A multi-channel endoscopic image feature enhancement method, apparatus, and medium

Abstract

A multi-channel endoscope image feature enhancement method, device and medium, medical image processing technology field, the main steps include: endoscope image acquisition; image preprocessing; based on the corresponding relationship model of tissue spectral characteristics-pathology, the corresponding biological tissue feature enhancement analysis is carried out on each spectral channel; multi-channel fusion; and self-adaptive enhancement based on histopathology guidance, realize the mapping from physical image attribute to pathological significance. The method forms the whole process optimization of "color perception-detail enhancement-feature quantization" through systematic image preprocessing, specific enhancement and post-processing optimization, etc., provides an integrated solution from image enhancement to quantitative analysis for endoscope intelligent diagnosis system, and can realize high-quality integration of multi-channel images and effective improvement of lesion saliency.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method, apparatus and medium for enhancing the features of multi-channel endoscopic images. Background Technology

[0002] Multispectral endoscopy utilizes light of different wavelengths to acquire more information about tissues, making it more advanced than traditional white light endoscopy. However, some issues exist in image processing technology.

[0003] Multispectral images may come from different light sources or sensors, and accurately registering these images is a challenge. Registration errors can lead to information misalignment, affecting doctors' judgment of lesion location. Poor fusion techniques may lose important details or introduce artifacts. Noise can further interfere with the signal, affecting diagnostic accuracy.

[0004] Furthermore, variations in tissue optical properties can lead to significant differences in these properties between individuals, such as pigmentation and blood flow. Existing processing techniques may not be well adapted to these variations, resulting in biases in the interpretation of spectral data. In addition, uneven illumination can also affect image quality; for example, uneven distribution of endoscopic light sources can lead to inaccurate local spectral data.

[0005] In clinical applications, doctors need specialized training to interpret multispectral images, and existing visualization methods are not intuitive enough, leading to low diagnostic efficiency, and so on.

[0006] The existence of these problems limits the widespread application of multispectral endoscopy technology. Summary of the Invention

[0007] This disclosure provides a multi-channel endoscopic image feature enhancement method and overall process, aiming to achieve high-quality integration of multi-channel images and improve the significance of lesions through systematic image preprocessing, specific enhancement and deep fusion.

[0008] The multispectral endoscopic image enhancement method disclosed herein mainly includes the following steps:

[0009] S1, Multi-channel endoscopic image acquisition;

[0010] S2, Image preprocessing, including: multi-channel image alignment and registration; noise suppression and motion artifact correction; reflectance and illuminance correction for each channel image;

[0011] S3, based on the correspondence model between tissue spectral characteristics and pathology, performs corresponding biological tissue feature enhancement analysis on each spectral channel;

[0012] S4, fuse the input images from each channel;

[0013] S5, Post-processing optimization: Adaptive enhancement based on histopathology guidance to achieve mapping from physical image attributes to pathological meaning.

[0014] Furthermore, the multi-channel endoscopic image in step S1 includes: original white light, red light, green light and blue light four-channel data images.

[0015] Furthermore, in step S2, the multi-channel image alignment and registration steps specifically include:

[0016] (1) Scale-space extremum detection:

[0017] First, based on the improved SIFT algorithm, the input image is processed... Perform multi-scale Gaussian blurring to generate a Gaussian pyramid. The specific formula is as follows:

[0018]

[0019] Among them, parameter α is introduced to control the anisotropy of the Gaussian kernel, and γ is a tissue reflectance correction factor, which is used to adapt to the nonlinear illumination response of the endoscopic image. The scale parameter determines the degree of ambiguity;

[0020] The two-dimensional Gaussian function is:

[0021] Secondly, construct the difference Gaussian pyramid:

[0022]

[0023] Where k is the scale multiplication factor; β(σ) is the scale adaptive compensation factor, used to improve the sensitivity of feature extraction of microvascular structures and glandular textures in endoscopic images:

[0024] λ and τ are set to 0.15 and 1.8 respectively;

[0025] Then, extreme value detection is performed: for each pixel in the DoG pyramid, it is compared with 26 points, including 8 neighboring points at the same scale and 9×2 neighboring points at adjacent scales. If the point is a maximum or minimum value, it is considered a candidate keypoint.

[0026] (2) Precise key point positioning to eliminate low contrast or edge response points and improve stability:

[0027] The DoG function is differentiated to calculate the precise offset of the extreme point. If the offset is less than a threshold, the low-contrast point is discarded. Specifically, for low-texture regions of the endoscope, the contrast threshold criterion is modified to an adaptive local threshold.

[0028]

[0029] Where VAR(P) is the local variance of the neighborhood of key point P, and η is the weight coefficient (with a value of 0.8), thus preserving weak feature points in low-contrast regions.

[0030] For image edges, response suppression is applied, and edge points are removed using the Hessian matrix:

[0031]

[0032] Calculate the curvature ratio:

[0033]

[0034] Where r takes the value of 10; if this condition is not met, it is determined to be an edge point and is removed.

[0035] (3) Assign directions to key points to achieve rotational invariance:

[0036] Divide the 0°~360° range into 36 intervals and calculate a weighted histogram of gradient magnitudes, with the weights being Gaussian window weights;

[0037] Take the peak value of the histogram as the main direction. If there is a secondary peak, generate multi-directional key points.

[0038] (4) Perform descriptor generation to generate feature vectors with illumination, scale, and rotation invariance:

[0039] The neighborhood of the key point is divided into 4×4 sub-regions, and the gradient histogram of each sub-region is calculated in 8 directions, for a total of 4×4×8=128-dimensional vectors.

[0040] The descriptors are normalized, and large values ​​of light sensitivity are truncated, limiting each component to no more than 0.2, and finally normalized again.

[0041] Furthermore, in step S2, the noise suppression and motion artifact correction steps specifically include:

[0042] Noise suppression was performed on the red and blue light channel images. For the red light channel, which represents deep tissue, a nonlocal averaging method was used for noise reduction to preserve vascular texture. For the blue light channel, which represents the superficial mucosa, wavelet thresholding was used for noise reduction to protect high-frequency details.

[0043] Furthermore, in step S2, the steps for correcting the reflectance and illuminance of each channel image specifically include:

[0044] Based on Retinex theory, the reflection component and illumination component of each channel are separated and normalized to eliminate the influence of non-uniform illumination.

