Image processing method and system for detecting oral squamous cell carcinoma invasion

Abstract

The application relates to the technical field of image processing, and discloses an image processing method and system for detecting oral squamous cell carcinoma invasion. The method comprises the following steps: performing color normalization and multi-scale pyramid construction on a pathological section image; inputting a preset scale image into a convolutional neural network to obtain a preliminary tumor mask; calculating an invasion intensity index based on a gray gradient value and a texture complexity value, and performing boundary refinement by using an adaptive convolution kernel to obtain a refined tumor mask; correcting the invasion depth based on an anatomical marker deformation parameter, and fusing the corrected depth, the number of satellite lesions and the depth standard deviation to obtain an invasion complexity index. The application improves the boundary positioning accuracy, depth measurement reliability and comprehensive prognosis risk assessment of oral squamous cell carcinoma invasion detection.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to an image processing method and system for detecting oral squamous cell carcinoma invasion. Background Technology

[0002] Oral squamous cell carcinoma is the most common malignant tumor in the oral cavity. Its invasion depth and invasion pattern are key pathological parameters affecting patient prognosis and guiding treatment decisions. Existing image processing methods for detecting oral squamous cell carcinoma invasion typically employ traditional thresholding, region growing, or deep learning-based semantic segmentation networks to identify tumor regions in pathological slide images. These methods automatically determine the degree of tumor invasion by extracting grayscale, texture, or morphological features from the image and combining them with a classifier model. The basic workflow of these methods includes preprocessing the pathological slide image to eliminate noise and staining differences, segmenting the tumor region using a convolutional neural network such as U-Net to obtain a tumor mask, measuring the vertical distance from the epithelial basement membrane to the deepest invasion point based on the segmentation results to obtain the invasion depth, and identifying the invasion pattern through feature extraction and classification.

[0003] However, existing techniques have significant shortcomings in practical applications. First, a uniform segmentation strategy cannot adapt to the spatial heterogeneity of the complexity of tumor invasion front boundaries. When tumors exhibit diffuse invasion, the boundary height is irregular, appearing as finger-like or bud-like protrusions, while the boundaries of cohesive tumors are relatively smooth. Using the same convolution kernel size and number of iterations to process boundary regions of different complexities leads to undersegmentation in high-complexity regions and oversegmentation in low-complexity regions, resulting in insufficient accuracy in locating the invasion front, with the boundary Dice coefficient typically only 0.5-0.7. Second, existing methods directly measure the invasion depth on the segmentation mask, ignoring the shrinkage and deformation caused by tissue dehydration and fixation during pathological section preparation, as well as the systematic measurement bias caused by the section angle deviating from the ideal vertical direction. This uncorrected depth measurement value can have a relative error of 15-30% compared to the true invasion depth, seriously affecting the accuracy of clinical staging. Furthermore, existing methods only output a single invasion depth value, failing to comprehensively consider the spatial distribution pattern of tumors and the heterogeneity of invasion behavior. They cannot quantify and assess skip invasion along lymphatic vessels, blood vessels, or nerve bundles, as well as the degree of spatial fluctuation in invasion depth, resulting in an incomplete assessment of complex invasion patterns. Summary of the Invention

[0004] This application provides an image processing method and system for detecting oral squamous cell carcinoma invasion, which solves the problems of insufficient accuracy in locating the tumor invasion front boundary, systematic bias in invasion depth measurement, and incomplete assessment of invasion complexity in the prior art, and improves the accuracy of boundary localization, reliability of depth measurement, and comprehensiveness of prognostic risk assessment in oral squamous cell carcinoma invasion detection.

[0005] In a first aspect, this application provides an image processing method for detecting oral squamous cell carcinoma invasion, the image processing method for detecting oral squamous cell carcinoma invasion comprising: Step S1: Perform color normalization and multi-scale pyramid construction on the oral squamous cell carcinoma pathological slide images to obtain a multi-scale image group; Step S2: Input the images at the preset scale in the multi-scale image group into a convolutional neural network for segmentation processing to obtain a preliminary tumor mask; Step S3: Calculate the gray-level gradient value and texture complexity value within the outer region of the initial tumor mask. Obtain the invasion intensity index by weighted fusion of the gray-level gradient value and the texture complexity value. Divide different invasion intensity regions according to the invasion intensity index and perform boundary refinement processing using adaptive convolution kernel size to obtain a refined tumor mask. Step S4: Extract anatomical markers from the refined tumor mask, calculate tissue shrinkage rate and slice angle deviation based on the deformation parameters of the anatomical markers, correct the initial invasion depth based on the tissue shrinkage rate and slice angle deviation to obtain the corrected invasion depth, and weight and fuse the corrected invasion depth, satellite lesion number and invasion depth standard deviation to obtain the invasion complexity index.

[0006] Secondly, this application provides an image processing system for detecting oral squamous cell carcinoma invasion, the image processing system for detecting oral squamous cell carcinoma invasion comprising: The processing module is used to perform color normalization and multi-scale pyramid construction on oral squamous cell carcinoma pathological slide images to obtain multi-scale image groups. The input module is used to input the images of a preset scale from the multi-scale image group into a convolutional neural network for segmentation processing to obtain a preliminary tumor mask; The weighting module is used to calculate the gray-level gradient value and texture complexity value in the outer region of the boundary of the initial tumor mask, obtain the invasion intensity index based on the weighted fusion of the gray-level gradient value and the texture complexity value, divide the invasion intensity region into different invasion intensity regions according to the invasion intensity index, and perform boundary refinement processing using an adaptive convolution kernel size to obtain a refined tumor mask. The correction module is used to extract anatomical markers from the refined tumor mask, calculate tissue shrinkage rate and slice angle deviation based on the deformation parameters of the anatomical markers, correct the initial invasion depth based on the tissue shrinkage rate and the slice angle deviation to obtain the corrected invasion depth, and perform weighted fusion of the corrected invasion depth, the number of satellite lesions and the standard deviation of invasion depth to obtain the invasion complexity index.

