A high-recall network optimization method for breast cancer target detection

Abstract

The application discloses a high-recall network optimization method for breast cancer target detection, comprising the following steps: collecting and pre-processing breast nuclear magnetic resonance images to obtain an image dataset; introducing a YOLOv8 network and replacing the original backbone network thereof with a constructed high-recall backbone network to obtain an improved YOLOv8 network; the high-recall backbone network comprises an initial convolutional layer and N feature enhancement layer levels which are sequentially cascaded, each layer level comprises a feature extraction module, an attention recalibration module and a down-sampling module which are sequentially connected; the improved YOLOv8 network is trained by using the image dataset, network parameters are optimized by minimizing a bounding box regression loss, a classification loss and a target confidence loss, and finally a breast cancer image detection model is obtained; the detection result can be output by inputting the image to be detected into the model to assist in diagnosis; the application significantly improves the feature perception ability of the model to small-scale regions to be detected, and realizes multi-level step-by-step optimization of shallow detail features and deep semantic features.

Description

Technical Field

[0001] This invention relates to the field of computer-aided diagnostic technology for medical imaging, and in particular to a high-recall network optimization method for target detection of breast cancer. Background Technology

[0002] In recent years, with the rapid development of deep learning technology, general object detection frameworks such as YOLO, Faster R-CNN, and SSD have achieved great success in the field of natural image detection. The success of these technologies has inspired research enthusiasm for applying them to image analysis in professional fields (such as medical imaging and industrial non-destructive testing), with the aim of achieving efficient and automatic localization and recognition of specific targets in images.

[0003] However, directly transferring these general object detection frameworks to high-resolution, highly complex image analysis tasks in specialized fields faces a series of inherent challenges that limit detection performance: 1. Insufficient recall capability for small and low-contrast targets: Key targets in specialized images (such as certain types of medical images) are often small in size, have low contrast with the background, and have blurred boundaries. Existing general frameworks' backbone network designs are mostly optimized for salient objects in natural scenes, lacking feature extraction and enhancement mechanisms for such "difficult targets." This results in low sensitivity (i.e., recall) of the model for these targets, making it prone to missed detections.

[0004] 2. Limited ability to perceive and utilize subtle image features: Professional images contain a large number of subtle but crucial discriminative features. The feature extraction process of general backbone networks may not be able to fully capture or emphasize these subtle features, making it difficult for the model to distinguish the target from similar backgrounds, or to accurately identify atypical target variants.

[0005] 3. Feature selection bias leads to weak generalization ability: During training, the model may tend to rely on some of the most prominent but non-universal feature patterns in the image for judgment. When faced with targets with atypical features and heterogeneous performance, this bias will cause the model to fail, the generalization ability will decrease, and the reliability in practical applications will be insufficient.

[0006] In summary, while existing technologies have attempted to incorporate attention mechanisms into target detection networks—for example, by adding SE or CBAM modules at the end of the backbone network or after the neck network—to enhance feature representation capabilities, these methods represent "shallow" or "single-time" attention enhancements. They cannot progressively filter and optimize multi-level feature maps in the early and middle stages of the network. For breast MRI images of variable-sized and complex-shaped regions (e.g., lesions), this single-time attention mechanism struggles to effectively capture and convey comprehensive key information from low-level details to high-level semantics. This results in limited recall capabilities for small or atypical regions, low detection accuracy, and a high risk of missed detections. Summary of the Invention

[0007] This invention provides a high-recall network optimization method for breast cancer target detection to overcome the above-mentioned technical problems.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows: A high-recall network optimization method for breast cancer target detection includes: S1: Acquire breast MRI images and preprocess them to obtain an image dataset; S2: Introduce the YOLOv8 network and construct a high-recall backbone network to replace the original backbone network of the YOLOv8 network, resulting in an improved YOLOv8 network; the high-recall backbone network includes an initial convolutional layer and N sequentially cascaded feature enhancement layers, each of which includes a feature extraction module, an attention recalibration module, and a downsampling module connected in sequence. S3: Input the image dataset into the improved YOLOv8 network for training, and optimize the network parameters by minimizing the bounding box regression loss, classification loss and target confidence loss to obtain the breast cancer image detection model; S4: Input breast MRI images into the breast cancer image detection model for image detection, output image detection results, and assist doctors in diagnosis.

