Colorectal cancer diagnosis method based on multi-granularity cascaded perception

Abstract

The application discloses a kind of based on multi-granularity cascading perception's colorectal cancer diagnosis method.The specific implementation process is: this method constructs discriminant feature from fine granularity and coarse granularity dual perspective.In fine granularity perspective, regular adaptive sampling convolution is designed, the shift of sampling point is guided using gland boundary, and the sub-pixel level feature extraction of disordered gland is realized by combining hyperbolic tangent constraint limit and bilinear interpolation.In the coarse granularity perspective, group residual linear encoder is constructed, class features are linearly modeled in multiple subspace through kernel mapping, and the inter-class difference is strengthened using residual path, to realize the efficient representation of tissue structure.To realize cross-granularity collaboration, a multi-granularity cascading feature coupling module is constructed, discriminant response is re-scaled in channel and spatial dimension, low response noise is inhibited through content perception gating, and cross-granularity semantic conflict is alleviated.Experiments show that the method has significant diagnostic performance and clinical interpretability, can reduce the workload of pathologists and improve the efficiency of colorectal cancer diagnosis.

Description

Technical Field

[0001] This invention relates to a method for diagnosing colorectal cancer based on multi-granularity cascade sensing, which belongs to the field of colorectal cancer diagnosis. Background Technology

[0002] Colorectal cancer is the third most common cancer and the second leading cause of cancer-related deaths worldwide. Current diagnosis primarily relies on colonoscopy and histopathological evaluation. However, in the early stages of colorectal cancer, newly developed tumors or small polyps often present as flat or low-lying lesions with low optical contrast to the surrounding mucosa. This makes lesion visualization insufficient when relying solely on colonoscopy for screening and diagnosis, leading to a significant increase in missed or misdiagnosed rates. Therefore, histopathological examination is the "gold standard" for colorectal cancer detection. Due to the irregularity of the staining process and variations in texture, structure, and heterogeneity, colorectal cancer tissue sections always exhibit high variability and complexity in their appearance. This variability and complexity of histopathological images makes visual examination under a microscope typically time-consuming, and colorectal cancer diagnosis requires a high level of pathological expertise, posing a challenge to pathologists. Furthermore, the current lack of objective diagnostic standards means that different pathologists may differ in identifying morphological features, determining lesion boundaries, and applying grading standards, resulting in observations significantly influenced by individual expert heterogeneity and inter-expert heterogeneity. In recent years, while single-granularity deep learning-based assisted diagnostic methods have improved detection efficiency to some extent, the fixed sampling mechanism of convolutional kernels limits their adaptability in irregular glandular regions, making them inadequate when faced with morphologically heterogeneous lesions. Furthermore, the quadratic complexity related to sequence length in the self-attention operator and the lack of modeling ability for fine-grained local priors limit their scalability in colorectal cancer diagnosis. In addition, single-granularity models cannot simultaneously capture fine-grained microscopic features and coarse-grained macroscopic tissue structures in CRC pathological sections, easily leading to the loss of discriminative lesion features. To address these issues, this invention proposes a colorectal cancer diagnostic method based on multi-granularity perspective cascaded perception. Through the synergy of fine-grained and coarse-grained dual perspectives and multi-granularity dynamic coupling, it achieves high-precision, highly robust, and clinically interpretable colorectal cancer diagnosis. Summary of the Invention

[0003] This invention addresses the common problems in colorectal cancer tissue samples, such as disordered glandular structure, inter-class similarity, and the tendency of existing diagnostic methods, which are based on a single granular perspective and easily lead to the loss of discriminative features. It proposes a colorectal cancer diagnostic method based on multi-granular perspective cascaded sensing, which simultaneously captures the microscopic features of cell nuclei and the macroscopic structural layout of colorectal cancer tissue sections. The specific implementation steps of this invention are as follows: 1. Data Acquisition and Non-overlapping Windowing Preprocessing: Whole-section images of three types of colorectal cancer tissue—cancer, inflammation, and normal—are acquired from hospitals. High-resolution sections are then divided into non-overlapping windowed sections. Therefore, this invention performs non-overlapping windowing on the high-resolution whole-section images, decomposing the entire section into multiple image blocks to reduce computational complexity while maintaining spatial structural integrity. A structure-preserving color normalization method is then used to normalize these image blocks, aligning the color distribution of different sections while maintaining consistency in tissue structure and morphological features. This effectively reduces inter-domain shifts caused by staining differences, resulting in the input feature map. ,in For the number of channels, , The height and width of the feature map.

