Mammary gland X-ray image risk layering method based on masking perception

Abstract

The invention discloses a mammary gland X-ray intelligent screening and risk stratification system based on masking perception, and provides a double-flow neural network fusing CNN and Swin Transform for solving the problem of high-risk cancer missed diagnosis caused by neglect of a'masking effect 'of an existing AI model, and local lesions and global structure features are synchronously extracted in combination with a regional perception attention mechanism. And the system executes focus detection and masking potential level prediction (1-8 levels) in parallel, and fuses prediction uncertainty to generate personalized clinical suggestions. For a high-masking or high-uncertainty negative result, MRI screening is actively recommended and supplemented, normal form transformation from'disease detection 'to'risk stratification' is achieved, the early detection rate of interval cancer and large invasive cancer is increased, and medical resource allocation is optimized.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and artificial intelligence-assisted diagnosis technology, specifically involving an intelligent screening method and system for breast X-ray images that integrates masking effect modeling and multimodal deep learning, which is used to improve the early detection rate of breast cancer and achieve personalized risk stratification. Background Technology

[0002] Breast cancer is one of the most common malignant tumors among women worldwide, and early screening is a key means of reducing mortality. Digital mammography, as a mainstream screening tool, has been widely used in clinical practice. However, despite the continuous improvement in screening coverage, a considerable proportion of breast cancers are still missed during screening, manifesting as interval cancer (cancer clinically discovered between two screenings) or large invasive cancer (typically defined as tumors ≥2 cm in diameter). These two types of cancer often have poor prognoses and significantly affect patient survival rates.

[0003] One of the main reasons for missed diagnoses is the masking effect—the tumor is "hidden" by the surrounding dense breast tissue, making it difficult to identify on X-ray images. Studies have shown that up to 50% of septal carcinomas are closely associated with a high masking effect. Traditionally, BI-RADS breast density classification (e.g., fatty, scattered fibroglandular, heterogeneous dense, extremely dense) has been used clinically as a proxy for assessing masking risk and determining whether supplemental screening (e.g., MRI) should be recommended. However, recent studies have indicated that breast density is not equivalent to masking potential. Density only reflects the overall proportion of fibroglandular tissue in the breast, while the masking effect is also influenced by various factors such as tissue distribution patterns, texture complexity, and contrast. In actual image interpretation, radiologists often judge whether an image is prone to missed diagnoses based on comprehensive visual perception, rather than relying solely on density grades.

[0004] In 2021, Sorkhei et al. proposed the CSAW-M dataset (CSAW-M: An Ordinal Classification Dataset for Benchmarking Mammographic Masking of Cancer). For the first time, five senior radiologists directly labeled the masking potential of over 10,000 cancer-free mammograms, quantifying the masking level into an ordered 1–8 level (level 1 being the easiest to interpret, and level 8 the most difficult), and linking it to the clinical endpoint of whether the cancer would subsequently develop into septal or highly invasive cancer. The results showed that the masking level predicted by the deep learning model trained on CSAW-M was more effective than traditional breast density indicators (such as percentage density calculated by Libra) in predicting high-risk cancers (significantly improved AUC), and the cancer risk in the high-masking group showed a monotonically increasing trend. This finding provides a solid data and algorithmic foundation for personalized screening strategies "centered on the masking effect."

[0005] Meanwhile, significant progress has been made in recent years in AI-assisted mammography diagnostic systems based on deep learning. Mainstream methods employ convolutional neural networks (CNNs) or Vision Transformers (such as the Swin Transformer) for end-to-end lesion detection or benign / malignant classification of images. Some systems incorporate attention mechanisms, multi-scale fusion, and uncertainty estimation to improve performance. However, existing AI systems generally suffer from a fundamental flaw: they focus solely on "the presence of lesions," neglecting the crucial clinical dimension of "whether the image is prone to missed diagnoses." In other words, even if the AI ​​outputs a "negative" result in a highly masked image, its confidence level may be artificially high, while the actual risk of missed diagnoses is extremely high. This "high-confidence false negative" is highly misleading in clinical applications, potentially causing high-risk individuals to miss opportunities for supplemental screening.

[0006] Furthermore, existing systems often suffer from insufficient sensitivity to small or low-contrast lesions when processing highly masked images due to factors such as inadequate coupling between local texture and global structural information, lack of region-prior guidance in attention mechanisms, and data augmentation strategies not optimized for the characteristics of medical images. Although some studies have attempted to use breast density as an input feature or post-processing rule, as mentioned earlier, the correlation between density and masking potential is limited, making it difficult to fundamentally solve the problem of missed diagnoses in highly masked scenarios.

