Peripheral blood image data feature extraction method based on double-channel feature pyramid

Abstract

The application discloses a peripheral blood image data feature extraction method based on a double-channel feature pyramid, and comprises the following steps: S1, a MobileNetV3-Large network is used to extract a feature map from an original data image; S2, a double-channel feature pyramid network composed of a small target feature pyramid network and a large target feature pyramid network is constructed; and S3, a Faster-RCNN target detection network is constructed, wherein the Faster-RCNN target detection network comprises a feature extraction network, a region proposal network RPN, an ROI pooling layer and a classification layer. The double-channel feature pyramid network of the application designs a special feature extraction channel for different sizes of targets, effectively reduces information interference between targets, and avoids the loss of feature information of small size targets when a traditional feature extraction network extracts features of large size targets.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and relates to a method for feature extraction of peripheral blood image data based on a dual-channel feature pyramid. Background Technology

[0002] With advancements in computer vision and image processing technologies, artificial intelligence has begun to demonstrate its potential in medical image analysis. Machine learning or deep learning techniques can significantly assist doctors in quickly identifying targets and locating regions of interest in image data. However, in peripheral blood image data, multiple targets involved in disease diagnosis are often overlapping. Inaccurate feature extraction for different targets can lead to reduced target recognition accuracy. This is especially true when multiple targets differ in size; larger targets occupy a larger proportion of the image area. Traditional feature extraction methods often focus too much on the features of large targets, easily resulting in the loss of features from smaller targets. This neglect of feature extraction for targets of both large and small sizes leads to poor performance in detecting targets of different sizes simultaneously. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a feature extraction method for peripheral blood image data based on a dual-channel feature pyramid. This dual-channel feature pyramid network effectively reduces information interference between targets by designing dedicated feature extraction channels for targets of different sizes, and avoids the loss of feature information of small targets when traditional feature extraction networks extract features from large targets.

[0004] The objective of this invention is achieved through the following technical solution: a method for feature extraction from peripheral blood image data based on a dual-channel feature pyramid, comprising the following steps:

[0005] S1. Acquire peripheral blood images and scale them to a fixed size;

[0006] S2. The MobileNetV3-Large network is used as the backbone network to extract feature maps from the original data images. The MobileNetV3-Large network includes two two-dimensional convolutional layers and 15 bneck blocks located between the two two-dimensional convolutional layers. The two-dimensional convolutional layers and the 15 bneck blocks are connected sequentially. Each convolutional layer and bneck block outputs a feature map, resulting in 17 feature maps.

[0007] The 1st, 2nd, 3rd, 4th, 8th, 9th, 10th, and 11th bneck blocks respectively include 1×1 convolution, DW convolution, and 1×1 convolution;

[0008] The 5th, 6th, 7th, 12th, 13th, 14th, and 15th bneck blocks consist of a 1×1 convolution, a DW convolution, an SE module, and another 1×1 convolution, respectively. Each bneck block first uses a 1×1 convolution to increase the dimensionality of the input tensor, then uses a DW convolution to extract features; subsequently, the SE module processes the extracted features, and finally, a 1×1 convolution is used to reduce the dimensionality of the features output by the SE module to obtain the output tensor. The SE module consists of a pooling layer, a fully connected layer with a ReLU activation function, and a fully connected layer with an H-Swish activation function, connected in sequence.

[0009] S3. Construct a dual-channel feature pyramid network consisting of a small target feature pyramid network and a large target feature pyramid network. The input to the small target feature pyramid network is the feature maps of layers 2, 4, and 16 of the MobileNetV3-Large network, which are denoted as S1, S2, and S3, respectively. The input to the large target feature pyramid network is the feature maps of layers 8, 13, and 16 of the MobileNetV3-Large network, which are denoted as K1, K2, and K3, respectively.

[0010] The small target feature pyramid network and the large target feature pyramid network have the same structure, both including an inverted pyramid structure. That is, S1 / K1, which is at the top level in the MobileNetV3-Large network, is used as the bottom layer C1 / C1* of the pyramid, S2 / K2, which is in the middle level, is used as the middle layer C2 / C2* of the pyramid, and S3 / K3, which is at the bottom level, is used as the top layer C3 / C3* of the pyramid. First, S3 / K3 is downsampled to generate a new feature map S4 / K4, and a new top layer C4 / C4* of the pyramid is constructed. Then, S1 / K1~S4 / K4 are input into 1×1 convolutional layers for dimensionality compression.

