Colorectal cancer pathological image segmentation system and method

Abstract

The invention discloses a colorectal cancer pathological image segmentation system and method. The system comprises an image input unit, a pathological image preprocessing module, a microenvironment feature sensing module, a dynamic feature adaptation module, a multi-constraint segmentation module and a result output module. The modules cooperate to realize targeted pathological image processing, microenvironment feature extraction, dynamic feature adaptation and multi-constraint segmentation. The scheme fully considers the pathological microenvironment feature coupling relationship, solves the problem of inaccurate segmentation in the prior art, realizes accurate segmentation of the tumor area, provides reliable data support for colorectal cancer pathological diagnosis, and is suitable for a digital pathological section analysis scene.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence image analysis technology, and in particular to a system and method for segmenting pathological images of colorectal cancer. Background Technology

[0002] In the pathological diagnosis of colorectal cancer, accurate identification of tumor regions is the core foundation for subsequent quantitative analysis and prognostic prediction. Existing pathological image segmentation techniques often focus only on single pathological features, failing to fully consider the interaction mechanisms between microenvironmental features such as glandular structural abnormalities, cellular atypia, and stroma density. The pathological tissue microenvironment is an organic whole; each feature does not function independently. Abnormalities in glandular structure directly affect the manifestation of cellular atypia, changes in stroma density restrict the distribution of immune cells, and the presence of necrotic areas alters the morphological characteristics of glands and cells. This coupling relationship between features results in traditional single-feature segmentation models being extremely unsuitable for complex pathological scenarios, leading to inaccurate tumor region boundary delineation and frequent instances of missed or over-segmentation. This deficiency results in a lack of reliable data support for subsequent key applications such as tumor burden calculation and immune cell infiltration analysis, directly affecting the accuracy of diagnostic results and potentially leading to deviations in clinical treatment planning.

[0003] Therefore, there is an urgent need for a technical solution that can fully consider the coupling relationship of pathological microenvironment characteristics and achieve precise segmentation of tumor regions in order to solve the core defects of existing technologies. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a colorectal cancer pathological image segmentation system. This system includes an image input unit for acquiring digital pathological images of colorectal cancer. Its key feature is that it further includes a pathological image preprocessing module, a microenvironment feature perception module, a dynamic feature adaptation module, a multi-constraint segmentation module, and a result output module. The pathological image preprocessing module is connected to the image input unit and performs image denoising, grayscale normalization, and region enhancement processing on the acquired digital pathological images. The microenvironment feature perception module is connected to the pathological image preprocessing module and extracts microenvironment features related to colorectal cancer from the digital pathological images. These microenvironment features include glandular structural abnormality, cellular atypia, and stromal dysplasia. The system includes: quantitative indicators of density, immune cell distribution density, and the proportion of necrotic areas; a dynamic feature adaptation module connected to both the pathological image preprocessing module and the microenvironment feature perception module; an adaptation model constructed based on the extracted microenvironment feature quantitative indicators; a feature extraction strategy dynamically adjusted by modifying the convolution kernel parameters, stride, and activation function type of the feature extraction layer; and adaptation weights for each feature extraction channel determined. A multi-constraint segmentation module connected to both the dynamic feature adaptation module and the microenvironment feature perception module uses the adapted feature extraction strategy and adaptation weights for each feature extraction channel, combined with constraint rules, to segment the tumor region of the pathological image. A result output module connected to the multi-constraint segmentation module outputs the tumor region segmentation results.

[0005] Preferably, the targeted processing of the pathological image preprocessing module includes image denoising, grayscale normalization, and region enhancement. The image denoising adopts an adaptive median filtering algorithm, the grayscale normalization adopts a linear stretching transformation, and the region enhancement adopts a multi-scale contrast-limited adaptive histogram equalization algorithm.

[0006] More preferably, the microenvironment features extracted by the microenvironment feature perception module include glandular structural abnormality, cell atypia, interstitial density, immune cell distribution density, and necrotic area ratio. The microenvironment feature perception module extracts quantitative indicators of the above features through a convolutional neural network, which includes four feature extraction layers and two pooling layers.

[0007] More preferably, the dynamic feature adaptation module constructs an adaptation model based on the quantitative indicators extracted by the micro-environment feature perception module. The adaptation model achieves dynamic adaptation of the feature extraction strategy by adjusting the convolution kernel parameters, stride, and activation function type of the feature extraction layer. The activation function types include LeakyReLU, GELU, and Mish.

[0008] More preferably, the microenvironment feature sensing module quantifies the interaction between features through a microenvironment feature coupling strength calculation formula, which is:

[0009] ;

[0010] in, The coupling strength is a characteristic of the microenvironment. This is a quantification of glandular structural abnormalities. This is a quantification value for cell atypia. This represents the quantitative value of interstitial density. This is a quantitative value for the distribution density of immune cells. This is a quantitative value representing the percentage of necrotic area. , , These are the feature weight coefficients. It is the minimum constant. The attenuation coefficient is... For correction factor, As a baseline value for cellular atypia, It is a symbolic function.

[0011] More preferably, the dynamic feature adaptation module determines the weight of each feature extraction channel through a dynamic feature adaptation weight calculation formula, which is:

[0012] ;

[0013] in, Let be the adaptation weight for the j-th feature extraction channel. Extraction channels corresponding to glandular structural abnormality, cellular atypia, stroma density, immune cell distribution density, and the proportion of necrotic areas, respectively. As the baseline weight for the channel, For channel characteristic responsivity, These are the variance weighting coefficients. Let be the variance of the quantized value of the j-th feature. This is the coupling strength correction factor. The coupling strength of the microenvironment features as described in claim 5.

[0014] More preferably, the multi-constraint segmentation module optimizes the segmentation boundary through a segmentation boundary constraint calculation formula, which is:

[0015] ;

[0016] in, These are the boundary constraint values ​​for the partition. The adaptation weights for the j-th feature extraction channel as described in claim 6, The gradient value of the region corresponding to the j-th feature. The maximum gradient value in the region corresponding to the j-th feature. Let be the neighborhood consistency coefficient corresponding to the j-th feature. Let be the average neighborhood consistency coefficient corresponding to the j-th feature. For gradient constraint weights, For neighborhood constraint weights, To avoid a compensation constant with a denominator of zero.