[0004] 2. Constructing a fine-grained sensing network (2a) Regularized adaptive sampling convolution ① Offset Amplitude Prediction An offset generation subnet was constructed. This subnet has an output channel of of Convolutions are used to generate a corresponding two-dimensional offset for each sampling point. Then, the input feature map is... transported to The original two-dimensional offset field is obtained. : Where H′ and W′ are the spatial dimensions of the offset map; ②Offset Amplitude Regularization Mechanism We employ a learnable scalar α and define a positive constraint upper bound based on it. m To dynamically adjust the sampling radius for different glandular structures: in, It is a smooth activation function; subsequently, it is applied through a hyperbolic tangent mapping. Original offset Mapping to a bounded interval to ensure the geometric stability of the offset yields the actual sampling offset. : in Ensure the offset is continuous Shape changes cause the offset of adjacent pixels to have Continuity, in theory, avoids the explosion of offset gradients.

[0005] ③ Continuous Domain Sampling Reconstruction Each relative sampling position Projected onto continuous domain coordinates : And for each channel Bilinear interpolation is used to approximate the sampling, thereby obtaining the continuous domain characteristic response. : in, for The set of nearest-neighbor integer points. Features of the continuous domain obtained based on bilinear interpolation. , get location Central Channel Sampling response : in For the convolution kernel at the relative sampling position The weights are then assigned. Finally, the responses from all positions and channels form the complete output feature map: (2b) Fine-grained sensing module: First, the input features are processed using a regularized adaptive sampling convolution. Perform spatial adaptive sampling to capture deformed glandular structures, and obtain : in, This represents a regularized adaptive sampling convolution; subsequently, A linear mapping is performed in the channel dimension, and batch normalization is applied. Standardizing the feature distribution mitigates internal covariate bias and improves training stability. Next, a non-linear activation function is applied. This is introduced to enhance the nonlinear expressive power of feature mapping, thereby capturing more complex inter-channel interactions. Finally, residual connections are used to achieve an identity mapping between input and output features, effectively mitigating degradation and gradient vanishing problems in deep structures, resulting in the final module output. : in This represents pointwise convolution, used to stabilize feature distribution and accelerate convergence. This represents a nonlinear activation function, which introduces nonlinear transformation capabilities into the feature space of the module, thereby improving the expressive flexibility and discriminative power in the channel dimension.

[0006] 3. Construct a coarse-grained sensing network (3a) Group of residual linear encoders ① Subspace grouping Let the input feature map be ,in for First, a linear mapping of query, key, and value is obtained through channel projection operations. and along the channel dimension Grouped into G independent complementary subspaces, in different subspaces The subspace vectors focus on learning features such as gland morphology, cell nuclear texture, or matrix structure, thereby enhancing the separability of each category in the feature space. ,in .

[0007] ② Depth Residual Path Designing deep residual paths preserves the detailed differences between categories, thereby enhancing the consistency of category features and obtaining discriminative details. : ③ Linear kernel mapping Design kernel function mapping , will the Subspace Internal query AND key The attention mechanism is achieved by projecting the data onto the kernel feature space, thereby realizing an approximately linear attention mechanism. The mapping function is expressed as: .pass The explicit similarity calculation in traditional attention is transformed into a linear operation within the kernel feature space, thus reconstructing the attention mechanism into a linear form, reducing the complexity from... Reduce to , obtain attention output : ④ Coarse-grained feature reconstruction Will Reconstructed into a two-dimensional feature map form And by using depth residuals to determine details With global reconstruction features The components are interconnected, and finally mapped to obtain the final output. : in, This represents pointwise convolution.

[0008] (3b) Coarse-grained sensing module An integrated layer normalization and dual residual structure construct a coarse-grained encoder. First, the input features are layer-normalized and then fed into a group residual linear encoder module to model the global dependencies of the coarse-grained organizational structure with linear complexity. Then, residual connections preserve the original semantic information and stabilize gradient propagation. in, The group represents a residual linear encoder; then, X is normalized twice and then nonlinearly enhanced in the channel dimension through a feedforward network to obtain a more discriminative semantic representation, thus obtaining the final output.

