Prostate cancer image classification method based on axial consistency enhancement and threshold guide selection

Abstract

The invention discloses a prostate cancer image classification method based on axial consistency enhancement and threshold guide selection, and the method comprises the steps: designing an axial consistency context enhancement module ACE, combining strip pooling with axial self-attention, and enhancing the correlation expression between different spatial positions in a feature map through explicit modeling cross-row / cross-column long-distance structural dependence; a threshold guide semantic selection module TGSS is proposed to fuse horizontal, vertical and channel three-view weight information, differential processing of foreground focus enhancement and background tissue reconstruction is realized through hard threshold segmentation, and morphological gradient modeling is performed on a boundary transition region. Compared with the prior art, the method can further improve the precision of the algorithm model on related tasks of prostate cancer image classification.

Description

Technical Field

[0001] This invention relates to the field of accurate classification of prostate cancer images, specifically to a prostate cancer image classification method that enhances axial consistency and uses threshold-guided selection. Background Technology

[0002] Prostate cancer, one of the most common malignant tumors in men worldwide, is experiencing a continuous rise in incidence globally. With an aging population and changing lifestyles, the disease burden of prostate cancer is expected to increase further in the coming decades, making early diagnosis and accurate grading crucial for improving patient prognosis. While deep learning technology has achieved initial success in prostate cancer image diagnosis in recent years, compared to mature applications in other medical imaging fields, the boundary between the prostate and normal tissue is often unclear, especially in the peripheral zone. Conventional image recognition methods, such as boundary modeling, primarily focus on the individual extraction and processing of boundary information. Although significant progress has been made in boundary localization and detail preservation, the lack of comprehensive consideration of the synergistic effects between the foreground, background, and boundary regions leads to misjudgments of boundary regions and incomplete representation of target features, making accurate lesion identification difficult and further increasing the challenge of precise segmentation. Summary of the Invention

[0003] Purpose of the invention: To address the problems mentioned in the background art, this invention discloses a prostate cancer image classification method with enhanced axial consistency and threshold-guided selection. By designing the ACE module and TGSS module in synergy, it achieves unified optimization of foreground, background, and boundary regions, solves the classification uncertainty of the boundary fuzzy region unique to prostate cancer image recognition, and achieves accurate identification of key regions.

[0004] Technical solution:

[0005] This invention discloses a prostate cancer image classification method with enhanced axial consistency and threshold-guided selection, the method comprising the following steps:

[0006] S1: Obtain the prostate cancer image dataset to be processed, and divide the dataset into training set and test set respectively;

[0007] S2: Constructing a prostate cancer image classification model:

[0008] S2.1: Extract deep features from prostate cancer images using backbone network features;

[0009] S2.2: The Deep Feature Input Axial Consistency Context Enhancement Module (ACE) explicitly aggregates long-range structural cues through strip guidance and axial self-attention, and introduces relative positional bias to enhance output features;

[0010] S2.3: Enhanced feature input threshold guided semantic selection module TGSS, introduces morphological boundary refinement path to stabilize transition region response, and fuses with the original feature in a residual manner;

[0011] S3: Use the S1 dataset for training and testing, and use the final model for prostate cancer image classification.

[0012] Furthermore, in the ACE module, the explicit aggregation of long-range structural cues through strip guidance and axial self-attention is specifically as follows:

[0013] For the input deep feature map X, perform strip pooling operations along the horizontal and vertical directions respectively to obtain the horizontal average stripes. Horizontal maximum band Vertical average strip With the vertical maximum strip , where SP represents strip pooling operation, and C, H, and W are the number of channels, height, and width of the input feature map, respectively;

[0014] The average and maximum horizontal stripes are element-wise added and fused to obtain the horizontal axial feature representation S. h (X); Element-wise addition and fusion are performed on the average and maximum stripes in the vertical direction to obtain the vertical axial feature representation Sᵥ(X), where S h (X)∈R^C×1×W, Sᵥ(X)∈R^C×H×1.

[0015] Furthermore, the horizontal axial feature represents S h Linear projections are applied to both (X) and the vertical axis feature representation Sᵥ(X) to generate a horizontal query matrix Q. h Key matrix K h Value matrix V h And the vertical query matrix Qᵥ, key matrix Kᵥ, and value matrix Vᵥ, which are expanded into a multi-head structure, where the number of attention heads is H. a The dimension of each head of the query / key is d. k The value is dᵥ for each head dimension.