[0045] Furthermore, the specific method of step S3 includes:

[0046] S30, based on large-scale clinical annotation data, obtains the mapping relationship between reflectance at different wavelengths and tissue pathological state, and establishes a correspondence model between tissue spectral characteristics and pathology.

[0047] S31, White Light Channel Enhanced Analysis: The feature vector of the white light reflectance curve is extracted by principal component analysis, and the deviation is quantified relative to the standard healthy tissue template using the tissue color mapping function. This is used to efficiently detect abnormal color areas and is sensitive to the slight color difference caused by early cancer.

[0048] The mapping function is:

[0049] Among them, w i For RGB channel weights, T healthy A template for healthy organizations;

[0050] I white (x,y) i Let represent the i-th color component value at position (x, y) in the white light channel image, where i = 1, 2, 3 correspond to the three RGB color channels respectively:

[0051] When i=1, it represents the value of the red (R) channel;

[0052] When i=2, it represents the value of the green (G) channel;

[0053] When i=3, it represents the value of the blue (B) channel;

[0054] S32, for enhancing vascular features in the red light channel, is used for deep tissue analysis:

[0055] Constructing a blood vessel connectivity graph G v :

[0056]

[0057] Where V is the set of blood vessel nodes, E is the set of connecting edges, and W is the weight function based on depth d and direction θ;

[0058] Analysis of G v The topological properties are used to detect abnormal blood vessel growth patterns;

[0059] Design a depth-sensing blood vessel recognition function to distinguish between vascular networks at different depths:

[0060]

[0061] Among them, H frangiFor the improved Frangi filter, D tissue This is a tissue depth estimation map, where κ is the depth attenuation coefficient;

[0062] S33 enhances the capillary density characteristics of the green light channel:

[0063] Introducing a multi-scale vessel enhancement filter optimized with Hessian matrix:

[0064]

[0065] In the formula, λ1 and λ2 are the eigenvalues ​​of the Hessian matrix, RB is the measure of vessel circularity, and S is the measure of structure. and Set them to 0.5 and 15 respectively;

[0066] Based on a multi-scale vascular enhancement filter, the capillary density index V is calculated. density It is used to assess the microcirculatory status, which is highly sensitive to inflammation and early tumors.

[0067]

[0068] Where A is the area of ​​the region of interest, τ v This refers to the vascular threshold.

[0069] S34, for the blue light channel, enhances the surface mucosal texture features:

[0070] Design the fusion feature vector of LBP local binary mode and GLCM gray-level co-occurrence matrix:

[0071]

[0072] Among them, F texture (x, y) is the texture feature vector at position (x, y); LBP r，p (x, y) represents a local binary pattern feature with radius r and number of points p; Contrast GLCM The contrast feature of the gray-level co-occurrence matrix; Correlation GLCM Correlation characteristics of the gray-level co-occurrence matrix; Energy GLCM The energy characteristic of the gray-level co-occurrence matrix; Homogeneity GLCM This represents the homogeneity characteristic of the gray-level co-occurrence matrix;

[0073] LBP r,p A rotation-invariant LBP operator with a radius of 2 and a number of points of 8 is adopted;

[0074] Based on this fused feature vector, a mucosal gland morphology analysis model is constructed:

[0075]

[0076] in, The i-th component of the texture feature vector; parameter α i The values ​​were obtained by comparing samples of normal glands and atypical glands, with values ​​of [0.25, 0.3, 0.15, 0.1, 0.2] respectively; this model is used to identify early intestinal metaplasia and glandular structural disorders.

[0077] Furthermore, step S5 specifically includes:

[0078] S51, based on a tissue-lesion association model, performs pseudo-color mapping to enhance lesion expression using medical priors:

[0079]

[0080] Among them, C pseudo This is the RGB color image after pseudo-color synthesis, where (x,y) represents the pixel coordinates in the image; f HSV2RGB This is a conversion function from HSV color space to RGB color space;

[0081] Among them, H semantic Hue assignment function based on organizational semantics:

[0082]

[0083] In the formula, H is the balance parameter. base Based on hue, H pathology For pathology-guided color mapping, T(x,y) represents the tissue classification result;

[0084] S enhance As a saturation enhancement function, an adaptive strategy based on lesion significance is adopted:

[0085]

[0086] In the formula, S base Based on saturation, L is the magnification factor. abnormal Probability map of abnormal regions;

[0087] V balanced The brightness balance function is achieved through a piecewise S-curve:

[0088]

[0089] Among them, V raw(x, y) represents the brightness channel values ​​of the original image, i.e., the original data of the V channel in the HSV color space; the brightness balance function V balanced The original brightness is non-linearly mapped using the Sigmoid function, with parameters... The steepness of the control curve, i.e., contrast adjustment, and θ(x,y) control the position of the center point, i.e. brightness adjustment, are both parameters that will be adaptively adjusted according to tissue depth and endoscope lens distance, so that tissues at different depths can obtain appropriate brightness display.

[0090] S52, adaptive enhancement of pathological regions through multi-scale, multi-feature fusion:

[0091] Generating lesion probability maps through multi-model ensemble:

[0092]

[0093] Among them, M vessel M texture M color and M depth These are detection models for vascular abnormalities, texture abnormalities, color abnormalities, and depth abnormalities, respectively.

[0094] Based on this lesion probability map, a lesion-specific enhancement filter is constructed:

[0095]

[0096] In the formula, I input The original input image, To enhance the coefficient, 2 G σ(x,y) For the adaptive scaling Laplacian operator, σ(x,y) is dynamically adjusted according to the lesion size:

[0097]

[0098] σ base The base scale, δ is the adjustment factor, Size lesion For local lesion size estimation, parameter σ base With δ set to 1.0 and 0.3 respectively, this mechanism ensures optimal enhancement for lesions of different sizes.

[0099] Furthermore, the method also includes: a step of performing configurable goal-oriented feature enhancement based on different clinical diagnostic goals, specifically including:

[0100] Early cancer detection modality: Enhanced detection of irregular microvascular distribution and mucosal gland atypicality.

[0101] ;

[0102] Inflammation assessment model: Enhanced vasodilation and mucosal hyperemia features

[0103] ;

[0104] Bleeding monitoring mode: Optimized display of deep tissue bleeding points

[0105] ;

[0106] Among them, I cancer (x,y) represents the optimized image for early cancer detection, which enhances cancer-related features;

[0107] V enhance (x,y) is a vascular enhancement image generated by a multi-scale vascular enhancement filter optimized by the Hessian matrix mentioned earlier, highlighting the irregular distribution of microvessels, which is an important feature of early carcinogenesis.