[0017] Further preferably, the system also includes a feature calibration module, which is connected to both the dynamic feature adaptation module and the multi-constraint segmentation module. The feature calibration module calibrates the adapted features based on the segmentation boundary constraint values. The feature calibration is achieved by adjusting the dimension weights of the feature vectors, and the dimension weights are positively correlated with the segmentation boundary constraint values.

[0018] A method for segmenting pathological images of colorectal cancer includes the following steps:

[0019] S1. Obtain digital pathological images of colorectal cancer through the image input unit, and perform image denoising, grayscale normalization and region enhancement processing on the acquired digital pathological images using the pathological image preprocessing module.

[0020] S2. Quantitative indicators such as glandular structural abnormality, cell atypia, stroma density, immune cell distribution density, and necrotic area ratio in the processed pathological images are extracted through the microenvironment feature perception module.

[0021] S3. The feature extraction strategy is dynamically adjusted based on the extracted quantitative indicators through the dynamic feature adaptation module to determine the adaptation weight of each feature extraction channel.

[0022] S4. The tumor region of the pathological image is segmented by the multi-constraint segmentation module using the adapted feature extraction strategy combined with the segmentation boundary constraint value to obtain the tumor region segmentation result; S5. The tumor region segmentation result is output through the result output module.

[0023] Technical Effects: This invention creatively quantifies the coupling relationships of multiple features in the pathological microenvironment by setting up a collaborative structure of a microenvironment feature perception module, a dynamic feature adaptation module, and a multi-constraint segmentation module. It dynamically adjusts the feature extraction strategy and optimizes the segmentation boundary. This solution specifically addresses the segmentation inaccuracies caused by neglecting the influence of feature coupling in existing technologies, achieving precise segmentation of tumor regions in colorectal cancer pathological images. This provides reliable underlying data support for subsequent clinical applications such as tumor burden calculation and immune cell infiltration analysis, improving the objectivity and accuracy of pathological diagnosis. Attached Figure Description

[0024] Figure 1 This is a block diagram of the colorectal cancer pathological image segmentation system of this application;

[0025] Figure 2 This is a flowchart of the colorectal cancer pathological image segmentation method of this application. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0027] The core technical problem with existing technologies is that in the process of segmenting colorectal cancer pathological images, the coupling influence between microenvironmental features such as glandular structural abnormality and cellular atypia is ignored, resulting in poor adaptability of the segmentation model to complex pathological scenes and inaccurate delineation of tumor region boundaries.

[0028] Based on this, please refer to Figures 1-2 This embodiment provides a colorectal cancer pathological image segmentation system, including an image input unit for acquiring digital pathological images of colorectal cancer. The system further includes a pathological image preprocessing module, a microenvironment feature perception module, a dynamic feature adaptation module, a multi-constraint segmentation module, and a result output module. The pathological image preprocessing module is connected to the image input unit and performs image denoising, grayscale normalization, and region enhancement processing on the acquired digital pathological images. The microenvironment feature perception module is connected to the pathological image preprocessing module and extracts microenvironment features related to colorectal cancer from the digital pathological images. These microenvironment features include glandular structural abnormality, cellular atypia, stroma density, and immune cell distribution. The system includes: quantification indicators of density and the proportion of necrotic areas; a dynamic feature adaptation module connected to both the pathological image preprocessing module and the microenvironment feature perception module; an adaptation model constructed based on the extracted microenvironment feature quantification indicators; a feature extraction strategy dynamically adjusted by modifying the convolution kernel parameters, stride, and activation function type of the feature extraction layer; and adaptation weights for each feature extraction channel determined. A multi-constraint segmentation module connected to both the dynamic feature adaptation module and the microenvironment feature perception module uses the adapted feature extraction strategy, adaptation weights for each feature extraction channel, and constraint rules to segment the tumor region of the pathological image. A result output module connected to the multi-constraint segmentation module outputs the tumor region segmentation results.

[0029] After the multi-constraint segmentation module outputs the preliminary segmentation results, a post-processing submodule is added to perform small connected component removal and cavity filling operations. Small connected component removal: The area threshold method is used, and the area threshold of the connected component is set to 50 pixels. This threshold is based on the statistics of the smallest effective area of ​​the tumor in clinical pathological sections, and isolated noise areas with an area smaller than the threshold are removed. Cavity filling: Morphological closing operation is used, and the structuring element is set to a 3×3 rectangular kernel to fill cavities with an area of ​​less than 100 pixels inside the tumor area to ensure the integrity of the tumor area.

[0030] The specific implementation of this technical solution is as follows: The image input unit uses a high-definition scanner, model Hamamatsu NanoZoomer S60, with a scanning resolution of 0.25μm / pixel. It supports full-slide scanning and storage of digital pathological slides. The scanned images are in TIFF format and are transmitted to the processing terminal via a USB 3.0 interface. The processing terminal uses an Intel Core i9-12900K processor, 64GB of memory, and an NVIDIA RTX 3090 graphics card to ensure fast transmission and processing of image data. The pathological image preprocessing module is directly connected to the image input unit via a data bus. It receives the scanned digital pathological images and performs targeted processing. The processed image data is transmitted to the microenvironment feature perception module via memory sharing. The microenvironment feature perception module is based on a convolutional neural network architecture and extracts microenvironment features related to colorectal cancer. The extracted quantified indicators are transmitted to the dynamic feature adaptation module via a high-speed serial interface. The dynamic feature adaptation module receives the preprocessed image data and the microenvironment feature quantified indicators, constructs an adaptation model, and dynamically adjusts the feature extraction strategy. The adjusted strategy parameters are transmitted to the multi-constraint segmentation module via a control bus. The multi-constraint segmentation module combines the adapted feature extraction strategy and constraint rules to perform pixel-level segmentation of pathological images. The segmented results are transmitted to the results output module, which outputs them as image files and statistical data. Communication between modules adopts a standardized data transmission protocol to ensure the stability and real-time performance of data transmission. Collaboration between modules is achieved through a synchronous clock signal with a clock frequency set to 100MHz to ensure consistent operation timing across modules.

[0031] This solution constructs a complete technical chain from image acquisition to segmentation result output through the collaborative design of multiple modules. Each module performs its own function and works closely together, realizing full consideration and dynamic adaptation of the characteristics of the pathological microenvironment. It solves the core problem of inaccurate segmentation in existing technologies and provides systematic technical support for the accurate identification of tumor regions.