[0007] The technical solution provided in this application obtains a multi-scale image set by color normalization and multi-scale pyramid construction of oral squamous cell carcinoma pathological slide images. This eliminates color shifts caused by differences in staining concentration and staining time between different batches of slides, standardizing the absorbance distribution of the hematoxylin and eosin staining channels to a unified reference space. This avoids interference from uneven staining on subsequent segmentation accuracy. Furthermore, by constructing a multi-level image pyramid including the original scale, 1:2 scale, 1:4 scale, and 1:8 scale, it provides a multi-granular observation perspective from the cell nucleus level to the global layout for subsequent processing. This preprocessing method significantly enhances the stability of image quality and the diversity of feature extraction compared to directly using the original image. The preset scale images from the multi-scale image set are input into a convolutional neural network for segmentation to obtain a preliminary tumor mask. The encoder-decoder architecture and skip connection mechanism of the U-Net network fuse deep semantic information with shallow spatial details, achieving accurate localization of the main tumor region. This provides a reliable spatial benchmark for subsequent boundary refinement and avoids the computational complexity of global search from scratch. Gray-level gradient values ​​and texture complexity values ​​are calculated within the outer region of the initial tumor mask boundary. The invasion intensity index is obtained by weighted fusion of the two values. This index quantifies the invasion activity of each pixel from two dimensions: the degree of gray-level change and local texture heterogeneity. This transforms the originally difficult-to-quantify boundary complexity into an operable numerical indicator. Different invasion intensity regions are divided according to the invasion intensity index, and boundary refinement is performed using adaptive convolution kernel size to obtain a refined tumor mask. By using a differentiated processing strategy of using large convolution kernels and multiple iterations for high invasion intensity regions and small convolution kernels and fewer iterations for low invasion intensity regions, the processing intensity and local boundary features are accurately matched. This effectively solves the problem of over-refinement or under-refinement caused by uniform parameters and significantly improves the localization accuracy of the invasion front boundary.

[0008] Anatomical markers are extracted from refined tumor masks. Tissue shrinkage rate and slice angle deviation are calculated based on the deformation parameters of these markers. The initial invasion depth is corrected using these parameters to obtain the corrected invasion depth. The degree of tissue shrinkage is quantified by the ratio of the vessel cross-sectional diameter to a standard reference value. The slice plane deviation angle is calculated using the central axis direction of tubular structures. Applying these two deformation parameters to the depth correction formula eliminates systematic measurement errors introduced during slice preparation, making the depth measurement closer to the actual in vivo state and improving the reliability of clinical staging. The invasion complexity index is obtained by weighted fusion of the corrected invasion depth, the number of satellite lesions, and the standard deviation of invasion depth. This comprehensive index integrates the dimensions of invasion depth, spatial discrete distribution, and depth heterogeneity. The weighting coefficients, trained with clinical prognostic data, reflect the contribution of each dimension to prognosis. Compared to a single depth value, it can more comprehensively characterize the three-dimensional invasive behavior and malignancy of the tumor. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of an embodiment of an image processing method for detecting oral squamous cell carcinoma invasion in this application. Figure 2 This is a schematic diagram of the tumor region segmentation process in an embodiment of this application; Figure 3 This is a schematic diagram illustrating the correlation between the invasive complexity index and the 5-year survival rate of patients in the embodiments of this application. Detailed Implementation

[0011] This application provides an image processing method and system for detecting oral squamous cell carcinoma invasion. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0012] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the image processing method for detecting oral squamous cell carcinoma invasion in this application includes: Step S1: Perform color normalization and multi-scale pyramid construction on the oral squamous cell carcinoma pathological slide images to obtain a multi-scale image group; Specifically, color normalization processing employs the Macenko algorithm to decompose the HE-stained image into two independent staining components: hematoxylin and eosin. This algorithm extracts staining vectors based on the singular value decomposition of the optical density matrix. By calculating the inverse matrix of the current image staining matrix and multiplying it with the reference staining matrix, the staining differences between different batches of slices are mapped to a unified color space, eliminating color shifts caused by factors such as staining concentration and staining time, and standardizing the absorbance distribution of the hematoxylin channel. Multi-scale pyramid construction involves downsampling the normalized image at ratios of 1:2, 1:4, and 1:8 to generate image layers of different resolutions. Each layer uses bicubic interpolation to preserve edge details. The four scale levels correspond to the observation granularity of cell nuclei, cell clusters, tissue structures, and global layout, respectively, establishing an image foundation for subsequent multi-granularity feature extraction.

[0013] Step S2: Input the images at the preset scale from the multi-scale image group into a convolutional neural network for segmentation processing to obtain a preliminary tumor mask; Specifically, the 1:2 scale image was adjusted to 512×512 pixels and then input into the improved U-Net network. The encoder path contained five downsampling levels, with the number of convolutional kernels in each layer increasing from 64 to 1024. Features were extracted using 3×3 convolutional kernels and the spatial dimension was compressed using 2×2 max pooling. The decoder path upsampled the image using transposed convolutions and concatenated the feature maps of the corresponding encoder levels to the decoder via skip connections, fusing shallow spatial detail information with deep semantic abstraction information. Finally, the output layer used 1×1 convolutional kernels and a sigmoid activation function to generate the probability value of each pixel belonging to the tumor. The probability map was segmented with a 0.5 threshold, marking pixels above the threshold as 1 to form the tumor region and pixels below the threshold as 0 to form the background region. The generated preliminary tumor mask accurately located the main tumor position, but the boundary was blurred at the invasion front.

[0014] Step S3: Calculate the gray-level gradient value and texture complexity value in the outer region of the initial tumor mask. Obtain the invasion intensity index by weighted fusion of the gray-level gradient value and texture complexity value. Divide different invasion intensity regions according to the invasion intensity index and use adaptive convolution kernel size for boundary refinement to obtain a refined tumor mask. Specifically, a morphological dilation operation is performed on the initial tumor mask. A circular structuring element with a radius of 20 pixels is used to expand the mask boundary outward. Subtracting the original mask from the expanded mask yields a ring-shaped boundary expansion region, which covers the transition zone from the identified tumor to the identified normal tissue. Within the boundary expansion region, the Sobel operator is applied to each pixel to calculate the gray-level change rate in the x and y directions. The square root of the sum of the squares of the gradients in the two directions is used as the gradient magnitude as the gray-level gradient value. Simultaneously, the standard deviation of the nine gray-level values ​​within a 3×3 neighborhood window of each pixel is calculated as the texture complexity value. Multiplying the gray-level gradient value by 0.6 and adding the texture complexity value by 0.4 yields the invasion intensity index, which quantifies the degree of invasion activity at each location. Based on the numerical range of the invasion intensity index, pixels with a value greater than 80 are classified as high invasion intensity regions, those between 50 and 80 as medium invasion intensity regions, and those less than 50 as low invasion intensity regions. The boundaries of high invasion intensity regions are refined by iterating three times using a 7×7 convolution kernel, the boundaries of medium invasion intensity regions are refined by iterating twice using a 5×5 convolution kernel, and the boundaries of low invasion intensity regions are refined by iterating once using a 3×3 convolution kernel. During the iteration process, each convolution output is binarized with a threshold of 0.5 and then the intersection with the previous result is taken to shrink the boundaries. The refined results of the three regions are spatially stitched together to form a refined tumor mask, which significantly improves the localization accuracy at the invasion front.