[0007] In conclusion, there is an urgent need for a novel intelligent mammography screening system capable of simultaneously modeling both "lesion presence" and "image masking potential," and generating interpretable and actionable clinical risk stratification recommendations based on the combined analysis of these two factors. This system should not only possess high-precision lesion detection capabilities but also proactively identify individuals at high risk of masking and recommend supplementary screenings such as MRI, thereby truly achieving the clinical goals of "precise screening, risk stratification, and resource optimization." Summary of the Invention

[0008] The technical problem this invention aims to solve is that existing intelligent screening systems for breast X-ray images only focus on the presence of lesions, neglecting the masking effect of the image itself—the risk of missed diagnoses due to tumors being obscured by dense breast tissue. Traditional methods use breast density as a proxy for masking risk, but studies have shown that density has limited correlation with actual masking potential. Furthermore, existing deep learning models are prone to producing high-confidence false negatives in highly masked images, lacking the ability to model whether an image is easy to interpret, and thus failing to provide personalized screening recommendations for clinicians, making it difficult to effectively prevent the occurrence of septal carcinoma or highly invasive carcinoma.

[0009] This invention uses mammogram images as input. After adaptive breast region cropping, text erasure, CLAHE contrast enhancement, and 632×512 normalization preprocessing, the images are fed into a dual-stream backbone network consisting of EfficientNet-B5 and Swin Transformer: the CNN branch captures local details such as microcalcifications, while the Transformer branch models long-range structural relationships. Features are fused through channel-level adaptive weighting, and a breast region-aware attention mechanism is introduced to suppress interference from the chest wall and background. The fused unified features are fed in parallel into two task heads: a lesion detection head outputs the probability of cancer presence; and a masking assessment head outputs a masking potential level of 1–8 using multi-hot ordered regression. During the inference phase, Monte Carlo Dropout is used for T forward samplings to obtain prediction uncertainty. The system integrates lesion probability, masking level, and uncertainty to generate personalized clinical recommendations according to predetermined threshold logic. Attached Figure Description

[0010] Figure 1 The above is a flowchart of the overall system described in this invention.

[0011] Figure 2 This is a schematic diagram of the hybrid feature extraction network and the dual-task prediction head. Detailed Implementation

[0012] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings:

[0013] This invention mainly studies a method for intelligent screening and risk stratification of mammogram images based on masking effect perception. Figure 1 This is the overall flowchart, employing the following technical solutions:

[0014] Step 1, Image Preprocessing and Standardization, specifically:

[0015] Suppose the input image is a unilateral mammogram. Its pixel values ​​are in 16-bit DICOM format. To ensure the consistency, robustness, and clinical interpretability of subsequent deep learning model inputs, this invention performs the following standardized preprocessing steps:

[0016] 1. Adaptive cropping of the breast region: First, the original image is binarized using the Otsu thresholding algorithm to generate a breast mask. Where a pixel value of 1 represents breast tissue and 0 represents the background; Calculate The minimum bounding rectangle region The original image was cropped to obtain a sub-image containing only breast tissue. ,Right now

[0017]

[0018] Where ⊙ represents a region clipping operation. and This represents the height and width of the cropped image.

[0019] 2. Automatic text area erasure: Clinical X-ray images typically contain the technician's initials, breast side (L / R), and view identifiers (such as MLO) in the upper right corner. This text area mask is denoted as... Median of background grayscale Fill the text area, where

[0020]

[0021] That is, take the median pixel value of the non-breast region; to obtain the text-free image. for:

[0022]

[0023] 3. Enhanced contrast: For Contrast-limited adaptive histogram equalization (CLAHE) is applied with the following parameters: clipping threshold = 2.0, grid size = (8, 8), and the output enhanced image is denoted as [image description missing]. .

[0024] 4. Geometric Normalization and Size Alignment: To maintain complete consistency with the CSAW-M dataset (facilitating model transfer and evaluation), the following operations are performed:

[0025] (1) Flip all images horizontally to face left (left-posed);

[0026] (2) Calculate the center of mass of the breast Shift it horizontally to the center column of the image;

[0027] (3) Scale the image to a fixed resolution using zero-padding. ,Right now

[0028]

[0029] This size (632×512) is the input specification officially recommended by CSAW-M.

