Improved convolution integral neural network model for leukocyte calculation

Abstract

The invention discloses an improved convolution neural network model for leukocyte calculation, and relates to the field of artificial intelligence and medical image processing. Aiming at the problems of inaccurate feature extraction, strong background noise interference and insufficient multi-scale feature fusion in a traditional leukocyte calculation method, the model realizes fine processing and feature enhancement of a medical image through a multi-level modular design; the image preprocessing module is used for completing white blood cell positioning and image adaptive division and enhancement; multi-scale edge, texture and gray features are fused by means of an image feature extraction module, and a composite feature vector is constructed; through multi-scale convolution, a double-layer attention mechanism and convolution layer calculation of the improved convolution neural network model module, precise focusing of white blood cell contours and cell nucleus features, background noise suppression and calculation efficiency optimization are realized; and finally, a calculation result is output through a visual terminal, so that the accuracy and robustness of leukocyte counting are remarkably improved, and efficient technical support is provided for medical microscopic image analysis.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and medical image processing, and in particular to an improved convolutional neural network model for white blood cell calculation. Background Art

[0002] In the field of medical diagnosis, white blood cell count and classification, as key clinical test indicators, are crucial for the diagnosis and monitoring of inflammation, infection, and hematologic diseases. Traditional white blood cell counting methods rely primarily on manual microscopic observation and counting. This method is not only time-consuming and labor-intensive, with low detection efficiency, but is also susceptible to subjective factors of the examiner and carries significant individual errors, making it difficult to meet the demands of modern medicine for rapid and accurate diagnosis. With the development of computer vision and deep learning technologies, white blood cell counting methods based on convolutional neural networks (CNNs) have gradually become a research hotspot. However, existing CNN models still face many challenges in practical applications. In the image preprocessing stage, due to the problems of cell overlap, complex background, and uneven illumination in medical microscopic images, traditional image segmentation and enhancement methods are difficult to accurately locate white blood cells, resulting in deviations in subsequent feature extraction. In terms of feature extraction, single-scale feature extraction cannot effectively capture the key edge and texture features of white blood cells at different resolutions, and lacks the fusion of semantic information and spatial structure of cell membrane contours and cell nuclear texture, which reduces the model's accuracy in recognizing white blood cells. In terms of model structure, the traditional convolution module has low computational efficiency and cannot balance computational accuracy and speed. At the same time, when faced with interference from red blood cell-irrelevant areas, it lacks an effective feature suppression mechanism, which causes background noise to have a serious impact on the expression of white blood cell features, resulting in insufficient robustness of the model. Furthermore, existing models suffer from irrational allocation of computing resources when processing high-resolution medical images, making it difficult to achieve efficient processing while maintaining computational accuracy, limiting their feasibility for large-scale clinical applications. Therefore, an improved model that can overcome these shortcomings and achieve accurate and efficient WBC counting is urgently needed to meet the growing demands of the medical image processing field. Summary of the Invention

[0003] In response to the shortcomings of existing technologies, this paper discloses an improved convolutional neural network model and method for white blood cell counting, which involves the fields of medical image processing and artificial intelligence technology. This model addresses the problems of inaccurate cell positioning, incomplete feature extraction, strong background noise interference, and low computational efficiency in traditional white blood cell counting methods. Through multi-level modular collaborative design, it achieves refined processing of medical images and in-depth mining of white blood cell features. Utilizing the adaptive multi-dimensional enhancement technology in the image preprocessing module, combined with the white blood cell positioning module and the overlapping area compensation algorithm, accurate identification and segmentation of white blood cells in complex backgrounds are achieved. With the help of the multi-scale fusion architecture of the image feature extraction module, a composite feature vector containing edge, texture, and grayscale co-occurrence matrices is constructed. Through the parallel heterogeneous convolution kernel group and double-layer attention mechanism of the improved convolutional neural network model module, the focus enhancement of key white blood cell features and the effective suppression of background noise are achieved. Finally, through the dynamic resource allocation and computational acceleration strategy of the convolutional hierarchical computing module, the accuracy and efficiency of white blood cell counting are significantly improved, providing reliable technical support for medical diagnosis.

[0004] In order to achieve the above technical effects, the present invention adopts the following technical solutions: An improved convolutional neural network model for white blood cell counting, comprising: An image preprocessing module is used to acquire a medical image containing leukocytes, wherein the image preprocessing module includes a leukocyte positioning module, an image segmentation module, a window size adaptation module, an overlapping area compensation module, and an adaptive image enhancement module; The image feature extraction module generates a multi-scale representation of the image using a Gaussian pyramid, extracting edge Canny operator, texture, and gray-level co-occurrence matrix features at different resolutions. The preprocessed sub-images are then normalized and fed into the bottom layer of the improved convolutional neural network model module. Basic visual features of cell membrane contours or cell nuclear textures are extracted and concatenated along the channel dimension to form a composite feature vector containing both semantic information and spatial structure. Improved convolutional neural network model module, including: Multi-scale convolution module, including parallel heterogeneous convolution kernel group, computation acceleration module, adaptive adjustment module, global average pooling module, global maximum pooling module and spatial attention mask module; A dual-layer attention mechanism module, including a channel attention mechanism and a spatial attention mechanism. The dual-layer attention mechanism module extracts contextual information in the channel dimension through global average pooling and global maximum pooling, and generates a channel attention weight of 1×1×C through a two-layer fully connected network to suppress the feature response of red blood cell-irrelevant areas. Secondly, a spatial dimension convolution operation (7×7 convolution) is performed to generate a spatial attention mask H×W×1, focusing on the white blood cell outline and cell nucleus area. Finally, the two attention weights are superimposed on the feature map through element-wise multiplication to enhance the key features of white blood cells, reduce background noise interference, and improve the discriminability of feature expression. A convolution layer calculation module, comprising a calculation acceleration module and a convolution expansion module connected to the calculation acceleration module, a channel correlation coefficient calculation module, and an implicit feature conversion module; A visualization terminal, used to display the data information output by the improved convolutional neural network model module; The output end of the image preprocessing module is connected to the input end of the image feature extraction module, the output end of the image feature extraction module is connected to the input end of the improved convolutional neural network model module, and the output end of the improved convolutional neural network model module is connected to the input end of the visualization terminal.

[0005] As a further embodiment of the present invention, the white blood cell positioning module pre-trains three semantic labels of white blood cells, red blood cells, and background through the lightweight semantic segmentation model U-NetLite, removes red blood cell adhesion interference and connected domain analysis through morphological opening operation to screen cell nucleus features, and achieves accurate positioning of white blood cell ROI, by fusing the grayscale histogram threshold OTSU and texture feature LBP operator with the semantic segmentation results; The image segmentation module constructs a variable window library based on the 5-20 μm white blood cell size distribution and corresponding pixels of 50-200 px @ 1000× magnification. It calculates the minimum circumscribed circle diameter of white blood cells using the Canny operator edge detection algorithm and automatically matches 32×32 small cells, 64×64 medium cells, and 128×128 large cell windows. It also performs edge pixel mirroring compensation on cells across windows and restores the grayscale continuity of the cut area through bilinear interpolation. The window size adaptation module uses an intelligent adjustment algorithm based on diameter-pixel mapping. It establishes a mapping relationship between the actual diameter of white blood cells and the pixel size through microscopic image calibration to achieve physical adaptation of the window size. When the cell density in the ROI area is greater than 50 cells / mm², the window step size is automatically reduced from 40px to 20px to avoid missing high-density cells. The overlapping area compensation module calculates the gradient amplitude distribution of the overlapping area based on the weighted gradient fusion algorithm. When the gradient is greater than the threshold, a weight of 0.3 is applied to the edge pixels, and when the gradient is less than or equal to the threshold, a weight of 0.7 is applied to the center pixels to suppress the generation of artifacts. The adaptive image enhancement module calculates texture complexity through local binary pattern (LBP). When the noise density detects that the salt and pepper noise is greater than 10% or the Gaussian noise σ is greater than 15, a dynamic selection enhancement algorithm based on bilateral filtering is used to automatically match elliptical or polygonal structural elements according to the morphological parameters of white blood cells for expansion and corrosion. The roundness of 0.8-1.2 is for lymphocytes, and 0.6-0.8 is for neutrophils.

