, drug dosage mismatch with lesions, excessive drug stimulation of normal tissues, low drug utilization, reliance on experience in operation, and poor consistency. This approach achieves high-precision and stable identification of vocal cords and lesion regions, completes objective quantitative grading and accurate three-dimensional targeted positioning of lesions, ensures that drug delivery location and dosage are strictly matched to the lesion, avoids damage to normal tissues, improves drug utilization and drug delivery consistency, reduces human error, and enhances treatment safety and effectiveness, thus meeting the clinical needs for automated and accurate targeted drug delivery for vocal cord lesions. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating an AI-based visual recognition-based automatic drug delivery determination method for vocal cord targeting, provided by an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of an AI-based visual recognition-based automatic drug delivery system for vocal cord targeting, provided by an embodiment of the present invention. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0019] like Figure 1 As shown, an embodiment of the present invention proposes a method for automatically determining vocal cord-targeted drug delivery based on AI visual recognition, the method comprising the following steps: Step 1: Real-time acquisition of laryngeal video image data; image enhancement processing and wavelet transform denoising preprocessing are performed on the laryngeal video image data to obtain the preprocessed laryngeal image; Step 2: The trained convolutional neural network AI visual recognition model is used to coarsely segment the vocal cord candidate region of the preprocessed laryngeal image to obtain the initial vocal cord region mask. Step 3: Based on the initial vocal cord region mask, the vocal cord edges of the preset active contour model are extracted in a refined manner, and a geometric energy functional is constructed. By iteratively optimizing the geometric energy functional, a sub-pixel level vocal cord contour curve is obtained. Based on the sub-pixel level vocal cord contour curve, geometric topological connectivity analysis is performed to obtain the vocal cord region recognition result. Step 4: Combine the lesion region coordinate range, lesion area ratio, and lesion type classification information from the vocal cord region identification results with the preset vocal cord lesion judgment criteria to quantitatively assess the vocal cord health status and obtain the assessment results. Step 5: Based on the evaluation results, when it is determined that there is a lesion area that needs to be administered, the pixel coordinate range, lesion area size and severity level of the lesion area in the laryngeal image are mapped to three-dimensional spatial coordinates to determine the final action point coordinates of the targeted drug delivery nozzle relative to the center of the lesion area. The spray flow rate and spray duration of the nebulized drug are determined according to the lesion area and severity level to generate a targeted drug delivery determination instruction.

[0020] In this embodiment of the invention, by employing laryngeal image enhancement and wavelet transform denoising preprocessing, convolutional neural network vocal cord region coarse segmentation, active contour model combined with geometric energy functional iterative optimization to obtain sub-pixel level vocal cord contours, and geometric topological connectivity analysis to accurately identify vocal cord regions, and performing multi-dimensional quantitative evaluation of lesion regions in terms of coordinates, area, and type, mapping two-dimensional pixel coordinates to three-dimensional spatial coordinates, automatically determining drug delivery points and injection parameters, and generating targeted drug delivery judgment instructions, this invention overcomes the technical problems of manual laryngoscopy in accurately identifying vocal cords and lesion boundaries, large three-dimensional spatial positioning deviations, lack of lesion quantitative evaluation, resulting in uneven drug delivery coverage, excessive drug stimulation of normal vocal cord tissue, insufficient drug dosage in lesion areas, low drug utilization, reliance on physician experience, poor drug delivery consistency, and easy induction of laryngeal irritation. This achieves high-precision identification and objective quantitative evaluation of vocal cords and lesion regions, completes automatic and accurate matching of targeted drug delivery points and dosages, avoids excessive drug stimulation of normal tissues, and improves drug utilization and drug delivery stability.

[0021] In a preferred embodiment of the present invention, step 1 above may include: Step 1.1: Continuously capture dynamic video streams of the laryngeal cavity using the miniature imaging module of the wireless laryngeal endoscope. Extract discrete video frames at preset frame intervals and combine them into raw laryngeal video image data. Specifically, this involves aligning the built-in miniature CMOS imaging module of the wireless laryngeal endoscope with the target area of ​​the patient's laryngeal cavity, activating continuous imaging mode, and continuously capturing dynamic video streams of the laryngeal cavity. This dynamic video stream can completely record the morphological changes of the vocal cords during physiological movements such as breathing and slight vocalization, comprehensively reflecting the dynamic anatomical structure of the vocal cords and surrounding laryngeal tissues. To balance the real-time performance of image processing with data validity, discrete video frames are extracted from the continuously captured dynamic video stream of the laryngeal cavity at preset fixed frame intervals. Redundant frame data with high repetition and no obvious morphological changes in adjacent frames are removed. All extracted discrete video frames are sequentially combined and arranged according to their chronological order in the dynamic video stream to obtain raw laryngeal video image data with uniform format and standardized data.

[0022] Step 1.2 involves performing multi-scale illumination compensation and local contrast stretching on the original laryngeal video image data to obtain an enhanced laryngeal image with balanced illumination. Specifically, this includes: performing multi-scale illumination compensation and local contrast stretching on the original laryngeal video image data to improve image quality. Through multi-scale illumination compensation, each discrete video frame in the original laryngeal video image data is decomposed into luminance and chrominance components. A Gaussian multi-scale decomposition method is used to perform global luminance correction and local luminance weighted adjustment on the luminance component at different scales. Brightness suppression is applied to areas with excessive illumination, while luminance compensation is applied to areas with insufficient illumination, eliminating image brightness differences caused by light source angle deviation and laryngeal tissue occlusion. Local contrast stretching is then performed. For the illumination-compensated image, a piecewise linear transformation algorithm is used to non-linearly stretch the local grayscale range of the image, highlighting the grayscale difference between the vocal cord mucosa and surrounding laryngeal tissue, and enhancing the detailed features of the vocal cord edges, mucosal texture, and potential lesion areas. The mathematical expression for local contrast stretching is: ; In the formula: After local contrast stretching, the image at pixel coordinates The grayscale value at that location; Original throat image in pixel coordinates The grayscale value at that location; In pixel coordinates The minimum grayscale value of the image within a preset local window centered on the image. In pixel coordinates The maximum grayscale value of the image within a preset local window centered on the image; This is the preset lower limit for image output grayscale, usually set to 0; The preset upper limit of grayscale for image output is usually set to 255. Through a combination of multi-scale illumination compensation and local contrast stretching, the local narrow-band grayscale distribution in the original image is dynamically stretched to the full grayscale range of 0-255, which effectively improves the visual recognition of vocal cord edges, mucosal texture and potential lesion areas, and finally obtains an enhanced laryngeal image with balanced illumination, prominent details and appropriate contrast.

[0023] Step 1.3: Perform discrete wavelet transform decomposition on the enhanced laryngeal image to obtain low-frequency approximate subbands and high-frequency detail subbands. Based on the coefficient distribution characteristics of the high-frequency detail subbands, an adaptive threshold shrinkage strategy is used to perform nonlinear filtering on the high-frequency coefficients to obtain the filtered wavelet coefficient set. Specifically, the enhanced laryngeal image contains random interference noise such as electronic noise and Gaussian noise generated during endoscopic imaging. This noise can interfere with the accuracy of vocal cord region segmentation and lesion feature extraction. Discrete wavelet transform decomposition and adaptive threshold shrinkage denoising are performed on the enhanced laryngeal image. Two to three levels of discrete wavelet transform decomposition are performed on the enhanced laryngeal image, and each level of decomposition yields 1... The image consists of one low-frequency approximation sub-band and three high-frequency detail sub-bands. The low-frequency approximation sub-band mainly preserves global information such as the overall structure and brightness distribution of the image, reflecting the overall morphology of the vocal cords and laryngeal tissues. The high-frequency detail sub-bands mainly contain local details such as the image edges and textures, and also contain most of the random interference noise. The noise in the high-frequency detail sub-bands is represented by random coefficients with small amplitudes. To achieve denoising without losing effective details such as vocal cord edges and textures, this invention uses an adaptive threshold shrinkage strategy based on the coefficient distribution characteristics of the high-frequency detail sub-bands to perform nonlinear filtering on the coefficients of each high-frequency detail sub-band, and calculates the adaptive denoising threshold corresponding to each high-frequency detail sub-band. The calculation formula is as follows: In the formula, σ is the adaptive denoising threshold corresponding to the high-frequency detail subband in the i-th row and j-th column. Different high-frequency coefficients at different positions correspond to different denoising thresholds to achieve adaptive denoising. σ is the noise standard deviation estimate of the high-frequency detail subband coefficients, which is obtained by calculating the standard deviation of all coefficients in the high-frequency detail subband. N is the total number of pixels in a single high-frequency detail subband, i.e., the size of the high-frequency subband.