[0016] Furthermore, in the ACE module, the specific steps for introducing the relative position bias output enhancement feature are as follows:

[0017] Introduce the relative position offset of ALiBi into each attention head, and calculate the horizontal offset matrix respectively. and vertical offset matrix The bias value is determined by the slope parameter. Decide;

[0018] Calculate the horizontal self-attention using the bias matrix. and vertical self-attention And obtain horizontal attention output. and vertical attention output ⁾;

[0019] The outputs from each head are spliced ​​together and the strip shape is restored through linear projection to obtain the horizontal projection output. and vertical projection output Then, through upsampling operation, and Restore to the original feature map space size.

[0020] Furthermore, the Axial Consistency Context Enhancement (ACE) module also includes a feature enhancement module, with the following specific structure:

[0021] After upsampling and Perform spatial-level fusion to obtain the spatial context field U∈R^C×H×W;

[0022] The spatial context field U is sequentially subjected to 3×3 convolutional local integration, normalization and nonlinear combination, 1×1 convolutional channel compression, and then pixel-wise attention weights A are generated through the Sigmoid function;

[0023] The input feature map X is multiplied pointwise with the attention weights A, then residually connected to the deep feature map X, and finally activated by ReLU to obtain the context-enhanced feature map. .

[0024] Furthermore, the threshold-guided semantic selection module TGSS has the following structure:

[0025] The input features are sequentially processed through three-view weight map generation and hard threshold segmentation, foreground-background dual-path parallel modeling, and morphological boundary refinement path processing. Finally, a complete feature representation is output through weighted fusion, achieving effective separation of foreground and background and targeted optimization of blurred boundary regions.

[0026] Furthermore, the three-view weighted graph generation and hard thresholding segmentation specifically include:

[0027] Weight maps in the horizontal, vertical, and channel directions are generated by 1×1 convolution, and the complementarity is ensured by softmax normalization of the corresponding dimensions. After averaging the weights of the three views, they are mapped to the range of [0,1] by spatial min-max calibration per sample and per channel. A threshold is set to perform hard threshold segmentation on the calibrated weights to generate a binary weight matrix M for distinguishing the foreground and background.

[0028] Furthermore, the foreground-background dual-path parallel modeling includes: processing the input features in parallel using a foreground enhancement path and a background reconstruction path, both paths being Conv3×3→BN→ReLU structures; performing point-by-point multiplication selection on the outputs of the two paths using a binary weight matrix M and its complement 1-M respectively, and then adding the residuals with the input features to obtain the preliminary fused feature Y.

[0029] Furthermore, the introduced morphological boundary refinement path includes:

[0030] Boundary regions are identified through the boundary refinement path U-path, and boundary weights are obtained. Multi-scale boundary features are extracted using 5×5 depth convolution and 3×3 depth convolution with an inflation rate of 2. Boundary optimization features are obtained through feature fusion and bottleneck structure refinement. The boundary optimization features are then weighted and fused with the preliminary fusion features Y to output the final feature representation Z. The weight coefficients of the boundary paths are adjustable.

[0031] Beneficial effects:

[0032] 1. This invention designs an Axial Consistency Context Enhancement (ACE) module, which combines strip pooling with axial self-attention. By explicitly modeling long-distance structural dependencies across rows / columns, it enhances the association expression between different spatial locations in the feature map, making the boundary between prostate cancer lesions and normal tissue clearer and improving classification accuracy.

[0033] 2. This invention designs a Threshold-Guided Semantic Selection (TGSS) module that integrates horizontal, vertical, and channel-based weight information. Through hard threshold segmentation, it achieves differentiated processing of foreground lesion enhancement and background tissue reconstruction. It also introduces dual-path parallel modeling and morphological boundary refinement path processing, significantly concentrating feature activation on key lesion regions, thus achieving accurate semantic selection and foreground-background separation. Notably, the model exhibits good discriminative ability on benign samples, avoiding erroneous high activation responses, while accurately locating regions with significant lesion features on malignant samples. Furthermore, it shows corresponding changes in focus intensity as the pathological grade increases. Attached Figure Description

[0034] Figure 1 This is an overall flowchart of the present invention;

[0035] Figure 2 This is a diagram illustrating the overall architecture of the prostate cancer image classification model of the present invention.