[0009] 4. Construct a multi-granularity cascaded feature fusion mechanism (4a) Cascaded feature refinement: The prior knowledge of the previous level is added to each fusion module in a cascade manner. The multi-granularity cascaded feature fusion mechanism adds the output of the previous stage to the next multi-granularity cascaded feature fusion module in turn to further refine the features.

[0010] (4b) Feature recalibration unit ① Feature alignment strategy First, let's look at the previous stage. joint priors The core semantic distribution is obtained by using channel linear projection and then spatial downsampling with windowed average pooling. : in This is a spatially position-by-spatial linear projection operator that performs linear channel recombination for each spatial position while preserving the spatial topology. Indicates window average pooling; ② Fine-grained spatial recalibration The aggregation phase uses global average pooling and global max pooling at the channel level. Spatial information of fine-grained feature maps This generates two different context descriptors. and Then, these two descriptors are correlated and the weight of each spatial location is obtained by combining the spatial context. : in Indicates global average pooling. Indicates global max pooling; This indicates a feature map concatenation operation. for activation, Spatial convolutions with large receptive fields provide a wider receptive field to aggregate neighborhood context. Finally, utilizing... The product is based on spatial location weights. right The corresponding spatial location is recalibrated, and the following is inferred: : in, This indicates element-wise multiplication.

[0011] ③ Coarse-grained feature recalibration Employing spatial dimension-based global average pooling and global max pooling convergence stages Global context information describing the channel Then, utilize the shared fully connected layer. Adaptive learning of channel position weights : in, Indicates global average pooling. Indicates global max pooling; This represents element-wise addition. Finally, the features are analyzed through element-wise multiplication. Adaptive recalibration is performed on the significant channels to obtain : ④ Multi-granularity feature coupling The previous stage Core semantic distribution With multi-granularity features and Interconnected features along the channel dimension to integrate cross-layer and cross-granularity semantic information, aiming to construct multi-level semantic associations in the feature space. Then, channel dimensionality reduction and linear remapping are used to compress redundant dimensions and enhance the coupling between features, thereby obtaining a more compact and discriminative joint representation. .

[0012] in, Indicates size is Pointwise convolution. Finally, the joint representation... Compared with the improved and The components are further interconnected to enrich the multi-scale contextual representation of the joint representation. A gating mechanism is used to filter and reconstruct high-response semantic components, giving the output representation stronger discriminative power. Then, residual connections are introduced to ensure the stability of the coupled representations and alleviate gradient decay and semantic information degradation problems in deep training, thereby achieving deep coupling of multi-granularity features and obtaining the final feature representation. : in This indicates content-aware gating; Layer normalization normalizes the feature distribution, unifying features from different sources on the same numerical scale, thereby promoting subsequent stable fusion.

[0013] (4c) Content-aware gating aggregation Let the input features be ,in and These represent the number of spatial locations and the number of channels, respectively. First, a linear transformation operator is used... The input features are projected into the feature space and divided into gated branches along the channel dimension. With modulation To simultaneously accommodate fine-grained modeling and coarse-grained structure perception, It is divided into two complementary sub-branches along the channel dimension. Then, depthwise separable convolutions of different scales are used to expand the receptive field and capture the overall tissue structure layout of the lesion region from a near-global perspective, while focusing on fine-grained structures and texture boundaries, thus enhancing the perception of fine-grained gland morphology and coarse-grained tissue structure from a multi-granular perspective: in, This indicates a proportional grouping operation of the channels. Indicates the kernel size as Depth-separable convolutions are then used. Subsequently, to mitigate the impact of features of different granularities... and The effects of statistical bias and scale differences between them were analyzed using batch normalization. With Gaussian error linear unit The correlated features are standardized and nonlinearly transformed to mitigate statistical biases across different scales and introduce a smooth nonlinear response. Finally, through the gating mechanism This approach achieves dynamic feature selection and adaptive reconstruction, further modulates multi-granularity coupled features based on semantic relevance, and removes low-response background noise from the joint representation, thereby effectively reducing background noise interference and obtaining the final output. : in Represents the element-wise gate. This indicates element-wise multiplication. Representation layer normalization.