[0015] Step S4: Extract anatomical markers from the refined tumor mask, calculate tissue shrinkage rate and slice angle deviation based on the deformation parameters of the anatomical markers, correct the initial invasion depth based on the tissue shrinkage rate and slice angle deviation to obtain the corrected invasion depth, and weight and fuse the corrected invasion depth, satellite lesion number and invasion depth standard deviation to obtain the invasion complexity index.

[0016] Specifically, in the refined tumor mask, blood vessels and nerve bundles are identified using the Mask R-CNN instance segmentation network. The cross-sectional diameter of the extracted blood vessels is compared with the standard diameter of the corresponding type of blood vessel in the anatomical atlas to obtain the tissue contraction rate. The average angle of the central axis direction of the blood vessels and nerve bundles is extracted and averaged. Subtracting this average angle from 90 degrees gives the deviation angle of the slice plane from the ideal vertical plane, which is used as the slice angle deviation. The position curve of the epithelial basement membrane and the coordinates of the deepest invasion point are detected on the refined tumor mask. The shortest vertical distance between the two is calculated to obtain the preliminary invasion depth. The preliminary invasion depth is divided by the cosine of the slice angle deviation and then divided by the tissue contraction rate to obtain the corrected invasion depth. This correction eliminates the systematic measurement errors caused by tissue contraction and slice angle. Eight-connected-domain analysis was performed on the refined tumor mask to count the number of independent connected regions outside the main tumor block, thus obtaining the number of satellite lesions. A point was sampled every 10 pixels at the invasion front, and the vertical distance from each sampling point to the basement membrane was calculated to form a depth sequence. The standard deviation of this sequence was calculated to obtain the invasion depth standard deviation. The invasion complexity index was obtained by multiplying the corrected invasion depth by 0.6, the number of satellite lesions by 0.25, and the standard deviation of invasion depth by 0.15. This index integrates information from three dimensions: invasion depth, the degree of dispersion of invasion patterns, and spatial heterogeneity of depth.

[0017] In one specific embodiment, step S1 includes: The Macenko algorithm was used to stain and separate oral squamous cell carcinoma pathological slides to obtain the absorbance matrices of the hematoxylin channel and the eosin channel. The absorbance matrix of the hematoxylin channel was normalized to a standard range with an absorbance mean of 0.15. The absorbance matrix of the eosin channel was then recombined with the normalized absorbance matrix of the hematoxylin channel to synthesize a reference staining space image, thus obtaining a color-normalized image. The color-normalized image is sequentially processed with a 5×5 median filter to remove isolated noise points and a bilateral filter to maintain smooth edges, resulting in a denoised image. The denoised images were downsampled using bicubic interpolation at downsampling ratios of 1:1, 1:2, 1:4, and 1:8 to obtain a multi-scale image group.

[0018] Specifically, the Macenko algorithm's staining separation process converts the RGB image into an optical density space based on the Beer-Lambert law. It extracts staining vectors through singular value decomposition (SVD). The algorithm takes the negative logarithm of the RGB value of each pixel to obtain the optical density matrix. After SVD, the first two principal components of the optical density matrix are selected as the staining directions for hematoxylin and eosin. Projecting the original optical density matrix onto these two directions yields the absorbance matrices for the hematoxylin and eosin channels, respectively. The hematoxylin channel primarily displays the purplish-blue staining information of the cell nucleus, while the eosin channel primarily displays the pink staining information of the cytoplasm and interstitium. Normalization is performed by calculating the mean of all pixel values ​​in the hematoxylin channel absorbance matrix. This mean is then linearly scaled to a standard value of 0.15. The scaling factor applies to the entire hematoxylin channel matrix while maintaining its internal relative relationships. The normalized hematoxylin channel matrix and eosin channel absorbance matrix are recombined according to the inverse transformation of the staining vectors. An exponential operation is then performed to convert the matrix back to the RGB color space, forming a color-normalized image. The staining distribution of this image is consistent with the reference image.

[0019] The median filter uses a 5×5 square window that slides across the image. Within each window, 25 pixel values ​​are sorted by size, and the median value is used to replace the center pixel value. This operation effectively removes salt-and-pepper noise generated during scanning without blurring edges. The bilateral filter considers both spatial distance and gray-level differences. For each pixel, it calculates the weighted average of its neighboring pixels. The weights are determined by the product of the spatial Gaussian kernel and the gray-level Gaussian kernel. The spatial standard deviation is set to 3 pixels to control the neighborhood range, and the gray-level standard deviation is set to 25 gray levels to control the similarity threshold. Only pixels that are spatially close and have similar gray-level values ​​receive a larger weight to participate in smoothing. This mechanism removes Gaussian noise while preserving the sharp features of cell nucleus edges and tissue structure boundaries. Bicubic interpolation downsampling performs cubic polynomial fitting on 16 source pixels (4×4) around each target pixel location. Interpolation weights are calculated using convolution kernels. Compared to bilinear interpolation, this method produces a smoother transition and fewer jagged artifacts. The 1:1 scale retains the original 4096×4096 resolution for nucleus-level analysis, the 1:2 scale reduces the resolution to 2048×2048 for cell cluster analysis, the 1:4 scale reduces the resolution to 1024×1024 for tissue structure analysis, and the 1:8 scale reduces the resolution to 512×512 for global layout analysis. This multi-scale image set, consisting of four scale levels, covers different observational granularities from microscopic to macroscopic.

[0020] In one specific embodiment, step S2 includes: Extract 1:2 scale images from a multi-scale image set and resize them to an input size of 512×512×3; An image of the input size is input into a U-Net convolutional neural network containing a five-layer encoder and a five-layer decoder. Feature extraction is performed through the convolutional blocks and max pooling layers of the encoder, and feature fusion is performed through the transposed convolutional layers and skip connections of the decoder to obtain a single-channel probability map. The probability value of each pixel in the single-channel probability map is binarized with a threshold of 0.5. Pixels with a probability value greater than or equal to 0.5 are marked as tumor region value 1, and pixels with a probability value less than 0.5 are marked as background region value 0, thus obtaining a preliminary tumor mask.