[0030] 5. Medical-specific data augmentation during training (for training only): To improve the model's generalization ability, standardized images are augmented. The following enhancement strategies will be applied randomly:

[0031] (1) CutMix: Randomly select another training sample and its cover level labels Generate a mixed sample:

[0032]

[0033] in The input image is enhanced by CutMix. This refers to the masking level label after blending. For random rectangular masks, This represents the percentage of the mask area.

[0034] (2) MixUp: MixUp for sample pairs and Perform linear interpolation, interpolation coefficients ,generate:

[0035]

[0036] (3) Geometric perturbation: Apply a random rotation of ±10° with a 50% probability to enhance the model's robustness to pose changes. Final output: preprocessed image. This is used for subsequent feature extraction.

[0037] Step 2: Hybrid Feature Extraction and Region-Aware Attention Fusion

[0038] The preprocessed standardized image Input the hybrid neural network backbone and perform two-stream feature extraction, adaptive fusion, and region-aware attention enhancement, such as... Figure 2 As shown, the specific steps include the following:

[0039] 1. Dual-stream feature extraction:

[0040] (1) CNN Branch: The EfficientNet-B5 network, pre-trained on ImageNet, is used as the local feature extractor. This branch captures local texture details such as microcalcifications and edge sharpness through progressive downsampling via depthwise separable convolutions. Let its output feature map be...

[0041]

[0042] Among the number of channels Spatial dimension , (Corresponding to input downsampling by 32 times);

[0043] (2) Transformer Branch: The Swin Transformer (window size 7, layers 4, heads 16) is used as the global structural modeler. This branch models long-range dependencies such as breast tissue distribution patterns and dense region topology through a shifted window self-attention mechanism. Let its output feature map be...

[0044]

[0045] The two branches share the same spatial resolution and channel dimensions, which facilitates subsequent fusion.

[0046] 2. Adaptive Feature Fusion

[0047] To dynamically balance the contributions of local details and global semantics, a channel-level adaptive weighted fusion mechanism is designed:

[0048] (1) To and Perform global average pooling (GAP) separately to obtain channel statistics vectors. ,in

[0049]

[0050] Similar definitions;

[0051] (2) and The input is concatenated along the channel dimension, fed into a two-layer fully connected network (MLP), with the intermediate layer having a dimension of C / 2. The output is activated by a sigmoid function to generate a fused weight vector. :

[0052]

[0053] in Indicates channel splicing. For the Sigmoid function;

[0054] (3) Calculate complementary weights And perform channel-by-channel weighted fusion to obtain fusion features:

[0055]

[0056] Where ⊙ represents element-wise multiplication of the channel dimension. This mechanism allows the network to adaptively emphasize CNN or Transformer features in different image regions.

[0057] 3. Breast Region Sensory Attention Mechanism (BRPA):

[0058] To guide the model to focus on the breast parenchyma and suppress interference from irrelevant areas such as the chest wall and background, a BRPA module was designed:

[0059] (1) Channel attention: for Perform GAP and two fully connected layers (compression ratio 16) to generate channel weights. ;

[0060] (2) Spatial attention: Features weighted by channels Max pooling and average pooling are performed along the channel dimension. The results are concatenated and then subjected to a 7×7 convolution followed by sigmoid activation to generate a spatial weight map. ;

[0061] (3) Area Prior Guidance (Optional): If the system has already generated a masking heatmap through the masking evaluation head. (Indicating high-risk areas for concealment), then it is combined with Multiplication enhances attention response in highly concealed areas:

[0062]

[0063] (4) Final output enhanced features:

[0064]

[0065] This module significantly improves the model's ability to perceive key areas of the breast, especially enhancing the robustness of lesion search in highly masked images.