[0009] Furthermore, in the high-recall backbone network, N is an integer greater than or equal to 1; the initial convolutional layer is connected to the first feature enhancement layer; The breast MRI images in the image dataset are input into the initial convolutional layer for convolution to obtain the convolutional image. Then, the convolutional image is enhanced by N sequentially cascaded feature enhancement layers to output a multi-scale feature map. Furthermore, the feature extraction module is used to perform feature transformation on the output image of the previous stage through convolution to obtain the transformed feature map, and input the transformed feature map into the attention recalibration module; The attention recalibration module is used to recalibrate the attention of the input transformed feature map, adaptively adjust the number of channels, and input the output result to the downsampling module. The downsampling module is used to perform downsampling pooling on the input image and output the downsampled feature map.

[0010] Furthermore, the attention recalibration module includes an input layer, a global average pooling layer, a first transformation layer, a second transformation layer, a multiplier, and an output layer; The input layer is used to receive the transformed feature map output by the feature extraction module and output it to the global average pooling layer; The global average pooling layer is connected to the input layer and is used to perform global average pooling on the transformed feature map of the input to generate a feature vector. The first transformation layer, connected to the global pooling layer, is used to perform the first convolution operation on the feature vector and process it through an activation function; The second transformation layer, connected to the first transformation layer, is used to perform a second convolution operation on the features processed by the first transformation layer, and then process them through a normalized activation function to generate a channel weight vector corresponding to the number of channels in the input feature map. The multiplier is used to perform channel-by-channel multiplication on the channel weight vector and the transformed feature map in the channel dimension, and output the recalibrated feature map.

[0011] Furthermore, the image dataset is fed into the improved YOLOv8 network for training, including: Breast MRI images from the image dataset are input into a high-recall backbone network for feature extraction to obtain multi-scale feature maps. Multi-scale feature maps are input into the neck network for feature fusion to obtain a fused feature map; The fused feature map is input into the detection head network for prediction, classification, and localization, and the image detection result is output.

[0012] Beneficial effects: This invention provides a high-recall network optimization method for breast cancer target detection, which has the following advantages: (1) This invention introduces an attention recalibration module in each feature enhancement level to adaptively allocate weights to the channels of the feature map, which can dynamically enhance the channels sensitive to the region to be inspected and suppress background noise, thereby significantly improving the model’s ability to perceive features of small-scale regions to be inspected, and ultimately resulting in a significant improvement in recall (experimental data show an improvement of 3.9% compared to the baseline model). (2) The present invention adopts N sequentially cascaded feature enhancement layers to achieve multi-level progressive optimization of shallow detail features and deep semantic features, so that the model can maintain high accuracy while significantly optimizing the overall detection performance F1-score. (3) The high recall backbone network constructed in this invention can be seamlessly integrated into existing detection frameworks such as YOLOv8 as a replaceable module, without the need to redesign the entire detection system, and has good practicality and portability. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This invention provides a flowchart of a high-recall network optimization method for breast cancer target detection. Figure 2 An improved overall system architecture diagram for YOLOv8; Figure 3 This is an architecture diagram of the high-recall backbone network; Figure 4 Workflow diagram for the attention recalibration module; Figure 5 This is a flowchart of the image processing for a breast cancer image detection model. Figure 6 This is a comparison chart of experimental data from an embodiment of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0016] This embodiment provides a high-recall network optimization method for breast cancer target detection, such as... Figure 1 As shown, it includes: S1: Acquire breast MRI images and preprocess them to obtain an image dataset; S2: Introduce the YOLOv8 network and construct a high-recall backbone network to replace the original backbone network of the YOLOv8 network, resulting in an improved YOLOv8 network; the high-recall backbone network includes an initial convolutional layer and N sequentially cascaded feature enhancement layers, each of which includes a feature extraction module, an attention recalibration module, and a downsampling module connected in sequence. S3: Input the image dataset into the improved YOLOv8 network for training, and optimize the network parameters by minimizing the bounding box regression loss, classification loss and target confidence loss to obtain the breast cancer image detection model; S4: Input breast MRI images into the breast cancer image detection model for image detection, output image detection results, and assist doctors in diagnosis.