[0032] In existing technologies, pathological image preprocessing often uses a single fixed algorithm, which cannot simultaneously achieve the effects of denoising, normalization, and enhancement, resulting in residual noise or blurred features in the preprocessed images.

[0033] Based on this, the targeted processing of the pathological image preprocessing module includes image denoising, grayscale normalization, and region enhancement. The image denoising adopts an adaptive median filtering algorithm, the grayscale normalization adopts a linear stretching transformation, and the region enhancement adopts a multi-scale contrast-limited adaptive histogram equalization algorithm.

[0034] The specific implementation of this technical solution is as follows: The adaptive median filtering algorithm for image denoising is implemented based on the Python 3.9 programming language and the OpenCV 4.5.5 library. The algorithm first defines the initial size of the filtering window as 3×3. It determines whether the current pixel is a noise point by calculating the mean and variance of the pixels within the window. If the difference between the current pixel value and the mean exceeds 1.5 times the variance, it is determined to be a noise point. In this case, the filtering window is expanded to 5×7, and the median of the non-noise pixels within the window is used to replace the current pixel value. If the difference does not exceed a threshold, the original pixel value is retained, and the filtering window remains unchanged. The linear stretching transformation for grayscale normalization is performed using the formula... Implementation, in which These are the original image pixel values. The minimum grayscale value of the original image. The maximum grayscale value of the original image. The normalized pixel values ​​are mapped to a uniform grayscale value between 0 and 255, eliminating grayscale differences caused by different scanning conditions. The multi-scale contrast-constrained adaptive histogram equalization algorithm for region enhancement uses three scale windows: 16×16, 32×32, and 64×64. The algorithm first decomposes the image into sub-images of three scales, performing histogram equalization on each sub-image separately. During equalization, the upper limit of contrast gain is set to 4 to avoid over-enhancement leading to noise amplification. Then, the equalization results of the three scales are fused using a weighted summation method, with weight coefficients set to 0.3, 0.5, and 0.2, respectively. The fused image is then normalized to obtain the final enhanced image.

[0035] The preprocessing module is implemented using a Xilinx XC7K325T FPGA chip. The adaptive median filtering algorithm and linear stretching transformation are implemented in parallel within the FPGA via hardware circuitry, achieving a processing speed of up to 1GB / s. The multi-scale contrast-limited adaptive histogram equalization algorithm is implemented collaboratively by the FPGA and the processor. The FPGA handles image segmentation and parallel processing, while the processor handles weight fusion and normalization operations, ensuring real-time performance of the preprocessing process. This scheme achieves synergistic optimization of denoising, normalization, and enhancement through the combined application and parameter optimization of these three algorithms. The preprocessed image exhibits less residual noise, uniform grayscale distribution, and clear pathological features, laying a solid foundation for subsequent feature extraction and segmentation.

[0036] Existing technologies often focus on a single pathological feature in feature extraction, failing to fully capture the microenvironment information related to colorectal cancer, resulting in insufficient feature expression and affecting segmentation accuracy.

[0037] Based on this, the microenvironment features extracted by the microenvironment feature perception module include glandular structural abnormality, cell atypia, interstitial density, immune cell distribution density, and necrotic area ratio. The microenvironment feature perception module extracts quantitative indicators of the above features through a convolutional neural network, which contains 4 feature extraction layers and 2 pooling layers.

[0038] The specific implementation of this technical solution is as follows: The convolutional neural network of the micro-environment feature perception module is built based on the PyTorch 1.12 framework. The network structure includes an input layer, four feature extraction layers, two pooling layers, and an output layer. The input layer receives a preprocessed 256×256 pixel image, and the output has three channels. The first feature extraction layer has 64 convolutional kernels, a kernel size of 3×3, a stride of 1, and uses SAME padding with ReLU activation. This layer is used to extract basic texture features of the image. The first pooling layer uses max pooling with a 2×2 pooling window and a stride of 2, used to reduce the dimensionality of the feature map while retaining key features. The second feature extraction layer has 128 convolutional kernels, a kernel size of 3×3, a stride of 1, and uses SAME padding with ReLU activation, used to extract more complex local features. The third feature extraction layer has 256 convolutional kernels, a kernel size of 3×3, a stride of 1, and uses SAME padding. The ReLU activation function is employed to extract morphological features of glands and cells. The second pooling layer also uses max pooling, with a pooling window of 2×2 and a stride of 2. The fourth feature extraction layer has 512 convolutional kernels, a kernel size of 3×3, a stride of 1, and uses SAME padding. The ReLU activation function is employed to extract the correlation information between features. The output layer uses global average pooling to convert each feature map into a single numerical value, outputting five quantitative indicators corresponding to glandular structural abnormality, cellular atypia, stroma density, immune cell distribution density, and the proportion of necrotic areas.

[0039] The quantification values ​​for glandular structural abnormalities are obtained by measuring the irregularity of the glandular contour, calculating the ratio of the gland's perimeter to its area, and then normalizing the result. The range is between 0 and 1; a higher value indicates a more abnormal glandular structure. The quantification values ​​for cellular atypia are obtained by analyzing parameters such as cell size, morphology, and nucleocytoplasmic ratio. Principal component analysis is used to reduce multiple parameters to a single quantification value, ranging from 0 to 1. A higher value indicates higher cellular atypia. The quantification values ​​for interstitial density are obtained by calculating the pixel proportion of the interstitial region, ranging from 0 to 1. A higher value indicates denser interstitium. The quantification values ​​for immune cell distribution density are obtained by counting the number of immune cells per unit area and normalizing the result, ranging from 0 to 1. A higher value indicates a denser distribution of immune cells. The quantification values ​​for the proportion of necrotic areas are obtained by calculating the ratio of the pixel area of ​​the necrotic area to the pixel area of ​​the entire image region, ranging from 0 to 1. A higher value indicates a higher proportion of necrotic areas. This scheme achieves full capture of pathological microenvironment information through comprehensive extraction of multiple features and deep processing of convolutional neural networks, providing rich feature support for subsequent dynamic adaptation and accurate segmentation.

[0040] In existing technologies, feature extraction strategies are mostly fixed settings and cannot be dynamically adjusted according to the specific feature distribution of pathological images, resulting in weak feature extraction targeting and affecting segmentation results.