[0066] Step 3: Parallel execution of dual-task prediction for lesion detection and masking potential assessment

[0067] Enhanced features based on the output of step two (in , , This invention performs two prediction tasks in parallel: lesion presence detection and image masking potential assessment, specifically including the following sub-steps:

[0068] 1. Lesion detection head

[0069] (1) To Perform global average pooling (GAP) to obtain the global feature vector. ;

[0070] (2) Input fully connected layer (weights) bias After activation by the Sigmoid function, the probability of lesion presence is output:

[0071]

[0072] 2. Concealment potential assessment head

[0073] (1) Similarly to Perform GAP to obtain feature vectors ;

[0074] (2) Input fully connected layer (weight matrix) bias Output a 7-dimensional logits vector. ;

[0075] (3) Multi-hot encoding is used to transform the 8 masking levels (1–8) into 7 binary targets: for the true masking level Its multi-hot tag is

[0076] ,

[0077] Among them, the former The bit is 1, then The bit is 0;

[0078] (4) Apply Sigmoid activation to each element of logits z to obtain the binary classification probabilities. , ;

[0079] The final predicted cover level is

[0080] ;

[0081] Where M represents the final predicted shielding potential level, with a value ranging from 1 (easiest to interpret) to 8 (most difficult to interpret).

[0082] 3. Training Strategies

[0083] The masking evaluation head can be pre-trained on the CSAW-M dataset and then jointly fine-tuned with the lesion detection head on the target dataset (such as INBreast). The total loss function is a weighted sum:

[0084] ;

[0085] in This is the true label for the lesion (0 = negative, 1 = positive). For the first The real labels of the masked subtasks (from multi-hot encoding); Represents the binary cross-entropy loss function; , This is the multi-task loss balance coefficient, used to adjust the relative importance of lesion detection and masking assessment.

[0086] Step 4: Uncertainty Perception and Risk Decision Generation

[0087] Based on the lesion presence probability output in step three With concealment potential level This invention introduces an uncertainty quantification mechanism and integrates the two to generate personalized clinical recommendations, specifically including the following sub-steps:

[0088] 1. Uncertainty estimation

[0089] (1) During the inference phase, Monte Carlo Dropout is enabled, and forward propagation is performed T times (default T=30), randomly discarding a portion of neurons in the network each time; thus obtaining the lesion probability set. With cover level set ;

[0090] (2) Calculate the predicted mean of lesions With variance :

[0091] , ,

[0092] in This indicates the uncertainty in the model's predictions (the larger the variance, the lower the confidence level).

[0093] (3) Calculate the median of the shielding level

[0094] ,

[0095] in The final masking level to be adopted.

[0096] 2. Risk Decision-Making Logic

[0097] based on , and Follow these clinical guidelines:

[0098] High concealment, high risk: If and (where the threshold) If the output is positive, then the following suggestions will be provided:

[0099]

[0100] Low-risk negative pathway: If and (in )and (in If the value is 0.05, then the following suggestions will be output:

[0101]

[0102] Uncertainty-dominated path: if Then regardless and Regardless of the value chosen, a suggestion will be output:

[0103]

[0104] The core of this invention, a mask-aware-based intelligent mammography screening system, lies in its parallel prediction of lesion probability and masking potential levels (1–8), rather than simply determining whether cancer is present. The system employs a two-stream network (EfficientNet-B5 + Swin Transformer) to extract local and global features, and enhances key region responses through adaptive fusion and region-aware attention mechanisms. Unlike traditional AI models that rely solely on single lesion detection or coarse breast density assessment, this invention innovatively uses "image interpretation difficulty" as an explicit learning objective and achieves fine-grained masking quantification through multi-hot ordered regression. During the inference phase, the system integrates lesion probability, masking level, and the predictive uncertainty of Monte Carlo Dropout estimation to construct a multi-dimensional decision logic: for negative results with high masking or high uncertainty, it proactively suggests "recommended MRI" or "recommends manual review," thereby achieving a crucial shift from "disease detection" to "risk stratification," effectively avoiding high-confidence false negatives and improving the early detection capability of septal carcinoma.

[0105] Compared with the prior art, the present invention has the following technical effects:

[0106] 1. Achieving Joint Modeling of "Lesion Detection" and "Mask Potential": This invention breaks through the limitation of traditional AI systems that only focus on "whether there is a lesion". It innovatively introduces the 1-8 level mask potential defined by CSAW-M as an explicit prediction target. By outputting the lesion probability and mask level in parallel through dual task heads, it realizes the quantitative modeling of whether an image is "easy to miss", filling the gap in the existing technology in the assessment of missed diagnosis risk.

[0107] 2. Significantly improves the robustness of lesion detection in highly masked images. By fusing channel and spatial attention through the Breast Region Perceptual Attention (BRPA) mechanism and optionally introducing a masking heatmap as a region prior, the model is guided to enhance its lesion search ability in high-density, low-contrast regions, effectively alleviating the problem of traditional models missing small lesions in highly masked images.