[0011] The compressed S4 / K4 is input into a 3×3 convolutional layer, and then into a prediction layer to obtain the feature map P4 / P4*.

[0012] The compressed S4 / K4 and S3 / K3 are fused together using element-wise addition, then fed into a 3×3 convolutional layer, and then into a prediction layer to obtain the feature map P3 / P3*.

[0013] The fused S3 / K3 and S2 / K2 are fused using element-wise addition, then fed into a 3×3 convolutional layer, and then into a prediction layer to obtain the feature map P2 / P2*.

[0014] The fused S2 / K2 and S1 / K1 are fused using element-wise addition, then fed into a 3×3 convolutional layer, and then into a prediction layer to obtain feature maps P1 / P1*.

[0015] S4. Constructing the Faster-RCNN object detection network: The Faster-RCNN object detection network includes a feature extraction network, a region proposal network, a ROI pooling layer, and a classification layer. The feature extraction network uses the dual-channel feature pyramid network constructed in step S3. The feature map extracted by the feature extraction network is input into the RPN to generate object candidate boxes, which are then input into the ROI pooling layer. The pooling layer converts candidate boxes of different sizes into a fixed size. The classification layer is used to classify the object category and regress the bounding boxes based on the features output by the ROI pooling layer.

[0016] The regional proposal network specifically includes the following processing steps:

[0017] S41. Input the 8 feature maps output by the dual-channel FPN into the region proposal network respectively;

[0018] S42. Extract features from the input feature map using a 3×3 convolutional layer;

[0019] S43. Input the feature map output from the 3×3 convolutional layer into two 1×1 convolutional layers respectively. The first 1×1 convolutional layer is used for the binary classification task of anchor points to determine whether the anchor point region contains the foreground target. Then, the output of the first 1×1 convolutional layer is transposed once, then a softmax operation is performed, and then it is transposed again to obtain the final classification score.

[0020] The second 1×1 convolutional layer is used for the bounding box regression task, which is to predict the precise location of each anchor point;

[0021] S44. Generate final region suggestions based on classification scores and regression results. These suggestion boxes will be sent to the ROI pooling layer for processing.

[0022] The beneficial effects of this invention are as follows: First, a dual-channel feature pyramid network is constructed based on the feature map generated by the MobileNetV3-Large network. By designing a dedicated feature extraction channel for small targets, the loss of feature information in the feature extraction of large targets is effectively avoided. Simultaneously, the independent channel design reduces feature interference between targets of different sizes, improving the accuracy of feature extraction. Finally, the dual-channel feature pyramid network is used to extract features from peripheral blood smear images, and the recognition accuracy of the selected features is verified using the Faster-RCNN target detection algorithm. The results of the verification experiments show that, due to the improved accuracy of feature extraction, the target detection accuracy for both small and large targets is significantly improved. Attached Figure Description

[0023] Figure 1 A flowchart of the overall process for the dual-channel feature pyramid method for feature extraction from peripheral blood image data;

[0024] Figure 2 This is a diagram of the MobileNetV3-Large architecture;

[0025] Figure 3 This is a diagram of a dual-channel FPN architecture;

[0026] Figure 4 This is a feature pyramid architecture diagram;

[0027] Figure 5 This is a functional diagram of the Faster-RCNN object detection algorithm;

[0028] Figure 6 This is a diagram showing the target detection results for peripheral blood image data. Detailed Implementation

[0029] This invention first uses the MobileNetV3-Large network, based on a deep convolutional neural network, as the backbone, primarily responsible for extracting effective information from images and obtaining feature maps. Then, it constructs a multi-scale and hierarchical pyramid structure for the feature maps, building two dual-channel Feature Pyramid Network (FPN) architectures that fuse high-level feature semantic information. Finally, it employs a Faster-RCNN object detection model to process the features extracted by the dual-channel FPN network. Experimental results on a peripheral blood image dataset are used to select appropriate FPN feature map parameters and verify the effectiveness of the invented method. The technical solution of this invention is further illustrated below with reference to the accompanying drawings.