[0036] Figure 3 A schematic diagram of the axial consistency context enhancement module designed for this invention;

[0037] Figure 4 A schematic diagram of the threshold-guided semantic selection module designed for this invention;

[0038] Figure 5 This is a schematic diagram of the attention visualization results of the prostate cancer image classification model of the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] like Figure 1 As shown, this invention discloses a prostate cancer image classification method with enhanced axial consistency and threshold-guided selection. The method steps are as follows:

[0041] Step 1: Divide the private HY Prostate dataset and the public SICAPv2 dataset into training and test sets, respectively.

[0042] Step 1.1: HY Prostate Dataset: This dataset was constructed by the First People's Hospital of Huai'an City, Jiangsu Province. The dataset consists of CT images of the prostate from 888 different patients, totaling 888 images. Among them are 431 images of benign prostatic hyperplasia and 457 images of prostate cancer. (2) SICAPv2 Dataset: A public dataset, which selected 12081 prostate tissue pathology images according to the official documentation, including 4417 NC (no cancer), 2222 GG3 (Gleason level 3), 4494 GG4 (Gleason level 4), and 948 GG5 (Gleason level 5);

[0043] Step 1.2: HY Prostate Dataset: Divided in a 4:1 ratio, the training set contains 709 images, of which 344 are prostate hyperplasia and 365 are prostate cancer. The test set contains 179 images, of which 87 are prostate hyperplasia and 92 are prostate cancer. (2) SICAPv2 Dataset: Divided according to the official documentation, the ratio is also approximately 4:1. The training set contains 9959 images, of which 3773 are NC, 1829 are GG3, 3641 are GG4 and 716 are GG5. The test set contains 2122 images, of which 644 are NC, 393 are GG3, 853 are GG4 and 232 are GG5, of which the label for NC is 0, the label for GG3 is 1, the label for GG4 is 2 and the label for GG5 is 3.

[0044] Step 1.3: Perform scale adjustment only on the samples in the multiple datasets, uniformly modifying all samples to 256*256.

[0045] Step 2: As Figure 3 As shown, deep features are extracted from prostate cancer images using a backbone network. An Axial Consistency Context Enhancement (ACE) module is designed, employing strip pooling to obtain horizontal and vertical feature representations. Axial self-attention is used to establish long-distance associations across rows / columns, maintaining consistent feature responses and strengthening structural correlations. This effectively alleviates the problems of insufficient correlation between distant regions and interference from local similarity. The specific structure of the module is as follows:

[0046] First, the input deep feature map By performing axial strip pooling operations to average and maximize convergence in the horizontal and vertical directions respectively, four strip feature maps are obtained, including the horizontally averaged strip. Horizontal maximum band Vertical average strip With the vertical maximum strip For example, the following formula:

[0047] ;

[0048] ;

[0049] SP represents strip pooling operation. , , These represent the number of channels, height, and width of the input feature map, respectively. For the first Channel 1 Line number The characteristic values ​​of the column.

[0050] For the four strip features obtained, the average stripe will be calculated within each axis. With the largest strip The process involves pairwise fusion, which combines complementary information from "overall trend + significant response" across two axes to form a single and stable axial feature representation, as shown in the following formula:

[0051] ;

[0052] in Indicates element-wise addition; horizontal fusion Preserve column orientation (width) (sequence, vertical fusion) Preserve line direction (height) ) sequence.

[0053] To characterize long-range relationships across columns / rows while preserving strip priors, a linear projection is applied to the axial representation and expanded into a multi-head structure, forming a sequential attention along the axial direction. Let the number of attention heads be . The head dimension of the query / key is The value of each head dimension is For example, the following formula:

[0054] ;

[0055] in, , , (Expand along width W) , , (Expand along height H)

[0056] Next, an ALiBi relative position bias is introduced into each attention head to characterize the positional difference with a monotonically linear bias, ensuring stable response under inputs of different lengths and resolutions. The biases of the m-th head in the horizontal and vertical directions are calculated, as shown in the following formula:

[0057] ;

[0058] in This is the slope parameter of the head. Representation matrix express The element in the i-th row and i′-th column.

[0059] To explicitly bind structural dependencies at distant locations within column and row sequences, strengthen long-distance relationships with consistent directions, and suppress confusion caused by "local similarity," self-attention calculations are performed along both axes to obtain the two-axis attention and axial output. and For example, the following formula:

[0060] ;

[0061] in , and This represents a query, key, and value matrix in the horizontal direction. , and This represents a query, key, and value matrix in the vertical direction. and These are attention outputs in the horizontal and vertical directions, respectively.