[0024] Based on the aforementioned adaptive threshold, the original wavelet coefficients of each high-frequency detail subband are subjected to soft thresholding shrinkage, retaining effective detail coefficients with amplitudes greater than the threshold and suppressing noise coefficients with amplitudes less than or equal to the threshold. The mathematical expression for this soft thresholding shrinkage process is as follows: ; In the formula, The original wavelet coefficients of the high-frequency detail subband in the i-th row and j-th column after discrete wavelet transform decomposition; These are the wavelet coefficients of the high-frequency detail subband in the i-th row and j-th column after soft threshold shrinkage filtering. The adaptive denoising threshold corresponding to the high-frequency detail subband in the i-th row and j-th column is calculated. After performing the above adaptive threshold shrinkage filtering on all high-frequency detail subbands, the processed high-frequency wavelet coefficients are integrated to obtain the filtered wavelet coefficient set.

[0025] Step 1.4 involves performing inverse wavelet transform reconstruction on the wavelet coefficient set and the low-frequency approximate subband to obtain a preprocessed throat image with noise suppression and clear texture. Specifically, this includes: based on the filtered wavelet coefficient set, combined with the low-frequency approximate subband obtained from discrete wavelet transform decomposition, performing inverse discrete wavelet transform reconstruction; recombining the decomposed low-frequency approximate subband with the denoised high-frequency detail subband to restore the complete throat image, thus completing the image preprocessing process and obtaining a preprocessed throat image with noise suppression and clear texture. The core of inverse discrete wavelet transform reconstruction is to reverse the operation of the subband coefficients from discrete wavelet transform decomposition to restore the original size and details of the image. Its mathematical expression is: In the formula, The preprocessed throat image after inverse discrete wavelet transform reconstruction is shown in pixel coordinates. The gray value at that location; IDWT is the inverse discrete wavelet transform operator; c A c represents the low-frequency approximate subband coefficients obtained from discrete wavelet transform decomposition; H c represents the horizontal high-frequency detail subband coefficients after adaptive threshold shrinkage filtering. V c represents the vertical high-frequency detail subband coefficients after adaptive threshold shrinkage filtering. D The diagonal high-frequency detail subband coefficients are processed by adaptive threshold shrinkage filtering. During the reconstruction process, the global structure of the low-frequency approximate subband is organically integrated with the edge and texture information of the denoised high-frequency detail subband through the inverse superposition of wavelet basis functions. This ensures that the reconstructed preprocessed laryngeal image eliminates random interference noise and fully preserves key details such as vocal cord edges and mucosal texture.

[0026] In this embodiment of the invention, by employing dynamic video stream acquisition and frame-by-frame capture, the dynamic anatomical information of the vocal cords is fully preserved, while reducing data redundancy, improving processing efficiency, avoiding repeated calculations of consecutive frames, ensuring system real-time performance, and using multi-scale illumination compensation and local contrast stretching to effectively eliminate uneven illumination, local dark areas, and reflection interference in the endoscope, improve the grayscale contrast between the vocal cords and surrounding tissues, highlight mucosal texture and potential lesion areas, and reduce the difficulty of AI recognition. Discrete wavelet transform and adaptive threshold shrinkage denoising can accurately suppress image noise without blurring vocal cord edge details, unlike the edge loss caused by traditional global filtering, achieving a balance between noise suppression and preservation of tissue details.

[0027] In a preferred embodiment of the present invention, step 2 above may include: Step 2.1 involves inputting the preprocessed laryngeal image into the encoder network of a convolutional neural network (CNN) AI visual recognition model to extract multi-scale deep semantic feature maps of the laryngeal anatomy. This process compresses the spatial dimension and aggregates tissue contextual information to obtain a high-level semantic feature representation. Specifically, this includes: inputting the noise-suppressed and texture-clear preprocessed laryngeal image frame by frame into the pre-trained CNN AI visual recognition model's encoder network. The CNN AI visual recognition model prepares and labels the image data through a recognition task, performing preprocessing such as normalization and enhancement; constructing the network structure by sequentially stacking convolutional and pooling layers to extract image features, followed by a fully connected layer to output the recognition result; and finally... After testing and optimization on the validation set, a usable visual recognition model was finally obtained. The encoder network adopts a structure of alternating stacked convolutional layers, batch normalization layers, and max pooling layers to extract multi-scale deep semantic feature maps of the laryngeal anatomy layer by layer. The first convolutional layer uses a 3×3 convolutional kernel to perform preliminary feature extraction on the preprocessed laryngeal image, resulting in a shallow feature map. This shallow feature map mainly contains basic details such as the edges and textures of the laryngeal tissue. The batch normalization layer normalizes the shallow feature map to accelerate the model's convergence speed. The max pooling layer downsamples the normalized feature map to compress the spatial dimension of the feature map, retain key feature information, and reduce computational cost.

[0028] Repeating the convolution, batch normalization, and pooling operations described above, as the network depth increases, mid-to-high-level features of the laryngeal anatomy are gradually extracted, resulting in deep semantic feature maps at different scales. These multi-scale deep semantic feature maps are then mapped to local details, regional structures, and overall morphology of the laryngeal tissues, respectively. A global average pooling layer is used to further compress the spatial dimension of all multi-scale deep semantic feature maps, aggregating feature information at different scales. Simultaneously, contextual information from various regions of the laryngeal tissue is integrated to eliminate the isolation of local features. Finally, a high-level semantic feature representation that can fully characterize the global features of the vocal cords and surrounding laryngeal tissues is obtained. This high-level semantic feature representation can effectively distinguish vocal cord tissues from other non-vocal cord tissues in the laryngeal cavity, mainly involving feature extraction and aggregation.

[0029] Step 2.2 involves inputting the high-level semantic feature representation into the decoder network to gradually restore the spatial resolution of the multi-scale deep semantic feature map and performing cross-layer feature fusion to obtain a fused feature map. Specifically, this includes: inputting the high-level semantic feature representation into the first deconvolution layer; upsampling the feature map through deconvolution to expand its spatial size and initially restore its spatial resolution; refining the deconvolutioned feature map through convolution layers to remove jagged artifacts generated during upsampling and improve its clarity; and repeating the deconvolution and convolution operations. As the network depth increases, the spatial resolution of the feature map gradually recovers to match that of the preprocessed throat image. During resolution restoration, a cross-level feature fusion operation is performed simultaneously. This involves concatenating and fusing multi-scale deep semantic feature maps extracted from different levels of the encoder network with the feature maps restored from the corresponding levels of the decoder network channel by channel. The shallow feature maps of the encoder provide detailed information about the laryngeal tissue, while the feature maps from the corresponding levels of the decoder provide global semantic information. By combining global semantic information with local detailed information through feature fusion, the limitations of single-scale features are overcome, and the ability of the feature maps to represent vocal cord tissue is improved. After fusion, the fused feature maps are integrated and optimized through the last convolutional layer to obtain a fused feature map that can accurately represent the features of the vocal cord tissue region.

[0030] Step 2.3: Input the fused feature map into a pixel-level classification convolutional layer, calculate the probability confidence of each image coordinate point belonging to the vocal cord tissue region, and obtain the vocal cord region probability distribution map. Specifically, this includes: performing pixel-by-pixel feature mapping on the fused feature map through the pixel-level classification convolutional layer, mapping the high-dimensional feature vector of each pixel point to a two-dimensional vector, which corresponds to the two categories of vocal cord tissue region and non-vocal cord tissue region respectively; normalizing the two-dimensional vector through an activation function, and calculating the probability confidence of each pixel point belonging to the vocal cord tissue region, the expression of which is: ; In the formula, For the preprocessed throat image, pixel coordinates The probability confidence level of a pixel belonging to the vocal cord tissue region ranges from [0,1]. The closer the value is to 1, the greater the probability that the pixel belongs to the vocal cord tissue region; the closer the value is to 0, the greater the probability that the pixel belongs to a non-vocal cord tissue region. pixel coordinates The feature vector is mapped to the output value of the vocal cord tissue region category. pixel coordinates The output value of the feature vector is mapped to the non-vocal cord tissue region category. After performing the above probability calculation operation on all pixels in the fused feature map, a vocal cord region probability distribution map with the same size as the preprocessed laryngeal image is obtained. The gray value of each pixel in the probability distribution map corresponds to its probability confidence of belonging to the vocal cord tissue region, which intuitively reflects the approximate distribution range of the vocal cord tissue region in the image.