[0041] Based on this, the dynamic feature adaptation module constructs an adaptation model based on the quantitative indicators extracted by the micro-environment feature perception module. The adaptation model achieves dynamic adaptation of the feature extraction strategy by adjusting the convolution kernel parameters, stride and activation function type of the feature extraction layer. The activation function types include LeakyReLU, GELU and Mish.

[0042] The specific implementation of this technical solution is as follows: The adaptation model of the dynamic feature adaptation module is built based on the deep learning framework TensorFlow 2.8. The model input consists of five quantized indicators extracted by the micro-environment feature perception module, and the output consists of the adjustment parameters of the feature extraction layer. The adjustment of the convolution kernel parameters includes the number and size of the convolution kernels. The number of convolution kernels can be adjusted from 64 to 1024, and the selectable values ​​for the kernel size are 3×3, 5×5, and 7×7. The stride can be adjusted from 1 to 2. When the variance of the quantized feature values ​​is large, the stride is set to 1 to retain more detailed features; when the variance is small, the stride is set to 2 to improve processing efficiency. The choice of activation function is based on a comprehensive judgment of feature coupling strength and channel responsivity. When the feature coupling strength is greater than 0.6 and the channel responsivity is greater than 0.8, the Mish activation function is chosen because it has better gradient flow characteristics and is suitable for extracting complex features. When the feature coupling strength is between 0.3 and 0.6 and the channel responsivity is between 0.5 and 0.8, the GELU activation function is chosen to balance feature extraction capability and computational efficiency. When the feature coupling strength is less than 0.3 and the channel responsivity is less than 0.5, the LeakyReLU activation function is chosen to avoid the gradient vanishing problem.

[0043] The training process of the adaptation model employs the Adam optimizer with a learning rate of 0.001, 500 epochs, and a batch size of 32. The training dataset contains feature quantification metrics and corresponding optimal tuning parameters for 10,000 colorectal cancer pathological images. An early stopping strategy is used during model training: training is stopped and the current optimal model is saved when the loss function on the validation set shows no decrease for 10 consecutive epochs. In practical applications, the dynamic feature adaptation module receives micro-environment feature quantification metrics in real time, inputs them into the trained adaptation model, and outputs tuning parameters, which are sent to the feature extraction layer via a control interface to achieve dynamic adjustment of the feature extraction strategy. This scheme, through the adaptation model and dynamic adjustment mechanism, enables the feature extraction strategy to accurately match the feature distribution of different pathological images, improving the targeting and effectiveness of feature extraction and providing optimized feature input for subsequent accurate segmentation.

[0044] Existing technologies cannot quantify the coupling relationships between features of the pathological microenvironment, resulting in a lack of consideration for such interactions during feature extraction and segmentation, which affects segmentation accuracy.

[0045] Based on this, the microenvironment feature sensing module quantifies the interaction between features through a microenvironment feature coupling strength calculation formula, which is:

[0046] ;

[0047] in, The coupling strength is a characteristic of the microenvironment. This is a quantification of glandular structural abnormalities. This is a quantification value for cell atypia. This represents the quantitative value of interstitial density. This is a quantitative value for the distribution density of immune cells. This is a quantitative value representing the percentage of necrotic area. , , These are the feature weight coefficients. It is the minimum constant. The attenuation coefficient is... For correction factor, As a baseline value for cellular atypia, It is a symbolic function.

[0048] The logical derivation and specific implementation of the calculation formula are as follows: the coupling relationship between various features in the pathological microenvironment exhibits nonlinear characteristics; therefore, a combination of multiple nonlinear functions is used to quantify this relationship. The coupling between glandular structural abnormality and cellular atypia shows a logarithmic growth relationship. As cellular atypia increases, its influence on glandular structural abnormality gradually decreases; therefore, a combination of nonlinear functions is used... A function is used for modeling. The coupling between interstitial density and immune cell distribution density exhibits a square root growth relationship; the rate of increase in the effect of immune cell distribution density on interstitial density gradually slows down. Therefore, a function is adopted... function, The value is 1e-6, to avoid When the value is 0, the value inside the square root is negative, ensuring the validity of the calculation. The coupling between glandular structural abnormality and the proportion of necrotic area exhibits an exponential decay relationship; as the glandular structural abnormality increases, its influence on the proportion of necrotic area gradually decreases. Therefore, a method is adopted... function, The value is set to 2, and experiments have verified that this value can accurately reflect this decay pattern.

[0049] Feature weight coefficients , , The values ​​obtained through optimization using the training set data are 0.4, 0.3, and 0.3, respectively, which satisfy... This ensures that the weight allocation of each feature is reasonable and conforms to pathophysiological principles. Correction coefficient. A value of 0.1 is used to correct for deviations in cell atypia relative to a baseline value. This serves as the baseline value for cellular atypia, set at 0.5, corresponding to the critical value between normal and abnormal cells. Greater than When the sign function returns 1, it increases the correction amount for the coupling strength; when Less than When this happens, the sign function returns -1, reducing the amount of correction for the coupling strength and making the calculation of the coupling strength more consistent with the actual pathological situation.

[0050] It should be noted that: feature weight coefficients , , Selection criteria: Based on a control experiment using 5000 labeled colorectal cancer pathological sections, the contribution of each feature to segmentation accuracy was calculated using the random forest algorithm. Glandular structural abnormalities contributed the most, at 42%. =0.4; the contribution rates of interstitial density and necrotic areas were 31% and 27%, respectively, therefore =0.3、 =0.3, which satisfies + + =1; 2. Parameter attribute description: β values ​​of 0.1 and θ values ​​of 2 are both training optimization results. The 1000 candidate parameters were iteratively optimized through 5-fold cross-validation, and the parameter combination that maximizes the segmentation IoU was finally selected.

[0051] The hardware implementation of this computation uses a TITMS320C6748 DSP chip. An optimized computation program is written in assembly language to ensure real-time performance, with a computation cycle not exceeding 1ms. All quantized values... , , , , All are dimensionless relative values, ranging from 0 to 1. Parameters such as weighting coefficients, attenuation coefficients, and correction coefficients are also dimensionless adjustment coefficients. The calculation results are also dimensionless values, ranging from 0 to 1. Larger values ​​indicate stronger coupling between features. This formula accurately models the coupling relationships between features, providing a quantitative basis for subsequent dynamic feature adaptation and multi-constraint segmentation. This allows the technical solution to fully consider this interaction and improve segmentation accuracy.