[0062] Then, the individual ends are stitched together and projected back to the channel dimension to restore the strip shape. See the following formula:

[0063] ;

[0064] in It is a linear projection of the strip output. It is a strip output after horizontal splicing and projection, with the shape of... , It is a strip output after vertical splicing and projection, with the shape of... .

[0065] Next, the two-axis outputs are upsampled to their original spatial dimensions and then spatially fused, taking into account both column-to-column and row-to-row consistency relationships to form a spatial context field U with complementary directions. See the following formula:

[0066] ;

[0067] in This refers to the upsampling operation.

[0068] Subsequently through 3 3. Convolutional Local Integration and 1 1. Convolutional compression of channels generates pixel-wise attention weights A. As shown in the following formula:

[0069] ;

[0070] in This represents a combination of normalization and nonlinearity. It is Sigmoid.

[0071] Finally, the input feature maps are weighted and residuals are used to obtain the context enhancement result. For example, the following formula:

[0072] ;

[0073] in This indicates point-by-point multiplication. This is the activation function.

[0074] Step 3: As Figure 4 As shown, a Threshold-Guided Semantic Selection (TGSS) module is proposed. It integrates horizontal, vertical, and channel-based weight information, and uses hard thresholding segmentation to achieve differentiated processing of foreground lesion enhancement and background tissue reconstruction, while also performing morphological gradient modeling on the boundary transition region. The specific content of the module is as follows:

[0075] First, the input features A three-view weight map is generated to obtain a weight representation that is complementary to the axial and channel weights. , and These are the horizontal, vertical, and channel-oriented weighted graphs, respectively. See the following formula:

[0076] ;

[0077] Where * denotes a 1×1 convolution mapping, These represent the normalized competition along the three dimensions of height, width, and channel, ensuring the complementarity of the axis and channel.

[0078] Subsequently, the weights from the three perspectives are averaged and spatially min-max calibrated to map the weights to the range [0,1]. As shown in the following formula:

[0079] ;

[0080] in, This indicates a sample-by-sample, channel-by-channel spatial min-max calibration operation used to eliminate differences in numerical ranges.

[0081] Next, set the threshold. Hard thresholding is applied to the calibrated weights. If the average weight of a feature point is less than the threshold, its weight is replaced with 0; otherwise, it is replaced with 1, generating a binary weight matrix. For example, the following formula:

[0082] ;

[0083] Where c, These represent the coordinates of the channel, length, and width in the feature map. Hard thresholding, rather than soft weighting, is used to ensure clear and interpretable foreground / background segmentation.

[0084] Guided by weights, the module independently models the foreground and background regions. To fully utilize background information rather than simply discarding it, a dual-path parallel design is adopted to process the input features. Simultaneously, the data is processed through foreground enhancement and background reconstruction paths, then selectively fused using binary weights and added to the input residual. See the following formula:

[0085] ;

[0086] in, and These represent foreground enhancement and background reconstruction operations, respectively (both are Conv3×3→BN→ReLU structures). This represents pointwise multiplication, with weights. Select the foreground area. Then select the background area.

[0087] To further address the blurred region at the foreground-background boundary, the module introduces a boundary refinement path (U-path). This path specifically models the uncertainty of the transition region, identifying the boundary area through morphological gradient operations. See the following formula:

[0088] ;

[0089] in, The channel mean with binary weights. The morphological pooling kernel size (default is 3) is shown in the figure. For example.

[0090] After obtaining the boundary weights, multi-scale depthwise convolution is used for feature extraction, and the features are refined using a bottleneck structure. See the following formula:

[0091] ;

[0092] in, and These represent 5×5 depthwise convolution and 3×3 depthwise convolution with a dilation rate of 2, respectively, extracting complementary boundary features through different convolution methods. This is a feature fusion operation (1×1 convolution + BatchNorm + GELU). For the bottleneck structure, 1×1 convolution is used to implement dimensionality reduction and dimensionality increase operations (the number of channels is first compressed to 1 / 4 and then restored) for feature refinement.

