CN-122023388-A - Interpretable stage prediction method for colorectal cancer image of small sample

CN122023388ACN 122023388 ACN122023388 ACN 122023388ACN-122023388-A

Abstract

The invention relates to the technical field of medical image analysis, in particular to an interpretable stage prediction method of a small sample colorectal cancer image, which comprises the steps of carrying out two-domain cooperative variation bias field correction on an acquired original tissue pathology image so as to eliminate imaging artifacts. And carrying out multi-scale morphological gradient texture analysis on the corrected image to generate a structure consistency adjustment coefficient matrix, and carrying out differential enhancement on the image area according to the structure consistency adjustment coefficient matrix to obtain a visual enhancement image. And inputting the enhanced image into a pre-trained feature decoupling network to obtain an image feature coding vector. The vector is input into a small sample learning frame, a task-adaptive feature expression is generated by combining the support set samples, and an interpretable classifier is utilized to output a cancer stage prediction result and a corresponding visual interpretation graph. The method can effectively improve the accuracy of colorectal cancer stage prediction and the interpretability of model decision under the condition of small samples.

Inventors

ZHENG HAOXUAN
XU JINGLIN
LONG TIANDI

Assignees

南方医科大学南方医院

Dates

Publication Date: 20260512
Application Date: 20260326

Claims (10)