[0031] Step 2.4: Based on a preset confidence threshold, perform binarization thresholding on the vocal cord region probability distribution map. Pixel regions with probability values ​​higher than the threshold are marked as vocal cord candidate foreground regions, and pixel regions with probability values ​​lower than the threshold are marked as background regions, thus obtaining an initial vocal cord region mask. Specifically, this includes: setting a preset confidence threshold, which is determined through training and optimization based on laryngeal image samples, typically set to 0.7, to distinguish between vocal cord candidate foreground regions and background regions; subsequently, threshold judgment is performed on each pixel in the vocal cord region probability distribution map, using the following formula: In the formula, In the initial vocal cord region mask, pixel coordinates Mask value at the location; pixel coordinates The probability confidence level of a pixel belonging to the vocal cord tissue region; T: a preset confidence threshold, in this embodiment T=0.7, used to define the boundary between the vocal cord candidate foreground region and the background region. Pixel regions with a probability confidence level higher than the preset threshold are marked as vocal cord candidate foreground regions, with a corresponding mask value of 1; pixel regions with a probability confidence level lower than or equal to the preset threshold are marked as background regions, with a corresponding mask value of 0. Through the above binarization threshold segmentation operation, a binarized initial vocal cord region mask is obtained.

[0032] In this embodiment of the invention, by extracting multi-scale deep semantic features and aggregating contextual information through an encoder network, the global and local features of the laryngeal anatomy can be effectively captured, accurately distinguishing vocal cord tissue from non-vocal cord tissue. This solves the problem in the background technology where manual identification is difficult to accurately distinguish vocal cords from surrounding tissues and is easily interfered with, providing a reliable feature basis for segmentation. By restoring the spatial resolution of the feature map through a decoder network and performing cross-level feature fusion, the loss of details during feature extraction is compensated for, and the feature map's ability to represent vocal cord tissue details is improved, ensuring accurate pixel-level classification. Through pixel-level classification convolutional layers, accurate classification is achieved pixel by pixel, quantifying the probability that each pixel belongs to vocal cord tissue. Compared with subjective human judgment, this improves the objectivity and accuracy of vocal cord region identification, reduces the risk of missed or false judgments, and obtains an initial vocal cord region mask through binarized threshold segmentation, clearly marking the vocal cord candidate foreground region and effectively eliminating most of the interference from non-vocal cord tissue.

[0033] In a preferred embodiment of the present invention, step 3 above may include: Step 3.1: Using the boundary pixel set of the initial vocal cord region mask as the initial closed curve, construct a geometric energy functional that includes an internal geometric smoothing constraint term and an external image gradient driving term. Specifically, this includes: extracting the boundary of the initial vocal cord region mask, extracting the boundary pixel set of the initial vocal cord region mask, and connecting the boundary pixels in a clockwise direction to obtain a continuous and closed curve, which serves as the initial closed curve of the active contour model. This initial closed curve can roughly fit the edge contour of the vocal cord tissue.

[0034] To achieve accurate migration of the initial closed curve to the actual anatomical edge of the vocal cords, a geometric energy functional is constructed, comprising an internal geometric smoothing constraint term and an external image gradient driving term. Energy minimization controls the evolution of the closed curve, ensuring that the final contour curve is smooth and continuous, accurately conforming to the actual edge of the vocal cords. The complete expression of the geometric energy functional is: In the formula, The total energy value of the geometric energy functional; λ represents the internal geometric smoothing constraint term, used to constrain the shape of the closed curve; λ is the energy weight coefficient, used to balance the influence of the internal geometric smoothing constraint term and the external image gradient driving term. The external image gradient driving term provides the driving force for the closed curve to migrate towards the actual edge of the vocal cords, while the internal geometric smoothing constraint term... It consists of the curvature energy and length energy of the curve, and its specific expression is: In the formula, α represents the length-energy weighting coefficient; The length energy of a closed curve is used to control the stability of the curve length, and its expression is: Where C represents a closed curve, ds represents the arc length infinitesimal element of the curve, and β is the curvature energy weighting coefficient; The curvature energy of a closed curve is used to measure the smoothness of the curve, and its expression is: , where κ represents the curvature of the closed curve at a certain point, and ds represents the arc length infinitesimal element of the curve.

[0035] External image gradient driving term The gradient calculation based on the preprocessed throat image is specifically expressed as follows: In the formula, The preprocessed throat image in pixel coordinates The gradient vector at a given point is used to characterize the edge strength of the image; The squared magnitude of the gradient vector is used to emphasize the difference between edge and non-edge regions; C represents a closed curve, and ds represents the arc length infinitesimal element of the curve. The integral operation is used to calculate the overall external driving energy at the location of the closed curve. The complete geometric energy functional is constructed through the above formula.

[0036] Step 3.2: Based on the variational evolution mechanism of the preset active contour model, the geometric energy functional is solved for energy minimization, and the governing equations for the spatial position evolution of the contour curve are derived. The core logic of the variational evolution mechanism of the preset active contour model is that the evolution of the preset active contour model is essentially an energy-driven dynamic process. Its core is to find the contour curve shape that minimizes the total energy value of the geometric energy functional through the variational principle. That is, by taking the variation of the total energy functional with respect to the spatial position of the contour curve, setting the variational result to zero, the necessary condition for energy minimization is obtained, and then the governing equations for the spatial position evolution of the contour curve over time are derived, realizing the ordered and stable evolution of the contour curve. Based on the variational principle, the total energy functional is solved for energy minimization. The core of minimizing the energy lies in finding the variational expression of the total energy functional with respect to the profile curve C, and then setting the variational result to zero. Where δ is the variational operator, this equation is a necessary condition for minimizing the total energy, and the internal geometrical smoothness constraint term in the total energy functional is considered separately. With external image gradient driving terms Find the variational results, combine them to obtain the total variational results, and then solve the variational problem for the internal geometrical smoothness constraint term: By taking the variational properties of curve length energy and curvature energy, we can derive the following: In the formula, The variational result of the internal geometric smoothing constraint term reflects the trend of internal energy changing with the shape of the contour curve; The Laplace operator is used to calculate the second derivative of the curve's spatial position, characterizing the rate of change of the curve's shape. The unit normal vector of the closed contour curve points outward and is used to determine the direction of contour curve evolution; variational solution of the external image gradient-driven term: for By taking the variational value and combining it with the variational properties of image gradient, we can derive the following: In the formula, The variational result of the external image gradient driving term reflects the trend of external energy changing with the shape of the contour curve; The gradient of the squared magnitude of the image gradient is used to provide the driving force for the contour curve to migrate towards the vocal cord edge. Its direction points to the region of maximum image gradient. Combining the internal and external energy variations, the variational result of the total energy functional is obtained: ; Let the total variation result Since the integration interval C is an arbitrary closed curve, the integrand must be zero, thus deriving the governing equations for the spatial position evolution of the contour curve: ; In the formula, The evolution rate of the profile curve C with time t represents the spatial position change trend of the profile curve in each iteration; t is the iteration time step, which is a quantitative representation of the number of times the profile curve is iterated and updated, and each time step corresponds to one round of profile node position update. κ is the Laplacian operator used to calculate the second derivative of the spatial position of the contour curve, reflecting the degree of smoothness of the curve shape; κ is the curvature of the contour curve. It is the unit normal vector of the closed contour curve, pointing to the outside of the curve, and determines the direction of the contour curve's evolution; The gradient of the squared magnitude of the image gradient plays a crucial role in providing the driving force for the migration of the contour curve toward the actual edge of the vocal cords. Its direction points to the region with the largest image gradient, i.e., the vocal cord edge region. C is a closed contour curve, and its spatial position evolves with time t by this governing equation.

[0037] Step 3.3 involves continuously iterating and updating the contour node positions, causing the closed curve to migrate and fit towards the actual anatomical edge of the vocal cord under the control of the energy functional convergence condition, until the contour deformation displacement of adjacent iteration cycles is lower than the preset convergence threshold, thus obtaining a sub-pixel level vocal cord contour curve. Specifically, this includes: based on the new spatial coordinates of the contour nodes obtained after each iteration, continuously performing iterative update operations, causing the closed curve to continuously migrate and fit towards the actual anatomical edge of the vocal cord under the control of the geometric energy functional convergence condition, gradually optimizing the curve shape until the convergence condition is met, stopping the iteration, and finally obtaining a sub-pixel level vocal cord contour curve.