[0052] In existing technologies, the weights of feature extraction channels are mostly fixed and cannot be dynamically adjusted according to the feature coupling strength and channel characteristics, resulting in unreasonable channel weight allocation and affecting the feature extraction effect.

[0053] Based on this, the dynamic feature adaptation module determines the weight of each feature extraction channel through a dynamic feature adaptation weight calculation formula, which is:

[0054] ;

[0055] in, Let be the adaptation weight for the j-th feature extraction channel. Extraction channels corresponding to glandular structural abnormality, cellular atypia, stroma density, immune cell distribution density, and the proportion of necrotic areas, respectively. As the baseline weight for the channel, For channel characteristic responsivity, These are the variance weighting coefficients. Let be the variance of the quantized value of the j-th feature. This is the coupling strength correction factor. The coupling strength of the microenvironment features as described in claim 5.

[0056] The logical derivation and specific implementation of the calculation formula are as follows: the calculation of channel adaptation weights requires comprehensive consideration of four factors: feature coupling strength, channel baseline weights, channel feature responsivity, and variance of feature quantization values. Feature coupling strength... This reflects the degree of interaction between various features, which is positively correlated with channel weights, therefore... Introduced into the numerator as a multiplication factor. Channel baseline weight. Experiments determined that values ​​of 0.25, 0.25, 0.2, 0.15, and 0.15 respectively reflect the inherent importance of each channel. Channel characteristic responsivity The mean value of the channel output feature is calculated, ranging from 0 to 1. A larger value indicates a more sensitive channel response to that feature. The function is modeled to reflect its impact on the weights, making the effect of changes in responsiveness on the weights more consistent with actual needs.

[0057] variance of feature quantization values This reflects the volatility of the features; the larger the variance, the more drastic the changes in the features, requiring higher weights to accommodate such volatility. Therefore, a variance weighting coefficient is introduced. The value is 0.2, used to adjust the contribution of variance to the weights. The denominator is a normalization term to ensure... The value range is between 0 and 1, where This is the coupling strength correction coefficient, with a value of 0.3, which reflects the influence of the coupling strength of the microenvironment characteristics on the normalization process, making the weight calculation more accurate.

[0058] This calculation formula is implemented in Python and uses the NumPy library for vectorized computation to improve efficiency. 64-bit floating-point precision is employed during the calculation to ensure accuracy. In practical applications, the dynamic feature adaptation module receives data in real time. , , Parameters such as these are used to calculate the appropriate weights for each of the five feature extraction channels using this formula. After the weight calculation is completed, it is transmitted to the feature extraction layer via the control bus to adjust the feature contribution of each channel. This scheme, through dynamic weight calculation, allows the weights of each feature extraction channel to be adjusted in real time according to the actual situation, ensuring the rationality and relevance of the weight allocation and improving the quality of feature extraction.

[0059] It should be noted that, Weight design rationale: In clinical practice, cellular atypia is a core indicator for tumor diagnosis, but glandular structural abnormality is a fundamental characteristic for tumor regional localization. In the absence of glandular structural abnormalities, cellular atypia is mostly benign hyperplasia. The two are highly complementary in the segmentation task, so they are assigned the same baseline weight of 0.25. The baseline weights for stroma density, immune cell distribution density, and necrotic area proportion are 0.2, 0.15, and 0.15, respectively. These weights are based on the priority ranking of secondary indicators in clinical diagnosis. The source of the values ​​of γ (0.2) and δ (0.3): The parameters of the training set (8000 pathological images) are optimized using the gradient descent algorithm. The objective function is to minimize the segmentation boundary error, and the parameter combination is finally obtained by convergence.

[0060] The optimization of segmentation boundaries in existing technologies lacks an effective constraint mechanism, leading to inaccurate boundary delineation and affecting the accurate identification of tumor regions.

[0061] Based on this, the multi-constraint segmentation module optimizes the segmentation boundary through a segmentation boundary constraint calculation formula, which is:

[0062] ;

[0063] in, These are the boundary constraint values ​​for the partition. The adaptation weights for the j-th feature extraction channel as described in claim 6, The gradient value of the region corresponding to the j-th feature. The maximum gradient value in the region corresponding to the j-th feature. Let be the neighborhood consistency coefficient corresponding to the j-th feature. Let be the average neighborhood consistency coefficient corresponding to the j-th feature. For gradient constraint weights, For neighborhood constraint weights, To avoid a compensation constant with a denominator of zero.

[0064] The logical derivation and specific implementation of the calculation formula are as follows: the accuracy of the segmentation boundary depends on two key factors: the regional gradient and the neighborhood consistency. Therefore, a weighted combination of these two factors is used to construct the constraint calculation formula. Regional gradient Calculated using the Sobel operator, the value ranges from 0 to 255; a larger value indicates a clearer region boundary. Function modeling of its impact on constraint values The value is set to 255 to ensure that the gradient value is normalized and mapped to a range of 0 to 1, so that the influence of the gradient on the constraint value changes in a sinusoidal curve, and the constraint effect is more significant when the gradient is large.

[0065] Neighborhood consistency coefficient This is obtained by calculating the feature similarity between the pixel and its 8 neighboring pixels, ranging from 0 to 1. A higher value indicates higher neighborhood consistency. Function modeling of its impact on constraint values The value is 0.5. The value is set to 1e-6 to avoid the denominator being zero. This function makes the influence of neighborhood consistency on the constraint value grow non-linearly, and the constraint effect is stronger when the consistency is high.

[0066] Gradient constraint weights Neighborhood constraint weights The values ​​of are 0.6 and 0.4, respectively. Experiments have verified that this ratio can achieve the best balance between the gradient and the neighborhood consistency constraint on the segmentation boundary. We assign appropriate weights to each feature channel to ensure that the contribution of different features to the segmentation boundary constraints matches the importance of the features. The value ranges from 0 to 1. A larger value indicates a higher probability that the pixel belongs to the boundary of a tumor region. The multi-constraint segmentation module determines this based on... The value adjusts the pixel assignment of the segmentation boundary, when When the value is greater than 0.5, it is determined to be a pixel representing the boundary of the tumor region; when... When the value is less than or equal to 0.5, it is determined to be a non-boundary pixel.