1. An interpretable staging prediction method for a small sample colorectal cancer image, the method comprising: Obtaining original image data of a colorectal tissue pathological image, and performing two-domain cooperative variation bias field correction treatment on the original image data to obtain a corrected pathological image; performing texture analysis processing of multi-scale morphological gradient factors on the corrected pathological image to generate a structure consistency adjustment coefficient matrix; according to the structural consistency adjustment coefficient matrix, performing differential enhancement processing on the corrected pathological image to obtain an area-targeted optimized image; Performing combination processing of global color remapping and local brightness optimization on the region targeted optimized image to obtain a visually enhanced image to be analyzed; Inputting the image to be analyzed after visual enhancement into a pre-trained feature decoupling network to generate an image feature coding vector; inputting the image feature coding vector into a small sample learning frame, and generating a feature expression based on task adaptation by combining the marked samples in the support set; Based on the task adaptation-based feature expression, a stage prediction result and a corresponding visual interpretation graph are generated through an interpretable classifier.
2. The method of claim 1, wherein performing a two-domain collaborative variational bias field correction process on the raw image data to obtain a corrected pathology image comprises: Decomposing the original image data into a low-frequency offset field component and a high-frequency detail layer component, wherein the low-frequency offset field component is obtained through low-pass filtering extraction; applying L2 norm constraint to the low-frequency offset field component to inhibit global intensity variation of the low-frequency offset field component, and simultaneously applying total variation norm constraint to the high-frequency detail layer component to maintain local structure information of the high-frequency detail layer component; Calculating gradient self-adaptive detail retaining weight according to the gradient amplitude of the high-frequency detail layer component, wherein the gradient self-adaptive detail retaining weight and the gradient amplitude form positive correlation; Constructing a collaborative optimization objective function comprising a norm constraint term and the gradient self-adaptive detail retention weight term, wherein the norm constraint term comprises an L2 norm constraint term for a low-frequency bias field component and a total variation norm constraint term for a high-frequency detail layer component, and solving the collaborative optimization objective function through an iterative optimization algorithm to obtain an optimized low-frequency bias field component and an optimized high-frequency detail layer component; And carrying out weighted fusion on the optimized low-frequency bias field component and the optimized high-frequency detail layer component to obtain a corrected pathological image with uniform illumination and enhanced detail.
3. The method of claim 2, wherein the performing texture analysis of the multi-scale morphological gradient factor on the corrected pathology image to generate a structure consistency adjustment coefficient matrix comprises: performing morphological gradient operation on the corrected pathological image by adopting a plurality of structural elements with different sizes, and respectively acquiring a texture distribution feature map and an edge gradient feature map under each scale; Performing space alignment and feature stacking on the texture distribution feature graphs of all scales to form a multi-scale texture distribution feature tensor; Performing space alignment and feature stacking on the edge gradient feature graphs of all scales to form a multi-scale edge gradient feature tensor; The multi-scale texture distribution characteristic tensor and the multi-scale edge gradient characteristic tensor are input into a nonlinear mapping module together, wherein the nonlinear mapping module is formed by a multi-layer fully-connected network; And mapping the multi-scale texture distribution feature tensor and the multi-scale edge gradient feature tensor to the same feature representation space through the nonlinear mapping module, and outputting a coefficient matrix with the same size as the input image space as the structure consistency adjustment coefficient matrix, wherein the numerical value of each position in the matrix represents the structure consistency intensity of the region to which the pixel point belongs.
4. An interpretable staged prediction method for small sample colorectal cancer images according to claim 3, wherein the differential enhancement processing is performed on the corrected pathological images according to the structural consistency adjustment coefficient matrix to obtain region-specific optimized images, comprising: comparing each numerical value in the structure consistency adjustment coefficient matrix with a preset texture threshold value, judging the corresponding position as a texture area if the numerical value is larger than the texture threshold value, and judging the corresponding position as a non-texture area if the numerical value is not larger than the texture threshold value; For the texture region, smoothing the corresponding region of the corrected pathology image by adopting a low-pass filter to obtain a basic smooth image of the texture region, calculating a high-frequency difference image between the corrected pathology image and the basic smooth image, multiplying the Gao Pincha partial image by a weighting coefficient derived from the structure consistency adjustment coefficient matrix, and then superposing the result back to the basic smooth image to finish self-adaptive sharpening of the texture region; for the non-texture region, performing primary smoothing on the corresponding region of the corrected pathological image by adopting a Gaussian filter, performing edge protection smoothing on the same region by adopting a bilateral filter, and mixing the result of the primary smoothing with the result of the edge protection smoothing according to a preset proportion to finish the mixed smoothing of the non-texture region; and splicing and fusing the texture region which is subjected to self-adaptive sharpening and the non-texture region which is subjected to mixed smoothing according to the original spatial position, and generating the region-targeted optimized image.