[0038] The specific iterative process involves calculating the spatial displacement of the corresponding contour node in two adjacent iterations after each round of contour node update, specifically the displacement of the i-th contour node between round κ and round (κ-1). The calculation formula is as follows: In the formula, This represents the spatial displacement of the i-th contour node between the κ-th and κ-1-th iterations; , This represents the spatial coordinates of the i-th contour node in the x and y directions after the κ-th iteration; , This represents the spatial coordinates of the i-th contour node in the x and y directions after the (κ-1)-th iteration.

[0039] After calculating the displacement of all contour nodes, the average value of all displacements is taken and compared with a preset convergence threshold. The preset convergence threshold is set to 0.1 pixels based on the sub-pixel accuracy requirement. The convergence judgment condition is: when the average displacement of three consecutive iterations is lower than the preset convergence threshold, it indicates that the closed curve has stably conformed to the actual anatomical edge of the vocal cord, the geometric energy functional has been minimized, and the iteration stops. Sub-pixel interpolation processing is performed on the finally obtained contour node coordinates to supplement the sub-pixel coordinate points between contour nodes, making the contour curve smoother and more continuous, accurately conforming to the actual anatomical edge of the vocal cord, and finally obtaining a sub-pixel vocal cord contour curve. This contour curve can accurately represent the edge morphology of the vocal cord with sub-pixel accuracy.

[0040] In this embodiment of the invention, by constructing a geometric energy functional that includes internal smoothing constraints and external gradient driving, the smooth continuity of the contour curve is ensured, and the curve can be driven to migrate towards the actual edge of the vocal cord. This solves the problems of jagged edges and discontinuous contours that are easy to occur when manually identifying the vocal cord edge in the background technology, and lays the energy constraint foundation for accurate extraction. The control equation is solved by variational principle, and the contour node coordinates are calculated by discretization numerical iteration method, realizing the accurate evolution of the closed curve and avoiding the problem of difficulty in solving continuous equations. At the same time, by balancing gradient traction force and smoothing constraint force through vector superposition, the curve evolution is ensured to be stable, and the accuracy of edge extraction is improved. By setting a subpixel level convergence threshold and combining continuous iteration and subpixel interpolation processing, a subpixel level vocal cord contour curve is finally obtained, solving the pain point of insufficient accuracy and inability to achieve subpixel level edge extraction manually in the background technology.

[0041] In a preferred embodiment of the present invention, step 3 above may include: Step 3.4: Perform connected component traversal marking based on a preset neighborhood rule on the internal pixel region enclosed by the sub-pixel level vocal cord contour curve to identify the set of independent foreground regions that are interconnected at the pixel level in the image space. Specifically, this includes: mapping the sub-pixel level vocal cord contour curve to the pixel coordinate system of the preprocessed laryngeal image to obtain a binarized initial foreground mask; marking the internal pixel region enclosed by the sub-pixel level vocal cord contour curve as foreground pixels and assigning a value of 1; marking the pixel region outside the contour curve as background pixels and assigning a value of 0, thus obtaining a binarized foreground mask with the same size as the preprocessed laryngeal image; and performing connected component traversal marking based on a preset neighborhood rule on the binarized foreground mask. In this embodiment, the preset neighborhood rule is the 8-neighborhood rule, that is, for any pixel coordinate... Its neighborhood range includes four adjacent pixels in the horizontal and vertical directions and four adjacent pixels in the diagonal direction, for a total of eight adjacent pixels. This rule can completely identify the continuously distributed vocal cord tissue area and avoid the continuous area being mistakenly divided into multiple independent areas due to the neighborhood range being too small.

[0042] Traverse all pixels in the binarized foreground mask. When an unlabeled foreground pixel is detected, set it as a seed point. Simultaneously search all unlabeled foreground pixels in its 8-neighborhood. Assign the same unique connected component number to all interconnected foreground pixels to complete the labeling of an independent foreground region. Repeat the above seed point search, neighborhood traversal, and number assignment process until all foreground pixels in the mask are assigned the corresponding connected component number. Finally, identify the set of all pixel-level interconnected independent foreground regions in the image space. Each independent foreground region corresponds to a unique connected component number. At the same time, record the set of all pixel coordinates and the total number of pixels contained in the region.

[0043] Step 3.5: By calculating the geometric center coordinates, outer envelope size, and pixel filling area of ​​each independent foreground region, the target connected region located in the glottic region below the epiglottis and satisfying the condition of left-right symmetry is obtained. Specifically, this includes calculating the core geometric feature parameters of each independent foreground region, including the geometric center coordinates, outer envelope size, and pixel filling area. The calculation formulas and meanings of each parameter are as follows: Formula for calculating the coordinates of the geometric center: In the formula, is the geometric center coordinates of the independent foreground region, in pixels; S is the pixel filling area of ​​the independent foreground region, i.e., the total number of foreground pixels contained in this region; R is the set of pixel coordinates of the independent foreground region. : The coordinates of a single foreground pixel within the set R. The outer envelope size is the width W and height H of the minimum bounding rectangle of the independent foreground region, where the width W is the difference between the maximum and minimum values ​​of the pixel x coordinates in the region, the height H is the difference between the maximum and minimum values ​​of the pixel y coordinates in the region, and the pixel filling area S is the total number of foreground pixels contained in the independent foreground region, used to characterize the overall size of the region.

[0044] The first round of screening is performed based on prior knowledge of laryngeal anatomy. In the human laryngeal anatomy, the epiglottis is located in the upper part of the laryngeal cavity, and the glottis is located in the lower middle part of the laryngeal cavity below the epiglottis. In the corresponding image coordinate system, the y-coordinate of the glottis ranges from 1 / 2 to 3 / 4 of the total image height. The geometric center y-coordinate is then set to... c Independent foreground regions within the preset range whose pixel filling area S is within the preset vocal cord area threshold range are retained as candidate connected regions. Noise regions with excessively small areas, non-target regions with excessively large areas, and regions whose locations do not conform to the anatomical position of the glottis are removed. A second round of left-right symmetry distribution condition screening is performed on the candidate connected regions: the vocal cords on both sides of the human body are symmetrically distributed about the vertical midline of the laryngeal cavity. Therefore, the symmetry of each candidate connected region about the vertical midline of the image is calculated. The symmetry calculation formula is: In the formula, η is the symmetry of the candidate connected domain, and its value ranges from [0, 1]. This represents the total width of the preprocessed throat image, in pixels. Let x be the symmetric coordinate of the pixel coordinates about the vertical midline of the image; S be the pixel filling area of ​​the candidate connected region; R be the set of pixel coordinates of the candidate connected region. The candidate connected regions with symmetry η greater than the preset symmetry threshold are determined as the target connected regions that meet the condition of left-right symmetry distribution, thus completing the screening of target connected regions.

[0045] Step 3.6: Remove the interfering regions of the ventricular folds that are irregularly clustered above the target connected domain, as well as the interfering regions of the rigid tissue of the arytenoid cartilage with a small radius of curvature on the posterolateral side of the vocal cords, to obtain the set of the main connected domains of the vocal cords after geometric topological filtering. Specifically, this includes: removing the interfering regions of the ventricular folds. In the anatomical structure of the human larynx, the ventricular folds are located above the vocal cords and are irregularly clustered. The y-coordinate of their geometric center is smaller than the y-coordinate of the geometric center of the target connected domain of the vocal cords, and their shape is irregular and their symmetry is lower than that of the vocal cord tissue. Therefore, using the y-coordinate of the geometric center of the target connected domain as a reference, the independent foreground regions above the reference line, with a symmetry less than 0.7 and exhibiting irregular cluster distribution, are marked as ventricular zone interference regions and removed from the target connected domain set. The arytenoid cartilage rigid tissue interference region is also removed. The arytenoid cartilage is located on the posterolateral side of the vocal cords and is a rigid tissue with sharp edges, high curvature, and a relatively small radius of curvature. It differs significantly from the smooth edge morphology of the vocal cords. The curvature and radius of curvature of each sub-region contour in the target connected domain are calculated. The curvature calculation formula is: In the formula, This represents the curvature of the contour curve at the arc length parameter s. The larger the curvature value, the sharper the contour edge. , The parametric equation representing the profile curve, with arc length s as the parameter; , This represents the first derivative of the contour curve coordinates with respect to the arc length s; , The second derivative of the contour curve coordinates with respect to the arc length *s* is given by the radius of curvature, which is the reciprocal of the curvature. The formula for calculating this radius is: In the formula, Let be the radius of curvature of the profile curve at the arc length parameter s.