[0006] As a further embodiment of the present invention, an intelligent adjustment algorithm based on diameter-pixel mapping adjusts the feature information of the image by a weighted fusion coefficient in image preprocessing. The weighted fusion coefficient is calculated by the following formula: In formula (1), represents the edge strength of the i-th pixel in the high-frequency detail layer, represents the regional grayscale change of the i-th pixel in the low-frequency background layer, and N is the total number of image pixels. As a further embodiment of the present invention, the adaptive image enhancement module is based on a multi-scale feature collaborative optimization architecture, adding a coding module and a data classification module to fuse a generative adversarial network (cGAN), a spatial attention mechanism, and a feature pyramid network to achieve intelligent improvement of the clarity of white blood cell images; In high-frequency detail layer processing, a conditional generative adversarial network (cGAN) is introduced, using prior data on white blood cell morphology as conditional input to drive the generator to learn and produce ideal edge-enhanced images that conform to real cell morphology. The unsharp mask parameters are jointly optimized using a feature matching loss function and an adversarial loss function, enabling the edge enhancement process to both enhance white blood cell contour details and effectively suppress noise and artifacts. In processing the low-frequency background layer, a spatial attention-weighted histogram equalization strategy is adopted. Based on a convolutional neural network, feature maps of background subregions are extracted. A dual-channel attention mechanism is used to calculate subregion weights from both the channel and spatial dimensions. The histogram equalization strength is dynamically adjusted to accurately suppress over-enhancement of non-critical background areas and highlight the contrast of potential white blood cell areas. In the multi-scale fusion link, the multi-resolution processing framework based on the feature pyramid network FPN performs adaptive enhancement on the high-frequency detail layer and low-frequency background layer at different scales respectively. Through the top-down and horizontally connected feature fusion paths, the organic unity of local detail enhancement and global structure optimization is achieved, effectively balancing noise suppression and detail retention, and significantly improving the visual distinction between white blood cells and the background.

[0007] As a further embodiment of the present invention, in the parallel heterogeneous convolution kernel group, the convolution layers with different convolution kernel sizes include 3×3, 5×5 and 7×7 convolution layers, and each convolution layer is connected to a batch normalization layer and an activation function layer; The computing acceleration module includes a tensor processing unit of the NPU and a convolution kernel dimensionality reduction processing unit connected to the tensor processing unit of the NPU, an 8-channel parallel computing unit, and a fixed-point conversion module, wherein the tensor processing unit of the NPU is a multi-scale convolution kernel processor based on the AI-SSC369G chip; The adaptive adjustment module includes a data training classifier, a weight dynamic monitoring module, a weight dynamic adjustment module and a computing power dynamic allocation module; wherein the output end of the data training classifier is connected to the input end of the weight dynamic monitoring module, the output end of the weight dynamic monitoring module is connected to the input end of the weight dynamic adjustment module, and the output end of the weight dynamic adjustment module is connected to the input end of the computing power dynamic allocation module; The global average pooling module calculates the global mean of each channel, retaining the average grayscale of the leukocyte cytoplasm and the average temperature of the device surface in the feature map to suppress local noise interference; The global maximum pooling module extracts the global maximum value of each channel, highlights the high grayscale value area of the cell nucleus or the abnormal hot spot of the device in the feature map, and strengthens the discriminative power of key features; The spatial attention mask module generates an attention mask H×W×1 through a convolution operation in the spatial dimension, and the focus target area contains at least: Edge enhancement: A 7×7 convolution with a dilationrate of 2 is used to expand the receptive field to capture the edges of white blood cell contours or the location of device terminals of the target structure. Saliency detection: Combining Sobel edge detection with grayscale threshold segmentation and Otsu algorithm, a binary saliency map is generated. After Gaussian blurring with a kernel size of 5×5, the map is converted into a continuous mask with values between 0 and 1. Mask fusion: The mask is multiplied element-by-element with the feature map, which increases the feature response of the target area by 1.5 times and attenuates the response of the background area to less than 0.6 times, thereby improving the recognition rate of abnormal areas by 22%.

[0008] As a further embodiment of the present invention, the working method of the global average pooling module is: (S1) Perform average pooling and maximum pooling on the input feature map in the channel dimension to obtain two different two-dimensional feature maps; (S2) The two 2D feature maps are concatenated in the channel dimension and subjected to a convolutional layer for feature fusion. (S3) The fused feature map is passed through the Sigmoid activation function to obtain the spatial weight matrix.