5. The method for interpretive stage prediction of small sample colorectal cancer image according to claim 1, wherein the combined processing of global color remapping and local brightness optimization is performed on the region-specific optimized image to obtain a visually enhanced image to be analyzed, comprising: Converting the region-targeted optimized image from an original color space to a color space, and respectively calculating global statistical properties of a brightness channel, a first chromaticity channel and a second chromaticity channel under the color space; Based on the global statistical characteristics, mapping functions are respectively constructed for the brightness channel, the first chromaticity channel and the second chromaticity channel, the mapping functions map original pixel values of the channels to a target distribution interval, and the mapping functions are independently executed for each channel to obtain a color balanced intermediate image; Calculating a local brightness average value of each pixel point in the area targeted optimized image, and constructing an exponential decay function or a gain function according to a difference value between the local brightness average value and preset target brightness, wherein the exponential decay function is used for reducing the brightness of an over-bright area, and the gain function is used for improving the brightness of an over-dark area; introducing a boundary distance weighting factor when applying the exponential decay function or the gain function, the boundary distance weighting factor ensuring a smooth transition of brightness adjustment at a region boundary; and merging the brightness channel subjected to brightness adjustment with the mapped color channel, converting the brightness channel into an original color space, and outputting the image to be analyzed after visual enhancement.
6. The method of claim 1, wherein inputting the visually enhanced image to be analyzed into a pre-trained feature decoupling network to generate an image feature encoding vector comprises: The pre-trained characteristic decoupling network comprises a shared encoder and a plurality of parallel decoupling branches, wherein the shared encoder is formed by alternately a convolution layer and a pooling layer; firstly, extracting the image to be analyzed after visual enhancement through the shared encoder to obtain a shared feature map; The shared feature map is simultaneously sent into a morphological feature decoupling branch, a texture feature decoupling branch and a color feature decoupling branch, wherein the morphological feature decoupling branch is focused on extracting features related to gland structures and cell nucleus morphology, the texture feature decoupling branch is focused on extracting features related to cell arrangement and matrix textures, and the color feature decoupling branch is focused on extracting features related to dyeing depth and color distribution; Each decoupling branch outputs a decoupled feature vector, all the feature vectors output by the decoupling branches are spliced and fused, the dimension reduction and integration are carried out through a full-connection layer, and finally a vector with fixed dimension is output as the image feature coding vector.
7. The method of claim 6, wherein inputting the image feature encoding vector into a small sample learning framework, in combination with the labeled samples in the support set, generates a task-adaptive based feature expression, comprising: The small sample learning framework adopts a learning method based on measurement, and the support set comprises a small number of marked sample images from a plurality of different stages and corresponding stage labels thereof; Respectively extracting sample feature coding vectors of each marked sample image in the support set by using the pre-trained feature decoupling network; Constructing a task feature prototype, wherein the task feature prototype is obtained by solving the average value of all sample feature coding vectors belonging to the same stage in the support set; Calculating a feature distance between the image feature coding vector and the task feature prototype of each stage; Using an attention mechanism, distributing weights to task feature prototypes of each stage according to the feature distance, wherein the prototype weights are higher when the distance is closer; And aggregating all weighted task feature prototypes to generate a task context-aware feature representation, wherein the task context-aware feature representation is the task adaptation-based feature representation.
8. The method of claim 7, wherein the task-adaptive feature expression-based generation of the staging prediction results and the corresponding visual interpretation map by an interpretable classifier comprises: The interpretable classifier comprises a fully-connected classification layer and a gradient return interpretation module; Inputting the feature expression based on task adaptation into the fully-connected classification layer, outputting a prediction probability value of each stage class after nonlinear transformation, and taking the stage class corresponding to the highest prediction probability value as the stage prediction result; After the fully-connected classification layer obtains a prediction result, calculating the gradient of the feature expression based on task adaptation to the final prediction category through the gradient feedback interpretation module; Mapping the calculated gradient back to the original image space, wherein the specific process is that the gradient value is reversely propagated along the decoding path of the pre-trained characteristic decoupling network until the pixel space of the image to be analyzed after visual enhancement; And obtaining a thermodynamic diagram consistent with the size of the input image in a pixel space, wherein the intensity value of each pixel point in the thermodynamic diagram represents the contribution degree of the pixel to the final prediction result, and the thermodynamic diagram is the visual interpretation diagram.
9. An interpretable staging prediction method for small sample colorectal cancer images according to claim 2, characterized in that the process of calculating gradient adaptive detail preserving weights from the gradient magnitudes of the high frequency detail layer components includes: On the high-frequency detail layer component, calculating gradient components of each pixel point in the horizontal direction and the vertical direction by adopting a sobel operator; calculating the gradient amplitude of each pixel point according to the horizontal gradient component and the vertical gradient component; normalizing the gradient amplitude value to enable the value range to be between zero and one; and inputting the normalized gradient amplitude value into a monotonically increasing conversion function, wherein the output value of the conversion function is the gradient self-adaptive detail preserving weight of the pixel point, and the pixel point with the larger gradient amplitude value is provided with the corresponding gradient self-adaptive detail preserving weight so as to better preserve the edge and detail information of the corresponding pixel position in the subsequent fusion.
10. The method of claim 4, wherein the step of determining the base smoothed image by multiplying the Gao Pincha th order image by the weighting coefficients derived from the matrix of structural consistency adjustment coefficients comprises: Normalizing the structure consistency adjustment coefficient matrix to ensure that the numerical value range in the matrix is between zero and one; setting a basic sharpening intensity coefficient and a maximum sharpening intensity coefficient; For each pixel position in the texture region, multiplying the normalized structure consistency adjustment coefficient value of the corresponding position by the difference value between the maximum sharpening intensity coefficient and the basic sharpening intensity coefficient, and adding the basic sharpening intensity coefficient to obtain a final weighting coefficient of the corresponding pixel position; the weighting coefficient is used for controlling the intensity of the high-frequency differential image when the high-frequency differential image is superimposed on the basic smooth image at the corresponding pixel position, and the higher the coefficient is, the stronger the sharpening effect is.