[0046] Based on the above calculation results, the independent foreground region located on the outer side of the target connected domain of the vocal cord and whose average radius of curvature is less than the preset radius of curvature threshold is marked as a rigid tissue interference region of the arytenoid cartilage and removed from the target connected domain set. After the above two types of interference regions are removed, the vocal cord main connected domain set after geometric topology filtering is obtained. This set only contains the connected domains corresponding to the vocal cord tissues on the left and right sides, without interference from other non-vocal cord tissues.

[0047] Step 3.7: Based on the set of connected components of the vocal cord body, perform contour boundary smoothing fitting and fusion processing with adjacent pixel regions to obtain vocal cord region recognition results that fully cover the vocal cord tissues on both sides. Specifically, this includes: based on the set of connected components of the vocal cord body, performing contour boundary smoothing fitting and fusion processing with adjacent pixel regions to obtain vocal cord region recognition results that fully cover the vocal cord tissues on both sides. The specific implementation process is as follows: For the contour boundaries of each connected region in the vocal fold's main connected region set, a cubic curve smoothing fitting process is performed. This process effectively eliminates minute jagged edges and discrete noise on the contour boundaries, improving the smoothness and continuity of the contour. Specifically, the contour nodes of the vocal fold's main connected region are used as control points. Through a weighted cubic curve fitting method, discrete contour nodes are connected into a smooth and continuous contour curve, replacing the original discrete contour nodes. This makes the contour curve more closely match the actual anatomical edge shape of the vocal fold. Adjacent pixel region fusion processing is performed. For any small gaps that may exist between the left and right vocal fold connected regions, as well as any small holes that may appear inside the connected regions, morphological closing operations are used for filling. The structural element used is a 3x3 circular structural element. This structural element can fill gaps and holes without destroying the original shape of the vocal cord region, ensuring the integrity of the vocal cord region. The connected domains of the left and right vocal cords are integrated to obtain a unified binary vocal cord region mask. In this mask, the area assigned a value of one is the complete vocal cord tissue region, and the area assigned a value of zero is the non-vocal cord background region, realizing a unified label for the vocal cord tissues on both sides. The integrated binary vocal cord region mask, the vocal cord contour curve after smooth fitting, and various geometric feature parameters of the vocal cord region are integrated to form a complete vocal cord region recognition result. This result can completely cover the vocal cord tissues on both sides, with smooth boundaries and accurate regions.

[0048] In this embodiment of the invention, addressing the limitations of artificial laryngoscopy, such as susceptibility to interference from surrounding larynx tissues, difficulty in accurately distinguishing between vocal cords and non-vocal cord tissues, and the tendency for missed or incorrect vocal cord region identification, the geometric topological connectivity analysis process achieves accurate identification and interference removal of vocal cord regions. Through a series of operations, including traversal of 8-neighborhood rule connected domains, anatomical prior combined with geometric feature screening, targeted removal of interference from the ventricular folds and arytenoid cartilage, contour smoothing, and region fusion, discrete noise and continuous vocal cord tissue are effectively separated, accurately locating vocal cord regions in the glottis that conform to anatomical features. This solves the core pain points of easy misjudgment and numerous interferences in manual identification, ultimately yielding a smooth, complete, and accurate vocal cord region identification result.

[0049] In a preferred embodiment of the present invention, step 4 above may include: Step 4.1 involves pixel-level pathological feature analysis of the vocal cord region identification results. This includes extracting the pixel coordinate boundary range of the lesion area, the area ratio of the lesion area to the total vocal cord area, and lesion type classification information based on color distribution and texture features. Specifically, this includes: using the vocal cord region identification results as spatial constraints, locking the complete vocal cord tissue pixel interval in the preprocessed laryngeal image; performing pixel-level pathological feature analysis on all pixels within this interval; sequentially extracting the spatial location information, area ratio information, and pathological type classification information of the lesion area; extracting the pixel coordinate boundary range of the lesion area; and traversing all pixels within the vocal cord region. Image pixels were compared to normal vocal cord mucosa and lesion tissue to identify differences in grayscale distribution and color. Normal vocal cord mucosa exhibits uniform grayscale distribution and a uniform light red color, while lesion tissue shows abnormal grayscale, dark color, or whitish color. Based on this difference, all pixels belonging to lesion tissue were marked. The maximum and minimum values ​​of the horizontal, vertical, and lateral coordinates of all lesion pixels were retrieved. Using these four extreme values ​​as boundaries, the minimum circumscribed boundary of the lesion region in the image pixel coordinate system was determined. This boundary was then used as the pixel coordinate boundary of the lesion region to clarify its spatial location.

[0050] To calculate the percentage of the vocal cord area occupied by the lesion region, the total number of pixels in the lesion region and the total number of pixels in the entire vocal cord area are counted separately. The ratio of these two values ​​yields the percentage of the lesion area. The calculation formula is as follows: In the formula, This indicates the percentage of the total area of ​​the vocal cords that is affected by the diseased area. This indicates the total number of pixels contained in the lesion area; This indicates the total number of pixels contained in the entire vocal cord area.

[0051] Lesion type classification information based on color distribution and texture features is extracted. Color feature statistics are performed on pixels within the lesion area, mainly focusing on key indicators such as mean grayscale and grayscale variance. The mean grayscale reflects the overall color depth of the lesion area, while the grayscale variance reflects the uniformity of color within the lesion area. Simultaneously, texture features of the lesion area are calculated, primarily focusing on indicators such as texture roughness and texture contrast. Texture roughness reflects the smoothness of the lesion area surface, while texture contrast reflects the clarity of the texture within the lesion area. The obtained color and texture features are then matched one-to-one with a pre-established lesion feature database containing typical color and texture features of common lesion types such as vocal cord inflammation, vocal cord polyps, and vocal cord nodules. Through feature matching, the specific type of the current lesion is accurately distinguished, yielding the corresponding lesion type classification information.

[0052] Step 4.2 involves comparing the lesion region coordinate range, area proportion, and lesion type classification information with the preset vocal cord lesion judgment criteria, including morphological size thresholds, area proportion thresholds, and pathological type grading rules, to obtain the deviation of each lesion indicator. Specifically, this includes comparing the extracted lesion region pixel coordinate boundary range, lesion area proportion, and lesion type classification information with the built-in preset vocal cord lesion judgment criteria. These criteria include threshold ranges for corresponding morphological sizes, threshold ranges for corresponding lesion area proportions, and grading rules for different pathological types. For the lesion region's morphological size, the size parameters corresponding to the actual extracted pixel coordinate boundary range are compared with the preset morphological size threshold to obtain the morphological size deviation. For the lesion area proportion, the actually calculated area proportion value is compared with the preset area proportion threshold to obtain the area proportion deviation. For the lesion type, the actually classified lesion type is matched with the preset pathological type grading rules to obtain the pathological type deviation. The formula for calculating the deviation of a single-dimensional indicator is as follows: In the formula, This represents the deviation of the i-th pathological indicator. The larger the deviation value, the greater the difference between the indicator and the normal physiological state. This represents the actual measured value of the i-th lesion indicator; This represents the threshold value of the i-th lesion indicator in the preset vocal cord lesion judgment criteria. Through the above calculation, the deviation of the three lesion indicators, namely morphological size, area ratio, and pathological type, is obtained respectively, providing a data basis for the quantification of health status.