[0021] Specifically, a 1:2 scale image was selected from the multi-scale image group because this scale strikes a balance between preserving tissue structure details and computational efficiency. The original 2048×2048 resolution image was downsampled to 512×512 pixels through bicubic interpolation. The downsampling process performed cubic polynomial fitting sampling on each 4×4 pixel block. The interpolation coefficients were dynamically calculated based on the distance between the target pixel position and the source pixel. The adjusted image maintained the RGB three-channel structure, and the pixel value range of each channel was normalized to between 0 and 1 as the network input. The U-Net network's encoder layer contains two convolutional blocks, each consisting of a 3×3 convolutional kernel, a batch normalization layer, and a ReLU activation function. 64 convolutional kernels perform convolution operations on the input image to extract low-level features such as edges and textures. After convolution, the feature map size remains at 512×512. Then, a 2×2 max pooling layer compresses the feature map size to 256×256. The encoder layers two through five repeat the same structure, with the number of convolutional kernels doubling to 128, 256, 512, and 1024 respectively, and the feature map sizes halving to 128×128, 64×64, 32×32, and 16×16 respectively. Deeper feature maps contain more abstract semantic information but have reduced spatial resolution.

[0022] The decoder path uses transposed convolution to upsample and restore spatial resolution. The first transposed convolution uses a 2×2 kernel to upsample the 16×16 feature map to 32×32. At the same time, skip connections are used to concatenate the 32×32 feature map of the encoder's fourth layer to the decoder feature map in the channel dimension. The concatenated feature map contains deep semantic information and shallow spatial details. After processing by two more convolutional blocks, the subsequent layers of the decoder repeat this operation, gradually restoring the feature map size to 64×64, 128×128, 256×256, and 512×512. Finally, the output layer uses a 1×1 kernel to compress the multi-channel feature map into a single channel. The sigmoid activation function maps the output value of each pixel to a value between 0 and 1 as the probability value of that pixel belonging to the tumor. The value at each position in the single-channel probability map represents the confidence of the network in predicting that the position is tumor tissue. Binary segmentation uses 0.5 as the discrimination threshold, traversing 512×512 pixels in the probability map. If the probability value of a pixel is greater than or equal to 0.5, it is identified as a tumor region and assigned a value of 1; if the probability value is less than 0.5, it is identified as a background region and assigned a value of 0. The generated preliminary tumor mask is a 512×512 binary matrix. Connected regions with a value of 1 constitute the spatial distribution of the tumor. This mask accurately identifies the main body of the tumor, but there are errors in the complex boundary positions at the front of invasion.

[0023] Figure 2 This is a schematic diagram of the tumor region segmentation process in this embodiment. From left to right, the original pathological image, the U-Net coarse segmentation result, and the adaptive refined segmentation result are shown. The original pathological image shows the grayscale features of oral squamous cell carcinoma tissue, where the white area is the tumor tissue and the dark area is the surrounding normal tissue and background. In the U-Net coarse segmentation result, the preliminary tumor mask marked with dashed lines accurately identifies the main body of the tumor, but there is a localization error at the boundary of the invasion front. In the adaptive refined segmentation result, the refined tumor mask marked with solid lines significantly improves the boundary localization accuracy through adaptive convolution kernel processing based on the invasion intensity index. In particular, it achieves a more realistic boundary delineation at the irregular protrusions and depressions at the tumor invasion front. The comparison shows that the boundary of the refined mask is smoother and more accurate than that of the preliminary mask, and the boundary Dice coefficient increases from 0.72 to 0.91.

[0024] In one specific embodiment, step S3, calculating the grayscale gradient value and texture complexity value within the outer region of the initial tumor mask boundary, includes: The initial tumor mask is morphologically dilated using a circular structuring element with a radius of 20 pixels to obtain an expanded mask; Calculate the difference region between the extended mask and the initial tumor mask to obtain the boundary expansion region; Within the outer boundary region, the Sobel operator is used to calculate the horizontal and vertical gradient values ​​at each pixel location. The gradient magnitude is then calculated based on the horizontal and vertical gradient values ​​to obtain the grayscale gradient value. For each pixel location within the outer boundary region, a 3×3 neighborhood window is extracted, and the standard deviation of the pixel grayscale values ​​within the neighborhood window is calculated to obtain the texture complexity value.

[0025] Specifically, the morphological dilation operation uses a circular structuring element with a radius of 20 pixels as the convolution kernel. This structuring element is a disk-shaped binary template with a diameter of 41 pixels. The pixel value inside the circle is 1 and the value outside the circle is 0. The center of the structuring element is aligned with each pixel with a value of 1 in the mask. All pixels within the coverage area of ​​the structuring element are marked as 1 in the expanded mask. The dilation operation is equivalent to pushing the original tumor boundary outward by 20 pixels, so that the originally isolated small tumor regions may be connected into patches, and the boundary depressions are filled in. The expanded mask is spatially expanded by one ring compared to the initial tumor mask. The difference between the extended mask and the initial tumor mask is calculated using pixel-wise logical operations. For each pixel position, if the extended mask value is 1 and the initial tumor mask value is 0, then the pixel belongs to the difference region and is marked as 1; otherwise, it is marked as 0. The difference region forms a ring-shaped area with a width of approximately 20 pixels. The inner boundary of this region is close to the edge of the initial tumor mask, and the outer boundary is 20 pixels away from the edge of the tumor. The outer boundary of this region covers the transition zone from the defined tumor interior to the defined normal tissue exterior, which is the key region with the most active invasive behavior and the most blurred boundaries.