Description

Interpretable stage prediction method for colorectal cancer image of small sample Technical Field The invention relates to the technical field of medical image analysis, in particular to an interpretable stage prediction method for a colorectal cancer image of a small sample. Background The performance of the existing colorectal cancer pathological image stage prediction method is greatly dependent on the quality of image preprocessing. In the process of digitizing pathological sections, nonuniform brightness bias fields are often introduced due to factors such as dyeing concentration difference, uneven section thickness, scanning illumination conditions and the like, and the artifacts can mask real tissue morphological characteristics and interfere subsequent characteristic extraction and classification. Conventional preprocessing techniques mostly employ a single-domain correction strategy, such as histogram matching or homomorphic filtering in the spatial domain, or background estimation and subtraction in the transform domain. The methods are often difficult to simultaneously and effectively eliminate the influence of large-range smooth brightness gradual change and local high-frequency noise, and the corrected image can still remain artifacts or lose important texture details, so that the image quality cannot meet the requirement of high-precision analysis. In the feature enhancement and extraction stage, the existing method generally adopts a unified image enhancement algorithm or a fixed convolutional neural network structure. However, histological features corresponding to different stages of colorectal cancer differ in texture, contrast and spatial distribution. The general enhancement mode cannot adaptively enhance the local structures with discriminant, and may excessively smooth key areas or excessively enhance irrelevant backgrounds. This is particularly disadvantageous in small sample learning scenarios, because limited marker data requires that feature expressions must be highly refined and discriminative, and any noise introduced by mishandling or feature enhancement inaccuracy can reduce the generalization ability and prediction reliability of the model. Disclosure of Invention The invention aims to solve the defects in the prior art, and provides an interpretable stage prediction method for colorectal cancer images of small samples. In order to achieve the above purpose, the invention adopts the following technical scheme that the method for the interpretable stage prediction of the colorectal cancer image of the small sample comprises the following steps: Obtaining original image data of a colorectal tissue pathological image, and performing two-domain cooperative variation bias field correction treatment on the original image data to obtain a corrected pathological image; performing texture analysis processing of multi-scale morphological gradient factors on the corrected pathological image to generate a structure consistency adjustment coefficient matrix; according to the structural consistency adjustment coefficient matrix, performing differential enhancement processing on the corrected pathological image to obtain an area-targeted optimized image; performing combination processing of global color remapping and local brightness optimization on the region-targeted optimized image to obtain a visually enhanced image to be analyzed; Inputting the image to be analyzed after visual enhancement into a pre-trained feature decoupling network to generate an image feature coding vector; inputting the image feature coding vector into a small sample learning frame, and generating a feature expression based on task adaptation by combining the marked samples in the support set; based on the feature expression based on task adaptation, a stage prediction result and a corresponding visual interpretation graph are generated through an interpretability classifier. As a further aspect of the present invention, a dual-domain cooperative variation bias field correction process is performed on original image data to obtain a corrected pathology image, including: Decomposing the original image data into a low-frequency offset field component and a high-frequency detail layer component, wherein the low-frequency offset field component is obtained through low-pass filtering extraction; applying L2 norm constraint to the low-frequency offset field component to inhibit global intensity change of the low-frequency offset field component, and simultaneously applying total variation norm constraint to the high-frequency detail layer component to maintain local structure information of the high-frequency detail layer component; Calculating gradient self-adaptive detail retention weight according to the gradient amplitude of the high-frequency detail layer component, wherein the gradient self-adaptive detail retention weight and the gradient amplitude form positive correlation; constructing a collaborative optimizati