[0017] Specifically, this invention provides a high-recall network optimization method for breast cancer target detection. The core of this method is the design of an innovative backbone network architecture, namely a high-recall backbone network. By embedding a multi-level attention mechanism, it significantly improves the detection capability of feature regions in medical images, assisting doctors in making more accurate judgments about the condition.

[0018] In a specific embodiment, the scheme for acquiring breast MRI images and performing preprocessing to obtain the image dataset is as follows: With the patient's informed consent, breast MRI images were collected. Medical experts performed detailed contour annotation on the regions to be identified in each of the collected images, generating professional medical contour annotation files that correspond one-to-one with each image, thus forming high-quality annotated raw data. Considering that computer vision models typically use rectangular bounding boxes as the standard input annotation format, the generated medical contour annotations need to be converted. This solution uses the maximum bounding rectangle method for conversion, specifically as follows: For each region to be identified, annotated by experts, calculate the minimum bounding rectangle of all its contour points. This rectangle is the smallest rectangle that completely encloses the contour and whose sides are parallel to the image coordinate axes. This bounding box is the transformed annotation, which contains the category and location information of the region to be identified (usually represented by the coordinates of the top-left and bottom-right corners of the rectangle). Together with the image, it forms the image dataset. Based on the pathological examination results corresponding to each image, namely the immunohistochemical (IHC) expression level of HER-2 (human epidermal growth factor receptor 2), the labeled image data were divided into four predefined categories for subsequent model training and optimization. The specific classification results are as follows: HER-2(0): Negative expression; HER-2(1+): Weakly positive expression; HER-2(2+): Moderately positive expression; HER-2(3+): Strong positive expression.

[0019] In a specific embodiment, a YOLOv8 network is introduced, and a high-recall backbone network is constructed to replace the original YOLOv8 network, resulting in an improved YOLOv8 network. The high-recall backbone network includes an initial convolutional layer and N sequentially cascaded feature enhancement layers. Each feature enhancement layer includes a sequentially connected feature extraction module, attention recalibration module, and downsampling module. like Figure 2 As shown, a high-recall backbone network is constructed as a replaceable module in the YOLOv8 framework, replacing the original standard backbone network. The input breast MRI image is sequentially processed by the high-recall backbone network for feature extraction, the neck network for multi-scale feature fusion, and the detection head for target localization and classification, and finally outputs the detection result. like Figure 3 As shown, the high-recall backbone network includes: an initial convolutional layer and N sequentially cascaded feature enhancement layers, where N is an integer greater than or equal to 1; the initial convolutional layer is connected to the first feature enhancement layer; The breast MRI images in the image dataset are input into the initial convolutional layer for convolution to obtain the convolutional image. Then, the convolutional image is enhanced by N sequentially cascaded feature enhancement layers to output a multi-scale feature map. In this scheme, the initial convolutional layer has a stride of 4. Convolutional layer; the value of N is increased or decreased according to the actual detection results. In this embodiment, N is set to 5 to balance model performance and computational complexity.