[0093] Finally, a complete feature representation is obtained through weighted fusion. For example, the following formula:

[0094] ;

[0095] in, This is the weighting coefficient for the boundary path (default is 0.2), used to balance the contribution of boundary refinement and main features. Through this design, the TGSS module not only achieves effective separation of foreground and background, but also performs targeted optimization for blurred boundary regions, improving the completeness and discriminativeness of feature representation.

[0096] Step 4: As Figure 2 As shown, all proposed modules are integrated to construct a complete algorithm model. In the blue dashed section, the ACE strategy is used to combine strip pooling with axial self-attention, explicitly modeling long-distance structural dependencies across rows / columns to enhance the correlation between different spatial locations in the feature map. In the green dashed section, the TGSS strategy is used to fuse horizontal, vertical, and channel weight information, achieving differentiated processing of foreground lesion enhancement and background tissue reconstruction through hard thresholding, and performing morphological gradient modeling on the boundary transition region. Specifically:

[0097] Step 4.1: Perform scale adjustment only on the samples in multiple datasets, uniformly modifying all samples to 256*256;

[0098] Step 4.2: By establishing axial priors with consistent orientation, then explicitly pulling together long-range dependencies within the axis, and maintaining multi-scale stability with length-independent relative biases, the input feature maps at different stages are adapted; after spatial alignment and pixel-wise attention-weighted residual fusion, the axial context is injected into the original representation in a fine-grained manner, which both strengthens the consistency of long-range structures and avoids destroying the original distribution.

[0099] Step 4.3: Threshold-guided semantic selection and boundary refinement of deep features highlight discriminative regions and stabilize the response of the transition zone between the inside and outside of the gland. TGSS achieves effective separation of foreground and background and also performs targeted optimization for blurred boundary regions, improving the completeness and discriminability of feature expression.

[0100] Step 5: The proposed algorithm model will be compared with the current state-of-the-art classification algorithm models, and a series of ablation experiments and attention visualizations will be conducted to verify the feasibility and superiority of the proposed algorithm.

[0101] The proposed algorithm was compared with several state-of-the-art (SOTA) models on two prostate cancer datasets. The classification results of each model on the HY Prostate dataset and the SICAPv2 dataset are shown in Tables 1 and 2, respectively.

[0102] Table 1

[0103]

[0104] Table 2

[0105]

[0106] The algorithm model proposed in this embodiment of the invention has a reasonable construction for both the axial consistency context enhancement module and the threshold-guided semantic selection module, thus achieving the best prediction results on both datasets.

[0107] Attention visualization:

[0108] like Figure 5 As shown, where Figure 5 (a) The column shows the original image; (b) The column shows the visualization results after using the ACE strategy module; (c) The column shows the visualization results after using the TGSS strategy module. Figure 5The first row in the image shows benign prostatic hyperplasia (BPH) samples from the HY Prostate dataset; the second row shows prostate cancer (PC) samples from the HY Prostate dataset; the third row shows G3 PC cancer samples from the SICAPv2 dataset; the fourth row shows G4 PC cancer samples from the SICAPv2 dataset; and the fifth row shows G5 PC cancer samples from the SICAPv2 dataset. The visualization results show that the ACE module effectively establishes long-distance spatial dependencies, enhancing the correlation between different regions while maintaining broad feature activation, providing rich contextual information for subsequent semantic selection. After processing by the TGSS module, feature activation is significantly concentrated in key lesion areas, achieving accurate semantic selection and foreground / background separation. Notably, the algorithm exhibits good discriminative ability on benign samples, avoiding erroneous high activation responses, while accurately locating regions with significant lesion features on malignant samples, and showing corresponding changes in focus intensity as the pathological grade increases. Visual analysis verifies the synergistic effect of the ACE and TGSS modules, demonstrating that the proposed algorithm can effectively capture key features in prostate cancer images, achieving a feature optimization process from coarse-grained context enhancement to refined semantic selection, providing reliable technical support for accurate disease diagnosis.