[0009] As a further embodiment of the present invention, the global average pooling module calculates the global mean of each channel, and the temperature coefficient adjustment factor is calculated by the following formula: In formula (2), are the preset adjustment parameters. is a complexity measure of the white blood cell image, obtained by calculating the degree of overlap and edge blur of different cells in the image. As a further embodiment of the present invention, (1) the working principle of the data training classifier is: the 128-dimensional edge features obtained by OTSU image segmentation, the control features of the 32-dimensional fuzzy logic output and the 64-dimensional original sensor data are mapped to a 256-dimensional unified feature space through a fully connected layer, and an improved cross entropy loss function is used to enhance the learning of difficult samples; the loss function formula is: In formula (3), is the category balance factor For the focus parameters, abnormal samples is the predicted probability, is the true label; (2) The working method of the weight dynamic monitoring module is: to monitor the distribution characteristics of the neural network weight matrix in real time through a multi-dimensional weight anomaly detection mechanism, and to identify weight anomalies caused by noise interference or data drift; the core includes: calculating the information entropy of the weight gradient of each layer , the function reflecting the degree of chaos of gradient distribution is: In formula (4), when hour, is the historical average, is the standard deviation, and the weight update is judged to be abnormal; is a gradient element; is the normalization term, and the absolute value of the gradient Normalized to probability distribution; L2 norm of monitoring weight matrix norm , when the dynamic threshold Initial norm, identifying weight explosion caused by overfitting; generating weight anomaly signals , triggering the weight dynamic adjustment module to perform parameter correction; (3) The working method of the weight dynamic adjustment module is: first, the multi-objective optimization objective function is constructed: a three-dimensional benefit function including feature extraction accuracy, computational efficiency, and overfitting suppression is defined to achieve the global optimal scheduling of weight updates: In formula (5), the characteristic benefit term : Based on significant differences in leukocyte nuclear and cytoplasmic characteristics; component benefit value , benefit term based on component analysis; regularized benefit term : benefit term based on regularization method; They are The weight coefficient of ; Design channel-specific benefit function: In formula (6), is the characteristic channel of the cell nucleus, It is a cytoplasmic characteristic channel. is the channel weight, The grayscale contrast of the nucleus / cytoplasm area is calculated in real time by the leukocyte positioning results of the preprocessing module and is set as Prioritize feature extraction to ensure that the weight of key features such as the edge of the cell nucleus and cytoplasm granularity is increased by 30%; computational efficiency : Suppress the weight update of redundant feature channels, definition: In formula (7), is the number of inactivated feature channels, The computing power threshold of the edge computing unit is ≥80%. When the computing power utilization rate is greater than 80%, the weight update step size of non-critical channels is automatically reduced by 50%. is the total number of all feature channels in the model, used as the denominator to normalize FreeChan to a ratio , represents the proportion of inactivated channels; The number of floating-point operations, the total number of floating-point operations required for the model to perform one forward / backward propagation; overfitting suppression term : Introducing weight norm constraints and feature correlation penalties: In formula (8), is the regularization coefficient, by calculating the Pearson correlation coefficient between channels , penalizing highly correlated redundant features, thus improving the independence of feature channels by 40%; designing a weight update rule with timing constraints: if the current layer is the cell nucleus contour extraction layer, and the cell nucleus position confidence output by the previous layer is greater than 0.9, then the priority promotion mechanism is triggered: In formula (9), is the gain coefficient, which increases the weight update speed of the cell nucleus feature channel by 50%, ensuring the rapid convergence of key edge features; As the baseline update amount, adjustments are made on this basis; is the adjusted weight update amount, which is used to update the actual parameters, increase the weight update amplitude of the cell nucleus feature channel, and accelerate convergence; The confidence score of channel C for the cell nucleus feature is usually [0, 1]. A higher value indicates a more significant response of the channel to the cell nucleus feature. Delay-sensitive weight correction: When a white blood cell classification delay of > 20ms is detected, the edge computing unit's computing power monitoring module provides real-time feedback and automatically activates the lightweight adjustment mode. In formula (10), only the gradient update of the key feature channel is retained, and the weight of the non-key channel is decayed by 5%. In the scenario of limited computing power, the classification accuracy loss is still kept less than 3%; is the weight of channel c, and different update rules are executed depending on whether the channel belongs to the key channel; is the learning rate, which controls the magnitude of the key channel weight update; is the gradient of channel c, which determines whether the key channel performs enhanced update; 0.95 is the attenuation coefficient, which attenuates the weight of non-key channels and weakens their influence; to avoid feature extraction failure caused by abnormal weights, three layers of resource capacity constraints are established: Channel capacity constraint: The weight update range of each feature channel does not exceed the medical prior threshold defined in the preprocessing stage: ; When the lower limit of the weight of the cell nucleus edge feature channel is set to 0.8; the upper limit of the background noise channel weight is set to 0.3; computing power capacity constraint: the floating-point operation amount of the single-layer weight update does not exceed the edge computing power threshold: In formula (11), The computing power utilization safety threshold is set. When the threshold is approached, it automatically switches to fixed-point calculation, replacing 32-bit with 16-bit calculation, which increases the calculation speed by 3 times and the accuracy loss is less than 1%. Measures the computational complexity of model training and the number of floating-point operations during model update. Timing constraints: weight updates for dependent features must be initiated after the previous features converge. In formula (12), To reduce feature dependency delays, timestamps are used to ensure orderly optimization of hierarchical features and avoid classification errors caused by feature misalignment. To mark the starting moment of cytoplasm feature training, the cytoplasm feature channel Timestamp when optimization started; To mark the completion time of cell nucleus feature training, End timestamp optimized for the cell nucleus feature channel; Microscope resolution adaptation: Dynamically adjust the weight update step size according to the objective lens magnification 40×-400×: In formula (13), To address the insensitivity of weight optimization caused by the reduction of feature size under high-magnification microscopes, the accuracy of cell nucleus segmentation under a 400× objective lens was improved by 9%. The step size for updating the dynamically adjusted weights is adaptively adjusted according to the objective lens magnification to ensure optimization efficiency at different resolutions; As the basic step length, as the benchmark value, through Zoom in or out; (4) The working method of the dynamic allocation module of computing power is as follows: The working principle of the dynamic allocation module of computing power is to build a multi-dimensional decision model based on task complexity and resource load, and integrate the n-dimensional task parameters of image resolution R, feature map size S, number of network layers L and CPU usage. , GPU memory occupancy The m-dimensional resource parameters are mapped to the resource allocation weight vector W through the linear weighting function F; the weight calculation adopts the dynamic adjustment formula: In formula (14), Assign a weight to the resource corresponding to the i-th parameter, is the parameter importance adjustment factor, is the normalized value of the i-th parameter, is the nonlinear adjustment coefficient; when it is detected that the task parameter exceeds the threshold When resource parameters reach the warning state, the dynamic adjustment mechanism is activated to adjust Strengthen the allocation of key task resources to achieve dynamic balance and efficiency optimization of computing resources. As a further embodiment of the present invention, the computing acceleration module includes a convolution expansion module, a channel correlation coefficient calculation module, and an implicit feature conversion module; the convolution expansion module uses a dilated convolution to dynamically adjust the receptive field of the convolution kernel through the expansion rate d, and the calculation formula is: In formula (15), is the convolution output feature map, is the input feature map, is the convolution kernel, n is the input feature map size, m is the convolution kernel size, s is the convolution step size, and d is the dilation rate. Through dynamic adjustment, the network's perception range of white blood cell morphological characteristics is expanded without increasing the number of parameters, reducing repeated calculations.

[0010] The channel correlation coefficient calculation module uses the Pearson correlation coefficient to evaluate the correlation between feature channels and accelerates the calculation by penalizing redundant channels. The calculation formula is: In formula (16), For channel The correlation coefficient between Channel The i-th eigenvalue, is the mean of the corresponding channel eigenvalues, and N is the number of eigenvalues. By calculating the correlation coefficients of all channel pairs, the weights of highly correlated channels are identified and reduced to reduce the amount of computation. The implicit feature conversion module decomposes the high-dimensional feature tensor T into a combination of low-dimensional tensors based on tensor decomposition. The formula is as follows: In formula (17), T is the original high-dimensional feature tensor, are the elements of the core tensor, is a factor matrix, is the rank of the corresponding dimension, and ∘ represents the tensor outer product operation. By reducing the tensor rank and the feature dimension, the feature calculation and transmission process is accelerated, thereby improving the overall computational efficiency of the model.

[0011] Positive beneficial effects During the image preprocessing stage, the image preprocessing module integrates white blood cell positioning, window adaptation, and overlapping area compensation functions, which can accurately identify white blood cells in complex backgrounds, effectively solving the problems of cell overlap and uneven illumination, and improving image segmentation accuracy by over 30%, laying a solid foundation for subsequent calculations. The image feature extraction module, using Gaussian pyramids and multi-dimensional feature fusion strategies, can comprehensively capture the key edge and texture features of white blood cells at different resolutions. The constructed composite feature vector contains rich semantic and spatial structural information. Compared with traditional methods, the feature expression capability is improved by approximately 40%, significantly enhancing the model's ability to recognize white blood cells. In the improved convolutional neural network model module, the multi-scale convolution module takes into account the extraction of local details and global structural features of white blood cells through parallel heterogeneous convolution kernel groups and adaptive adjustment mechanisms. At the same time, the computing acceleration module optimizes the calculation process, which improves the overall computing efficiency by more than 50%; the two-layer attention mechanism module uses a dual attention mechanism of channel and space to accurately focus on the key areas of white blood cells, suppress irrelevant interference from red blood cells, significantly enhance the discriminability of feature expression, and effectively reduce the misjudgment rate; the convolutional layered calculation module uses channel correlation coefficient calculation and implicit feature conversion to reduce redundant calculations while ensuring calculation accuracy, further improve computing efficiency, and reduce model operation resource consumption. The visualization terminal intuitively presents the calculation results, allowing medical experts to quickly obtain information on white blood cell count, classification, and morphological characteristics. Compared with traditional manual analysis, diagnostic efficiency is increased by over 80%, and the results are more stable and reliable. Through multi-module collaborative innovation, this model significantly improves the accuracy, efficiency, and reliability of white blood cell counting, providing a more efficient and precise technical means for medical diagnosis. It is of great significance to the development of medical image processing technology and the improvement of clinical diagnostic capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is a schematic diagram of the overall module flow of an improved convolutional neural network model for white blood cell counting according to the present invention; Figure 2 This is a workflow diagram of a global average pooling module of an improved convolutional neural network model for white blood cell counting according to the present invention; Figure 3 This is a workflow diagram of an adaptive adjustment module of an improved convolutional neural network model for white blood cell counting according to the present invention; Figure 4This is a workflow diagram of a calculation acceleration module of an improved convolutional neural network model for white blood cell calculation according to the present invention; DETAILED DESCRIPTION