[0020] like Figure 3 As shown, each feature enhancement layer includes a feature extraction module, an attention recalibration module, and a downsampling module connected in sequence: The feature extraction module is used to perform feature transformation on the image output from the previous stage through convolution to obtain the transformed feature map, and input the transformed feature map into the attention recalibration module. The attention recalibration module is used to recalibrate the attention of the input transformed feature map, adaptively adjust the number of channels, and input the output result to the downsampling module. The downsampling module is used to perform downsampling pooling on the input image and output the downsampled feature map; Specifically, each feature enhancement level first passes through a feature extraction unit. Convolutional processing transforms features, followed by adaptive adjustment of channel importance via a core "attention recalibration module," and finally, pooling downsampling as needed. This repetitive structure, embedding an attention mechanism, is key to achieving high recall. The attention recalibration module includes an input layer, a global average pooling layer, a first transformation layer, a second transformation layer, a multiplier, and an output layer. The input layer is used to receive the transformed feature map output by the feature extraction module and output it to the global average pooling layer; The global average pooling layer is connected to the input layer and is used to perform global average pooling on the transformed feature map of the input to generate a feature vector. The first transformation layer, connected to the global pooling layer, is used to perform the first convolution operation on the feature vector and process it through an activation function; In this embodiment, the first transformation layer is composed of The convolutional layer is composed of a ReLU activation function; The second transformation layer, connected to the first transformation layer, is used to perform a second convolution operation on the features processed by the first transformation layer, and then process them through a normalized activation function to generate a channel weight vector corresponding to the number of channels in the input feature map. In this embodiment, the second transformation layer is composed of The convolutional layer is composed of a sigmoid activation function; The multiplier is used to perform channel-by-channel multiplication on the channel weight vector and the transformed feature map in the channel dimension, and output the recalibrated feature map.

[0021] Figure 4 The core innovation—the "attention recalibration module"—was demonstrated through its three-step workflow of "compression-stimulation-recalibration": Compression: Spatial information is compressed into channel descriptors through global average pooling; Incentive: Through a nonlinear transformation of a bottleneck structure (dimensionality reduction and then dimensionality increase), the nonlinear relationship between channels is learned and weights are generated; Relabeling: The learned weights are multiplied by the original feature map to complete the relabeling of the features.

[0022] This design achieves enhancements to key features with relatively low computational cost.

[0023] In a specific embodiment, the image dataset is input into the improved YOLOv8 network for training, and the network parameters are optimized by minimizing the bounding box regression loss, classification loss, and target confidence loss. The resulting breast cancer image detection model is as follows: Breast MRI images from the image dataset are input into a high-recall backbone network for feature extraction to obtain multi-scale feature maps. Multi-scale feature maps are input into the neck network for feature fusion to obtain a fused feature map; The fused feature map is input into the detection head network for prediction, classification and localization, and the image detection result is output, namely the bounding box coordinates and classification confidence of the region to be detected; The network parameters were optimized by minimizing the bounding box regression loss, classification loss, and target confidence loss to obtain a breast cancer image detection model. In this scheme, loss functions commonly used by those skilled in the art for optimizing YOLOv8 networks are used, namely bounding box regression loss, classification loss, and target confidence loss. The three are used to optimize the model parameters according to the set weights, which is common knowledge to those skilled in the art. Therefore, the specific optimization process and loss function will not be described in detail.

[0024] In a specific embodiment, the method of inputting breast MRI images into a breast cancer image detection model for image detection and outputting the final image detection results to assist doctors in diagnosis is as follows: like Figure 5 As shown, the image is input into the breast cancer image detection model, and the image undergoes feature preprocessing, including normalization, denoising, and resizing. Multiple feature enhancement levels are executed in a loop. It is determined whether this is the last level. If it is not the last level, the process returns to continue processing the next level, repeating the "feature extraction" and "attention recalibration" processes. If it is the last level, it means that feature extraction and optimization are complete, and the process enters the next stage, namely neck network feature fusion and detection head prediction, until the final output result.

[0025] Example 1: Performance Verification Experiment for Breast Cancer Medical Imaging Detection Experimental setup: Dataset: The dataset used in this experiment is a high-quality breast medical image dataset constructed in close collaboration with radiologists. A total of 2,894 images with a resolution of 512×512 were collected from 135 patients and divided into four categories according to HER-2 expression levels: HER-2(0) 318 images, HER-2(1+) 855 images, HER-2(2+) 931 images, and HER-2(3+) 790 images.