[0109] Ablation experiment:

[0110] (1) The impact of two sets of thresholds in the ACE and TGSS modules on the algorithm. By setting different thresholds, feature points are selected for inclusion or exclusion, seeking a balance between retaining important feature information and filtering noise interference. The thresholds are set in units of 0.1 within the range of 0 to 1. The specific results are shown in Tables 3 and 4:

[0111] Table 3

[0112]

[0113] Table 4

[0114]

[0115] As shown in Tables 3 and 4, the threshold setting has a significant impact on the performance of the TGSS module. On both datasets, the optimal accuracy is achieved at T=0.3. When the threshold is too low (T=0.1, T=0.2), although more feature points are retained and the foreground region is expanded, a large amount of noisy features are introduced, resulting in insufficient distinction between the foreground and background, thus affecting model performance. When the threshold is too high (T=0.5, T=0.6, T=0.7), although noise interference can be effectively filtered, excessive feature selection will also lose important discriminative information, narrowing the foreground range and weakening the effect of the foreground enhancement path, also leading to performance degradation. This phenomenon is consistent on both datasets, indicating that T=0.3 can find the optimal balance between retaining important feature information and filtering noise interference, allowing foreground enhancement, background reconstruction, and boundary refinement paths to obtain appropriate feature allocation, thereby achieving optimal classification performance.

[0116] The impact of the ACE and TGSS modules on the classification accuracy of the model. Each module was evaluated individually. Using DenseNet121 as a baseline, the ACE module was added after the 3rd and 4th dense blocks, and the TGSS module was added before the classification layer. Specific results are shown in Tables 5 and 6.

[0117] Table 5

[0118]

[0119] Table 6

[0120]

[0121] The experimental results shown in Tables 5 and 6 demonstrate that adding either the ACE module or the TGSS module alone significantly improves the classification accuracy of the DenseNet baseline model, validating the independent contribution of each module. When the two modules are used together, the model performance is further improved, indicating a good synergistic effect between the two modules. Notably, on the SICAPv2 dataset, which faces the challenge of relatively small inter-class differences, the proposed module shows a more significant improvement in the performance of the baseline model, fully demonstrating the good performance of the algorithm in handling fine-grained classification tasks.

[0122] The impact of the two strategies on multiple backbone networks. To further verify the feasibility of the proposed strategies, we applied them to DenseNet, as well as to two other backbone models, AlexNet and ResNet. Specific results are shown in Tables 7 and 8:

[0123] Table 7

[0124]

[0125] Table 8

[0126]

[0127] As shown in Tables 7 and 8, all three backbone networks achieved improved accuracy after adding the proposed module, with DenseNet showing the most significant improvement. This may be related to its densely connected structure, which is more conducive to feature reuse and information transfer. Furthermore, consistent improvement was observed on two different modality datasets, further validating the module's stability. The results demonstrate that the constructed ACE and TGSS modules possess good versatility and can be effectively combined with deep learning models of different architectures, proving the feasibility and practical value of the proposed module.

[0128] This invention can be combined with computer-aided medical diagnostic systems to provide more accurate diagnostic results and assist doctors in making more informed decisions. This combination not only reduces the possibility of misdiagnosis but also optimizes the allocation of medical resources and improves the overall quality of medical services.

[0129] The axial consistency enhancement and threshold-guided selection prostate cancer image classification method of the present invention can also be set in a computer program. When the computer program is loaded into the processor, the above-mentioned axial consistency enhancement and threshold-guided selection prostate cancer image classification method is implemented.

[0130] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent transformations or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A prostate cancer image classification method with enhanced axial consistency and threshold-guided selection, characterized in that, The method includes the following steps: S1: Obtain the prostate cancer image dataset to be processed, and divide the dataset into training set and test set respectively; S2: Constructing a prostate cancer image classification model: S2.1: Extract deep features from prostate cancer images using backbone network features; S2.2: The Deep Feature Input Axial Consistency Context Enhancement Module (ACE) explicitly aggregates long-range structural cues through strip guidance and axial self-attention, and introduces relative positional bias to enhance output features; S2.3: Enhanced feature input threshold guided semantic selection module TGSS, introduces morphological boundary refinement path to stabilize transition region response, and fuses with the original feature in a residual manner; S3: Use the S1 dataset for training and testing, and use the final model for prostate cancer image classification.

2. The prostate cancer image classification method with enhanced axial consistency and threshold-guided selection according to claim 1, characterized in that, In the ACE module, the explicit aggregation of long-range structural cues through strip guidance and axial self-attention is specifically as follows: For the input deep feature map X, perform strip pooling operations along the horizontal and vertical directions respectively to obtain the horizontal average stripes. Horizontal maximum band Vertical average strip With the vertical maximum strip , where SP represents strip pooling operation, and C, H, and W are the number of channels, height, and width of the input feature map, respectively; The average and maximum horizontal stripes are element-wise added and fused to obtain the horizontal axial feature representation S. h (X); Element-wise addition and fusion are performed on the average and maximum stripes in the vertical direction to obtain the vertical axial feature representation Sᵥ(X), where S h (X)∈R^C×1×W, Sᵥ(X)∈R^C×H×1.