[0108] G pattern (x,y) represents the morphological analysis results of mucosal glands, which are used to detect glandular atypia, an important characteristic of early cancer.

[0109] I base (x,y) is the base image, usually a reference image of the preprocessed white light image, providing a basic anatomical reference;

[0110] V density (x,y) is a capillary density index map, which quantitatively represents the abundance of blood vessels per unit area;

[0111] f color (x,y) represents the tissue color mapping function result, which is used to detect tissue color abnormalities, especially mucosal congestion characteristics, and is an important basis for inflammation judgment;

[0112] I red (x,y) is the image after red light channel enhancement, which has high sensitivity to deep tissue bleeding points;

[0113] f threshold This is a thresholding function used to further enhance the salience of bleeding points;

[0114] f threshold (I ratio (x,y)) is a ratio image processed by a threshold function, where I ratio A ratio plot of different spectral channels is used to better identify bleeding areas;

[0115] The parameter vectors [λ1,λ2,λ3], [μ1,μ2,μ3], and [ν1,ν2,ν3] were clinically validated to be [0.4,0.3,0.3], [0.5,0.2,0.3], and [0.6,0.3,0.1], respectively.

[0116] A multispectral endoscopic image enhancement device applying the above method mainly includes:

[0117] Multi-channel endoscopic image acquisition module;

[0118] The image preprocessing module is used for alignment and registration of multi-channel images, noise suppression and motion artifact correction, as well as reflectance and illuminance correction of each channel image.

[0119] The multispectral endoscopic image enhancement module is used to enhance and analyze the corresponding biological tissue features of each spectral channel based on the tissue spectral characteristics-pathology correspondence model.

[0120] The multi-channel fusion module is used to fuse the input images from different channels;

[0121] The post-processing optimization module is used for histopathology-guided adaptive enhancement to achieve the mapping from physical image attributes to pathological meaning.

[0122] A storage medium storing an executable program, which, when invoked, executes the aforementioned multispectral endoscopic image enhancement method.

[0123] Compared with the prior art, the beneficial effects of this disclosure are: ① By synchronously acquiring multispectral data, high-dimensional input can be provided for subsequent spectral decoupling and information fusion;

[0124] ② Image preprocessing effectively eliminates spatial misalignment caused by time difference, providing a consistent spatial reference for channel fusion and comparison; removes motion artifacts caused by endoscope jitter, making the image clearer and more stable; and counteracts the effects of non-uniform lighting during the shooting process, ensuring that the subsequent enhancement steps process real tissue reflection information, rather than false brightness changes caused by lighting differences.

[0125] ③ By using a spectral characteristic decomposition model, a specific enhancement strategy is applied to the channel images to maximize the diagnostic value of different channels in the region of interest;

[0126] ④ Through deep fusion, enhanced images containing red, green, blue, and white light information can be generated, preserving the advantageous features of each channel and achieving collaborative alignment in spatial structure;

[0127] ⑤ Through post-optimization processing, the readability for doctors is improved, the salience of lesion areas is enhanced, and potential lesion areas are highlighted to assist in early clinical detection and precise localization. Attached Figure Description

[0128] The above and other objects, features and advantages of this disclosure will become more apparent from the more detailed description of exemplary embodiments of this disclosure taken in conjunction with the accompanying drawings, in which the same reference numerals generally represent the same components.

[0129] Figure 1 This is an overall flowchart of the multispectral endoscopic image enhancement described in this disclosure;

[0130] Figure 2 A detailed flowchart for an exemplary embodiment. Detailed Implementation

[0131] Preferred embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While preferred 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 the present disclosure will be thorough and complete, and will fully convey the scope of the present disclosure to those skilled in the art.

[0132] This disclosure provides a method for enhancing multispectral endoscopic images. The overall process is shown in the attached figure. Figure 1 As shown, the main steps include: endoscopic image acquisition, image preprocessing, multispectral feature analysis, multichannel fusion, and post-processing optimization.

[0133] (a) Acquisition of endoscopic images

[0134] The specific operation for acquiring endoscopic images involves acquiring raw white light, red light, green light, and blue light four-channel data images from the endoscope. White light LED cold light source, red light LED light source, green light LED light source, and blue light LED light source are respectively connected into the endoscope light cone orifice to obtain four-channel endoscopic images.

[0135] Dynamic frequency modulation technology is introduced in the endoscopic image acquisition stage, employing differentiated modulation strategies for different light source characteristics: white LEDs use 4800Hz high-frequency PWM modulation to reduce flicker; red LEDs (620-700nm) use a 2500Hz modulation frequency with precise control of the peak wavelength at 650±5nm; green LEDs use 3200Hz modulation with the wavelength locked at 540±3nm (the maximum absorption peak of hemoglobin); and blue LEDs use 3800Hz modulation with the wavelength concentrated at 470±3nm (the optimal reflection band for surface tissue). Furthermore, an optical bandpass filter array is added at the front end of the light source to effectively suppress spectral overlap and ensure the spectral purity of multi-channel data.

[0136] (II) Image Preprocessing Stage

[0137] The image preprocessing stage includes aligning and registering the four-channel images to eliminate minor displacements caused by shooting time differences; suppressing noise in the red and blue light channel images to eliminate endoscope motion artifacts; and separating the reflection and illumination components of each channel to eliminate non-uniform illumination.

[0138] For its operation, the Scale Invariant Feature Transform (SIFT) algorithm is used to eliminate the small displacements caused by the shooting time difference. SIFT specifically includes scale space extremum detection, key point precise localization, orientation assignment and descriptor generation.

[0139] (1) Scale-space extremum detection first involves constructing a Gaussian pyramid, i.e., processing the input image. Perform multi-scale Gaussian blurring to generate a Gaussian pyramid.

[0140] To adapt to the high reflectivity and low contrast characteristics unique to endoscopic images, this embodiment specifically optimizes the traditional SIFT algorithm by introducing a nonlinear response function and adaptive contrast normalization. First, the Gaussian pyramid construction in the scale-space extremum detection step is improved as follows:

[0141]

[0142] An additional parameter α is introduced to control the anisotropy of the Gaussian kernel; γ is a tissue reflectance correction factor (the optimal value was determined to be 1.2) to adapt to the nonlinear illumination response of endoscopic images. The scale parameter determines the degree of ambiguity.