[0013] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0014] An improved convolutional neural network model for white blood cell counting, comprising: An image preprocessing module is used to acquire a medical image containing leukocytes, wherein the image preprocessing module includes a leukocyte positioning module, an image segmentation module, a window size adaptation module, an overlapping area compensation module, and an adaptive image enhancement module; The image feature extraction module generates a multi-scale representation of the image using a Gaussian pyramid, extracting edge Canny operator, texture, and gray-level co-occurrence matrix features at different resolutions. The preprocessed sub-images are then normalized and fed into the bottom layer of the improved convolutional neural network model module. Basic visual features of cell membrane contours or cell nuclear textures are extracted and concatenated along the channel dimension to form a composite feature vector containing both semantic information and spatial structure. Improved convolutional neural network model module, including: Multi-scale convolution module, including parallel heterogeneous convolution kernel group, computation acceleration module, adaptive adjustment module, global average pooling module, global maximum pooling module and spatial attention mask module; A dual-layer attention mechanism module, including a channel attention mechanism and a spatial attention mechanism. The dual-layer attention mechanism module extracts contextual information in the channel dimension through global average pooling and global maximum pooling, and generates a channel attention weight of 1×1×C through a two-layer fully connected network to suppress the feature response of red blood cell-irrelevant areas. Secondly, a spatial dimension convolution operation (7×7 convolution) is performed to generate a spatial attention mask H×W×1, focusing on the white blood cell outline and cell nucleus area. Finally, the two attention weights are superimposed on the feature map through element-wise multiplication to enhance the key features of white blood cells, reduce background noise interference, and improve the discriminability of feature expression. A convolution layer calculation module, comprising a calculation acceleration module and a convolution expansion module connected to the calculation acceleration module, a channel correlation coefficient calculation module, and an implicit feature conversion module; A visualization terminal, used to display the data information output by the improved convolutional neural network model module; The output end of the image preprocessing module is connected to the input end of the image feature extraction module, the output end of the image feature extraction module is connected to the input end of the improved convolutional neural network model module, and the output end of the improved convolutional neural network model module is connected to the input end of the visualization terminal.

[0015] This model optimizes the white blood cell counting process from multiple aspects, specifically addresses the shortcomings of traditional methods, and achieves high-precision, high-efficiency, and intelligent medical image analysis. Specific examples are as follows: 1. Solutions to the Problem of Medical Image Preprocessing The image preprocessing module employs a multi-level adaptive enhancement strategy. The white blood cell localization module, based on a deep learning semantic segmentation algorithm, performs pixel-level recognition of medical images containing white blood cells. When red blood cell interference blurs white blood cell boundaries, the module calculates morphological gradients to improve the accuracy of white blood cell outline localization to sub-pixel levels. The window size adaptation module dynamically adjusts the processing window based on the image entropy. When the local entropy value of the image exceeds 0.8 (range 0-1), the window size is automatically reduced to 60% of its original size to ensure complete capture of tiny white blood cells. The overlap compensation module employs a weighted average fusion strategy to process overlapping portions of segmented images, retaining over 95% of edge information and effectively addressing information loss caused by image segmentation. The adaptive image enhancement module automatically enhances image contrast when the image brightness standard deviation is below 15, significantly improving the distinction between white blood cells and background and providing high-quality image data for subsequent processing.

[0016] 2. Solution to the problem of insufficient multi-scale feature extraction The image feature extraction module constructs a multidimensional feature fusion system. Five different image resolution levels are generated using a Gaussian pyramid. Edge features are extracted at each level using the Canny operator. Edge pixels with gradient magnitudes greater than a threshold of 0.3 are considered valid edges, improving edge extraction accuracy by 25% compared to traditional methods. Texture feature extraction uses the gray-level co-occurrence matrix (GLCM) to calculate energy, entropy, contrast, and correlation features at 0°, 45°, 90°, and 135°, generating a 64-dimensional texture feature vector. The preprocessed sub-images are normalized to the range [-1, 1] before being fed into an improved convolutional neural network model. The underlying network extracts basic visual features of the cell membrane outline and nuclear texture, which are then concatenated across the channel dimension to form a 256-dimensional composite feature vector containing both spatial structure and semantic information. Compared to single feature extraction methods, this improves feature representation by 40%, providing rich feature information for white blood cell identification. 3. Solution to the problem of balancing convolution computation efficiency and accuracy The improved convolutional neural network model module utilizes a multi-strategy collaborative optimization scheme. The multi-scale convolution module deploys a parallel heterogeneous convolution kernel group consisting of three sizes: 3×3, 5×5, and 7×7. When processing white blood cell images, small kernels capture local cell details, while large kernels capture overall morphological features. An adaptive adjustment module dynamically allocates computing resources based on feature response strength, improving computational efficiency by 50%. In the two-layer attention mechanism module, the channel attention mechanism extracts contextual information in the channel dimension through global average pooling and global maximum pooling operations. When the feature response value of irrelevant red blood cell regions exceeds that of white blood cell features by 30%, a two-layer fully connected network generates channel attention weights, suppressing the irrelevant region response by more than 60%. The spatial attention mechanism utilizes 7×7 convolution operations to generate a spatial attention mask, increasing the weighting of white blood cell contours and nuclei by 1.5 times, focusing on key feature regions. The computational acceleration module of the convolution layer calculation module uses the Winograd algorithm to optimize convolution operations. Combined with the channel correlation coefficient calculation module, when the correlation coefficient between channels is greater than 0.7, redundant channels are merged, reducing the amount of calculation by 30%. At the same time, the implicit feature conversion module maps high-dimensional features to low-dimensional space, greatly improving computational efficiency while ensuring recognition accuracy. 4. Solution to the problem of unintuitive visualization of calculation results The visualization terminal adopts an interactive multi-dimensional display solution. The white blood cell calculation results output by the improved convolutional neural network model module are not only presented in the form of traditional quantitative statistical tables, but also visualized through three-dimensional reconstruction technology to display the white blood cell morphology. When the user selects a white blood cell sample, the system automatically displays its cell membrane contours, cell nuclear texture details, and annotates key parameters such as cell diameter and nuclear-cytoplasmic ratio. At the same time, it supports comparative analysis of multiple groups of samples, and users can freely switch between images and parameters of different samples, making it easier for medical experts to quickly observe the laws of white blood cell morphological changes. In addition, the visualization terminal also provides a data export function, which can export calculation results in CSV and JSON formats to facilitate subsequent scientific research analysis and clinical report writing. Compared with the traditional text result display method, the information transmission efficiency is increased by 70%, significantly improving the auxiliary decision-making ability of medical diagnosis.

[0017] The above content has described in detail the solutions of the present invention to the problems in the prior art. The following will further detail the specific embodiments of the present invention.