[0014] The method of the present invention has the following advantages: (1) Fine-grained perception enhances heterogeneous gland capture: Construct regular adaptive sampling convolution, and achieve sub-pixel level fine modeling of disordered gland structure through learnable offset generation, hyperbolic tangent amplitude constraint and continuous domain bilinear interpolation, thereby enhancing the model's ability to capture gland deformation, boundary distortion and local abnormal texture.

[0015] (2) Coarse-grained efficient modeling of organizational structure: Through subspace grouping, deep residuals and linear and approximate, the coarse-grained organizational structure is efficiently modeled in different subspaces with linear complexity and the category boundaries are amplified.

[0016] (3) Multi-granularity coupling reduces cross-granularity semantic conflicts: The discriminative feature weights are adaptively allocated through the dual recalibration mechanism of channel and spatial dimensions, and content-aware gating is combined to suppress low-response noise features, so as to achieve stable, compact and priority cross-granularity feature coupling, significantly improve the purity, effectiveness and global discriminative ability of the fusion representation and reduce semantic conflicts.

[0017] (4) Multi-granularity cascade perception improves diagnostic accuracy: Fine-grained sampling, coarse-grained structural modeling and multi-granularity feature fusion mechanism work together to enable the model to capture complex glandular microstructures and perceive macroscopic tissue layout, effectively improving the diagnostic accuracy of colorectal cancer. Attached Figure Description

[0018] Figure 1 This is the overall framework diagram of the present invention. Figure 2 This is a classification result image based on the confusion matrix on the Chaoyang dataset. Figure 3 This is a graph showing the feature dimensionality reduction and classification results on the Kather-5K dataset. Figure 4 This is a classification result of a confusion matrix on a colorectal cancer dataset. Detailed Implementation

[0019] 1. Data Acquisition and Non-overlapping Windowing Preprocessing Data Collection and Construction: Three types of WSL (Warning-Side Scale) slides of colorectal cancer tissue were collected from the hospital. These included cancerous, inflammatory, and normal tissue samples.

[0020] Slicing and Windowing: Due to the extremely high resolution of WSL images, directly inputting them into a deep network would be computationally too expensive. Therefore, the OpenSlide library is used to perform non-overlapping windowing at a magnification of 40x, dividing the entire slice into image blocks of fixed size.

[0021] Color normalization: Considering that the staining degree, color distribution and tissue preparation differences of different slices will significantly affect the model's generalization ability, this embodiment uses the Vahadane structure-preserving color normalization method to normalize all image blocks, so that slices from different sources are aligned with the reference template in terms of color distribution, thereby mitigating the inter-domain differences caused by staining bias and thus obtaining the input feature map.

[0022] 2. Constructing a fine-grained perspective-aware flow After downsampling the input feature map and feeding it into the fine-grained sensing stream, the following operations are performed sequentially: First, offset prediction is performed on the downsampled feature map, and the offset generation subnetwork outputs a two-dimensional offset field corresponding to the spatial location. Then, the original offset is amplitude-limited by learnable parameters and hyperbolic tangent mapping to prevent over-sampling. Subsequently, each sampling position of the convolutional kernel is adjusted according to the constrained offset and mapped to continuous domain coordinates. Bilinear interpolation is performed at these positions to obtain continuous domain feature responses, achieving sub-pixel-level modeling of gland boundaries, fracture regions, and deformed structures. Next, the responses obtained from continuous domain sampling are combined with the convolutional kernel weights to complete the spatial aggregation of fine-grained structural features. Finally, the spatially aggregated features are subjected to linear mapping, normalization, and residual fusion with nonlinear activation to robustly enhance fine-grained expression, alleviate deep training instability, and significantly improve the sensitivity modeling ability for differences in glandular microstructure and cell texture, thereby obtaining the most discriminative fine-grained semantic features.