[0030] like Figure 1 As shown, the present invention provides a method for feature extraction from peripheral blood image data based on a dual-channel feature pyramid, comprising the following steps:

[0031] S1. Acquire peripheral blood images and scale images of any size to a fixed size L×M (default is 800×800 pixels); this is because the input size of the convolutional neural network is fixed, and if the input image size is inconsistent, the convolution operation will not be performed correctly.

[0032] S2. Using the MobileNetV3-Large network as the backbone network, feature maps are extracted from the original data images; the structure of the MobileNetV3-Large network is as follows: Figure 2As shown, it includes two two-dimensional convolutional layers and 15 bneck blocks located between the two two-dimensional convolutional layers. The two-dimensional convolutional layers and the 15 bneck blocks are connected sequentially, and the 17-layer structure is numbered 0 to 16. Each convolutional layer and bneck block outputs a feature map, resulting in 17 feature maps numbered 0 to 16.

[0033] The 1st, 2nd, 3rd, 4th, 8th, 9th, 10th, and 11th bneck blocks respectively include 1×1 convolution, DW convolution, and 1×1 convolution.

[0034] The 5th, 6th, 7th, 12th, 13th, 14th, and 15th bneck blocks consist of a 1×1 convolution, a DW convolution, an SE module, and another 1×1 convolution. Each bneck block first uses a 1×1 convolution to increase the dimensionality of the input tensor, then uses a DW convolution (with a kernel size of 3×3 or 5×5) for feature extraction. Subsequently, the SE (Squeeze-and-Excitation) module processes the extracted features, and finally, a 1×1 convolution is used to reduce the dimensionality of the features output by the SE module, yielding the output tensor. This structure, small at both ends and large in the middle, constitutes a bneck block; this configuration is called an Inverted Residuals Block (IRB). The concept of IRB helps to further reduce the computational cost of the model, while the SE module indicates whether squeezing and excitation exist within the bneck block.

[0035] The SE module consists of a series of interconnected pooling layers, a fully connected layer with ReLU activation (FC ReLU), and a fully connected layer with H-Swish activation (FC h-swish). The pooling layer reduces the size of the feature map while preserving important feature information, thus reducing computational cost and preventing overfitting. The fully connected layer with ReLU activation connects all input and output nodes and is often used in the last few layers of the network to transform the features extracted by the convolutional layers into specific outputs (such as classification results). ReLU activation is a common activation function defined as ReLU(x) = max(0,x). It truncates all negative values ​​to 0, retaining positive values, which helps introduce non-linearity and accelerates training convergence. The fully connected layer with H-Swish activation improves the model's predictive ability while maintaining efficient computational performance.

[0036] The SE module slightly increases the size of the system architecture but also enhances its accuracy. Furthermore, the SE module incorporates the h-swish activation function. The definition of the swish activation function is:

[0037]

[0038] in, It is a trainable parameter. For the sigmoid function, that is, when When this is the case, the function is called the sigmoid weighted linear unit function. However, because calculating this function is computationally expensive, the h-swish function was proposed, which is:

[0039]

[0040] Here To modify the linear unit of the correction. Through the above operations, the bneck block generates a feature map through optimization by the SE module and IRB.

[0041] The network specifications of MobilenetV3-Large are shown in Table 1. `exp size` represents the number of channels after dimensionality increase in the first convolutional layer within a module. `#out` represents the number of output channels. `SE` indicates whether the Squeeze-and-Excitation (SE) module is integrated into the corresponding bneck unit; `NL` indicates the type of nonlinear activation function used, where `HS` represents the h-swish activation function and `RE` represents the ReLU activation function; `NBN` (No Batch Normalization) means no batch normalization is included; and `stride` is denoted by `s`. MobilenetV3-Large outputs a total of 17 feature maps. Higher-level feature maps have a coarser resolution, containing complete semantic and positional information of small targets. These higher-level feature maps then pass high-level semantic information to deeper feature maps through convolutional operations, thereby extracting richer features and containing more detailed information. Therefore, deeper feature maps in MobilenetV3-Large contain richer features of large targets.

[0042] Table 1 MobilenetV3-Large Network Specifications

[0043]

[0044] S3. Construct a dual-channel Feature Pyramid Network (FPN) consisting of a small target feature pyramid network and a large target feature pyramid network. The input to the small target feature pyramid network is the feature maps of layers 2, 4, and 16 of the MobileNetV3-Large network, which are denoted as S1, S2, and S3, respectively. The input to the large target feature pyramid network is the feature maps of layers 8, 13, and 16 of the MobileNetV3-Large network, which are denoted as K1, K2, and K3, respectively.