[0018] Furthermore, the white blood cell positioning module pre-trains three types of semantic labels for white blood cells, red blood cells, and background through the lightweight semantic segmentation model U-NetLite, removes red blood cell adhesion interference and connected domain analysis through morphological opening operation to screen cell nucleus features, and achieves accurate positioning of white blood cell ROI, by fusing the grayscale histogram threshold OTSU and texture feature LBP operator with the semantic segmentation results; The image segmentation module constructs a variable window library based on the 5-20 μm white blood cell size distribution and corresponding pixels of 50-200 px @ 1000× magnification. It calculates the minimum circumscribed circle diameter of white blood cells using the Canny operator edge detection algorithm and automatically matches 32×32 small cells, 64×64 medium cells, and 128×128 large cell windows. It also performs edge pixel mirroring compensation on cells across windows and restores the grayscale continuity of the cut area through bilinear interpolation. The window size adaptation module uses an intelligent adjustment algorithm based on diameter-pixel mapping. It establishes a mapping relationship between the actual diameter of white blood cells and the pixel size through microscopic image calibration to achieve physical adaptation of the window size. When the cell density in the ROI area is greater than 50 cells / mm², the window step size is automatically reduced from 40px to 20px to avoid missing high-density cells. The overlapping area compensation module calculates the gradient amplitude distribution of the overlapping area based on the weighted gradient fusion algorithm. When the gradient is greater than the threshold, a weight of 0.3 is applied to the edge pixels, and when the gradient is less than or equal to the threshold, a weight of 0.7 is applied to the center pixels to suppress the generation of artifacts. The adaptive image enhancement module calculates texture complexity through local binary pattern (LBP). When the noise density detects that the salt and pepper noise is greater than 10% or the Gaussian noise σ is greater than 15, a dynamic selection enhancement algorithm based on bilateral filtering is used to automatically match elliptical or polygonal structural elements according to the morphological parameters of white blood cells for expansion and corrosion. The roundness of 0.8-1.2 is for lymphocytes, and 0.6-0.8 is for neutrophils.

[0019] In a specific embodiment, this technology focuses on white blood cell (WBC) counting, with an image preprocessing module at its core, integrating multiple technologies to achieve precise positioning and image optimization. The WBC positioning module utilizes a lightweight U-NetLite model to pre-train three semantic labels. Morphological operations and feature screening are then combined with the OTSU threshold, LBP operator, and semantic segmentation results to accurately locate the WBC ROI. The image segmentation module constructs a variable window library based on WBC size distribution, matches windows of different sizes based on edge detection, and performs edge mirror compensation and interpolation for cells across windows. The window size adaptation module intelligently adjusts the window step size based on cell density using a diameter-pixel mapping algorithm. The overlapping region compensation module utilizes a weighted gradient fusion algorithm to assign weights based on gradient amplitude to suppress artifacts. The adaptive image enhancement module dynamically selects an enhancement algorithm based on texture complexity and noise type using bilateral filtering, combining WBC morphological parameters with structural elements. During implementation, these modules work together to first locate WBCs, then segment and adaptively adjust windows, compensate for overlapping regions, and finally enhance the image. This technology achieves pixel-level precision positioning of white blood cells, with a window size adaptation rate of 98%, complete edge information preservation, effective noise and artifact suppression, enhanced image contrast and detail, and a 60% improvement in the distinction between white blood cells and background, providing high-quality image data for subsequent white blood cell calculations and significantly improving the accuracy and reliability of white blood cell calculations. Furthermore, the intelligent adjustment algorithm based on diameter-pixel mapping adjusts the image feature information through a weighted fusion coefficient during image preprocessing. The weighted fusion coefficient is calculated using the following formula: In formula (1), represents the edge strength of the i-th pixel in the high-frequency detail layer, Represents the regional grayscale change of the i-th pixel in the low-frequency background layer, and N is the total number of image pixels. In a specific embodiment, the technical essence of the intelligent adjustment algorithm based on diameter-pixel mapping is to construct a weighted fusion coefficient by quantifying the characteristic differences between the high-frequency details and the low-frequency background in the image, so as to achieve adaptive adjustment of the image feature information. In formula (1), the numerator is the sum of the edge intensities of all pixels in the high-frequency detail layer of the image, which represents the key information of the white blood cell contour; the denominator combines the high-frequency edge intensity and the regional grayscale change of the pixels in the low-frequency background layer, and the weighted fusion coefficient is obtained by the ratio of the two. This coefficient can dynamically measure the ratio of edge information to background information in the image, and then guide the fusion and enhancement strategy of feature information in image preprocessing, so that the algorithm can accurately adjust the processing method according to the difference in image characteristics. In the image preprocessing stage, the algorithm first decomposes the input medical image into a high-frequency detail layer and a low-frequency background layer. The edge strength of each pixel in the high-frequency detail layer is obtained through the Canny operator edge detection algorithm. , and calculate the regional grayscale change of each pixel in the low-frequency background layer , and the total number of image pixels N is counted. Subsequently, these parameters are substituted into the weighted fusion coefficient formula to calculate the coefficient α. The image feature information processing strategy is dynamically adjusted according to the α value: when the α value is high, it indicates that the image has rich edge information, and the algorithm enhances edge sharpening processing; when the α value is low, that is, the background information accounts for a large proportion, the algorithm focuses on suppressing background noise and highlighting the white blood cell target area, ultimately achieving optimized fusion of image feature information. This algorithm significantly improves the quality of image preprocessing and effectively enhances the distinction between white blood cell targets and background. In actual testing, when processing medical images with complex backgrounds, compared with traditional methods, the clarity of white blood cell edges is improved by 40%, and the background noise suppression rate is improved by 35%. This lays a high-quality data foundation for subsequent white blood cell feature extraction and calculation, and significantly improves the overall accuracy and reliability of the white blood cell calculation model. Assume that a local area of a 10×10 pixel medical image is selected for calculation example. Table 1 shows the calculation process and results of the weighted fusion coefficients for three images with different total pixel counts using the intelligent adjustment algorithm based on diameter-pixel mapping. The data shows that despite the differences in the total pixel counts of Images 1, 2, and 3, as well as the differences in the sum of the edge intensities of the high-frequency detail layer and the sum of the grayscale variations in the low-frequency background layer, the calculated weighted fusion coefficients are relatively close. This indicates that when the ratio of high-frequency detail to low-frequency background is similar in an image, the algorithm can output similar weighted fusion coefficients, thereby adjusting the image's feature information using a similar strategy. This demonstrates the consistency and stability of the algorithm's quantization and adaptive adjustment of image feature information across images of varying sizes.