3. The prostate cancer image classification method with enhanced axial consistency and threshold-guided selection according to claim 2, characterized in that, The horizontal axial feature represents S h Linear projections are applied to both (X) and the vertical axis feature representation Sᵥ(X) to generate a horizontal query matrix Q. h Key matrix K h Value matrix V h And the vertical query matrix Qᵥ, key matrix Kᵥ, and value matrix Vᵥ, which are expanded into a multi-head structure, where the number of attention heads is H. a The dimension of each head of the query / key is d. k The value is dᵥ for each head dimension.

4. The prostate cancer image classification method with enhanced axial consistency and threshold-guided selection according to claim 3, characterized in that, In the ACE module, the specific steps for introducing relative position bias output enhancement features are as follows: Introduce the relative position offset of ALiBi into each attention head, and calculate the horizontal offset matrix respectively. and vertical offset matrix The bias value is determined by the slope parameter. Decide; Calculate the horizontal self-attention using the bias matrix. and vertical self-attention And obtain horizontal attention output. and vertical attention output ; The outputs from each head are spliced ​​together and the strip shape is restored through linear projection to obtain the horizontal projection output. and vertical projection output Then, through upsampling operation, and Restore to the original feature map space size.

5. The prostate cancer image classification method with enhanced axial consistency and threshold-guided selection according to claim 4, characterized in that, The Axial Consistency Context Enhancement (ACE) module also includes a feature enhancement module, the specific structure of which is as follows: After upsampling and Perform spatial-level fusion to obtain the spatial context field U∈R^C×H×W; The spatial context field U is sequentially subjected to 3×3 convolutional local integration, normalization and nonlinear combination, 1×1 convolutional channel compression, and then pixel-wise attention weights A are generated through the Sigmoid function; The input feature map X is multiplied pointwise with the attention weights A, then residually connected to the deep feature map X, and finally activated by ReLU to obtain the context-enhanced feature map. .

6. The prostate cancer image classification method with enhanced axial consistency and threshold-guided selection according to claim 1, characterized in that, The threshold-guided semantic selection module TGSS has the following structure: The input features are sequentially processed through three-view weight map generation and hard threshold segmentation, foreground-background dual-path parallel modeling, and morphological boundary refinement path processing. Finally, a complete feature representation is output through weighted fusion, achieving effective separation of foreground and background and targeted optimization of blurred boundary regions.

7. The prostate cancer image classification method with enhanced axial consistency and threshold-guided selection according to claim 6, characterized in that, The three-view weighted graph generation and hard threshold segmentation specifically include: Weight maps in the horizontal, vertical, and channel directions are generated by 1×1 convolution, and the complementarity is ensured by softmax normalization of the corresponding dimensions. After averaging the weights of the three views, they are mapped to the range of [0,1] by spatial min-max calibration per sample and per channel. A threshold is set to perform hard threshold segmentation on the calibrated weights to generate a binary weight matrix M for distinguishing the foreground and background.

8. The prostate cancer image classification method with enhanced axial consistency and threshold-guided selection according to claim 7, characterized in that, The foreground-background dual-path parallel modeling includes: processing input features in parallel using a foreground enhancement path and a background reconstruction path, both of which are Conv3×3→BN→ReLU structures; selecting the outputs of the two paths by performing point-by-point multiplication using a binary weight matrix M and its complement 1-M respectively, and then adding the residuals with the input features to obtain the preliminary fused feature Y.

9. The prostate cancer image classification method with enhanced axial consistency and threshold-guided selection according to claim 1 or 8, characterized in that, The introduced morphological boundary refinement path includes: Boundary regions are identified through the boundary refinement path U-path, and boundary weights are obtained. Multi-scale boundary features are extracted using 5×5 depth convolution and 3×3 depth convolution with an inflation rate of 2. Boundary optimization features are obtained through feature fusion and bottleneck structure refinement. The boundary optimization features are then weighted and fused with the preliminary fusion features Y to output the final feature representation Z. The weight coefficients of the boundary paths are adjustable.

Images