[0067] This computational formula is implemented on a GPU using CUDA parallel computing, achieving a processing speed of up to 2GB / s, ensuring real-time performance in the segmentation process. Through this constraint-based computation, the multi-constraint segmentation module can accurately optimize the segmentation boundaries, solving the problem of inaccurate boundary delineation in existing technologies and improving the accuracy of tumor region segmentation.

[0068] In existing technologies, the features after dynamic adaptation may still have biases, and the lack of further calibration mechanisms affects the segmentation results.

[0069] Based on this, the colorectal cancer pathological image segmentation system also includes a feature calibration module, which is connected to the dynamic feature adaptation module and the multi-constraint segmentation module respectively. The adapted features are calibrated based on the segmentation boundary constraint value. The feature calibration is achieved by adjusting the dimension weight of the feature vector, and the dimension weight is positively correlated with the segmentation boundary constraint value.

[0070] It should be noted that the calibration features are combined with the fully convolutional network in the following way: The 512-dimensional calibration features output by the feature calibration module are fused with the feature maps of each stage of the fully convolutional network decoder through channel concatenation (Concat) operation. The fusion ratio is set to calibration features: decoder features = 1:1. After fusion, it is used as the input of the next layer of the decoder. 2. The specific logic of adjusting the convolution kernel weights is as follows: Let the original convolution kernel weight of the transposed convolutional layer be W. After adjustment, the weight W' = W × (1 + Φ). The value range of Φ is [0,1]. When the boundary confidence is high, i.e., Φ > 0.5, an additional weight decay factor of 0.1 is added to avoid overfitting boundary noise. When the boundary confidence is low, i.e., Φ ≤ 0.5, the weight decay factor is set to 0.05 to retain more potential boundary information.

[0071] The specific implementation of this technical solution is as follows: The feature calibration module adopts an FPGA and processor co-architecture. The FPGA model is Intel Arria10, and the processor is Intel Core i7-11700K, ensuring the real-time performance and accuracy of the calibration process. The feature calibration module receives the adapted feature vector output by the dynamic feature adaptation module and the segmentation boundary constraint values ​​output by the multi-constraint segmentation module through a high-speed interface. The feature vector has 512 dimensions, with each dimension corresponding to the expression value of a pathological feature.

[0072] It should be noted that the calibration features are combined with the fully convolutional network in the following way: The 512-dimensional calibration features output by the feature calibration module are fused with the feature maps of each stage of the fully convolutional network decoder through channel concatenation (Concat) operation. The fusion ratio is set to calibration features: decoder features = 1:1. After fusion, it is used as the input of the next layer of the decoder. 2. The specific logic of adjusting the convolution kernel weights is as follows: Let the original convolution kernel weight of the transposed convolutional layer be W. After adjustment, the weight W' = W × (1 + Φ). The value range of Φ is [0,1]. When the boundary confidence is high, i.e., Φ > 0.5, an additional weight decay factor of 0.1 is added to avoid overfitting boundary noise. When the boundary confidence is low, i.e., Φ ≤ 0.5, the weight decay factor is set to 0.05 to retain more potential boundary information.

[0073] Dimension weights are calculated using the formula This formula ensures that the dimensional weights range from 0.2 to 1. The larger the value, the greater the dimensional weight, making the guiding role of the segmentation boundary constraint value more significant in feature calibration. The feature calibration process is achieved by multiplying each dimension value of the feature vector by its corresponding dimensional weight, i.e. ,in The value of the d-th dimension of the adapted feature vector. This represents the value of the d-th dimension of the calibrated feature vector.

[0074] It should be noted that the coefficients 0.8 and 0.2 in the dimension weight coefficients originate as follows: Based on the experiment on the impact of calibration features on segmentation accuracy, the optimal mapping relationship between Φ and dimension weights is fitted by linear regression. 0.8 is the contribution coefficient of Φ, and 0.2 is the basic weight, ensuring that the feature retains its basic expression even when Φ=0. In the calibration feature verification step: a new verification step is added to calculate the signal-to-noise ratio (SNR) of the calibrated features. The requirement is that SNR≥35dB, which is based on the effective threshold of pathological image feature expression. If SNR<35dB, the dimension weights are readjusted, and the coefficient 0.8 is increased to 1.0 until the SNR requirement is met.

[0075] The calibrated feature vectors are transmitted to the multi-constraint segmentation module via the data bus to optimize the segmentation process. The feature calibration module's timing is synchronized with other modules, controlled by a 100MHz clock signal to ensure that the calibrated features can be used for segmentation in a timely manner. To improve calibration robustness, the feature calibration module also includes an exception handling mechanism. When the value exceeds the range of 0 to 1, the default dimension weight of 0.5 is automatically used for calibration to avoid the impact of abnormal data on the segmentation results.

[0076] This scheme establishes a calibration link between feature adaptation and segmentation by introducing a feature calibration module. This effectively corrects the deviation of features after dynamic adaptation, making feature representation more accurate and providing higher quality feature input for the multi-constraint segmentation module, thereby further improving the accuracy of tumor region segmentation.

[0077] Existing pathological image segmentation methods have fixed processes, and the lack of effective coordination and optimization in each step makes it difficult to balance segmentation efficiency and accuracy.

[0078] Based on this, this embodiment provides a method for segmenting colorectal cancer pathological images, applied to the colorectal cancer pathological image segmentation system as described in any of the above embodiments, comprising the following steps:

[0079] S1. Obtain digital pathological images of colorectal cancer through the image input unit, and perform image denoising, grayscale normalization and region enhancement processing on the acquired digital pathological images using the pathological image preprocessing module.

[0080] S2. Quantitative indicators such as glandular structural abnormality, cell atypia, stroma density, immune cell distribution density, and necrotic area ratio in the processed pathological images are extracted through the microenvironment feature perception module.

[0081] S3. The feature extraction strategy is dynamically adjusted based on the extracted quantitative indicators through the dynamic feature adaptation module to determine the adaptation weight of each feature extraction channel.

[0082] S4. The tumor region segmentation result is obtained by using the adapted feature extraction strategy and the segmentation boundary constraint value in the pathological image through the multi-constraint segmentation module.