[0026] All images were precisely annotated by medical experts on a professional platform according to clinical standards, ensuring the authority and reliability of the data. To adapt to the training requirements of computer vision models, we uniformly converted professional medical contour annotations to maximum bounding rectangle annotations. Finally, the dataset was divided into a training set (2315 images), a validation set (289 images), and a test set (290 images) in an 8:1:1 ratio to comprehensively evaluate the model's recognition performance and generalization ability.

[0027] Baseline model: Standard YOLOv8 model Comparison Model: The YOLOv8 model with high recall backbone network replacement proposed in this invention The experimental results are compared with those of the baseline model, as shown in Table 1 and Figure 6 As shown: Table 1

[0028] Figure 6 The performance of the baseline model (YOLOv8) and the model of this invention was visually compared using clustered bar charts in three core performance metrics: precision, F1-score, and recall. The model of this invention outperformed the baseline model across all metrics, with the most significant improvement in recall.

[0029] Results analysis: 1. Significantly improved accuracy: The accuracy of the model in this invention reaches 97.9%, which is 1.7 percentage points higher than the baseline model's 96.2%. This indicates that the model further reduces the false positive rate in samples that are judged as positive, and the results are more reliable.

[0030] 2. Overall Detection Performance Optimization: The F1-score improved from 94.16% of the baseline model to 95.84% of the present invention (an absolute improvement of 1.68%). The F1-score is the harmonic mean of precision and recall. This significant improvement demonstrates that the model of the present invention has achieved a better balance between precision and recall, resulting in stronger overall detection capabilities.

[0031] 3. Significantly Enhanced Recall (Core Advantage): The recall rate is significantly improved from 92.40% in the baseline model to 96.30% in this invention (an absolute increase of 3.9%). This is the most outstanding technical effect of this invention, proving that the method can greatly reduce the false negative rate and is of vital value in identifying more easily overlooked abnormal areas (such as early or small lesions) in medical imaging.

[0032] 4. Improved detection accuracy: mAP@[.5:.95] / % increased from 67.10% to 67.90%.

[0033] Conclusion: This invention significantly improves recall and overall detection performance while maintaining high accuracy. It can effectively enhance the model's ability to perceive target features, especially difficult samples and small areas to be detected. It is particularly suitable for medical imaging diagnosis scenarios that are sensitive to missed detections, and greatly reduces the risk of missed detections while maintaining high accuracy.

[0034] The experimental results of the model of this invention are compared with those of other models, and the comparison results are shown in Table 2: Table 2

[0035] Results Analysis 1. Comprehensive performance superiority, establishing technological advantage. This invention achieved optimal performance across all four core evaluation metrics: The mAP@0.5 reached 97.9%, which is 1.4% higher than the second-best YOLOv11 (96.5%). The mAP@[.5:.95] reached 67.9%, which is 0.8% higher than the second-best baseline model YOLOv8 (67.1%). The recall rate reached 96.3%, which is 1.9% higher than the second-best, YOLOv5 (94.4%). The F1 score reached 95.8%, which is 1.23% higher than the second-best YOLOv10 (94.57%). The results show that the present invention has achieved a comprehensive breakthrough in terms of detection accuracy, comprehensive measurement, detection capability of the area to be detected, and overall performance balance.

[0036] 2. Significant advantage in recall rate: Among all the comparison models, the recall rate of this invention (96.3%) is the most outstanding, improving by 3.9% compared to the baseline model YOLOv8 (92.4%) and by 1.9% compared to the best-performing comparison model YOLOv5 (94.4%). This strongly demonstrates that the "high-recall backbone network" and its core "attention recalibration module" proposed in this invention can extremely effectively enhance the model's ability to perceive features of specific regions and significantly reduce the risk of missed diagnoses.

[0037] 3. This invention not only outperforms other YOLO models in the same series, but also demonstrates significant advantages over representative models from other technical approaches: Compared to the two-stage detector paradigm Faster R-CNN, this invention outperforms Faster R-CNN by 13.1% in mAP@0.5 and by 32.34% in F1-score.