[0023] 3. Constructing a coarse-grained perspective-aware flow After downsampling the input features, they are projected through linear channels and divided into multiple complementary subspaces along the channels. Each subspace focuses on coarse-grained tissue features such as gland morphology, cell nucleus distribution, or matrix texture to improve the separability of categories in the feature space. Subsequently, a deep residual path is introduced within each subspace to preserve key local differences in category discrimination and strengthen intra-class structural consistency. Next, attention computation is reconstructed through kernel function mapping, transforming global relation learning into a cross-regional aggregation with linear complexity, thereby efficiently capturing macroscopic tissue structures. Then, the global features output by linear attention are fused with the discriminative details preserved by the residual path to achieve co-modeling of local differences and macroscopic structures. Finally, the coarse-grained encoder stabilizes the feature distribution, maintains the original semantics, and strengthens the nonlinear expression of the channel dimension through a structure of double normalization, double residuals, and feedforward enhancement, thereby obtaining robust and discriminative coarse-grained structural features.

[0024] 4. Multi-granularity feature coupling After inputting the multi-source features from both fine-grained and coarse-grained streams into the multi-granularity feature coupling module, the following steps are executed sequentially: First, channel-level linear projection is performed on the multi-granularity joint features of the previous level, and spatial scale alignment is achieved through local window average pooling to extract the core semantic distribution. Then, global average pooling and max pooling are used to construct spatial context descriptions for fine-grained features, and a spatial attention module adaptively allocates weights at each location to enhance key region features. For coarse-grained features, channel-level global context aggregation is performed, and cross-channel dependencies are modeled through a channel attention module to highlight semantically significant channels. Next, the fine-grained, coarse-grained, and previous-level core semantic distributions are correlated along the channel dimension, fusing cross-scale and cross-granularity semantic information. Redundant features are compressed through channel dimensionality reduction and linear remapping to obtain compact and discriminative joint features. Subsequently, the joint features capture local boundary details and global structural layout through multi-scale deep convolution, and scale differences are eliminated by normalization and nonlinear activation, high-value semantics are filtered, and redundant and collinear features are suppressed. Finally, a gating mechanism is used to dynamically reconstruct the fused features channel by channel and position by position, which enhances the class discrimination ability and reduces background organization interference. At the same time, residual connections are used to ensure the stability of feature coupling and effectively alleviate gradient decay in deep networks, ultimately obtaining multi-granularity joint feature representations.

[0025] The effectiveness of this invention was further verified through the following experiments: The model achieved 88.27% and 98.20% accuracy on the publicly available colorectal cancer datasets Chaoyang and Kather-5K, respectively, and 96.99% accuracy on private colorectal cancer datasets, significantly outperforming comparative models (ResNet50, FasterViT, and HiFuse, etc.), highlighting its advantages in colorectal cancer diagnosis. Visualization results using Grad-CAM further demonstrate that the proposed colorectal cancer diagnostic model's discrimination criteria are highly consistent with pathological features across different tissue categories, validating the model's discriminative rationality and clinical interpretability. Module ablation experiments showed the independent and synergistic contributions of each module to the method's performance. Adding the proposed module to coarse-grained and fine-grained variants improved accuracy by 1.31% and 1.68%, respectively. The dynamically coupled module designed in multi-granularity ablation improved accuracy by 1.41%, validating the effectiveness of the proposed module and multi-granularity synergistic representation. Stage ablation experiments also demonstrated the necessity of multi-granularity fusion at each level. The above experiments further validated the diagnostic efficacy of this invention for colorectal cancer.