[0083] S5. Output the tumor region segmentation results through the result output module.

[0084] The specific implementation of this method is as follows: In S1, the image input unit uses a high-definition scanner to acquire digital pathological images of colorectal cancer. The scanning range covers the entire pathological slice, the scanning resolution is 0.25 μm / pixel, the acquired image format is TIFF, and the image size is 10000×10000 pixels. The pathological image preprocessing module divides the image into blocks, each with a size of 256×256 pixels. An adaptive median filtering algorithm is used for noise reduction, linear stretching transformation is used for grayscale normalization, and a multi-scale contrast-limited adaptive histogram equalization algorithm is used for region enhancement. The processed image blocks are cached in memory for subsequent processing.

[0085] In S2, the micro-environment feature perception module reads the preprocessed image blocks and inputs them into the convolutional neural network. Through the processing of 4 feature extraction layers and 2 pooling layers, it outputs quantized indicators of 5 features. The value of each quantized indicator is between 0 and 1. The processing speed of the feature extraction process is 10ms per block. The extracted quantized indicators are stored in the feature database.

[0086] In S3, the dynamic feature adaptation module reads quantitative indicators from the feature database and calculates the micro-environment feature coupling strength. The adaptation weights of each feature extraction channel are determined by the dynamic feature adaptation weight calculation formula. The convolution kernel parameters, stride and activation function type of the feature extraction layer are adjusted according to the adaptation weights. The adjusted policy parameters are stored in the policy configuration file.

[0087] In S4, the multi-constraint segmentation module reads the preprocessed image blocks, adjusted policy parameters, and quantization metrics, and uses a fully convolutional neural network for segmentation. The encoder part adopts the same convolutional neural network structure as the micro-environment feature perception module, while the decoder part uses transposed convolutional layers to gradually restore image resolution. Adapted features and segmentation boundary constraint values ​​are incorporated into each stage of the decoder. ,pass Adjust the kernel weights of the transposed convolutional layer. The segmentation process takes 15ms per block, and the segmented image blocks are stored in the result cache.

[0088] In S5, the result output module reads the segmented image blocks from the result cache, performs stitching processing, and restores the complete segmented image. The pixel value of the tumor region in the segmented image is 255, and the pixel value of the non-tumor region is 0. At the same time, it calculates statistical parameters such as the area, perimeter, and circularity of the tumor region and outputs them in the form of image files and text files. The image file format is TIFF, and the text file format is CSV.

[0089] It should be noted that the segmentation performance of different layer combinations, including 2-5 feature extraction layers and 1-3 pooling layers, shows that a combination of 4 feature extraction layers and 2 pooling layers achieves the optimal trade-off between segmentation IoU and processing speed. The 4 feature extraction layers progressively extract basic texture, morphological features, and related features, while the 2 pooling layers balance feature dimensionality reduction and detail preservation. This combination achieves a segmentation IoU of 89.7% and a processing speed of 10ms / block. If the number of feature extraction layers is increased to 5, the IoU only... Improvement of 0.3%, but processing speed decreased by 40%; 2. Hyperparameter impact explanation: - Convolution kernel size: 3×3 kernel segmentation IoU is 89.7%, which is better than 5×7 kernel's 86.2%, because smaller kernels are better at capturing the fine structure of glands and cells; - Pooling window: 2×2 pooling is better than 3×3 pooling, 3×3 pooling will lose 12% of the cell heterogeneity details; - Activation function: The use of ReLU in the feature extraction layer reduces the gradient vanishing probability by 60% compared to Sigmoid, improving the network convergence speed.

[0090] This method achieves fully automated processing from image acquisition to segmentation result output through step-by-step process design and collaborative optimization of each step. The processing parameters and strategies of each step can be dynamically adjusted according to the actual image features, taking into account both segmentation efficiency and accuracy, and is suitable for batch processing of large-scale pathological images.

[0091] The adjustment of feature extraction strategies in existing technologies lacks clear logic and quantitative basis, resulting in weak targeting of the adjusted strategies and affecting the segmentation effect.

[0092] Based on this, the specific process of dynamically adjusting the feature extraction strategy in step S3 is as follows: The micro-environment feature coupling strength is calculated based on the quantification index extracted in step S2. The adaptation weights of each feature extraction channel are determined according to the micro-environment feature coupling strength using the dynamic feature adaptation weight calculation formula. The convolution kernel parameters, stride, and activation function type of the feature extraction layer are adjusted based on the adaptation weights. The dynamic feature adaptation weight calculation formula is as follows: .

[0093] The specific implementation of this process is as follows: First, based on the quantitative indicators extracted in step S2, such as glandular structural abnormality, cellular atypia, stroma density, immune cell distribution density, and the proportion of necrotic areas, the process is calculated using a microenvironment feature coupling strength calculation formula. The calculation process uses 64-bit floating-point precision to ensure accurate results. Then, the baseline weights for each feature extraction channel are obtained. Channel characteristic responsivity and the variance of the quantized feature values , The values ​​are 0.25, 0.25, 0.2, 0.15, and 0.15 respectively. It is obtained by calculating the mean of the channel output features. This is obtained by calculating the variance of the feature in the current image patch.

[0094] Will , , , Substitute the parameters into the dynamic feature adaptation weight calculation formula, and calculate the adaptation weights for each of the five feature extraction channels. After the calculation is completed, Normalization is performed to ensure the sum of the five weights is 1. Based on... Adjusting the size of the feature extraction layer parameters when When the value is greater than 0.25, the number of convolutional kernels in the corresponding channel is increased by 50%, the kernel size is adjusted to 5×5, the stride is set to 1, and the activation function is selected as Mish; when Between 0.15 and 0.25, the number of kernels remains constant, the kernel size is 3×3, the stride is set to 1, and the activation function is GELU; when When the value is less than 0.15, the number of convolutional kernels is reduced by 30%, the kernel size is adjusted to 3×3, the stride is set to 2, and the activation function is LeakyReLU.