[0020] Furthermore, the adaptive image enhancement module is based on a multi-scale feature collaborative optimization architecture, and incorporates a coding module and a data classification module to fuse a generative adversarial network (cGAN), a spatial attention mechanism, and a feature pyramid network to achieve intelligent improvement of white blood cell image clarity. In high-frequency detail layer processing, a conditional generative adversarial network (cGAN) is introduced, using prior data on white blood cell morphology as conditional input to drive the generator to learn and produce ideal edge-enhanced images that conform to real cell morphology. The unsharp mask parameters are jointly optimized using a feature matching loss function and an adversarial loss function, enabling the edge enhancement process to both enhance white blood cell contour details and effectively suppress noise and artifacts. In processing the low-frequency background layer, a spatial attention-weighted histogram equalization strategy is adopted. Based on a convolutional neural network, feature maps of background subregions are extracted. A dual-channel attention mechanism is used to calculate subregion weights from both the channel and spatial dimensions. The histogram equalization strength is dynamically adjusted to accurately suppress over-enhancement of non-critical background areas and highlight the contrast of potential white blood cell areas. In the multi-scale fusion phase, based on the multi-resolution processing framework of the Feature Pyramid Network (FPN), adaptively enhances high-frequency detail layers and low-frequency background layers at different scales. Through a top-down and lateral feature fusion pathway, this method achieves an organic integration of local detail enhancement and global structure optimization, effectively balancing noise suppression and detail preservation, significantly improving the visual distinction between white blood cells and the background. In a specific embodiment, the adaptive image enhancement module, centered on multi-scale feature collaborative optimization, integrates cGAN, a spatial attention mechanism, and a feature pyramid network to revolutionize white blood cell image enhancement. The technology essentially constructs a multi-layered, multi-strategy image enhancement system. The high-frequency detail layer utilizes a cGAN and a joint loss function to generate an ideal edge-enhanced image based on prior white blood cell morphology data, accurately optimizes unsharp mask parameters, and balances outline enhancement and noise suppression. The low-frequency background layer utilizes a dual-channel attention mechanism and weighted histogram equalization to dynamically adjust the background enhancement intensity and highlight the white blood cell region. The multi-scale fusion phase utilizes a feature pyramid network to enable cross-scale feature interaction, achieving local and global collaborative optimization. During implementation, the module first decomposes the white blood cell image into a high-frequency detail layer and a low-frequency background layer. During high-frequency processing, the cGAN generator generates edge-enhanced images based on the white blood cell morphology prior, and the unsharp mask parameters are simultaneously optimized using a joint loss function. During the low-frequency processing stage, the convolutional neural network extracts feature maps of background sub-regions, and after calculating the weights using a dual-channel attention mechanism, the degree of histogram equalization is dynamically adjusted. During multi-scale fusion, the feature pyramid network performs top-down and horizontal feature fusion on the enhancement results of each layer to complete image enhancement. This module significantly improves the image enhancement effect. In actual measurements, the clarity of white blood cell contours increased by 60%, the background noise suppression rate reached 75%, the probability of artifacts decreased by 80%, and the image contrast increased by 50%. This greatly improves the visual distinction between white blood cells and the background, providing high-quality images for subsequent white blood cell calculations. The overall model's white blood cell recognition accuracy increased by 35%, effectively promoting the precision of medical image analysis.

[0021] Furthermore, in the parallel heterogeneous convolution kernel group, convolution layers with different convolution kernel sizes include 3×3, 5×5, and 7×7 convolution layers, and each convolution layer is connected to a batch normalization layer and an activation function layer; The computing acceleration module includes a tensor processing unit of the NPU and a convolution kernel dimensionality reduction processing unit connected to the tensor processing unit of the NPU, an 8-channel parallel computing unit, and a fixed-point conversion module, wherein the tensor processing unit of the NPU is a multi-scale convolution kernel processor based on the AI-SSC369G chip; The adaptive adjustment module includes a data training classifier, a weight dynamic monitoring module, a weight dynamic adjustment module and a computing power dynamic allocation module; wherein the output end of the data training classifier is connected to the input end of the weight dynamic monitoring module, the output end of the weight dynamic monitoring module is connected to the input end of the weight dynamic adjustment module, and the output end of the weight dynamic adjustment module is connected to the input end of the computing power dynamic allocation module; The global average pooling module calculates the global mean of each channel, retaining the average grayscale of the leukocyte cytoplasm and the average temperature of the device surface in the feature map to suppress local noise interference; The global maximum pooling module extracts the global maximum value of each channel, highlights the high grayscale value area of the cell nucleus or the abnormal hot spot of the device in the feature map, and strengthens the discriminative power of key features; The spatial attention mask module generates an attention mask H×W×1 through a convolution operation in the spatial dimension, and the focus target area contains at least: Edge enhancement: A 7×7 convolution with a dilationrate of 2 is used to expand the receptive field to capture the edges of white blood cell contours or the location of device terminals of the target structure. Saliency detection: Combining Sobel edge detection with grayscale threshold segmentation and Otsu algorithm, a binary saliency map is generated. After Gaussian blurring with a kernel size of 5×5, the map is converted into a continuous mask with values between 0 and 1. Mask fusion: The mask is multiplied element-by-element with the feature map, which increases the feature response of the target area by 1.5 times and attenuates the response of the background area to less than 0.6 times, thereby improving the recognition rate of abnormal areas by 22%.

[0022] In a specific embodiment, the multi-scale convolution module of the improved convolutional neural network model operates through the collaborative operation of multiple components, revolutionizing the extraction and calculation of white blood cell features. The essence of this technology lies in the construction of an integrated architecture of "heterogeneous convolution + accelerated computing + adaptive control + feature focusing": Parallel heterogeneous convolution kernel groups capture multi-scale features with convolution layers of different sizes, and are combined with batch normalization and activation functions to stabilize training; the computational acceleration module, relying on the NPU tensor processing unit, combines convolution kernel dimensionality reduction, parallel computing, and fixed-point conversion to break through computational bottlenecks; the adaptive adjustment module uses data training, weight monitoring, dynamic adjustment, and computing power allocation to achieve intelligent optimization of parameters and resources; global average pooling and maximum pooling modules suppress noise and highlight key features, respectively; and the spatial attention mask module strengthens the response of the target area through edge enhancement, saliency detection, and mask fusion. During the technical implementation process, the image data first passes through a parallel heterogeneous convolution kernel group, the 3×3 convolution layer extracts local details, the 5×5 and 7×7 convolution layers capture the global structure, and the data after batch normalization and activation function processing flows into the computing acceleration module. The NPU tensor processing unit collaborates with each sub-unit to achieve efficient computing. The adaptive adjustment module dynamically optimizes weights and allocates computing power based on the training results; global average pooling and maximum pooling synchronously process feature maps to suppress noise and highlight key areas; the spatial attention mask module enhances the target feature response through convolution, detection and fusion operations. This technology achieves a multi-dimensional performance breakthrough. Compared with traditional convolution modules, the computing efficiency is improved by 60%, the feature extraction integrity is improved by 55%, and redundant calculations are effectively reduced. The spatial attention mask increases the recognition rate of white blood cell contours and key areas by 22%, significantly enhances the model's feature discrimination of white blood cells in complex backgrounds, provides strong support for accurate white blood cell calculations, and promotes the development of medical image analysis technology towards efficiency and intelligence. Furthermore, the working method of the global average pooling module is: (S1) Perform average pooling and maximum pooling on the input feature map in the channel dimension to obtain two different two-dimensional feature maps; (S2) The two 2D feature maps are concatenated in the channel dimension and subjected to a convolutional layer for feature fusion. (S3) The fused feature map is passed through the Sigmoid activation function to obtain the spatial weight matrix.

[0023] In a specific embodiment, this technology generates a spatial weight matrix by performing channel-dimensional pooling, fusion, and activation operations on feature maps. Its essence lies in leveraging the global statistics and fusion of channel information to enhance the model's focus on the target area. In the case of white blood cell (WBC) counting in medical images, the morphology and spatial position of WBCs are complex and variable. This technology processes feature maps to explore the potential relationships between channels, highlight the characteristic responses of the WBC region, and suppress background interference, thereby improving the model's WBC recognition accuracy. The technical implementation process strictly follows three steps. First, in (S1), average pooling and maximum pooling operations are performed on the input feature map in the channel dimension. Average pooling preserves the overall mean information of the feature map, while maximum pooling extracts the most significant features in the channel. The two operations are combined to compress the channel information from different angles, resulting in two two-dimensional feature maps with different semantics. Next, (S2) these two-dimensional feature maps are concatenated in the channel dimension to merge the channel information. The convolutional layer further fuses the features, eliminating the information redundancy caused by the concatenation and enhancing the expressive power of the features. Finally, (S3) the fused feature map is input into the Sigmoid activation function, which maps the eigenvalues to the range of 0-1, resulting in a spatial weight matrix. Each element of this matrix represents the weight of the corresponding spatial location. A larger value indicates a higher probability that the location belongs to the target area. This technology significantly improves the model's recognition of white blood cells. In practical applications, by performing weighted operations on the original feature map through the generated spatial weight matrix, the model can pay more attention to the white blood cell area and effectively suppress the interference of irrelevant areas such as red blood cells and background impurities. Compared with the model that does not use this technology, the accuracy of white blood cell positioning is improved by about 30%, and the targeted feature extraction is enhanced, providing a more accurate feature basis for subsequent white blood cell classification and counting tasks, greatly improving the overall performance and reliability of the white blood cell calculation model.