Claims

1. A colorectal cancer diagnostic method based on multi-granularity cascade sensing, characterized in that, The method includes the following steps: (1) Data acquisition and non-overlapping windowing preprocessing: Full-slice images of three types of colorectal cancer tissues (cancer, inflammation and normal) were collected from the hospital, non-overlapping windowing was performed, and the image blocks were normalized using a structure-preserving color normalization method to obtain feature maps for input. (2) Construct a fine-grained perception network, use regularized adaptive sampling convolution to obtain spatial adaptive features of deformable gland morphology, and combine it with the fine-grained perception module to enhance the channel-level semantics of the input feature map in order to extract global structural dependencies and class discrimination features at the cell and gland levels in the image; the steps are as follows: (2a) Regularized adaptive sampling convolution Offset magnitude prediction: An offset generation subnet was constructed. This subnet has an output channel of of K × K Convolutions are used to generate a corresponding two-dimensional offset for each sampling point; then, the input feature map is used. transported to The original two-dimensional offset field is obtained. : Where H′ and W′ are the spatial dimensions of the offset map; Offset amplitude regularization mechanism: A learnable scalar α is used, and a positive constraint upper bound is defined based on this. m To dynamically adjust the sampling radius for different glandular structures: in, It is a smooth activation function; subsequently, it is applied through a hyperbolic tangent mapping. Original offset Mapping to a bounded interval to ensure the geometric stability of the offset yields the actual sampling offset. : in Ensure the offset is continuous Shape changes cause the offset of adjacent pixels to have Continuity, in theory, avoids the explosion of offset gradients; Continuous domain sampling reconstruction: reconstructing each relative sampling position Projected onto continuous domain coordinates : And for each channel Bilinear interpolation is used to approximate the sampling, thereby obtaining the continuous domain characteristic response. : in, for The set of nearest-neighbor integer points; continuous domain features obtained based on bilinear interpolation. , get location Central Channel Sampling response : in For the convolution kernel at the relative sampling position The weights are assigned; ultimately, the responses of all positions and channels form a complete output feature map: (2b) Fine-grained sensing module First, the input features are processed using a regularized adaptive sampling convolution. Perform spatial adaptive sampling to capture deformed glandular structures, and obtain : in, This represents a regularized adaptive sampling convolution; subsequently, A linear mapping is performed in the channel dimension, and batch normalization is applied. Standardize the feature distribution to mitigate internal covariate bias and improve training stability; then, use a non-linear activation function. This is introduced to enhance the nonlinear expressive power of feature mapping, thereby capturing more complex inter-channel interactions. Finally, residual connections are used to achieve an identity mapping between input and output features, effectively alleviating degradation and gradient vanishing problems in deep structures, and obtaining the final module output. : in This represents pointwise convolution, used to stabilize feature distribution and accelerate convergence; This represents a nonlinear activation function, which introduces nonlinear transformation capabilities into the feature space of the module, thereby improving the expressive flexibility and discriminative power in the channel dimension; (3) Construct a coarse-grained sensing network. Extract local microstructure features from the input feature map using a group residual linear encoder and a coarse-grained sensing module to establish coarse-grained structural features that reflect tissue morphology. The steps are as follows: (3a) Group of residual linear encoders Subspace grouping: Let the input feature map be... ,in for First, a linear mapping of query, key, and value is obtained through channel projection operations. and along the channel dimension Grouped into G independent complementary subspaces, in different subspaces The subspace vectors focus on learning features such as gland morphology, cell nuclear texture, or matrix structure, thereby enhancing the separability of each category in the feature space. ,in ; Deep residual paths: Designing deep residual paths preserves the detailed differences between categories, thereby enhancing the consistency of category features and obtaining discriminative details. : Linear kernel mapping: Designing kernel function mappings , will the Subspace Internal query AND key The attention mechanism is achieved by projecting the data onto the kernel feature space, thereby realizing an approximately linear attention mechanism. The mapping function is expressed as: ;pass The explicit similarity calculation in traditional attention is transformed into a linear operation within the kernel feature space, thus reconstructing the attention mechanism into a linear form, reducing the complexity from... Reduce to , obtain attention output : Coarse-grained feature reconstruction: Reconstructed into a two-dimensional feature map form And by using depth residuals to determine details With global reconstruction features The components are interconnected, and finally mapped to obtain the final output. : in, This represents pointwise convolution; (3b) Coarse-grained sensing module An integrated layer normalization and dual residual structure construct a coarse-grained encoder. First, the input features are layer normalized and then fed into a group residual linear encoder module to model the global dependencies of the coarse-grained organizational structure with linear complexity. Then, residual connections are used to preserve the original semantic information and stabilize gradient propagation. in This indicates a group of residual linear encoders. The representation layer is normalized; then, X is normalized twice and then nonlinearly enhanced in the channel dimension through a feedforward network to obtain a more discriminative semantic representation, thus obtaining the final output. (4) Construct a multi-granularity cascaded feature fusion mechanism. Through cascaded feature refinement, feature recalibration and content-aware gating aggregation, cross-layer and cross-granularity semantic coupling of fine-grained features and coarse-grained features is achieved. The fused features are dynamically selected and modulated to obtain the final results for colorectal cancer diagnosis.