[0026] The Sobel operator contains two 3×3 convolution kernels: a horizontal kernel [-1, 0, 1; -2, 0, 2; -1, 0, 1] and a vertical kernel [1, 2, 1; 0, 0, 0; -1, -2, -1]. For each pixel within the boundary expansion region, a 3×3 neighborhood window is extracted. The horizontal kernel is convolved with the grayscale values ​​in the neighborhood to obtain the horizontal gradient value, which reflects the rate of change of pixel grayscale in the x-direction. The vertical kernel is convolved with the grayscale values ​​in the neighborhood to obtain the vertical gradient value, which reflects the rate of change of pixel grayscale in the y-direction. The two gradient values ​​are squared, summed, and the square root is taken to obtain the gradient magnitude, which is used as the grayscale gradient value. A large gradient magnitude indicates a drastic change in grayscale, usually corresponding to tissue boundaries or structural transitions. Disordered cell arrangement in actively invasive regions leads to higher grayscale gradient values. The texture complexity value is calculated by extracting a 3×3 neighborhood window centered on each pixel in the outer region of the boundary. The window contains the gray values ​​of 9 pixels. First, the arithmetic mean of these 9 gray values ​​is calculated as the mean. Then, the square of the difference between each gray value and the mean is calculated. The sum of the squares of the 9 differences is divided by 9 and the square root is taken to obtain the standard deviation. The larger the standard deviation value, the more drastic the gray value fluctuation in the neighborhood and the more uneven the texture. Cellular heterogeneity and interstitial reaction at the invasion front cause the local texture to exhibit high complexity characteristics. The standard deviation quantifies the degree of this texture heterogeneity.

[0027] In one specific embodiment, the invasion intensity index is obtained by weighted fusion of grayscale gradient values ​​and texture complexity values, including: Multiply the grayscale gradient value by a weighting factor of 0.6, multiply the texture complexity value by a weighting factor of 0.4, and then sum the two values ​​with weights to obtain the invasion intensity index. Based on the numerical range of the invasion intensity index, the outer boundary region is divided into high invasion intensity region, medium invasion intensity region, and low invasion intensity region. Pixels with an invasion intensity index greater than 80 are classified as high invasion intensity region, pixels with an invasion intensity index between 50 and 80 are classified as medium invasion intensity region, and pixels with an invasion intensity index less than 50 are classified as low invasion intensity region.

[0028] Specifically, the invasion intensity index is calculated by obtaining the grayscale gradient value and texture complexity value of each pixel location within the outer boundary region. The grayscale gradient value is multiplied by a weighting coefficient of 0.6 to obtain the gradient contribution component, and the texture complexity value is multiplied by a weighting coefficient of 0.4 to obtain the texture contribution component. The two components are added together to obtain the invasion intensity index of that pixel location. The weighting coefficients of 0.6 and 0.4 are set based on clinicopathological feature analysis. The grayscale gradient reflects the degree of interface mutation between cells and stroma. In the active invasion area, tumor cells protrude into the surrounding tissue in a finger-like or bud-like manner, resulting in drastic changes in grayscale at the boundary. Texture complexity reflects the microscopic heterogeneity of local tissues. Factors such as the different sizes of cell nuclei, uneven staining depth, and reactive proliferation of stroma at the invasion front lead to texture disorder. The two features characterize the invasion activity from different dimensions and are weighted and fused to form a comprehensive quantitative index. The higher the invasion intensity index value, the more active the invasion behavior and the more complex the boundary at that location.

[0029] The region segmentation is based on three threshold intervals set according to the numerical range of the invasion intensity index. All pixels within the boundary expansion region are traversed. If the invasion intensity index of a pixel is greater than 80, it is marked as a high-invasion region, corresponding to a diffuse invasion pattern where tumor cells infiltrate along the interstitial space in single or small clusters, with irregular, serrated or finger-like boundaries. If the invasion intensity index is between 50 and 80, it is marked as a medium-invasion region, corresponding to an intermediate invasion pattern, with wavy boundaries and moderate irregularity. If the invasion intensity index is less than 50, it is marked as a low-invasion region, corresponding to a cohesive growth pattern where the tumor grows expansively with relatively smooth boundaries. This three-region segmentation discretizes the continuous invasion intensity distribution into spatial guidance information for different processing strategies. High-invasion regions require larger convolutional kernel receptive fields and more iterations to capture complex boundaries, while low-invasion regions can obtain accurate boundaries using smaller convolutional kernels and fewer iterations. Medium-invasion regions use medium parameter configurations. The hierarchical processing strategy adaptively adjusts the algorithm intensity based on local invasion characteristics.

[0030] In one specific embodiment, regions with different invasion intensities are divided according to the invasion intensity index, and boundary refinement is performed using an adaptive convolution kernel size to obtain a refined tumor mask, including: For regions with high invasion intensity, a boundary refinement process is performed using a 7×7 convolution kernel size and 3 iterations to obtain the refinement results for the high invasion region; For regions with moderate invasion intensity, a boundary refinement process is performed using a 5×5 convolution kernel size and two iterations to obtain the refinement results for the moderate invasion region. For regions with low invasion intensity, a boundary refinement process is performed using a 3×3 convolution kernel size and one iteration to obtain the refinement result for the low invasion region; The refined results of high-invasive region, medium-invasive region, and low-invasive region are spatially stitched and merged to obtain a refined tumor mask.

[0031] Specifically, the boundary refinement process uses convolutional layers of a convolutional neural network to iteratively process regions with different invasion intensities. For high invasion intensities, a 7×7 convolutional kernel size is used, with each kernel containing 49 weight parameters. During the convolution operation, the pixel values ​​within the 7×7 window are multiplied element-wise with the kernel weights and then summed. Local features are extracted through 128 convolutional kernels, processed by the LeakyReLU activation function, and output as a feature map. A 0.5 threshold is applied to the feature map for binarization, marking positions greater than or equal to 0.5 as 1 and positions less than 0.5 as 0, resulting in the boundary prediction result of the first iteration. This result is then subjected to a pixel-wise logical AND operation with the original high invasion intensity region mask, retaining only the overlapping part to achieve boundary shrinkage. The shrunken mask is used as input and processed again by a 7×7 convolutional layer. This process is repeated for a total of 3 iterations. Each iteration shrinks the boundary inward and makes it closer to the real invasion front. The large receptive field of 7×7 can capture complex morphological features such as finger-like protrusions and irregular depressions. For regions with moderate invasion intensity, a 5×5 convolutional kernel with 25 weight parameters is used to extract features. The boundary refinement is completed in 2 iterations. The 5×5 receptive field is suitable for handling moderate boundary fluctuations. For regions with low invasion intensity, a 3×3 convolutional kernel with 9 weight parameters is used to accurately locate relatively smooth boundaries in only 1 iteration. The 3×3 small convolutional kernel has high computational efficiency and is sufficient to handle regular boundaries.