[0024] Furthermore, the global average pooling module calculates the global mean of each channel, and the temperature coefficient adjustment factor is calculated by the following formula: In formula (2), are the preset adjustment parameters. is the complexity measure of the white blood cell image, which is obtained by calculating the degree of overlap and edge blur of different cells in the image. In a specific embodiment, the technical essence of the global average pooling module is to effectively suppress local noise interference by calculating the global mean of each channel, and at the same time introduce a temperature coefficient adjustment factor based on the complexity measure of the white blood cell image to achieve dynamic optimization of image feature processing. In formula (2), Complexity comprehensively considers the degree of overlap and edge blur of cells in the white blood cell image to quantify the image complexity; β is a preset adjustment parameter that can flexibly control the sensitivity of the temperature coefficient T to complexity changes. The temperature coefficient T maps the image complexity to the range of 0-1 through the Sigmoid function characteristics. The higher the complexity, the closer the T value is to 1, and then dynamically adjusts the subsequent processing strategy to balance feature extraction and noise suppression, so that the model can achieve efficient processing in images of different complexities. In the image feature processing stage, the global average pooling module first calculates the global mean for each channel of the input feature map to reduce the impact of local noise on feature representation. Subsequently, the image complexity measure Complexity is calculated by analyzing the area ratio of the overlapping cells in the white blood cell image and the gradient change of the edge pixels. For example, the cell overlapping area is identified by threshold segmentation, and its proportion to the total image area is calculated. The Complexity value is comprehensively determined by combining the edge blur index obtained by the edge detection algorithm. Complexity and the preset adjustment parameter β are substituted into formula (2) to calculate the temperature coefficient T. According to the value of T, the subsequent feature processing process is dynamically adjusted: when the T value is close to 1, it indicates that the image complexity is high, the feature extraction precision is increased, and the processing of overlapping cells and blurred edges is strengthened; when the T value is close to 0, the processing process is appropriately simplified to improve the computational efficiency. This technology significantly improves the stability and accuracy of white blood cell image analysis. The noise is suppressed by global average pooling, which effectively improves the quality of the feature map and reduces the influence of interference factors on white blood cell feature extraction. The introduction of the temperature coefficient adjustment factor enables the model to adaptively process images of different complexities. Experimental data show that when processing complex images with highly overlapping cells and blurred edges, compared with traditional methods, the accuracy of white blood cell recognition is improved by about 25%, and the completeness of feature extraction is improved by 30%, effectively avoiding analysis errors caused by differences in image complexity, providing a more reliable feature basis for subsequent white blood cell counting and classification tasks, and enhancing the adaptability and reliability of the model in complex medical image scenarios. Table 2 shows the calculation process and results of the temperature coefficient adjustment factor, T, for five different white blood cell images. The cell overlap ratio is determined by identifying overlapping regions using threshold segmentation and calculating their proportion of the total image area. Edge blur is quantified based on the gradient changes of edge pixels obtained by the edge detection algorithm. The two factors together form the image complexity measure, Complexity. β, a preset adjustment parameter, was set to two values, 2 and 3, for comparative calculations. The data shows that, for a fixed β value, the temperature coefficient, T, gradually increases with increasing image complexity, indicating that the more complex the image, the closer T approaches 1. Conversely, for the same Complexity value, the larger the β value, the greater the T value, demonstrating the β parameter's modulating effect on the sensitivity of T to changes. This table intuitively presents the relationship between each parameter and the temperature coefficient, T, and clearly illustrates the calculation logic and effectiveness of the formula, T = 1 / (1 + e^(-β*Complexity)), for different image features. Furthermore, (1) the working principle of the data training classifier is: the 128-dimensional edge features obtained by OTSU image segmentation, the control features of the 32-dimensional fuzzy logic output and the 64-dimensional original sensor data are mapped to a 256-dimensional unified feature space through a fully connected layer, and an improved cross entropy loss function is used to strengthen the learning of difficult samples; the loss function formula is: In formula (3), is the category balance factor For the focus parameters, abnormal samples is the predicted probability, is the true label; (2) The working method of the weight dynamic monitoring module is: to monitor the distribution characteristics of the neural network weight matrix in real time through a multi-dimensional weight anomaly detection mechanism, and to identify weight anomalies caused by noise interference or data drift; the core includes: calculating the information entropy of the weight gradient of each layer , the function reflecting the degree of chaos of gradient distribution is: In formula (4), when hour, is the historical average, is the standard deviation, and the weight update is judged to be abnormal; is a gradient element; is the normalization term, and the absolute value of the gradient Normalized to probability distribution; L2 norm of monitoring weight matrix norm , when the dynamic threshold Initial norm, identifying weight explosion caused by overfitting; generating weight anomaly signals , triggering the weight dynamic adjustment module to perform parameter correction; (3) The working method of the weight dynamic adjustment module is: first, the multi-objective optimization objective function is constructed: a three-dimensional benefit function including feature extraction accuracy, computational efficiency, and overfitting suppression is defined to achieve the global optimal scheduling of weight updates: In formula (5), the characteristic benefit term : Based on significant differences in leukocyte nuclear and cytoplasmic characteristics; component benefit value , benefit term based on component analysis; regularized benefit term : benefit term based on regularization method; They are The weight coefficient of ; Design channel-specific benefit function: In formula (6), is the characteristic channel of the cell nucleus, It is a cytoplasmic characteristic channel. is the channel weight, The grayscale contrast of the nucleus / cytoplasm area is calculated in real time by the leukocyte positioning results of the preprocessing module and is set as Prioritize feature extraction to ensure that the weight of key features such as the edge of the cell nucleus and cytoplasm granularity is increased by 30%; computational efficiency : Suppress the weight update of redundant feature channels, definition: In formula (7), is the number of inactivated feature channels, The computing power threshold of the edge computing unit is ≥80%. When the computing power utilization rate is greater than 80%, the weight update step size of non-critical channels is automatically reduced by 50%. is the total number of all feature channels in the model, used as the denominator to normalize FreeChan to a ratio , represents the proportion of inactivated channels; The number of floating-point operations, the total number of floating-point operations required for the model to perform one forward / backward propagation; overfitting suppression term : Introducing weight norm constraints and feature correlation penalties: In formula (8), is the regularization coefficient, by calculating the Pearson correlation coefficient between channels , penalizing highly correlated redundant features, thus improving the independence of feature channels by 40%; designing a weight update rule with timing constraints: if the current layer is the cell nucleus contour extraction layer, and the cell nucleus position confidence output by the previous layer is greater than 0.9, then the priority promotion mechanism is triggered: In formula (9), is the gain coefficient, which increases the weight update speed of the cell nucleus feature channel by 50%, ensuring the rapid convergence of key edge features; As the baseline update amount, adjustments are made on this basis; is the adjusted weight update amount, which is used to update the actual parameters, increase the weight update amplitude of the cell nucleus feature channel, and accelerate convergence; The confidence score of channel C for the cell nucleus feature is usually [0, 1]. A higher value indicates a more significant response of the channel to the cell nucleus feature. Delay-sensitive weight correction: When a white blood cell classification delay of > 20ms is detected, the edge computing unit's computing power monitoring module provides real-time feedback and automatically activates the lightweight adjustment mode. In formula (10), only the gradient update of the key feature channel is retained, and the weight of the non-key channel is decayed by 5%. In the scenario of limited computing power, the classification accuracy loss is still kept less than 3%; is the weight of channel c, and different update rules are executed depending on whether the channel belongs to the key channel; is the learning rate, which controls the magnitude of the key channel weight update; is the gradient of channel c, which determines whether the key channel performs enhanced update; 0.95 is the attenuation coefficient, which attenuates the weight of non-key channels and weakens their influence; to avoid feature extraction failure caused by abnormal weights, three layers of resource capacity constraints are established: Channel capacity constraint: The weight update range of each feature channel does not exceed the medical prior threshold defined in the preprocessing stage: ; When the lower limit of the weight of the cell nucleus edge feature channel is set to 0.8; the upper limit of the background noise channel weight is set to 0.3; computing power capacity constraint: the floating-point operation amount of the single-layer weight update does not exceed the edge computing power threshold: In formula (11), The computing power utilization safety threshold is set. When the threshold is approached, it automatically switches to fixed-point calculation, replacing 32-bit with 16-bit calculation, which increases the calculation speed by 3 times and the accuracy loss is less than 1%. Measures the computational complexity of model training and the number of floating-point operations during model update. Timing constraints: weight updates for dependent features must be initiated after the previous features converge. In formula (12), To reduce feature dependency delays, timestamps are used to ensure orderly optimization of hierarchical features and avoid classification errors caused by feature misalignment. To mark the starting moment of cytoplasm feature training, the cytoplasm feature channel Timestamp when optimization started; To mark the completion time of cell nucleus feature training, End timestamp optimized for the cell nucleus feature channel; Microscope resolution adaptation: Dynamically adjust the weight update step size according to the objective lens magnification 40×-400×: In formula (13), To address the insensitivity of weight optimization caused by the reduction of feature size under high-magnification microscopes, the accuracy of cell nucleus segmentation under a 400× objective lens was improved by 9%. The step size for updating the dynamically adjusted weights is adaptively adjusted according to the objective lens magnification to ensure optimization efficiency at different resolutions; As the basic step length, as the benchmark value, through Zoom in or out; (4) The working method of the dynamic allocation module of computing power is as follows: The working principle of the dynamic allocation module of computing power is to build a multi-dimensional decision model based on task complexity and resource load, and integrate the n-dimensional task parameters of image resolution R, feature map size S, number of network layers L and CPU usage. , GPU memory occupancy The m-dimensional resource parameters are mapped to the resource allocation weight vector W through the linear weighting function F; the weight calculation adopts the dynamic adjustment formula: In formula (14), Assign a weight to the resource corresponding to the i-th parameter, is the parameter importance adjustment factor, is the normalized value of the i-th parameter, is the nonlinear adjustment coefficient; when it is detected that the task parameter exceeds the threshold When resource parameters reach the warning state, the dynamic adjustment mechanism is activated to adjust Strengthen the allocation of key task resources and achieve dynamic balance and efficiency optimization of computing resources. In a specific embodiment, the white blood cell calculation model realizes accurate and efficient cell analysis through multi-module collaborative innovation. The essence of the technology lies in that the data training classifier integrates multi-source features and optimizes the loss function to strengthen the learning of difficult examples; the weight dynamic monitoring module uses gradient entropy and norm monitoring to accurately identify weight anomalies; the weight dynamic adjustment module builds a multi-objective optimization system, combining timing and capacity constraints to realize intelligent weight updates; the computing power dynamic allocation module builds a decision model based on task and resource parameters to realize dynamic resource allocation. During the technical implementation process, the data training classifier first integrates multi-dimensional features and maps to a unified space, and then trains with an improved loss function; the weight dynamic monitoring module calculates the gradient entropy and weight norm in real time, and triggers an abnormal signal when the threshold is exceeded; the weight dynamic adjustment module dynamically optimizes the weight update strategy according to the multi-objective function and constraint conditions; the computing power dynamic allocation module calculates weights based on multi-dimensional parameters and allocates resources on demand. The technical effect is significant. The data training classifier improves the accuracy of identifying difficult samples by 35%; the weight dynamic monitoring and adjustment module reduces the risk of model failure caused by weight abnormalities, reduces overfitting by 40%, improves the independence of feature channels by 40%, and increases the accuracy of cell nucleus segmentation at different resolutions by up to 9%; the computing power dynamic allocation module increases resource utilization by 60% and computing efficiency by 50% under complex tasks, effectively improving the accuracy, stability and computing efficiency of the white blood cell calculation model. Table 3 presents performance comparisons of the key modules of the WBC calculation model before and after optimization. The data-trained classifier significantly improves the accuracy of difficult examples and overall classification through multi-feature fusion and an improved loss function. The dynamic weight monitoring module effectively identifies weight anomalies, significantly reducing the incidence of overfitting. The dynamic weight adjustment module, through multi-objective optimization and constraint mechanisms, enhances the independence of feature channels, improves the accuracy of cell nucleus segmentation under high-power microscopy, and optimizes computing power utilization. The dynamic computing power allocation module improves computational efficiency and resource allocation balance. These data validate the effectiveness of the technical solutions of each module and provide reliable support for the practical application of the WBC calculation model.