[0143] The two-dimensional Gaussian function in the formula is:

[0144] .

[0145] Secondly, a difference Gaussian pyramid is constructed, approximating the Laplacian Gaussian (LoG) operator by the difference between Gaussian images of adjacent scales:

[0146] In this embodiment, the DoG pyramid construction is improved:

[0147]

[0148] Introducing the scale-adaptive compensation factor β(σ), its mathematical expression is:

[0149]

[0150] In the formula, λ and τ are set to 0.15 and 1.8, respectively. This compensation mechanism significantly improves the sensitivity of feature extraction for microvascular structures and glandular textures in endoscopic images; k is the scale multiplication factor (usually taken as...). ).

[0151] Then, extreme value detection is performed. For each pixel in the DoG pyramid, it is compared with 26 points, including 8 neighboring pixels at the same scale and 9×2 neighboring pixels at adjacent scales. If the pixel is a maximum or minimum value, it is considered a candidate keypoint.

[0152] (2) Precise location of key points. The purpose of this work is to eliminate low contrast or edge response points and improve stability. The precise offset of the extreme point is calculated by taking the derivative of the DoG function. If the offset is less than 0.3 (threshold), the low contrast point is eliminated.

[0153] In this step, for low-texture areas of the endoscope, the contrast threshold determination criterion is modified to an adaptive local threshold:

[0154]

[0155] Where VAR(P) is the local variance of the neighborhood of key point P, and η is the weight coefficient (with a value of 0.8), which enables the algorithm to better preserve weak feature points in low-contrast regions.

[0156] For image edges, response suppression is required, and edge points are removed using the Hessian matrix:

[0157]

[0158] Calculate the curvature ratio:

[0159]

[0160] The value of r is 10. If this condition is not met, the point is identified as an edge point and removed.

[0161] (3) Direction assignment is performed to assign a main direction to the key point and achieve rotation invariance. The 0°~360° range is divided into 36 intervals (each interval is 10°), and a weighted histogram of gradient magnitudes is calculated (the weights are Gaussian window weights). The peak value of the histogram is taken as the main direction. If there is a secondary peak (>80% of the main peak magnitude), then multi-directional key points are generated.

[0162] (4) Then, descriptor generation is performed to generate feature vectors that are invariant to illumination, scale, and rotation. The neighborhood of the keypoint is divided into 4×4 sub-regions, and gradient histograms in 8 directions are calculated for each sub-region (a total of 4×4×8=128-dimensional vectors). The descriptors are normalized, and large illumination-sensitive values ​​are truncated (usually each component is limited to no more than 0.2), and finally, they are normalized again.

[0163] Noise suppression is performed on the red and blue light channels. Non-local means (NLM) is used for denoising the red light (deep tissue) channel to preserve vascular texture. Wavelet thresholding is used for denoising the blue light (superficial mucosa) channel to preserve high-frequency details.

[0164] When performing endoscopic illumination, the captured images are unevenly distributed. The method used is to separate the reflection component (tissue characteristics) and illumination component of each channel based on Retinex theory to eliminate the influence of non-uniform illumination.

[0165] (III) Multispectral Feature Analysis Stage

[0166] Multispectral feature analysis introduces a tissue spectral property-pathological correspondence model, establishing a mapping relationship between reflectance at different wavelengths and tissue pathological states based on large-scale clinical labeled data. Specifically:

[0167] 1. White light (visible light composite band)

[0168] Wavelength range: approximately 400–700 nm.

[0169] Spectral characteristics: Composed of multiple bands, the spectral composition is complex but complete.

[0170] Tissue response: Simultaneously illuminates the surface layer and a certain depth structure of the mucosa, revealing the tissue's natural color.

[0171] Imaging characteristics: It has good overall morphological visibility and intuitive anatomical structure, making it suitable for basic observation and navigation.

[0172] Advantages: High color fidelity and wide field of view; Limitations: Not sensitive to contrast in fine blood vessels and early lesions.

[0173] White light channel enhancement analysis: The feature vector of the white light reflectance curve is extracted using principal component analysis and its deviation is quantified by comparing it with a standard healthy tissue template. The tissue chromaticity mapping function fcolor(x,y) is introduced, and the calculation formula is as follows:

[0174]

[0175] Where w i The RGB channel weights are 0.3, 0.5, and 0.2 respectively. T healthy For reference templates of healthy organizations, I white (x,y) i Let be the i-th color component value at position (x, y) in the white light channel image. Here, i = 1, 2, 3 correspond to the three RGB color channels respectively.

[0176] When i=1, it represents the value of the red (R) channel;

[0177] When i=2, it represents the value of the green (G) channel;

[0178] When i=3, it represents the value of the blue (B) channel.

[0179] This mapping can efficiently detect abnormal color areas, especially sensitive to subtle color differences caused by early cancerous changes.

[0180] 2. Red light (600–700 nm)

[0181] Wavelength characteristics: long wavelength and strong tissue penetration.

[0182] Tissue response: It can penetrate 1–2 mm and act on deeper tissues.

[0183] Imaging characteristics: Deep structures (such as large blood vessels and hemorrhage) are clearly visible, and the overall image is bright with lower contrast than green light.

[0184] Advantages and limitations: It can visualize deep information, but the resolution of surface details is relatively poor.

[0185] Red-light channel deep tissue analysis: A depth-corrected vascular enhancement algorithm is introduced into the red-light channel to construct a vascular connectivity graph G. v :

[0186]

[0187] Where V is the set of vessel nodes, E is the set of connecting edges, and W is a weighting function based on depth d and direction θ. Through analysis of G... v The topological properties of the blood vessels can be used to quantify and detect abnormal blood vessel growth patterns (such as the "spiky" structure of tumor neovascularization). A depth-sensing blood vessel recognition function is designed:

[0188]

[0189] Where H frangi For the improved Frangi filter, D tissue This is a tissue depth estimation map, where κ is the depth attenuation coefficient (value 0.6). This function can effectively distinguish between vascular networks at different depths.

[0190] 3. Green light (500–600 nm)

[0191] Wavelength characteristics: located at the hemoglobin absorption peak (540–580 nm).

[0192] Tissue response: Strong absorption makes blood vessels clearly visible.

[0193] Imaging features: High-contrast rendering of superficial capillary networks enhances the visibility of blood vessels in the tissue.