[0095] The adjusted parameters are stored in a configuration file and updated in real time to the hardware configuration registers of the feature extraction layer, ensuring that subsequent feature extraction processes use the adjusted strategy. To verify the adjustment effect, a feedback mechanism is also implemented: the adjusted strategy is applied to feature extraction, the signal-to-noise ratio (SNR) of the extracted features is calculated, and when the SNR is less than 30 dB, the parameters are readjusted until the SNR is greater than or equal to 30 dB. This process, through the combination of quantization calculation and dynamic adjustment, enables the feature extraction strategy to accurately match the feature distribution of the current image, improving the quality and specificity of feature extraction and providing strong support for subsequent accurate segmentation.

[0096] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A colorectal cancer pathological image segmentation system, comprising an image input unit for acquiring digital pathological images of colorectal cancer, characterized in that, It also includes a pathological image preprocessing module, a microenvironment feature perception module, a dynamic feature adaptation module, a multi-constraint segmentation module, and a result output module. The pathological image preprocessing module is connected to the image input unit and performs image denoising, grayscale normalization, and region enhancement processing on the acquired digital pathological images. The microenvironment feature perception module is connected to the pathological image preprocessing module and extracts microenvironment features related to colorectal cancer from the digital pathological images. These microenvironment features include quantitative indicators such as glandular structural abnormality, cellular atypia, stroma density, immune cell distribution density, and the proportion of necrotic areas. The dynamic feature adaptation module... The system is connected to both the pathological image preprocessing module and the microenvironment feature perception module. An adaptation model is constructed based on the extracted microenvironment feature quantification indicators. The feature extraction strategy is dynamically adjusted by modifying the convolution kernel parameters, stride, and activation function type of the feature extraction layer, while simultaneously determining the adaptation weights for each feature extraction channel. The multi-constraint segmentation module is connected to both the dynamic feature adaptation module and the microenvironment feature perception module. It uses the adapted feature extraction strategy and the adaptation weights for each feature extraction channel, combined with constraint rules, to segment the tumor region of the pathological image. The result output module is connected to the multi-constraint segmentation module and outputs the tumor region segmentation results.

2. The colorectal cancer pathological image segmentation system according to claim 1, characterized in that, The targeted processing of the pathological image preprocessing module includes image denoising, grayscale normalization, and region enhancement. The image denoising adopts an adaptive median filtering algorithm, the grayscale normalization adopts a linear stretching transformation, and the region enhancement adopts a multi-scale contrast-limited adaptive histogram equalization algorithm.

3. The colorectal cancer pathological image segmentation system according to claim 1, characterized in that, The microenvironment features extracted by the microenvironment feature perception module include glandular structural abnormality, cell atypia, interstitial density, immune cell distribution density, and necrotic area ratio. The microenvironment feature perception module extracts quantitative indicators of the above features through a convolutional neural network, which contains 4 feature extraction layers and 2 pooling layers.

4. The colorectal cancer pathological image segmentation system according to claim 1, characterized in that, The dynamic feature adaptation module constructs an adaptation model based on the quantitative indicators extracted by the micro-environment feature perception module. The adaptation model achieves dynamic adaptation of the feature extraction strategy by adjusting the convolution kernel parameters, stride, and activation function type of the feature extraction layer. The activation function types include LeakyReLU, GELU, and Mish.

5. The colorectal cancer pathological image segmentation system according to claim 3, characterized in that, The microenvironment feature sensing module quantifies the interaction between features through a formula for calculating the coupling strength of microenvironment features. This formula is: ； in, The coupling strength is a characteristic of the microenvironment. This is a quantification of glandular structural abnormalities. This is a quantification value for cell atypia. This represents the quantitative value of interstitial density. This is a quantitative value for the distribution density of immune cells. This is a quantitative value representing the percentage of necrotic area. , , These are the feature weight coefficients. It is the minimum constant. The attenuation coefficient is... For correction factor, As a baseline value for cellular atypia, It is a symbolic function.

6. The colorectal cancer pathological image segmentation system according to claim 5, characterized in that, The dynamic feature adaptation module determines the weight of each feature extraction channel through a dynamic feature adaptation weight calculation formula, which is: ； in, Let be the adaptation weight for the j-th feature extraction channel. These extraction channels correspond to the degree of glandular structural abnormality, cellular atypia, stroma density, immune cell distribution density, and the proportion of necrotic areas, respectively. As the baseline weight for the channel, For channel characteristic responsivity, These are the variance weighting coefficients. Let be the variance of the quantized value of the j-th feature. This is the coupling strength correction factor. The coupling strength of the microenvironment features as described in claim 5.

7. The colorectal cancer pathological image segmentation system according to claim 6, characterized in that, The multi-constraint segmentation module optimizes the segmentation boundary through a segmentation boundary constraint calculation formula, which is: ； in, These are the boundary constraint values ​​for the partition. The adaptation weights for the j-th feature extraction channel as described in claim 6, The gradient value of the region corresponding to the j-th feature. The maximum gradient value in the region corresponding to the j-th feature. Let be the neighborhood consistency coefficient corresponding to the j-th feature. Let be the average neighborhood consistency coefficient corresponding to the j-th feature. For gradient constraint weights, For neighborhood constraint weights, To avoid a compensation constant with a denominator of zero.

8. The colorectal cancer pathological image segmentation system according to claim 1, characterized in that, It also includes a feature calibration module, which is connected to the dynamic feature adaptation module and the multi-constraint segmentation module respectively. The feature calibration module calibrates the adapted features based on the segmentation boundary constraint value. The feature calibration is achieved by adjusting the dimension weights of the feature vectors. The dimension weights are positively correlated with the segmentation boundary constraint value.

9. A method for segmenting pathological images of colorectal cancer, applied to the colorectal cancer pathological image segmentation system as described in any one of claims 1-8, characterized in that, Includes the following steps: S1. Obtain digital pathological images of colorectal cancer through the image input unit, and perform image denoising, grayscale normalization and region enhancement processing on the acquired digital pathological images using the pathological image preprocessing module. S2. Quantitative indicators such as glandular structural abnormality, cell atypia, stroma density, immune cell distribution density, and necrotic area ratio in the processed pathological images are extracted through the microenvironment feature perception module. S3. The feature extraction strategy is dynamically adjusted based on the extracted quantitative indicators through the dynamic feature adaptation module to determine the adaptation weight of each feature extraction channel. S4. The tumor region segmentation result is obtained by using the adapted feature extraction strategy and the segmentation boundary constraint value in the pathological image through the multi-constraint segmentation module. S5. Output the tumor region segmentation results through the result output module.

Images