[0045] The MobilenetV3-Large architecture used in this invention outputs a total of 17 feature maps. Through operations such as convolution, features are continuously extracted from the input image data. Deeper feature maps contain richer semantic information and are more suitable for feature extraction from large targets. Because small targets lose semantic information during feature extraction via convolution, deeper feature maps may not necessarily contain complete semantic information. Conversely, higher-level feature maps will possess more complete semantic information, representing the features of small targets. Therefore, as... Figure 3 As shown, in the dual-channel feature pyramid network, the feature maps of layers 2-4-16 (i.e., layers S1, S2, and S3) from MobilenetV3-Large, which have the most upper layers, are selected to build the FPN architecture adapted to small target feature extraction, corresponding to layers C1, C2, and C3 of the small target channel FPN. Because layers 2 and 4 are upper-level feature maps with coarser resolution, they contain complete semantic and positional information of small targets. They also work with layer 16 (which has the richest detail information) to build the FPN, achieving multi-scale feature extraction and fusion. For feature extraction of large targets, the upper-level feature maps use convolution operations to progressively pass high-level semantic information to the deeper feature maps, thereby extracting richer features and containing much detailed information. Therefore, the deeper feature maps of MobilenetV3-Large, namely layers K1, K2, and K3 (where K3 and S3 are the same feature map layer, i.e., the 16th layer), are selected to build an FPN suitable for large-size target feature extraction, corresponding to layers C1*, C2*, and C3* of the large target channel FPN.

[0046] Whether it's a large-target channel FPN or a small-target channel FPN, after selecting the number of feature map layers for the backbone network output, further processing of the input feature map is required. Taking the small-target channel FPN as an example... Figure 4The internal working mechanism of the FPN for small target channels is shown. The FPN constructs an inverted pyramid structure based on the hierarchical order of the selected feature maps S1, S2, and S3 layers in the original backbone network.

[0047] After the feature maps are extracted from the backbone network, they are fused at multiple levels through FPN to generate feature maps of different resolutions for the detection of small and large targets. The small target feature pyramid network and the large target feature pyramid network have the same structure, both including an inverted pyramid structure. That is, S1 / K1, which is at the top level in the MobileNetV3-Large network, is used as the bottom layer C1 / C1* of the pyramid, S2 / K2, which is in the middle level, is used as the middle layer C2 / C2* of the pyramid, and S3 / K3, which is at the bottom level, is used as the top layer C3 / C3* of the pyramid. Considering that S3 / K3 is at the bottom level in the original backbone network and belongs to the low-resolution feature map, S3 / K3 is first downsampled to generate a new feature map S4 / K4 to improve its spatial resolution and construct a new top layer C4 / C4* of the pyramid. S1 / K1~S4 / K4 are then input into 1×1 convolutional layers for dimensionality compression.

[0048] The compressed S4 / K4 is input into a 3×3 convolutional layer, and then into a prediction layer to obtain the feature map P4 / P4*.

[0049] The compressed S4 / K4 and S3 / K3 are fused element-wise by addition, then fed into a 3×3 convolutional layer, and then into a prediction layer to obtain the feature map P3 / P3*. The 3×3 convolutional layer is used for feature extraction, feature map fusion, and reducing information redundancy.

[0050] The fused S3 / K3 and S2 / K2 are fused using element-wise addition, then fed into a 3×3 convolutional layer, and then into a prediction layer to obtain the feature map P2 / P2*.

[0051] The fused S2 / K2 and S1 / K1 are fused using element-wise addition, then fed into a 3×3 convolutional layer, and then into a prediction layer to obtain feature maps P1 / P1*.

[0052] A prediction layer typically consists of one or more convolutional layers, followed by an output layer to generate the prediction result. Generally, these layers include 1x1 convolutional layers (for channel compression) and Softmax or Sigmoid activation functions for classification or regression.