[0194] Advantages and limitations: Excellent contrast of blood vessels, suitable for dynamic observation, but limited penetration depth, limited to superficial tissues.

[0195] Green channel feature enhancement: A capillary density quantification analysis method based on the green channel is developed, and a multi-scale vessel enhancement filter optimized by the Hessian matrix is ​​introduced.

[0196]

[0197] In the formula, λ1 and λ2 are the eigenvalues ​​of the Hessian matrix, RB is the measure of vessel circularity, and S is the measure of structure. and The values ​​are set to 0.5 and 15 respectively. The capillary density index V is calculated using this filter. density :

[0198]

[0199] Where A is the area of ​​the region of interest, τ v The vascular threshold (0.35) is an index that can objectively assess the microcirculation status and is highly sensitive to inflammation and early tumors.

[0200] 4. Blue light (400–500 nm)

[0201] Wavelength characteristics: short wavelength, strong scattering.

[0202] Tissue response: Primarily reflected in superficial tissues (<0.5 mm).

[0203] Imaging characteristics: enhanced surface texture, gland openings, and fine structures. The image is rich in detail but has low brightness.

[0204] Advantages and limitations: High surface resolution, suitable for microstructure analysis but unable to observe deep tissues, and has a low signal-to-noise ratio.

[0205] Blue light channel surface analysis: An algorithm for extracting surface mucosal texture features for the blue light channel was developed, and a fusion feature vector of Local Binary Pattern (LBP) and Gray-Level Co-occurrence Matrix (GLCM) was designed.

[0206]

[0207] LBP r,p A rotation-invariant LBP operator with a radius of 2 and 8 points is used. Based on this, a mucosal gland morphology analysis model is constructed using fused feature vectors.

[0208]

[0209] Parameter α iThe values ​​were obtained by comparing samples of normal glands and atypical glands, with values ​​of [0.25, 0.3, 0.15, 0.1, 0.2], which can accurately identify early intestinal metaplasia and glandular structural disorders.

[0210] By using white light to provide an overall structural background, combined with red light for deep vascular information, green light for high-contrast imaging of capillaries, and blue light for enhanced surface microstructure, comprehensive diagnosis and early lesion identification of the digestive tract mucosa from the surface to the depths can be achieved. By combining a spectral reflectance model with image fusion strategies, multi-channel images not only improve lesion visualization but also provide high-quality input for intelligent assisted diagnosis.

[0211] (iv) Multi-channel integration stage

[0212] By fusing multimodal image information and preserving complementary information from multiple spectral channels, deep blood vessels and mucosal texture structures are clearly displayed, while cross-channel noise and registration residuals are suppressed, generating endoscopic images with color gamut and contrast that conform to clinical diagnostic intuition.

[0213] (v) Post-processing optimization stage

[0214] In the post-processing optimization stage of four-channel endoscopic image enhancement, pseudo-color synthesis, dynamic range expansion, and pathological area enhancement can further improve the clinical diagnostic value of endoscopic images through multi-dimensional synergistic processing.

[0215] Post-processing optimization introduces a histopathology-guided adaptive enhancement system to map physical image attributes to pathological meaning:

[0216] 1. Spectrum-specific pseudo-color synthesis:

[0217] Based on the physical characteristics of multispectral channels, the original single-channel images of red (R), green (G), and blue (B) light are mapped to the red, green, and blue channels of the RGB color space, respectively, to generate preliminary pseudo-color images. However, due to the differences in the reflection and absorption characteristics of biological tissues to different wavelengths of light (e.g., red light has strong penetrability, while blue light is sensitive to surface reflection), direct mapping may lead to color aliasing or insufficient contrast in pathological areas. Therefore, further fine-tuning is performed through HSV (hue, saturation, brightness) color space conversion.

[0218] In the HSV color space, H, S, and V represent Hue, Saturate, and Value, respectively. Hue represents the hue information that maintains the original RGB mapping, ensuring that the colors of tissue structures and lesion areas conform to medical priors. Saturate represents making the colors of inflamed or tumor areas more vibrant. Value is the value channel, and image processing is generally performed on the V channel. It can adaptively equalize and eliminate uneven lighting, thus improving the contrast of lesions and maintaining color consistency across devices.

[0219] In this embodiment, a pseudo-color mapping scheme based on a tissue-lesion association model is used to enhance lesion expression through medical priors:

[0220]

[0221] Among them, C pseudo The resulting image after pseudo-color synthesis is an RGB color image, where (x,y) represents the pixel coordinates in the image. This pseudo-color image enhances the visualization of the lesion area based on multispectral data features.

[0222] f HSV2RGB This is a conversion function from HSV color space to RGB color space. It takes three components of the HSV color space (hue H, saturation S, and lightness V) as input and converts them into pixel values ​​in the RGB color space. This conversion is the final step in the pseudo-color synthesis process, combining the processed HSV channels into the final color image. The conversion formula is a standard HSV to RGB conversion; given the H (hue), S (saturation), and V (lightness) values, it calculates the corresponding R, G, and B values.

[0223] H semantic Hue assignment function based on organizational semantics:

[0224]

[0225] In the formula, For the balance parameter (0.7), H base Based on hue, H pathology For pathology-guided hue mapping, T(x,y) represents the tissue classification result.

[0226] Saturation enhancement function S enhance Adaptive strategy based on lesion significance:

[0227]

[0228] in, L is the magnification factor (1.5). abnormal This is a probability map of abnormal regions.

[0229] Brightness balance function V balanced Achieved through piecewise S-curves:

[0230]

[0231] Among them, parameters The angle is dynamically adjusted according to the local tissue depth and the distance of the endoscope lens to ensure that tissue details at different depths are clearly visible.

[0232] V raw (x, y) represents the original image's luminance channel values, i.e., the raw data of the V channel in the HSV color space. These are the luminance values ​​before processing with the piecewise S-curve (Sigmoid function). The luminance balance function Vbalanced performs a non-linear mapping of the original luminance using the Sigmoid function. The parameter k controls the steepness of the curve (contrast adjustment), and θ(x, y) controls the center point position (brightness adjustment). These two parameters adaptively adjust based on tissue depth and lens distance, ensuring that tissues at different depths receive appropriate luminance display.

[0233] 2. Adaptive enhancement of pathological areas:

[0234] Design a multi-scale, multi-feature fusion lesion region enhancement system. First, generate a lesion probability map through multi-model ensemble:

[0235]

[0236] Where M vessel M texture M color and M depth These are detection models for vascular abnormalities, texture abnormalities, color abnormalities, and depth abnormalities, respectively. A lesion-specific enhancement filter is constructed based on this probability map.