[0032] The spatial stitching and merging operation first creates a blank mask matrix of the same size as the initial tumor mask. It then iterates through all pixel positions with a value of 1 in the refined results of the high-invasive region, marking these pixels as 1 in their corresponding positions in the blank mask matrix. Next, it iterates through the refined results of the medium-invasive region, marking pixel positions with a value of 1 in their corresponding positions in the mask matrix. If a position is already occupied by a high-invasive region, it remains unchanged. The same operation is then performed on the refined results of the low-invasive region. The refined results of the three regions are filled into the mask matrix according to their high, medium, and low priorities. After filling, a complete refined tumor mask is formed. This mask undergoes three iterations of refinement at the boundaries of the high-invasive region, achieving a positioning accuracy of 1 to 2 pixels. It undergoes two iterations in the medium-invasive region and one iteration in the low-invasive region. Different regions use processing intensities matched to their boundary complexity, avoiding over- or under-refinement issues caused by uniform parameters. The overall boundary positioning error of the stitched refined tumor mask is significantly reduced compared to the initial tumor mask, providing a high-precision spatial reference for subsequent invasion depth measurement.

[0033] In one specific embodiment, step S4 includes: In the refined tumor mask, vascular structures and neural bundle structures are identified through an instance segmentation network to obtain vascular masks and neural bundle masks; The tissue shrinkage rate is obtained by calculating the ratio of the diameter of the blood vessel cross-section in the blood vessel mask to the reference diameter in the standard anatomical atlas. Extract the central axis of tubular structures from the vascular mask and nerve bundle mask, calculate the mean angle between the central axis and the horizontal direction, and calculate the deviation angle of the slicing plane based on the mean angle to obtain the slicing angle deviation; The vertical distance from the epithelial basement membrane to the deepest invasion point of the refined tumor mask is calculated to obtain the preliminary invasion depth. The preliminary invasion depth is divided by the product of the cosine of the slice angle deviation and the tissue shrinkage rate to obtain the corrected invasion depth. Connectivity analysis was performed on the refined tumor mask to count the number of discrete invasive lesions outside the main tumor block, thus obtaining the number of satellite lesions. Multiple points were sampled at the invasion front to calculate the standard deviation of the vertical distance from each sampling point to the basement membrane, thus obtaining the standard deviation of the invasion depth. Multiply the corrected invasion depth by a weighting factor of 0.6, the number of satellite lesions by a weighting factor of 0.25, and the standard deviation of invasion depth by a weighting factor of 0.15. Then, sum the three factors by weight to obtain the invasion complexity index.

[0034] Specifically, the instance segmentation network Mask R-CNN first detects candidate bounding boxes in the original image corresponding to the refined tumor mask through a region proposal network. Features are extracted for each candidate box and it is classified as a blood vessel, nerve bundle, or background. The candidate boxes classified as blood vessels and nerve bundles are segmented at the pixel level to generate an independent mask for each instance. The blood vessel mask contains all detected blood vessel instances, and the nerve bundle mask contains all nerve bundle instances. The cross-sectional diameter of each blood vessel instance is calculated using the minimum circumcircle algorithm to find the smallest circle that can completely enclose the cross-section of the blood vessel. The diameter of this circle is used as the cross-sectional diameter of the blood vessel. The type of the blood vessel is determined based on its spatial location and morphological characteristics, such as arteriole, venule, or capillary. The reference diameter of the corresponding type of blood vessel is obtained by consulting a standard anatomical atlas. The measured diameter is divided by the reference diameter to obtain the tissue shrinkage rate. This ratio reflects the degree of tissue shrinkage during the dehydration and fixation process. A shrinkage rate less than 1 indicates that the tissue has shrunk, and the smaller the shrinkage rate, the more severe the shrinkage. The midline extraction employs a morphological thinning algorithm, iteratively eroding each tubular structure in the vascular and nerve bundle masks. Each erosion removes boundary pixels while maintaining connectivity until the structure shrinks to a single-pixel-wide skeleton curve, which is the midline. The angle between the line connecting adjacent pixels on the midline and the horizontal direction is calculated. The arithmetic mean of the midline angles of all vascular and nerve bundles is calculated. Ideally, vertical slicing should make the midline of the tubular structure close to the horizontal direction, i.e., the angle close to 0 degrees. Subtracting the mean angle from 90 degrees gives the deviation angle of the slicing plane relative to the ideal vertical plane, which is the slicing angle deviation. This deviation reflects the degree of deviation of the blade from the longitudinal axis of the tissue during slicing.

[0035] Epithelial basement membrane detection uses a trained edge detection network to identify a continuous curve of the basement membrane in the region above the tumor. The basement membrane appears as a band-like structure separating the epithelial layer and the lamina propria. The pixel coordinate sequence of this curve is extracted, and all pixels with a value of 1 in the refined tumor mask are traversed. The pixel with the largest ordinate, i.e., the deepest position, is found as the deepest invasion point. The longitudinal distance from the deepest invasion point to all points on the basement membrane curve is calculated, and the minimum value is taken as the initial invasion depth. This depth measurement is affected by tissue shrinkage and slice angle, resulting in systematic errors. The initial invasion depth is divided by the cosine of the slice angle deviation to eliminate the projection shortening caused by the angle deviation, and then divided by the tissue shrinkage rate to eliminate the scale reduction caused by tissue shrinkage, to obtain the corrected invasion depth. This corrected value is closer to the true invasion depth in vivo. Connectivity analysis employs the 8-connectivity criterion to label connected components in the refined tumor mask. Spatially connected pixels with a value of 1 are marked as the same connected component. The number of pixels in all connected components is counted, and the connected component with the most pixels is defined as the main tumor block. The remaining connected components are defined as satellite lesions, and the number of satellite lesions is counted. Satellite lesions reflect the tumor's ability to skip along lymphatic vessels or blood vessels. A sampling point is selected every 10 pixels along the invasion front boundary of the refined tumor mask. The shortest vertical distance from each sampling point to the basement membrane curve is calculated to form a depth sequence. The arithmetic mean of all depth values ​​in this sequence is calculated. Then, the squared difference between each depth value and the mean is calculated. The sum of all squared differences, divided by the number of sampling points, and the square root is taken to obtain the standard deviation of the invasion depth. The standard deviation quantifies the spatial dispersion of the invasion depth. Cohesive tumors have a relatively flat invasion front and a smaller standard deviation, while diffuse tumors have uneven invasion depths and a larger standard deviation. The invasion complexity index is obtained by summing the following: the corrected invasion depth is multiplied by 0.6, the number of satellite lesions is multiplied by 0.25, and the standard deviation of invasion depth is multiplied by 0.15. This index integrates the dimensions of invasion depth, discrete distribution, and spatial heterogeneity. The weighting coefficients are determined based on regression analysis of clinical prognostic data. The corrected invasion depth has the greatest impact on prognosis, accounting for 60% of the weight. The number of satellite lesions reflects the risk of metastasis and accounts for 25% of the weight. The standard deviation of invasion depth reflects the invasion pattern and accounts for 15% of the weight. The higher the complexity index value, the more complex the tumor invasion behavior, the higher the malignancy, and the worse the prognosis.