2. The colorectal cancer diagnosis method based on multi-granularity cascade sensing according to claim 1, characterized in that... The construction of the multi-granularity cascaded feature fusion mechanism described in step (4) achieves cross-layer and cross-granularity semantic coupling of fine-grained features and coarse-grained features through cascaded feature refinement, feature recalibration, and content-aware gating aggregation. It also dynamically selects and modulates the fused features to obtain the final result for colorectal cancer diagnosis. The steps are as follows: (4a) Refinement of cascaded features The multi-granularity cascaded feature fusion mechanism adds prior knowledge from the previous level to each fusion module in a cascaded manner, and adds the output of the previous stage to the next multi-granularity cascaded feature fusion module in turn to further refine the features. (4b) Feature recalibration unit Feature alignment strategy: First, align the features from the previous stage... joint priors The core semantic distribution is obtained by using channel linear projection and then spatial downsampling with windowed average pooling. : in This is a spatially position-by-spatial linear projection operator that performs linear channel recombination for each spatial position while preserving the spatial topology. Indicates window average pooling; Fine-grained spatial recalibration: using global average pooling and global max pooling along the channel dimension for the aggregation stage. Spatial information of fine-grained feature maps This generates two different context descriptors. and ; Then, these two descriptors are correlated and the weight of each spatial location is obtained by combining the spatial context. : in Indicates global average pooling. Indicates global max pooling; This indicates a feature map concatenation operation. for activation, Spatial convolutions with large receptive fields provide a wider receptive field to aggregate neighborhood context; finally, utilizing... The product is based on spatial location weights. right The corresponding spatial location is recalibrated, and the following is inferred: : in, This represents element-wise multiplication; Coarse-grained feature recalibration: Employing global average pooling and global max pooling convergence stages in the spatial dimension. Global context information describing the channel Then, utilize the shared fully connected layer. Adaptive learning of channel position weights : in, Indicates global average pooling. Indicates global max pooling; This represents element-wise addition; finally, the features are analyzed by element-wise multiplication. Adaptive recalibration is performed on the significant channels to obtain : Multi-granularity feature coupling: the previous stage Core semantic distribution With multi-granularity features and Interconnected along the channel dimension to integrate semantic information across layers and granularities, aiming to construct multi-level semantic associations in the feature space; then, channel dimensionality reduction and linear remapping are used to compress redundant dimensions and enhance the coupling between features, thereby obtaining a more compact and discriminative joint representation. : in, Indicates size is Pointwise convolution; finally, the joint representations are... Compared with the improved and The components are further interconnected to enrich the multi-scale contextual representation of the joint representation. A gating mechanism is used to filter and reconstruct high-response semantic components, giving the output representation stronger discriminative power. Then, residual connections are introduced to ensure the stability of the coupled representations and alleviate gradient decay and semantic information degradation problems in deep training, thereby achieving deep coupling of multi-granularity features and obtaining the final feature representation. : in This indicates content-aware gating; For layer normalization, the feature distribution is normalized so that features from different sources are unified on the same numerical scale, thereby promoting subsequent stable fusion. (4c) Content-aware gating aggregation Let the input features be ,in and Representing the number of spatial locations and the number of channels respectively; firstly, through linear transformation operators... The input features are projected into the feature space and divided into gated branches along the channel dimension. With modulation To simultaneously accommodate fine-grained modeling and coarse-grained structure perception, It is divided into two complementary sub-branches along the channel dimension. Then, depthwise separable convolutions of different scales are used to expand the receptive field and capture the overall tissue structure layout of the lesion area from a near-global perspective, while focusing on fine-grained structures and texture boundaries, thereby enhancing the perception of fine-grained gland morphology and coarse-grained tissue structure from a multi-granular perspective: in, This indicates a proportional grouping operation of the channels; Indicates the kernel size as Depth-separable convolutions are then used; subsequently, to mitigate the impact of features of different granularities... and The effects of statistical bias and scale differences between them were analyzed using batch normalization. With Gaussian error linear unit The correlated features are standardized and nonlinearly transformed to mitigate statistical biases across different scales and introduce a smooth nonlinear response. Finally, through the gating mechanism This approach achieves dynamic feature selection and adaptive reconstruction, further modulates multi-granularity coupled features based on semantic relevance, and removes low-response background noise from the joint representation, thereby effectively reducing background noise interference and obtaining the final output. : in Represents the element-wise gate. This indicates element-wise multiplication. Representation layer normalization.

Images