[0237]

[0238] In the formula The enhancement factor is 1.2. 2 G σ(x,y) For the adaptive scaling Laplacian operator, σ(x,y) is dynamically adjusted according to the lesion size:

[0239]

[0240] Parameter σ base With δ set to 1.0 and 0.3 respectively, this mechanism ensures optimal enhancement for lesions of different sizes.

[0241] Building upon this, to address the common dynamic range compression problem in endoscopic images, an S-curve (Sigmoid function) is used for nonlinear mapping, through parameters... k Controlling contrast enhancement intensity θ The overall brightness is adjusted, prioritizing the expansion of intermediate grayscale details in mucosal texture and tissue structure, while suppressing highlight overexposure and shadow noise. Its adaptive strategy, combined with histogram analysis, dynamically optimizes parameters, stabilizing the noise signal-to-noise ratio above 35dB while expanding details, avoiding artifacts caused by traditional linear stretching. Further enhancement of pathological features is achieved through computer vision algorithms: a Hessian matrix is ​​constructed by analyzing the second derivative of the image using a Frangi filter, and the similarity of tubular structures is quantified based on eigenvalues, highlighting deep microvascular networks. For specific lesions, adaptive segmentation is achieved by fusing HSV / CIE Lab color gamut thresholding, local texture features, and morphological post-processing, generating smooth lesion masks and quantifying parameters to provide structured data support for diagnostic reports.

[0242] 3. Diagnostic-Specific Image Optimization: As a preferred approach, a configurable target-oriented enhancement workflow has been developed based on different clinical diagnostic objectives, including:

[0243] Early cancer detection modality: Enhanced detection of irregular microvascular distribution and mucosal gland atypicality.

[0244]

[0245] Inflammation assessment model: Enhanced vasodilation and mucosal hyperemia features

[0246]

[0247] Bleeding monitoring mode: Optimized display of deep tissue bleeding points

[0248]

[0249] Among them, Icancer(x,y) is an image optimized for early cancer detection, with enhanced cancer-related features;

[0250] Venhance(x,y) is a vascular enhancement image generated by a multi-scale vascular enhancement filter optimized by the Hessian matrix mentioned earlier, highlighting the irregular distribution of microvessels, which is an important feature of early carcinogenesis.

[0251] Gpattern(x,y) represents the morphological analysis results of mucosal glands, which is used to detect glandular atypia, an important characteristic of early cancer.

[0252] Ibase(x,y) is the base image, usually a reference image for a preprocessed white light image, providing a basic anatomical reference.

[0253] Vdensity(x,y) is a capillary density index map, which quantitatively represents the abundance of blood vessels per unit area;

[0254] fcolor(x,y) is the result of the tissue color mapping function, which is used to detect abnormal tissue color, especially mucosal congestion characteristics, and is an important basis for inflammation judgment;

[0255] Ired(x,y) is the image enhanced by the red light channel, which has high sensitivity to deep tissue bleeding points;

[0256] f threshold This is a thresholding function used to further enhance the salience of bleeding points;

[0257] f threshold (I ratio (x,y)) is a ratio image processed by a threshold function, where I ratio A ratio plot of different spectral channels is used to better identify bleeding areas;

[0258] The parameter vectors [λ1,λ2,λ3], [μ1,μ2,μ3], and [ν1,ν2,ν3] were clinically validated to be [0.4,0.3,0.3], [0.5,0.2,0.3], and [0.6,0.3,0.1], respectively.

[0259] The parameter vectors [λ1,λ2,λ3], [μ1,μ2,μ3], and [ν1,ν2,ν3] were optimized to [0.4,0.3,0.3], [0.5,0.2,0.3], and [0.6,0.3,0.1], respectively, through clinical validation.

[0260] The three elements work together to form a complete process optimization of "color perception - detail enhancement - feature quantification". Clinical verification shows that this solution improves the visual contrast of early lesions by 30%-50% and increases the detection efficiency of doctors by 40%. At the same time, the dynamic range control and noise suppression balance ensures diagnostic specificity. Combined with the automated parameter measurement function, it provides an integrated solution for endoscopic intelligent diagnostic systems from image enhancement to quantitative analysis.

[0261] Application Examples

[0262] As attached Figure 2 As shown, it is divided into the following key stages:

[0263] Step S201: Multi-channel image acquisition

[0264] First, raw four-channel data images (white, red, green, and blue light) are acquired from the endoscopic system. This stage is the starting point of the entire processing flow and the foundation for ensuring the effectiveness of subsequent multispectral information. Each channel image reflects the reflection and absorption characteristics of tissue at different wavelengths, containing rich pathological information. By simultaneously acquiring these spectral data, high-dimensional input can be provided for subsequent spectral decoupling and information fusion.

[0265] Step S202: Image preprocessing stage

[0266] This stage mainly includes three sub-steps, with the goal of improving image quality and unifying spatiotemporal alignment:

[0267] S2021: Image Registration and Alignment. During the acquisition process, four-channel images may experience slight spatial displacement due to differences in the shooting time of the devices. Therefore, image alignment (registration) is required. Image registration technology can effectively eliminate spatial misalignment caused by time differences, providing a consistent spatial reference for inter-channel fusion and comparison.

[0268] S2022: Noise suppression and motion artifact correction are specifically designed for the red and blue light channels, as these two bands are more susceptible to motion interference and system noise. This step uses noise filtering algorithms to suppress random interference and blurred areas in the image, thereby removing motion artifacts caused by endoscope jitter and making the image clearer and more stable.

[0269] S2023: Reflectance and Illuminance Correction In order to counteract the effects of non-uniform lighting during the shooting process, it is necessary to separate the reflection component and illumination component of each channel image and perform normalization processing to ensure that the subsequent enhancement steps process the real tissue reflection information, rather than the false brightness changes caused by illumination differences.

[0270] Step S203: Specific enhancement based on spectral characteristics

[0271] After preprocessing, the four-channel images are enhanced in a targeted manner by utilizing the physical properties of the interaction between different wavelengths of light and biological tissue. For example, red light can enhance deep blood vessels, green light can improve the clarity of capillary networks, and blue light can highlight gland openings and mucosal surface microstructures. This step essentially applies a specific enhancement strategy to the channel images through a spectral characteristic decomposition model, maximizing the diagnostic value of different channels in the region of interest.