[0053] Step 4.3: Construct a multidimensional health status quantitative mapping model based on deviation, and perform weighted scoring and grading of the overall vocal cord health status to obtain an assessment result that includes the severity level of the lesion and the health status score. Specifically, this includes: constructing a multidimensional health status quantitative mapping model based on the deviation of various lesion indicators. This model achieves a weighted score of the overall vocal cord health status by assigning appropriate weight coefficients to the deviation of different dimensions. Then, based on the scoring results, the severity level of the lesion is graded, and a weighted score calculation is performed. The calculation formula is as follows: In the formula, This score represents the overall health of the vocal cords, with a maximum score of 100. The higher the score, the better the health of the vocal cords. This represents the weight coefficient corresponding to the i-th lesion indicator. The weight coefficient is preset according to the clinical diagnosis and treatment priority, and the sum of all weight coefficients is one. The deviation of the i-th lesion indicator is represented. After weighted scoring, the severity of the lesion is classified according to the health status score. The scoring range is divided into four levels: no lesion, mild lesion, moderate lesion, and severe lesion. Finally, the health status score and the corresponding severity level of the lesion are integrated to obtain the vocal cord health status assessment result containing complete quantitative information.

[0054] In this embodiment of the invention, by analyzing pathological features at the pixel level, comparing standardized indicators, and using multidimensional quantitative scoring, objective extraction of vocal cord lesion information and digital grading of health status are achieved. This effectively avoids subjective errors in human judgment, improves the accuracy of lesion identification and grading, and provides a reliable quantitative basis for determining the location and matching the dosage for targeted drug administration. This makes the drug administration determination more consistent with the actual degree of lesion, and improves the accuracy and standardization of overall diagnosis and treatment.

[0055] In a preferred embodiment of the present invention, step 5 above may include: Step 5.1: When the severity level of the lesion in the assessment result reaches the preset drug administration trigger threshold, extract the pixel coordinate range, lesion area size, and severity level data of the lesion region in the laryngeal image. Specifically, this includes: extracting the severity level of the lesion in the vocal cord health status assessment result, comparing this level with the preset drug administration trigger threshold built into the system, and determining whether the targeted drug administration process needs to be initiated. The preset drug administration trigger threshold is pre-set based on clinical diagnosis and treatment standards, specifically set to mild lesion and above. When the severity level of the lesion in the assessment result is mild, moderate, or severe, the targeted drug administration process is triggered; when the severity level of the lesion is no lesion, the drug administration process is not triggered, and only the vocal cord health monitoring result is output.

[0056] Once the targeted drug delivery process is triggered, core data of the lesion area is extracted to provide reliable support for subsequent drug delivery parameter calculations and site determination. The extracted data is derived from previously analyzed results, ensuring data consistency and accuracy. Specific extraction content includes: the pixel coordinate range of the lesion area in the laryngeal image (directly retrieved from the analyzed pixel coordinate boundary range of the lesion area, clarifying its specific location in the two-dimensional image pixel coordinate system); the size of the lesion area (based on the total number of pixels in the lesion area, combined with the pre-calibrated relationship between the pixel size of the preprocessed laryngeal image and the actual anatomical size of the laryngeal cavity, converting the number of pixels in the lesion area into the actual anatomical area, ensuring the data aligns with actual clinical diagnostic and treatment needs); and lesion severity level data (by classifying lesion severity levels). After the above three types of data are extracted, data normalization is performed, outlier data points are removed, and the data format is standardized to ensure data accuracy and consistency.

[0057] Step 5.2: Based on the pixel coordinate range, a transformation relationship is established from the two-dimensional image pixel coordinate system to the three-dimensional anatomical coordinate system of the laryngeal cavity using the pre-calibrated intrinsic parameter matrix and depth mapping parameters of the wireless laryngeal endoscope. The pixel coordinate range is then inversely projected and calculated into spatial point cloud distribution data of the lesion area in three-dimensional space. Specifically, this includes: based on the pixel coordinate range of the lesion area, an accurate transformation relationship is established between the two-dimensional image pixel coordinate system and the three-dimensional anatomical coordinate system of the laryngeal cavity using the pre-calibrated intrinsic parameter matrix and depth mapping parameters of the wireless laryngeal endoscope. Through inverse projection calculation, the two-dimensional pixel coordinate range is converted into spatial point cloud distribution data of the lesion area in three-dimensional space. The two-dimensional image pixel coordinate system takes the upper left corner of the preprocessed laryngeal image as the origin, the horizontal axis as the horizontal axis, and the vertical axis as the vertical axis. The coordinate values ​​are in pixels and are used to characterize the two-dimensional position of the lesion area in the image. The three-dimensional anatomical coordinate system of the laryngeal cavity takes the center of the lens of the wireless laryngeal endoscope as the origin, the horizontal axis as the vertical axis, and the vertical axis as the horizontal axis. The coordinate values ​​are in pixels and are used to characterize the two-dimensional position of the lesion area in the image. X The axis and the longitudinal direction are Y Axis, perpendicular to the lens direction Z The axes, with coordinate values ​​in millimeters, establish a transformation relationship based on a pre-calibrated intrinsic parameter matrix and depth mapping parameters of the wireless laryngeal endoscope. The intrinsic parameter matrix, obtained in advance through camera calibration experiments, characterizes the optical properties of the endoscope lens, and its expression is as follows: ; In the formula, K represents the intrinsic parameter matrix of the wireless laryngeal endoscope, which is a 3×3 matrix used to characterize the optical parameters of the lens; f x This represents the focal length of the endoscope lens along the horizontal axis of the two-dimensional image pixel coordinate system, in pixels; f y The focal length of the endoscope lens in the vertical axis of the two-dimensional image pixel coordinate system is expressed in pixels; u0 represents the coordinates of the principal point of the lens in the horizontal axis of the two-dimensional image pixel coordinate system, in pixels; v0 represents the coordinates of the principal point of the lens in the vertical axis of the two-dimensional image pixel coordinate system, in pixels. The depth mapping parameter is the correspondence between the pre-calibrated pixel depth value and the actual three-dimensional distance. d represents the actual three-dimensional depth corresponding to a pixel in the lesion area, that is, the vertical distance from that pixel to the center of the endoscope lens. The depth is collected in real time by a depth sensor and pre-calibrated. The conversion formula from two-dimensional image pixel coordinates to three-dimensional anatomical coordinates of the laryngeal cavity is as follows: ; In the formula, (W,D,Z) represents the three-dimensional coordinates of a pixel in the lesion region within the three-dimensional anatomical coordinate system of the laryngeal cavity, in millimeters; (u,v) represents the coordinates of that pixel in the two-dimensional image pixel coordinate system, in pixels; u0, v0, f x f yTo ensure consistency with the meaning of the parameters in the intrinsic parameter matrix, d represents the actual 3D depth corresponding to the pixel point, in millimeters. All pixels within the pixel coordinate range of the lesion area are inversely projected using the aforementioned conversion formula to obtain the 3D coordinates of each pixel. The set of all 3D coordinates constitutes the spatial point cloud distribution data of the lesion area in 3D space. This data accurately characterizes the actual location and morphology of the lesion area in the 3D anatomical space of the laryngeal cavity.

[0058] Step 5.3 involves performing 3D centroid calculation on the spatial point cloud distribution data to obtain the coordinates of the 3D geometric center. The spatial relative offset vector between the 3D geometric center coordinates and the current pose data is calculated to determine the final action point coordinates of the targeted drug delivery nozzle. The lesion area size and severity level are input into a preset drug delivery dose matching strategy to calculate and determine the unit-time spray flow rate and total spray duration parameters of the nebulized drug. Specifically, this includes performing 3D centroid calculation on the spatial point cloud distribution data. The 3D centroid is the 3D geometric center of the lesion area, accurately representing the core location of the lesion area, ensuring the drug delivery nozzle is aligned with the core lesion area, and improving the drug delivery effect. The calculation formula for the 3D centroid calculation is as follows: ; In the formula, The coordinates of the three-dimensional geometric center of the lesion area are represented by the three-dimensional centroid coordinates, in millimeters; N represents the total number of three-dimensional coordinate points contained in the spatial point cloud distribution data, that is, the number of three-dimensional point clouds corresponding to the lesion area. , , The coordinates of the i-th 3D point in the spatial point cloud distribution data are represented by the X, Y, and Z axes. After calculating the coordinates of the 3D geometric center, the current pose data of the wireless laryngeal endoscope is used. This pose data is collected in real time by the endoscope's built-in attitude sensor and includes the coordinates of the endoscope lens center in the 3D anatomical coordinate system of the laryngeal cavity and the lens orientation angle. The spatial relative offset vector between the 3D geometric center coordinates and the lens center coordinates in the current endoscope pose data is calculated. This offset vector is used to adjust the orientation and position of the targeted drug delivery nozzle. The formula for calculating the spatial relative offset vector is as follows: In the formula, The spatial relative offset vector between the three-dimensional geometric center coordinates and the endoscope lens center coordinates; The coordinates of the three-dimensional geometric center of the lesion area; This refers to the coordinates of the lens center in the three-dimensional anatomical coordinate system of the laryngeal cavity in the current pose data of the wireless laryngeal endoscope.