[0053] Through this series of processing steps, a final set of feature maps is generated, labeled {P1, P2, P3, P4} and {P1*, P2*, P3*, P4*}. These maps match the original feature maps {S1, S2, S3, S4} and {K1, K2, K3, K4} respectively, but have improved multi-scale fusion feature representations, making them more suitable for supporting subsequent Faster-RCNN algorithms.

[0054] S4. Constructing the Faster-RCNN object detection network: The Faster-RCNN object detection network includes a feature extraction network, a Region Proposal Network (RPN), ROI (Region of Interest) pooling layers, and a classification layer; its structure is as follows: Figure 5 As shown in the diagram, the feature extraction network employs the dual-channel feature pyramid network constructed in step S3. The feature map extracted by the feature extraction network is input into the RPN to generate target candidate boxes, which are then fed into the ROI pooling layer. The pooling layer converts candidate boxes of different sizes into a fixed size for further processing. The classification layer is used to classify the target category and regress the bounding boxes based on the features output by the ROI pooling layer. These components work together to achieve fast and accurate detection of potential targets in the image.

[0055] The Faster R-CNN model first feeds the feature map into the RPN network to generate Regions of Interest (ROIs), i.e., target candidate boxes. The RPN uses Softmax to distinguish between positive and negative anchor boxes, and then fine-tunes the anchor boxes using bounding box regression to improve location accuracy. Next, the ROI pooling layer extracts and aggregates information from the input feature map and the predicted proposal boxes to form a feature map for each proposal box. These feature maps are then fed into the classification layer to identify the object category. In the final stage, the Faster R-CNN algorithm determines the category based on the feature maps of the proposal boxes and simultaneously uses bounding box regression to determine the exact location of the detection boxes, achieving accurate target detection.

[0056] exist Figure 5 In this invention, the dual-channel FPN is ported to the Faster-RCNN model and used as the feature extraction network to replace the original feature extraction network. The feature maps output by the large target channel FPN and the small target channel FPN are passed in parallel to the RPN layer and the ROI pooling layer for further processing.

[0057] The region proposal network consists of two parts: one part uses Softmax and performs a transpose (Reshape) before and after Softmax to distinguish between positive and negative anchor points; the other part is responsible for calculating the bounding box regression of the anchor points. Then, the proposal layer integrates the positive sample anchor points and their corresponding bounding box regression offsets into the final proposal box, and filters out proposal boxes that are too small or exceed the image boundaries. Specifically, the processing includes the following steps:

[0058] S41. Input the 8 feature maps (P1, P2, P3, P4, P1*, P2*, P3*, P4*) output by the dual-channel FPN into the region proposal network respectively;

[0059] S42. A 3×3 convolutional layer is used to extract features from the input feature map to integrate information from local regions, preparing for subsequent anchor point classification and bounding box regression. The 3×3 convolutional layer is a standard convolutional operation, its main function being to further extract features from the input feature map to better generate region proposals. Through this convolutional layer, the network can integrate information from local regions, preparing for subsequent bounding box regression and classification tasks.

[0060] S43. Input the feature map output from the 3×3 convolutional layer into two 1×1 convolutional layers respectively;

[0061] The first 1×1 convolutional layer is used for binary classification at anchor points, determining whether the anchor point region contains a foreground object. This 1×1 convolutional layer generates the corresponding classification score. Then, the output of the first 1×1 convolutional layer is transposed (reshaped), followed by a softmax operation, and then transposed again. The purpose of the first reshape is to transform the dimension of the feature map, creating a separate dimension of 2 to facilitate the softmax operation on that dimension. The second reshape restores the original shape, yielding the final classification score.

[0062] The second 1×1 convolutional layer is used for the bounding box regression task, which predicts the precise location of each anchor. By regressing (outputting translation and scaling parameters), the network can adjust the position of the anchors to make them closer to the actual target box.

[0063] S44. Based on the classification score and the regression results output by the second 1×1 convolutional layer, the final region proposals are generated. These proposal boxes are then sent to the ROI Pooling layer for further processing.

[0064] ROI pooling transforms each generated proposal box into a fixed size through pooling. The specific steps are: 1. Transform the proposal boxes onto a feature map; 2. Divide each proposal box on the feature map into a grid; 3. Perform Max Pooling on each part of the grid.