[0038] Compared to SSD, an early representative of single-stage detectors, this invention leads by 5.3% in mAP@0.5.

[0039] Compared to the advanced Transformer-based detector RT-DETR, this invention outperforms RT-DETR by 2.7%, 3.0%, and 1.76% in mAP@0.5, recall, and F1-score, respectively.

[0040] This demonstrates that the technical advantages of this invention transcend the limitations of a specific framework, and its innovative points possess broad advancement and universality.

[0041] in conclusion Comprehensive comparative experiments demonstrate that the method described in this invention surpasses all mainstream YOLO series versions and representative models from other technical routes in key performance indicators for breast MRI imaging detection. It particularly establishes an absolute advantage in recall rate, a crucial indicator for medical diagnostic quality. This fully verifies the excellence, advancement, and inventiveness of the technical solution of this invention, indicating that it has reached an industry-leading level in this specific application field, possesses significant clinical application value, and can assist doctors in better diagnosing breast cancer.

[0042] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A high-recall network optimization method for breast cancer target detection, characterized in that, include: S1: Acquire breast MRI images and preprocess them to obtain an image dataset; S2: Introduce the YOLOv8 network and construct a high-recall backbone network to replace the original backbone network of the YOLOv8 network, resulting in an improved YOLOv8 network; the high-recall backbone network includes an initial convolutional layer and N sequentially cascaded feature enhancement layers, each of which includes a feature extraction module, an attention recalibration module, and a downsampling module connected in sequence. S3: Input the image dataset into the improved YOLOv8 network for training, and optimize the network parameters by minimizing the bounding box regression loss, classification loss and target confidence loss to obtain the breast cancer image detection model; S4: Input breast MRI images into the breast cancer image detection model for image detection, output image detection results, and assist doctors in diagnosis.

2. The high recall network optimization method for breast cancer target detection according to claim 1, characterized in that, In the high-recall backbone network, N is an integer greater than or equal to 1; the initial convolutional layer is connected to the first feature enhancement layer; The breast MRI images from the image dataset are input into the initial convolutional layer for convolution to obtain the convolutional image. Then, the convolutional image is enhanced by N sequentially cascaded feature enhancement layers to output a multi-scale feature map.

3. The high recall network optimization method for breast cancer target detection according to claim 1, characterized in that, The feature extraction module is used to perform feature transformation on the image output from the previous stage through convolution to obtain the transformed feature map, and input the transformed feature map into the attention recalibration module. The attention recalibration module is used to recalibrate the attention of the input transformed feature map, adaptively adjust the number of channels, and input the output result to the downsampling module. The downsampling module is used to perform downsampling pooling on the input image and output the downsampled feature map.

4. A high-recall network optimization method for breast cancer target detection according to claim 3, characterized in that, The attention recalibration module includes an input layer, a global average pooling layer, a first transformation layer, a second transformation layer, a multiplier, and an output layer. The input layer is used to receive the transformed feature map output by the feature extraction module and output it to the global average pooling layer; The global average pooling layer is connected to the input layer and is used to perform global average pooling on the transformed feature map of the input to generate a feature vector. The first transformation layer, connected to the global pooling layer, is used to perform the first convolution operation on the feature vector and process it through an activation function; The second transformation layer, connected to the first transformation layer, is used to perform a second convolution operation on the features processed by the first transformation layer, and then process them through a normalized activation function to generate a channel weight vector corresponding to the number of channels in the input feature map. The multiplier is used to perform channel-by-channel multiplication on the channel weight vector and the transformed feature map in the channel dimension, and output the recalibrated feature map.

5. A high-recall network optimization method for breast cancer target detection according to claim 1, characterized in that, The image dataset is fed into the improved YOLOv8 network for training, including: Breast MRI images from the image dataset are input into a high-recall backbone network for feature extraction to obtain multi-scale feature maps. Multi-scale feature maps are input into the neck network for feature fusion to obtain a fused feature map; The fused feature map is input into the detection head network for prediction, classification, and localization, and the image detection result is output.

Images