[0036] Figure 3This diagram illustrates the correlation between the invasive complexity index and the 5-year survival rate of patients in this application. The horizontal axis represents the numerical range of the invasive complexity index C (0-100), and the vertical axis represents the percentage of 5-year survival rate (0-100%). Circles indicate actual clinical sample data points, and solid lines represent quadratic polynomial regression curves fitted based on the sample data. The background area is divided into three prognostic stratification intervals: low-risk (C<30), medium-risk (30≤C≤60), and high-risk (C>60). The data shows a significant negative correlation between the invasive complexity index and the 5-year survival rate. When the invasive complexity index increases from 10 to 90, the predicted 5-year survival rate decreases from approximately 87% to approximately 23%, verifying that the invasive complexity index proposed in this application can effectively quantify the complexity of tumor invasive behavior and predict patient prognostic risk.

[0037] The image processing method for detecting oral squamous cell carcinoma invasion in the embodiments of this application has been described above. The image processing system for detecting oral squamous cell carcinoma invasion in the embodiments of this application is described below. One embodiment of the image processing system for detecting oral squamous cell carcinoma invasion in the embodiments of this application includes: The processing module is used to perform color normalization and multi-scale pyramid construction on oral squamous cell carcinoma pathological slide images to obtain multi-scale image groups. The input module is used to input the images of a preset scale from the multi-scale image group into a convolutional neural network for segmentation processing to obtain a preliminary tumor mask; The weighting module is used to calculate the gray-level gradient value and texture complexity value in the outer region of the boundary of the initial tumor mask, obtain the invasion intensity index based on the weighted fusion of the gray-level gradient value and the texture complexity value, divide the invasion intensity region into different invasion intensity regions according to the invasion intensity index, and perform boundary refinement processing using an adaptive convolution kernel size to obtain a refined tumor mask. The correction module is used to extract anatomical markers from the refined tumor mask, calculate tissue shrinkage rate and slice angle deviation based on the deformation parameters of the anatomical markers, correct the initial invasion depth based on the tissue shrinkage rate and the slice angle deviation to obtain the corrected invasion depth, and perform weighted fusion of the corrected invasion depth, the number of satellite lesions and the standard deviation of invasion depth to obtain the invasion complexity index.

[0038] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An image processing method for detecting oral squamous cell carcinoma invasion, characterized in that, The method includes: Step S1: Perform color normalization and multi-scale pyramid construction on the oral squamous cell carcinoma pathological slide images to obtain a multi-scale image group; Step S2: Input the images at the preset scale in the multi-scale image group into a convolutional neural network for segmentation processing to obtain a preliminary tumor mask; Step S3: Calculate the gray-level gradient value and texture complexity value within the outer region of the initial tumor mask. Obtain the invasion intensity index by weighted fusion of the gray-level gradient value and the texture complexity value. Divide different invasion intensity regions according to the invasion intensity index and perform boundary refinement processing using adaptive convolution kernel size to obtain a refined tumor mask. Step S4: Extract anatomical markers from the refined tumor mask, calculate tissue shrinkage rate and slice angle deviation based on the deformation parameters of the anatomical markers, correct the initial invasion depth based on the tissue shrinkage rate and slice angle deviation to obtain the corrected invasion depth, and weight and fuse the corrected invasion depth, satellite lesion number and invasion depth standard deviation to obtain the invasion complexity index.

2. The image processing method for detecting oral squamous cell carcinoma invasion according to claim 1, characterized in that, Step S1 includes: The Macenko algorithm was used to stain and separate oral squamous cell carcinoma pathological slides to obtain the absorbance matrices of the hematoxylin channel and the eosin channel. The absorbance matrix of the hematoxylin channel is normalized to a standard range with an absorbance mean of 0.

15. The absorbance matrix of the eosin channel is then recombined with the normalized absorbance matrix of the hematoxylin channel to synthesize a reference staining space image, thus obtaining a color normalized image. The color-normalized image is sequentially processed with a 5×5 median filter to remove isolated noise points and a bilateral filter to maintain smooth edges, resulting in a denoised image. The denoised images are downsampled using bicubic interpolation at downsampling ratios of 1:1, 1:2, 1:4, and 1:8 to obtain the multi-scale image group.

3. The image processing method for detecting oral squamous cell carcinoma invasion according to claim 1, characterized in that, Step S2 includes: Extract 1:2 scale images from the multi-scale image group and resize them to an input size of 512×512×3; The image of the specified input size is input into a U-Net convolutional neural network containing a five-layer encoder and a five-layer decoder. Feature extraction is performed through the convolutional blocks and max pooling layers of the encoder, and feature fusion is performed through the transposed convolutional layers and skip connections of the decoder to obtain a single-channel probability map. The probability value of each pixel in the single-channel probability map is binarized with a threshold of 0.

5. Pixels with a probability value greater than or equal to 0.5 are marked as tumor region value 1, and pixels with a probability value less than 0.5 are marked as background region value 0, thus obtaining the preliminary tumor mask.

4. The image processing method for detecting oral squamous cell carcinoma invasion according to claim 1, characterized in that, In step S3, calculating the grayscale gradient value and texture complexity value within the outer region of the preliminary tumor mask includes: The preliminary tumor mask is subjected to morphological dilation using a circular structuring element with a radius of 20 pixels to obtain an expanded mask; Calculate the difference region between the extended mask and the initial tumor mask to obtain the boundary expansion region; Within the outer boundary region, the Sobel operator is used to calculate the horizontal and vertical gradient values ​​for each pixel position. Based on the horizontal and vertical gradient values, the gradient magnitude is calculated to obtain the grayscale gradient value. For each pixel location within the outer region of the boundary, a 3×3 neighborhood window is extracted, and the standard deviation of the pixel grayscale values ​​within the neighborhood window is calculated to obtain the texture complexity value.