[0272] Step S204: Image Fusion

[0273] To integrate the structural and functional information contained in the four channels, the input processed four-channel image can be deeply fused to generate an enhanced image containing red, green, blue and white light information, retaining the advantageous features of each channel and achieving collaborative alignment in spatial structure.

[0274] Step S205: Visualization and lesion prominence of the fused image

[0275] After fusion, the resulting image undergoes pseudo-color synthesis to simulate the visual effect under real visible light, improving readability for doctors. Simultaneously, dynamic range expansion is performed on the fused image to fully express both dark details and bright tissues, enhancing the salience of lesion areas. Furthermore, image segmentation or saliency detection techniques can be combined to highlight potential lesion areas, aiding in early clinical detection and precise localization.

[0276] The above technical solutions are merely exemplary embodiments of the present invention. For those skilled in the art, based on the application methods and principles disclosed in the present invention, it is easy to make various types of improvements or modifications, and not limited to the methods described in the specific embodiments of the present invention. Therefore, the methods described above are merely preferred and not restrictive.

Claims

1. A multi-channel endoscopic image feature enhancement method, characterized by, The method comprises the following steps: S1, multi-channel endoscope image acquisition; S2, image preprocessing, including: multi-channel image alignment and registration; noise suppression and motion artifact correction; reflection and illumination correction of each channel image; S3, based on the tissue spectrum characteristic-pathology corresponding relationship model, corresponding biological tissue feature enhancement analysis is performed on each spectral channel; S4, input each channel image for fusion; S5, post-processing optimization: adaptive enhancement based on histopathology guidance, realizing the mapping from physical image attributes to pathological significance; In the step S2, the step of multi-channel image alignment and registration specifically comprises: (1) scale space extreme value detection: First, based on the improved SIFT algorithm, the input image is processed... Perform multi-scale Gaussian blurring to generate a Gaussian pyramid. The specific formula is as follows: where x, y represent the horizontal and vertical coordinates of the image respectively, parameter a is used to control the anisotropy degree of the Gaussian kernel, and g is a tissue reflectance correction factor, which is used to adapt to the nonlinear illumination response of the endoscopic image; is a scale parameter, which determines the blur degree; where the two-dimensional Gaussian function is: Secondly, a difference Gaussian pyramid is constructed: Wherein, k is a scale multiplication factor; β(σ) is a scale adaptive compensation factor, which is used to improve the feature extraction sensitivity of the microvascular structure and glandular texture in the endoscope image: ， λ and τ are set to 0.15 and 1.8 respectively; Then, extreme value detection is performed: at each pixel point of the DoG pyramid, compared with 8 neighborhood points of the same scale and 9*2 neighborhood points of the adjacent scale, if the point is the maximum or minimum value, it is considered as a candidate key point; (2) key point accurate positioning, to remove low-contrast or edge response points and improve stability: Derive the DoG function to calculate the accurate offset of the extreme point, if the offset is less than the threshold, the low-contrast point is removed; wherein, for the low-texture area of the endoscope, the contrast threshold judgment standard is modified to an adaptive local threshold: Wherein, VAR(P) is the local variance of the neighborhood of the key point P, and η is a weight coefficient, which is used to retain weak feature points in the low-contrast area; For image edges, response suppression is performed, and edge points are removed by using Hessian matrix: Calculate the curvature ratio: Wherein, r takes the value of 10; if the condition is not met, it is determined as an edge point and removed; (3) direction assignment is performed to assign a main direction to the key point, realizing rotation invariance: Divide 0°~360° into 36 intervals, and calculate the weighted histogram of the gradient amplitude, and the weight is the Gaussian window weight; Take the histogram peak value as the main direction, and if there is a secondary peak, generate a multi-direction key point; (4) descriptor generation is performed to generate a feature vector with illumination, scale and rotation invariance: Divide the key point neighborhood into 4*4 sub-regions, calculate the gradient histogram of 8 directions in each sub-region, and get a 4*4*8=128-dimensional vector; The descriptor is normalized, and the large value sensitive to illumination is truncated, and each component is limited to not more than 0.2, and finally it is re-normalized.

2. The method of claim 1, wherein, The multi-channel endoscope image in the step S1 comprises: original white light, red light, green light and blue light four-channel data images.

3. The method of claim 1, wherein, In the step S2, the step of noise suppression and motion artifact correction specifically comprises: Noise suppression is performed on the red light and blue light channel images, wherein the red light channel reflects the deep tissue, and the non-local mean method is used for denoising to retain the vascular texture; the blue light channel reflects the surface mucosa, and the wavelet threshold denoising is used to protect the high-frequency details.

4. The method of claim 1, wherein, In the step S2, the step of reflection and illumination correction of each channel image specifically comprises: Based on the Retinex theory, the reflection component and the illumination component of each channel are separated and normalized to eliminate the influence of non-uniform illumination.