[0059] Based on the aforementioned spatial relative offset vector, the position and orientation of the targeted drug delivery nozzle are adjusted to ultimately determine the final action point coordinates of the targeted drug delivery nozzle, which accurately correspond to the three-dimensional geometric center coordinates of the lesion area.

[0060] The size of the lesion area and the severity level of the lesion are input into a preset drug dosage matching strategy. The parameters of the nebulized drug flow rate per unit time and the total spray duration are determined through quantitative calculations. The preset drug dosage matching strategy is pre-set based on clinical dosing data. Its core logic is that the larger the lesion area and the higher the severity level, the larger the required drug dose, resulting in a larger flow rate per unit time and a longer total spray duration. The formula for calculating the flow rate per unit time is: In the formula, Q represents the spray flow rate of the nebulized drug per unit time; Q0 represents the base unit flow rate, which is preset based on the drug type; S represents the actual anatomical area of ​​the lesion; L represents the lesion severity level coefficient, which is preset to 1.0 for mild lesions, 1.5 for moderate lesions, and 2.0 for severe lesions, and is used to adjust the flow rate according to the severity of the lesion. The formula for calculating the total spray duration is: In the formula, T represents the total spray duration of the atomized drug; The total dosage required for the affected area, expressed in milliliters, is determined by both the size of the lesion and its severity level. The parameters of the atomized drug's unit time spray flow rate and total spray duration are determined by calculating using the two formulas mentioned above.

[0061] Step 5.4 involves encapsulating the final point of action coordinates, jet flow rate per unit time, and total jet duration parameters into control commands and formatting the communication protocol to obtain a targeted drug delivery determination command. Specifically, this includes encapsulating control commands and formatting the communication protocol based on the final point of action coordinates, jet flow rate per unit time, and total jet duration parameters to obtain a targeted drug delivery determination command that can be recognized and executed by the wireless laryngeal endoscope drug delivery mechanism. This ensures stable command transmission and accurate execution. The control command encapsulation operation integrates the three core parameters—final point of action coordinates, jet flow rate per unit time, and total jet duration—into a unified control command data packet. The data packet also contains auxiliary information such as command identifier, data checksum, and execution priority. The command identifier is used to distinguish the targeted drug delivery command from other control commands, the data checksum is used to verify the data integrity during command transmission to prevent data loss or tampering, and the execution priority is set to the highest level to ensure that the drug delivery command is executed first and that the lesion area is treated with drug delivery in a timely manner.

[0062] By performing communication protocol formatting, the encapsulated control command data packets are converted according to the preset communication protocol format between the wireless laryngeal endoscope and the drug delivery institution, ensuring that the commands can be correctly parsed by the drug delivery institution. The preset communication protocol conforms to the communication specifications of the wireless laryngeal endoscope and includes parts such as data frame header, data length, core parameter area, check code, and frame tail. The core parameter area is used to store core parameters such as the final action point coordinates, the injection flow rate per unit time, and the total injection duration. The data frame header and frame tail are used to identify the start and end of the data frame.

[0063] After formatting, a complete targeted drug delivery determination instruction is obtained. This instruction is transmitted to the drug delivery mechanism of the wireless laryngeal endoscope via a wireless communication module. The drug delivery mechanism is driven to adjust the nozzle position to the final point of action according to the parameters in the instruction. According to the set spray flow rate per unit time and total spray duration, accurate atomized targeted drug delivery is completed to achieve targeted treatment of the lesion area.

[0064] In this embodiment of the invention, by employing technical means such as extracting core lesion data when the severity of the lesion reaches a preset threshold, establishing a two-dimensional to three-dimensional coordinate system transformation relationship through a pre-calibrated wireless laryngeal endoscope intrinsic parameter matrix and depth mapping parameters to generate a spatial point cloud, determining the drug delivery point by performing three-dimensional centroid calculation on the spatial point cloud and calculating drug delivery parameters in conjunction with the lesion condition, and encapsulating control instructions and formatting communication protocols for the core parameters, the invention overcomes the technical problems in the background technology, such as blind manual drug delivery triggering, inaccurate three-dimensional positioning of the lesion area, large deviation of the drug delivery point, mismatch between the drug delivery dose and the lesion severity, and inability to generate directly executable automated drug delivery instructions, resulting in low drug delivery accuracy, drug waste, and poor treatment effects. This achieves the technical effects of accurately triggering targeted drug delivery, accurately locating the lesion area in three dimensions, accurately matching the drug delivery point with the dose, generating directly executable targeted drug delivery judgment instructions, realizing automated and accurate targeted drug delivery, improving the standardization and accuracy of the drug delivery process, and ensuring the treatment effect of vocal cord lesions.

[0065] like Figure 2 As shown, embodiments of the present invention also provide an AI-based visual recognition-based automatic drug delivery system for vocal cord targeting, comprising: The acquisition module is used to acquire laryngeal video image data in real time; it performs image enhancement processing and wavelet transform denoising preprocessing on the laryngeal video image data to obtain the preprocessed laryngeal image. The segmentation module is used to coarsely segment the vocal cord candidate region of the preprocessed laryngeal image by the trained convolutional neural network AI visual recognition model to obtain the initial vocal cord region mask. The module is used to refine the vocal cord edge extraction of the preset active contour model based on the initial vocal cord region mask, construct the geometric energy functional, and obtain the sub-pixel level vocal cord contour curve by iteratively optimizing the geometric energy functional; and to perform geometric topological connectivity analysis based on the sub-pixel level vocal cord contour curve to obtain the vocal cord region recognition result. The assessment module is used to combine the coordinate range of the lesion area, the proportion of the lesion area and the lesion type classification information in the vocal cord region identification results with the preset vocal cord lesion judgment criteria to quantitatively assess the health status of the vocal cords and obtain the assessment results. The mapping module is used to map the pixel coordinate range, lesion area size, and severity level of the lesion area in the laryngeal image to three-dimensional spatial coordinates based on the evaluation results when it is determined that there is a lesion area that needs to be drugged. This is to determine the final action point coordinates of the targeted drug delivery nozzle relative to the center of the lesion area, and to determine the spray flow rate and spray duration of the nebulized drug according to the lesion area and severity level, so as to generate a targeted drug delivery determination instruction.

[0066] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0067] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0068] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0069] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for automatically determining vocal cord-targeted drug delivery based on AI visual recognition, characterized in that, The method includes: Step 1: Real-time acquisition of laryngeal video image data; image enhancement processing and wavelet transform denoising preprocessing are performed on the laryngeal video image data to obtain the preprocessed laryngeal image; Step 2: The trained convolutional neural network AI visual recognition model is used to coarsely segment the vocal cord candidate region of the preprocessed laryngeal image to obtain the initial vocal cord region mask. Step 3: Based on the initial vocal cord region mask, the vocal cord edges of the preset active contour model are extracted in a refined manner, and a geometric energy functional is constructed. By iteratively optimizing the geometric energy functional, a sub-pixel level vocal cord contour curve is obtained. Based on the sub-pixel level vocal cord contour curve, geometric topological connectivity analysis is performed to obtain the vocal cord region recognition result. Step 4: Combine the lesion region coordinate range, lesion area ratio, and lesion type classification information from the vocal cord region identification results with the preset vocal cord lesion judgment criteria to quantitatively assess the vocal cord health status and obtain the assessment results. Step 5: Based on the evaluation results, when it is determined that there is a lesion area that needs to be administered, the pixel coordinate range, lesion area size and severity level of the lesion area in the laryngeal image are mapped to three-dimensional spatial coordinates to determine the final action point coordinates of the targeted drug delivery nozzle relative to the center of the lesion area. The spray flow rate and spray duration of the nebulized drug are determined according to the lesion area and severity level to generate a targeted drug delivery determination instruction.