[0065] The classification layer is responsible for determining the object category (e.g., red blood cells, white blood cells, platelets, etc.) corresponding to each proposal box. This process involves using the feature map generated from the proposal boxes. The classification layer consists of fully connected layers and a softmax function to produce the predicted probabilities for each category, i.e., cls_prob. Faster-RCNN further employs bounding box regression techniques to calculate the position adjustment bbox_pred for each proposal box, aiming to provide more accurate object localization boxes.

[0066] This invention uses MobileNetV3-Large as the backbone network to build a dual-channel FPN and uses Faster-RCNN as the object detection algorithm for verification. AP and AR are both performance metrics used to evaluate object detection algorithms, where AP (Average Precision) represents the average precision value at different IOU levels. S For the average accuracy of small targets, AP L This represents the average precision for a large target. Average Recall (AR) represents the average recall rate. Intersection over Union (IOU) represents the ratio of the overlap between predicted and ground truth boxes to the total overlap, and can be expressed as:

[0067]

[0068] The area of ​​overlap here refers to the area of ​​the intersection between the predicted bounding box and the actual bounding box, while the area of ​​union refers to the total area covered by both. Therefore, a higher Intersection over Union (IOU) value means a higher degree of overlap between the predicted and actual bounding boxes. When the predicted and actual bounding boxes are perfectly aligned and seamlessly overlapped, the IOU reaches its maximum value of 1, indicating that the model's prediction has achieved perfect accuracy. In this invention, the threshold for IOU detection is set from 0.5 to 0.95, and the average of ten IOUs with intervals of 0.05 is used as the evaluation metric AP. Faster-RCNN with MobileNetV3-Large as the backbone network is used as the model for testing on a peripheral blood image dataset.

[0069] The backbone network possesses feature maps at many different levels, and it's not necessary to use all of them when building the FPN. Choosing different feature maps significantly impacts the results. Therefore, this method conducted multiple experiments and tested the target detection accuracy using different levels of FPN, consistently extracting upper-layer feature maps to build the FPN feature extraction network for small targets, and using deeper feature maps to build the FPN feature extraction network for large targets. The target detection results of some feature maps used to build the FPN are shown in Table 2.

[0070] Table 2. Target detection results of FPN constructed from partial feature maps

[0071]

[0072] Experimental results validated the approach of this invention. The FPN built using upper-layer feature maps significantly improves the recognition of small targets, while building an FPN using deeper feature maps effectively improves the detection accuracy of large targets. Therefore, based on the experimental results of this step, this invention determines the feature map parameters of the aforementioned small target channel FPN to be 1-3-16 layers, and the feature map parameters of the aforementioned large target channel FPN to be 8-12-16 layers.

[0073] The following verifies the effectiveness of the proposed method in feature extraction using a dual-channel FPN on a peripheral blood image dataset, such as... Figure 1 As shown in step e on the right. Consider feature extraction network schemes based on different backbone networks, and compare single-channel FPN and dual-channel FPN schemes. To compare with other mainstream backbone networks, this invention re-experimentsed and obtained target detection results based on feature extraction networks of other backbone networks, and the dual-channel FPN feature extraction network also adopted the same backbone network accordingly, for example, replacing MobileNetV3-Large with ResNet series networks, etc. Table 3 shows all experimental results. For example, when using dual-channel FPN to replace the traditional feature extraction network, using the Faster-RCNN model for target detection, and using MobileNetV3-Large as the backbone network, AP... S AP L The improvements are 7% and 4% respectively. This method's model represents a comprehensive improvement in detecting both small and large targets, achieving competitive performance.

[0074] Table 3 Comparison of target detection results between dual-channel FPN and single-channel FPN

[0075]

[0076] at last, Figure 6This paper visually demonstrates the performance comparison between the dual-channel FPN feature extraction network and the traditional single-channel FPN network in object detection. When detecting both small and large targets, the dual-channel FPN produces satisfactory results, outperforming the traditional single-channel FPN, especially for small targets, where the performance improvement is more significant. This is because the dual-channel FPN provides global contextual information to the subsequent object detection model through multi-scale features and performs separate feature extraction for targets of different sizes, improving the discriminative ability of the feature extraction network. This allows the Faster R-CNN object detection network to learn better, resulting in better detection performance.