[0108] 3. Introducing uncertainty quantification to enhance clinical credibility and safety: The Monte Carlo Dropout mechanism is used to estimate the prediction uncertainty. For results with high uncertainty, a "manual review suggestion" is proactively provided to avoid misleading clinical judgment with high-confidence false negatives, which meets the safety and interpretability requirements of medical AI systems.

[0109] 4. Generate personalized clinical recommendations to achieve a closed-loop system for precise screening. The system integrates lesion probability, masking level, and uncertainty to automatically generate three types of recommendations: "recommend supplementary MRI", "low-risk negative", or "recommend manual review". It directly connects to the logic of clinical trials such as DENSE, and promotes the upgrade of AI from "assisted detection" to "risk stratification and resource optimization".

Claims

1. A breast X-ray image intelligent screening and risk stratification system based on masking effect perception, characterized in that, Includes the following steps: Step 1, Image Preprocessing and Standardization, specifically: Suppose the input image is a unilateral mammogram. Its pixel values ​​are in 16-bit DICOM format. To ensure the consistency, robustness, and clinical interpretability of subsequent deep learning model inputs, this invention performs the following standardized preprocessing steps: (1) Adaptive cropping of breast region: First, the original image is binarized using the Otsu threshold segmentation algorithm to generate a breast mask. Where a pixel value of 1 represents breast tissue and 0 represents the background; Calculate The minimum bounding rectangle region The original image was cropped to obtain a sub-image containing only breast tissue. ,Right now ， Where ⊙ represents a region clipping operation. and The height and width of the cropped image; (2) Automatic text area erasure: Clinical X-ray images typically contain the technician's initials, breast side (L / R), and view identifier (such as MLO) in the upper right corner. This text area mask is denoted as... Median gray level of the background Fill the text area, where ， That is, take the median value of pixels in the non-breast region; Get the text-free image for: ， (3) Contrast enhancement: for Contrast-limited adaptive histogram equalization (CLAHE) is applied with the following parameters: clipping threshold = 2.0, grid size = (8, 8), and the output enhanced image is denoted as [image description missing]. , (4) Geometric normalization and size alignment: To maintain complete consistency with the CSAW-M dataset (for easy model transfer and evaluation), the following operations are performed: Flip all images horizontally to face left (left-posed); Calculating the center of mass of the breast Shift it horizontally to the center column of the image; Scaling an image to a fixed resolution using zero-padding ,Right now ， This size (632×512) is the input specification officially recommended by CSAW-M; (5) Medical-specific data augmentation during the training phase (for training only): To improve the model's generalization ability, standardized images are augmented. The following enhancement strategies will be applied randomly: CutMix: Randomly select another training sample and its cover level labels Generate a mixed sample: ， in The input image after CutMix enhancement. This refers to the masking level label after blending. For random rectangular masks, This represents the percentage of the mask area. MixUp: For sample pairs and Perform linear interpolation, interpolation coefficients ,generate: ， Geometric perturbation: Apply random rotations of ±10° with a 50% probability to enhance the model's robustness to pose changes. Final output: preprocessed image. Used for subsequent feature extraction; Step 2: Hybrid Feature Extraction and Region-Aware Attention Fusion The preprocessed standardized image The input hybrid neural network backbone performs two-stream feature extraction, adaptive fusion, and region-aware attention enhancement, as shown in Figure 2. Specifically, it includes the following sub-steps: (1) Dual-stream feature extraction: a) CNN Branch: This branch uses an EfficientNet-B5 network pre-trained on ImageNet as a local feature extractor. It captures local texture details such as microcalcifications and edge sharpness through progressive downsampling via depthwise separable convolutions. Let its output feature map be... ， Among the number of channels Spatial dimension , (Corresponding to input downsampling by 32 times); b) Transformer branch: The Swing Transformer (window size 7, number of layers 4, number of heads 16) is used as the global structure modeler. This branch models long-range dependencies such as breast tissue distribution patterns and dense region topology through a shift-window self-attention mechanism. Let its output feature map be... ， The two branches share the same spatial resolution and channel dimensions, which facilitates subsequent fusion. (2) Adaptive feature fusion To dynamically balance the contributions of local details and global semantics, a channel-level adaptive weighted fusion mechanism is designed: right and Perform global average pooling (GAP) separately to obtain channel statistics vectors. ,in ， Similar definitions; Will and The input is concatenated along the channel dimension, fed into a two-layer fully connected network (MLP), with the intermediate layer