2. The method for automatically administering medication to vocal cords based on AI visual recognition according to claim 1, characterized in that, Step 1: Real-time acquisition of video image data of the larynx; Image enhancement and wavelet transform denoising preprocessing are performed on the throat video image data to obtain the preprocessed throat image, including: The micro-imaging module of the wireless laryngeal endoscope continuously captures dynamic video streams of the laryngeal cavity, extracts discrete video frames at preset frame intervals, and combines the discrete video frames into original laryngeal video image data. Multi-scale illumination compensation and local contrast stretching were performed on the original laryngeal video image data to obtain an enhanced laryngeal image with balanced illumination. Discrete wavelet transform decomposition is performed on the enhanced throat image to obtain a low-frequency approximate subband and a high-frequency detail subband. Based on the coefficient distribution characteristics of the high-frequency detail subband, an adaptive threshold shrinkage strategy is used to perform nonlinear filtering on the high-frequency coefficients to obtain the filtered wavelet coefficient set. By performing inverse wavelet transform on the wavelet coefficient set and the low-frequency approximate subband, a preprocessed throat image with noise suppression and clear texture is obtained.

3. The method for automatically administering medication to vocal cords based on AI visual recognition according to claim 2, characterized in that, Step 2: The trained convolutional neural network AI visual recognition model performs coarse segmentation of the vocal cord candidate region on the preprocessed laryngeal image to obtain an initial vocal cord region mask, including: The preprocessed laryngeal image is input into the encoder network of the convolutional neural network AI visual recognition model to extract multi-scale deep semantic feature maps of laryngeal anatomy, compress spatial dimensions and aggregate tissue contextual information to obtain high-level semantic feature representations. The high-level semantic feature representation is input into the decoder network to gradually restore the spatial resolution of the multi-scale deep semantic feature map, and cross-level feature fusion is performed to obtain the fused feature map. The fused feature map is input into a pixel-level classification convolutional layer, and the probability confidence of each image coordinate point belonging to the vocal cord tissue region is calculated to obtain the probability distribution map of the vocal cord region. Based on a preset confidence threshold, the probability distribution map of the vocal cord region is binarized and thresholded. Pixel regions with probability values ​​higher than the threshold are marked as candidate foreground regions of the vocal cord, and pixel regions with probability values ​​lower than the threshold are marked as background regions, thus obtaining the initial vocal cord region mask.

4. The method for automatically administering medication to vocal cords based on AI visual recognition according to claim 3, characterized in that, Step 3: Based on the initial vocal cord region mask, the preset active contour model is subjected to refined extraction of vocal cord edges, and a geometric energy functional is constructed. Through iterative optimization of the geometric energy functional, sub-pixel-level vocal cord contour curves are obtained, including: Using the boundary pixel set of the initial vocal cord region mask as the initial closed curve, a geometric energy functional containing an internal geometric smoothing constraint term and an external image gradient driving term is constructed. Based on the variational evolution mechanism of the preset active contour model, the geometric energy functional is solved by minimizing energy, and the control equation for the spatial position evolution of the contour curve is derived. The control equation is solved by the discretized numerical iteration method. During the iteration process, the spatial coordinates of the contour nodes are obtained based on the vector superposition result of the gradient potential field traction force and the internal smoothing constraint force. The contour node positions are continuously updated iteratively, so that the closed curve migrates and fits to the actual anatomical edge of the vocal cord under the control of the energy functional convergence condition, until the contour deformation displacement of the adjacent iteration cycle is lower than the preset convergence threshold, thus obtaining the sub-pixel level vocal cord contour curve.

5. The method for automatically administering medication to vocal cords based on AI visual recognition according to claim 4, characterized in that, Geometric topological connectivity analysis based on sub-pixel level vocal cord contour curves yields vocal cord region identification results, including: Perform connected component traversal marking based on preset neighborhood rules on the internal pixel region enclosed by the sub-pixel level vocal cord contour curve to identify the set of independent foreground regions that are interconnected at the pixel level in the image space. By calculating the geometric center coordinates, outer envelope size and pixel filling area of ​​each independent foreground region, the target connected region located in the glottic region below the epiglottis and satisfying the condition of left-right symmetry distribution is obtained. By removing the irregularly clustered ventricular zone interference region above the target connected domain, as well as the rigid tissue interference region of the arytenoid cartilage with a small radius of curvature on the posterolateral side of the vocal cords, the set of the main connected domains of the vocal cords after geometric topological filtering is obtained. Based on the set of connected domains of the vocal cord body, contour boundary smoothing and adjacent pixel region fusion processing are performed to obtain vocal cord region recognition results that fully cover the vocal cord tissues on both sides.

6. The method for automatically administering medication to vocal cords based on AI visual recognition according to claim 5, characterized in that, Step 4: Combine the lesion region coordinates, lesion area percentage, and lesion type classification information from the vocal cord region identification results with preset vocal cord lesion judgment criteria to quantitatively assess the vocal cord health status and obtain the assessment results, including: The results of vocal cord region identification are analyzed at the pixel level to extract the pixel coordinate boundary range of the lesion area, the area ratio of the lesion area to the total area of ​​the vocal cord, and the lesion type classification information based on color distribution and texture features. The coordinate range, area proportion, and lesion type classification information of the lesion area are compared and matched with the morphological size threshold, area proportion threshold, and pathological type grading rules in the preset vocal cord lesion judgment criteria to obtain the deviation of each lesion indicator. A multidimensional health status quantitative mapping model is constructed based on deviation, and the overall health status of the vocal cords is weighted and graded to obtain an assessment result that includes the severity level of the lesion and the health status score.

7. The method for automatically administering medication to vocal cords based on AI visual recognition according to claim 6, characterized in that, Step 5: Based on the evaluation results, when a lesion area requiring drug administration is determined, the pixel coordinate range, lesion area size, and severity level of the lesion area in the laryngeal image are mapped to three-dimensional spatial coordinates to determine the final action point coordinates of the targeted drug delivery nozzle relative to the center of the lesion area. The spray flow rate and spray duration of the nebulized drug are determined according to the lesion area and severity level to generate a targeted drug delivery determination instruction, including: When the severity level of the lesion in the assessment results reaches the preset drug administration trigger threshold, the pixel coordinate range, lesion area size, and severity level data of the lesion area in the laryngeal image are extracted. Based on the pixel coordinate range, the transformation relationship between the two-dimensional image pixel coordinate system and the three-dimensional anatomical coordinate system of the laryngeal cavity is established by using the pre-calibrated intrinsic parameter matrix and depth mapping parameters of the wireless laryngeal endoscope. The pixel coordinate range is then inversely projected and calculated into spatial point cloud distribution data of the lesion area in three-dimensional space. The three-dimensional centroid of the spatial point cloud distribution data is calculated to obtain the coordinates of the three-dimensional geometric center. The spatial relative offset vector between the three-dimensional geometric center coordinates and the current pose data is calculated to determine the final action point coordinates of the targeted drug delivery nozzle. The size of the lesion area and the severity level are input into the preset drug delivery dose matching strategy to calculate and determine the parameters of the unit time spray flow rate and total spray duration of the nebulized drug. The final point of action coordinates, injection flow rate per unit time, and total injection duration parameters are encapsulated into control commands and formatted with a communication protocol to obtain the targeted drug delivery determination command.

8. An AI-based visual recognition-based automatic drug delivery system for vocal cord targeting, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The acquisition module is used to acquire video image data of the larynx in real time; Image enhancement and wavelet transform denoising preprocessing are performed on the throat video image data to obtain the preprocessed throat image; The segmentation module is used to coarsely segment the vocal cord candidate region of the preprocessed laryngeal image by the trained convolutional neural network AI visual recognition model to obtain the initial vocal cord region mask. The module is used to refine the vocal cord edge extraction of the preset active contour model based on the initial vocal cord region mask, construct the geometric energy functional, and obtain the sub-pixel level vocal cord contour curve by iteratively optimizing the geometric energy functional. Geometric topological connectivity analysis based on sub-pixel level vocal cord contour curves yields vocal cord region identification results. The assessment module is used to combine the coordinate range of the lesion area, the proportion of the lesion area and the lesion type classification information in the vocal cord region identification results with the preset vocal cord lesion judgment criteria to quantitatively assess the health status of the vocal cords and obtain the assessment results. The mapping module is used to map the pixel coordinate range, lesion area size, and severity level of the lesion area in the laryngeal image to three-dimensional spatial coordinates based on the evaluation results when it is determined that there is a lesion area that needs to be drugged. This is to determine the final action point coordinates of the targeted drug delivery nozzle relative to the center of the lesion area, and to determine the spray flow rate and spray duration of the nebulized drug according to the lesion area and severity level, so as to generate a targeted drug delivery determination instruction.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images