[0077] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for feature extraction from peripheral blood image data based on a dual-channel feature pyramid, characterized in that, Includes the following steps: S1. Acquire peripheral blood images and scale them to a fixed size; S2. The MobileNetV3-Large network is used as the backbone network to extract feature maps from the original data images. The MobileNetV3-Large network includes two two-dimensional convolutional layers and 15 bneck blocks located between the two two-dimensional convolutional layers. The two-dimensional convolutional layers and the 15 bneck blocks are connected sequentially. Each convolutional layer and bneck block outputs a feature map, resulting in 17 feature maps. The 1st, 2nd, 3rd, 4th, 8th, 9th, 10th, and 11th bneck blocks respectively include 1×1 convolution, DW convolution, and 1×1 convolution; The 5th, 6th, 7th, 12th, 13th, 14th, and 15th bneck blocks consist of a 1×1 convolution, a DW convolution, an SE module, and another 1×1 convolution, respectively. Each bneck block first uses a 1×1 convolution to increase the dimensionality of the input tensor, then uses a DW convolution to extract features; subsequently, the SE module processes the extracted features, and finally, a 1×1 convolution is used to reduce the dimensionality of the features output by the SE module to obtain the output tensor. The SE module consists of a pooling layer, a fully connected layer with a ReLU activation function, and a fully connected layer with an H-Swish activation function, connected in sequence. S3. Construct a dual-channel feature pyramid network consisting of a small target feature pyramid network and a large target feature pyramid network; The input to the small target feature pyramid network is the feature maps of layers 2, 4, and 16 of the MobileNetV3-Large network, which are denoted as S1, S2, and S3, respectively; the input to the large target feature pyramid network is the feature maps of layers 8, 13, and 16 of the MobileNetV3-Large network, which are denoted as K1, K2, and K3, respectively. The small target feature pyramid network and the large target feature pyramid network have the same structure, both including an inverted pyramid structure. That is, S1 / K1, which is at the top level in the MobileNetV3-Large network, is used as the bottom layer C1 / C1* of the pyramid, S2 / K2, which is in the middle level, is used as the middle layer C2 / C2* of the pyramid, and S3 / K3, which is at the bottom level, is used as the top layer C3 / C3* of the pyramid. First, S3 / K3 is downsampled to generate a new feature map S4 / K4, and a new top layer C4 / C4* of the pyramid is constructed. Then, S1 / K1~S4 / K4 are input into 1×1 convolutional layers for dimensionality compression. The compressed S4 / K4 is input into a 3×3 convolutional layer, and then into a prediction layer to obtain the feature map P4 / P4*. The compressed S4 / K4 and S3 / K3 are fused together using element-wise addition, then fed into a 3×3 convolutional layer, and then into a prediction layer to obtain the feature map P3 / P3*. The fused S3 / K3 and S2 / K2 are fused using element-wise addition, then fed into a 3×3 convolutional layer, and then into a prediction layer to obtain the feature map P2 / P2*. The fused S2 / K2 and S1 / K1 are fused using element-wise addition, then fed into a 3×3 convolutional layer, and then into a prediction layer to obtain feature maps P1 / P1*. S4. Constructing the Faster-RCNN object detection network: The Faster-RCNN object detection network includes a feature extraction network, a region proposal network, a ROI pooling layer, and a classification layer. The feature extraction network uses the dual-channel feature pyramid network constructed in step S3. The feature map extracted by the feature extraction network is input into the RPN to generate object candidate boxes, which are then input into the ROI pooling layer. The pooling layer converts candidate boxes of different sizes into a fixed size. The classification layer is used to classify the object category and regress the bounding boxes based on the features output by the ROI pooling layer. The regional proposal network specifically includes the following processing steps: S41. Input the 8 feature maps output by the dual-channel FPN into the region proposal network respectively; S42. Extract features from the input feature map using a 3×3 convolutional layer; S43. Input the feature map output from the 3×3 convolutional layer into two 1×1 convolutional layers respectively. The first 1×1 convolutional layer is used for the binary classification task of anchor points to determine whether the anchor point region contains the foreground target. Then, the output of the first 1×1 convolutional layer is transposed once, then a softmax operation is performed, and then it is transposed again to obtain the final classification score. The second 1×1 convolutional layer is used for the bounding box regression task, which is to predict the precise location of each anchor point; S44. Generate final region suggestions based on classification scores and regression results. These suggestion boxes will be sent to the ROI pooling layer for processing.

Images