5. The method of claim 1, wherein, The specific method of the step S3 comprises: S30, based on large-scale clinical annotation data, a mapping relationship between reflectivity of different wavelengths and histopathological states is obtained, and a tissue spectral characteristic-pathology correspondence model is established; S31, white light channel enhancement analysis: for the white light reflectivity curve, a feature vector is extracted through principal component analysis, and a deviation from a standard healthy tissue template is quantified using a tissue chroma mapping function, which is used for efficient detection of abnormal color and shade areas and is sensitive to weak color differences caused by early cancer; The mapping function is: where w i is the RGB channel weight, T healthy is the healthy tissue reference template; I white (x,y) i denotes the value of the i-th color component of the white light channel image at position (x,y), where i = 1,2,3 correspond to the three color channels of RGB respectively: When i=1, it represents the value of the red (R) channel; When i=2, it represents the value of the green (G) channel; When i=3, it represents the value of the blue (B) channel; S32, for the red light channel, a blood vessel feature enhancement is performed for deep tissue analysis: Constructing a vascular connectivity graph G v : Wherein, V is a blood vessel node set, E is a connection edge set, and W is a weight function based on depth d and direction θ; Analysis G v topological properties of the vasculature for detecting abnormal vascular growth patterns; A depth-aware blood vessel recognition function is designed to distinguish blood vessel networks at different levels of depth: where H frangi is a modified Frangi filter, D tissue is a tissue depth estimation map, and K is a depth decay coefficient. S33, for the green light channel, a capillary density feature enhancement is performed: A multi-scale blood vessel enhancement filter optimized by a Hessian matrix is introduced: where λ1and λ2are the eigenvalues of the Hessian matrix, RBis the circularity measure of the blood vessel, S is the structure measure, and are set to 0.5 and 15, respectively; Based on a multi-scale blood vessel enhancement filter, a capillary density index V is calculated density for assessing the microcirculation status with high sensitivity to inflammation and early tumors: where A is the area of the region of interest, τ v is the blood vessel threshold; S34, for the blue light channel, a surface mucosa texture feature enhancement is performed: A fusion feature vector of LBP local binary pattern and GLCM gray level co-occurrence matrix is designed: where F texture (x, y) is the texture feature vector at position (x, y); LBP r，p (x, y) represents the local binary pattern feature with radius r and number of points p; Contrast GLCM is the contrast feature of the gray level co-occurrence matrix; Correlation GLCM is the correlation feature of the gray level co-occurrence matrix; Energy GLCM is the energy feature of the gray level co-occurrence matrix; Homogeneity GLCM is the homogeneity feature of the gray level co-occurrence matrix; LBP r,p A rotation invariant LBP operator with radius 2 and 8 sampling points is used. Based on the fusion feature vector, a mucosa gland morphology analysis model is constructed for identifying early intestinal metaplasia and disorder of gland structure: wherein, is the i-th component of the texture feature vector; parameter a i By comparing the normal gland and the sample of the atypical gland, the learning results are [0.25, 0.3, 0.15, 0.1, 0.2], respectively.

6. The method of claim 1, wherein, The step S5 specifically comprises: S51, pseudo-color mapping based on the tissue-lesion correlation model is performed to enhance lesion expression using medical priori: where C pseudo is the RGB color image after pseudo-color synthesis, where (x, y) represents the pixel coordinates in the image; f HSV2RGB is the conversion function from the HSV color space to the RGB color space; where H semantic is a hue assignment function based on the tissue semantics: wherein H is a balancing parameter, base H is a base hue, pathology T(x, y) is a pathology-oriented hue map, and T(x, y) is a tissue classification result. S enhance For the saturation enhancement function, an adaptive strategy of lesion conspicuity is adopted: In the formula, S base is a base saturation, is an amplification factor, L abnormal is an anomaly region probability map; V balanced For the brightness balance function, it is achieved by piecewise sigmoid curves: where V raw (x, y) is the value of the original image in the lightness channel, that is, the original data in the V channel of the HSV color space; the lightness balance function V balanced The original lightness is nonlinearly mapped by a Sigmoid function, and the parameters The steepness of the control curve, that is, the contrast adjustment, and θ(x, y) controls the center point position, that is, the brightness adjustment. These two parameters are adaptively adjusted according to the depth of the tissue and the distance of the endoscope lens, so that the tissues at different depths can be properly displayed. S52, multi-scale and multi-feature fusion pathological region adaptive enhancement: A lesion probability map is generated through multi-model integration: wherein M vessel , M texture , M color , and M depth are a blood vessel abnormality, a texture abnormality, a color abnormality, and a depth abnormality detection model, respectively. A lesion-specific enhancement filter is constructed based on the lesion probability map: where I input is the original input image, is the enhancement coefficient, 2 G σ(x,y) is the adaptive scale Laplacian operator, σ(x, y) is dynamically adjusted according to the lesion size: σ base is the base scale, δ is the adjustment factor, Size lesion is the local lesion size estimate, and the parameters σ base and δ are set to 1.0 and 0.3, respectively, The multi-scale and multi-feature fusion pathological region adaptive enhancement mechanism is used to ensure the best enhancement effect for different sizes of lesions.

7. The method according to any one of claims 1 to 6, characterized in that, Further comprising: According to different clinical diagnosis targets, a step of configurable target-oriented feature enhancement is performed, Specifically comprising: Early cancer detection mode: enhancing irregular distribution of microvessels and mucosa gland atypia ； Inflammation evaluation mode: strengthening features of blood vessel expansion and mucosa hyperemia ； Bleeding monitoring mode: optimizing display of deep tissue bleeding points ； where I cancer (x,y) is the image optimized for early cancer detection, with cancer-related features enhanced; V enhance (x,y) is the blood vessel enhanced image, generated by a multi-scale blood vessel enhancement filter optimized by Hessian matrix, highlighting the irregular distribution of microvessels, which is an important feature of early canceration; G pattern (x, y) is the result of mucosal gland morphology analysis, used to detect gland atypia, which is an important feature of early canceration; I base (x, y) is the base image, which is the reference image of the pre-processed white light image, providing a basic anatomical reference; V density (x, y) is the capillary density index map, which quantitatively represents the richness of blood vessels per unit area; f color (x, y) is the tissue chrominance mapping function result, which is used to detect tissue color abnormalities, including mucosal congestion features, and is an important basis for judging inflammation; I red (x, y) is the image of red channel enhancement, which has high sensitivity to deep tissue bleeding points; f threshold is a thresholding function for further enhancing the saliency of the bleeding points; f threshold (I ratio (x,y)) is a ratio image processed by a threshold function, where I ratio is a ratio image of different spectral channels for better identification of hemorrhagic areas; The parameter vectors [λ1, λ2, λ3], [μ1, μ2, μ3] and [ν1, ν2, ν3] are respectively taken as [0.4, 0.3, 0.3], [0.5, 0.2, 0.3] and [0.6, 0.3, 0.1] through clinical verification.

8. A multi-channel endoscope image feature enhancement apparatus applying the multi-channel endoscope image feature enhancement method according to any one of claims 1 to 7, characterized by, Comprise: A multi-channel endoscope image acquisition module; An image preprocessing module for aligning and registering the multi-channel images, suppressing noise and correcting motion artifacts, and correcting the reflection and illumination of each channel image; A multi-spectral endoscope image enhancement module for performing corresponding biological tissue feature enhancement analysis on each spectral channel based on the tissue spectral characteristic-pathology correspondence model; A multi-channel fusion module for fusing input channel images; A multi-channel endoscope image acquisition module; A post-processing optimization module for realizing mapping from physical image attributes to pathologically meaningful based on histopathology-guided adaptive enhancement.

9. A storage medium, characterized by A non-transitory computer-readable medium having stored thereon executable program instructions that, when executed by a computer, implement the multi-channel endoscopic image feature enhancement method of any of claims 1-7.

Images