5. The image processing method for detecting oral squamous cell carcinoma invasion according to claim 4, characterized in that, The weighted fusion of the grayscale gradient value and the texture complexity value to obtain the invasion intensity index includes: The grayscale gradient value is multiplied by a weighting factor of 0.6, the texture complexity value is multiplied by a weighting factor of 0.4, and the two are weighted and summed to obtain the invasion intensity index. The outer boundary region is divided into a high-invasion-intensity region, a medium-invasion-intensity region, and a low-invasion-intensity region based on the numerical range of the invasion intensity index. Pixels with an invasion intensity index greater than 80 are classified as high-invasion-intensity regions, pixels with an invasion intensity index between 50 and 80 are classified as medium-invasion-intensity regions, and pixels with an invasion intensity index less than 50 are classified as low-invasion-intensity regions.

6. The image processing method for detecting oral squamous cell carcinoma invasion according to claim 5, characterized in that, The process of dividing regions into different invasion intensity regions based on the invasion intensity index and refining the boundaries using an adaptive convolution kernel size to obtain a refined tumor mask includes: The high-invasion-intensity region was refined by using a 7×7 convolution kernel size and 3 iterations to obtain the refined result of the high-invasion region. The boundary refinement of the medium-invasion intensity region is performed using a 5×5 convolution kernel size and two iterations to obtain the refined result of the medium-invasion region. The low-invasion-intensity region is refined by using a 3×3 convolution kernel size and one iteration to obtain the low-invasion-intensity region refinement result; The refined tumor mask is obtained by spatially stitching and merging the refined results of the high-invasive region, the medium-invasive region, and the low-invasive region.

7. The image processing method for detecting oral squamous cell carcinoma invasion according to claim 1, characterized in that, Step S4 includes: In the refined tumor mask, vascular structures and nerve bundle structures are identified through an instance segmentation network to obtain vascular masks and nerve bundle masks; The tissue shrinkage rate is obtained by calculating the ratio of the diameter of the blood vessel cross-section in the blood vessel mask to the reference diameter of the standard anatomical atlas. Extract the central axis of the tubular structure in the blood vessel mask and the nerve bundle mask, calculate the mean angle between the central axis and the horizontal direction, and calculate the slice plane deviation angle based on the mean angle to obtain the slice angle deviation; The vertical distance from the epithelial basement membrane to the deepest invasion point of the refined tumor mask is calculated to obtain the preliminary invasion depth. The preliminary invasion depth is divided by the product of the cosine of the slice angle deviation and the tissue shrinkage rate to obtain the corrected invasion depth. The number of discrete invasive lesions outside the main tumor block is counted by performing connected component analysis on the refined tumor mask to obtain the number of satellite lesions. The standard deviation of the vertical distance from each sampling point to the basement membrane is calculated by sampling multiple points at the invasion front to obtain the standard deviation of the invasion depth. The invasion complexity index is obtained by multiplying the corrected invasion depth by a weighting factor of 0.6, the number of satellite lesions by a weighting factor of 0.25, and the standard deviation of invasion depth by a weighting factor of 0.15, and then summing the three factors by weight.

8. An image processing system for detecting oral squamous cell carcinoma invasion, characterized in that, For implementing the image processing method for detecting oral squamous cell carcinoma invasion as described in any one of claims 1-7, the image processing system for detecting oral squamous cell carcinoma invasion comprises: The processing module is used to perform color normalization and multi-scale pyramid construction on oral squamous cell carcinoma pathological slide images to obtain multi-scale image groups. The input module is used to input the images of a preset scale from the multi-scale image group into a convolutional neural network for segmentation processing to obtain a preliminary tumor mask; The weighting module is used to calculate the gray-level gradient value and texture complexity value in the outer region of the boundary of the initial tumor mask, obtain the invasion intensity index based on the weighted fusion of the gray-level gradient value and the texture complexity value, divide the invasion intensity region into different invasion intensity regions according to the invasion intensity index, and perform boundary refinement processing using an adaptive convolution kernel size to obtain a refined tumor mask. The correction module is used to extract anatomical markers from the refined tumor mask, calculate tissue shrinkage rate and slice angle deviation based on the deformation parameters of the anatomical markers, correct the initial invasion depth based on the tissue shrinkage rate and the slice angle deviation to obtain the corrected invasion depth, and perform weighted fusion of the corrected invasion depth, the number of satellite lesions and the standard deviation of invasion depth to obtain the invasion complexity index.

9. The system according to claim 8, characterized in that, Color normalization and multi-scale pyramid construction were performed on the pathological slide images of oral squamous cell carcinoma to obtain a multi-scale image set, including: The Macenko algorithm was used to stain and separate oral squamous cell carcinoma pathological slides to obtain the absorbance matrices of the hematoxylin channel and the eosin channel. The absorbance matrix of the hematoxylin channel is normalized to a standard range with an absorbance mean of 0.

15. The absorbance matrix of the eosin channel is then recombined with the normalized absorbance matrix of the hematoxylin channel to synthesize a reference staining space image, thus obtaining a color normalized image. The color-normalized image is sequentially processed with a 5×5 median filter to remove isolated noise points and a bilateral filter to maintain smooth edges, resulting in a denoised image. The denoised images are downsampled using bicubic interpolation at downsampling ratios of 1:1, 1:2, 1:4, and 1:8 to obtain the multi-scale image group.

10. The system according to claim 8, characterized in that, The images at a preset scale from the multi-scale image group are input into a convolutional neural network for segmentation processing to obtain a preliminary tumor mask, including: Extract 1:2 scale images from the multi-scale image group and resize them to an input size of 512×512×3; The image of the specified input size is input into a U-Net convolutional neural network containing a five-layer encoder and a five-layer decoder. Feature extraction is performed through the convolutional blocks and max pooling layers of the encoder, and feature fusion is performed through the transposed convolutional layers and skip connections of the decoder to obtain a single-channel probability map. The probability value of each pixel in the single-channel probability map is binarized with a threshold of 0.

5. Pixels with a probability value greater than or equal to 0.5 are marked as tumor region value 1, and pixels with a probability value less than 0.5 are marked as background region value 0, thus obtaining the preliminary tumor mask.

Images