having a dimension of C / 2. The output is activated by a sigmoid function to generate a fused weight vector. : ， in Indicates channel splicing. For the Sigmoid function; Calculate complementary weights And perform channel-by-channel weighted fusion to obtain fusion features: ， Where ⊙ represents element-wise multiplication along the channel dimension. This mechanism allows the network to adaptively emphasize CNN or Transformer features in different image regions; (3) Breast Region Sensory Attention Mechanism (BRPA): To guide the model to focus on the breast parenchyma and suppress interference from irrelevant areas such as the chest wall and background, a BRPA module was designed: Channel attention: Perform GAP and two fully connected layers (compression ratio 16) to generate channel weights. ; Spatial attention: Features weighted by channels Max pooling and average pooling are performed along the channel dimension. The results are concatenated and then subjected to a 7×7 convolution followed by sigmoid activation to generate a spatial weight map. ; Region Prior Guidance (Optional): If the system has already generated a masking heatmap through the masking evaluation head. (Indicating high-risk areas for concealment), then it is combined with Multiplication enhances attention response in highly concealed areas: ， Final output augmented features: ， This module significantly improves the model's ability to perceive key areas of the breast, especially enhancing the robustness of lesion search in highly masked images; Step 3: Parallel execution of dual-task prediction for lesion detection and masking potential assessment Enhanced features based on the output of step two (in , , This invention performs two prediction tasks in parallel: lesion presence detection and image masking potential assessment, specifically including the following sub-steps: (1) Lesion detection head right Perform global average pooling (GAP) to obtain the global feature vector. ; Will Input fully connected layer (weights) bias After activation by the Sigmoid function, the probability of lesion presence is output: ， (2) Assessment of concealment potential Similarly to Perform GAP to obtain feature vectors Input fully connected layer (weight matrix) bias Output a 7-dimensional logits vector. Multi-hot encoding is used to transform the 8 masking levels (1–8) into 7 binary target classifications: for the true masking level Its multi-hot tag is ， Among them, the former The bit is 1, then The bit is 0; Applying Sigmoid activation element-wise to logits z yields the binary classification probabilities. ， ; The final predicted cover level is ； Where M represents the final predicted shielding potential level, with a value ranging from 1 (easiest to interpret) to 8 (most difficult to interpret); (3) Training strategy The masking evaluation head can be pre-trained on the CSAW-M dataset and then jointly fine-tuned with the lesion detection head on the target dataset (such as INBreast). The total loss function is a weighted sum: ； in This is the true label for the lesion (0 = negative, 1 = positive). For the first The real labels of the masked subtasks (from multi-hot encoding); Represents the binary cross-entropy loss function; , This is a multi-task loss balancing coefficient used to adjust the relative importance of lesion detection and masking assessment; Step 4: Uncertainty Perception and Risk Decision Generation Based on the lesion presence probability output in step three With concealment potential level This invention introduces an uncertainty quantification mechanism and integrates the two to generate personalized clinical recommendations, specifically including the following sub-steps: (1) Uncertainty estimation During the inference phase, Monte Carlo Dropout is enabled, and forward propagation is performed T times (default T=30), randomly discarding a portion of neurons in the network each time; this yields a set of lesion probabilities. With cover level set ; Calculate the predicted mean of lesions With variance : ， ， in This indicates the uncertainty in the model's predictions (the larger the variance, the lower the confidence level). Calculate the median of the shielding level , ， in The final masking level to be adopted; (2) Risk decision-making logic based on , and Follow these clinical guidelines: High concealment, high risk: If and (where the threshold) If the output is positive, then the following suggestions will be provided: ， Low-risk negative pathway: If and (in )and (in If the value is 0.05, then the suggested output is: ， Uncertainty-dominated path: if Then regardless and Regardless of the value chosen, a suggestion will be output: ， In the masking perception-based intelligent mammography screening and risk stratification system proposed in this invention, the model not only outputs the probability of lesion presence, but also predicts the masking potential level of the image from 1 to 8 in parallel, and uses this as the key input for risk decision-making. The system uses lesion probability, masking level and prediction uncertainty to make a comprehensive judgment to generate personalized clinical recommendations. This decision-making logic is a multi-dimensional fusion strategy. That is, when the model detects high masking risk or high prediction uncertainty, even if the lesion probability is low, it will generate an active warning of "recommending supplementary MRI screening" or "recommending manual review".

Images