[0025] Furthermore, the computing acceleration module includes a convolution expansion module, a channel correlation coefficient calculation module, and an implicit feature conversion module; the convolution expansion module adopts dilated convolution and dynamically adjusts the receptive field of the convolution kernel through the expansion rate d, and the calculation formula is: In formula (15), is the convolution output feature map, is the input feature map, is the convolution kernel, n is the input feature map size, m is the convolution kernel size, s is the convolution step size, and d is the dilation rate. Through dynamic adjustment, the network's perception range of white blood cell morphological characteristics is expanded without increasing the number of parameters, reducing repeated calculations.

[0026] The channel correlation coefficient calculation module uses the Pearson correlation coefficient to evaluate the correlation between feature channels and accelerates the calculation by penalizing redundant channels. The calculation formula is: In formula (16), For channel The correlation coefficient between Channel The i-th eigenvalue, is the mean of the corresponding channel eigenvalues, and N is the number of eigenvalues. By calculating the correlation coefficients of all channel pairs, the weights of highly correlated channels are identified and reduced to reduce the amount of computation. The implicit feature conversion module decomposes the high-dimensional feature tensor T into a combination of low-dimensional tensors based on tensor decomposition. The formula is as follows: In formula (17), T is the original high-dimensional feature tensor, are the elements of the core tensor, is a factor matrix, is the rank of the corresponding dimension, and ∘ represents the tensor outer product operation. By reducing the tensor rank, the feature dimension is reduced, thereby accelerating the feature calculation and transmission process and improving the overall computational efficiency of the model. In a specific embodiment, the computation acceleration module optimizes the convolution computation process through multiple components to achieve efficient processing of white blood cell image analysis. The essence of the technology lies in that the convolution expansion module uses dilated convolution to dynamically adjust the receptive field, expanding the perception range of white blood cell morphology without increasing parameters; the channel correlation coefficient calculation module evaluates the correlation between channels and penalizes redundant channels to reduce the amount of computation; the implicit feature conversion module reduces the feature dimension based on tensor decomposition, accelerating computation and transmission. The three together construct an efficient computing architecture that balances accuracy and efficiency. During the technical implementation process, the convolution expansion module dynamically adjusts the dilation rate d according to formula (15), performs dilated convolution on the input feature map, and captures the multi-scale morphological characteristics of white blood cells; the channel correlation coefficient calculation module calculates the Pearson correlation coefficient of all channel pairs, identifies high-correlation channels and reduces their weights; the implicit feature conversion module decomposes the high-dimensional feature tensor into a low-dimensional tensor combination according to formula (16), reducing the feature dimension. The technology has achieved remarkable results. Dilated convolution increases the model's perception of white blood cell edges by 2.5 times, reducing repeated computations by 30%. The channel correlation penalty mechanism reduces redundant computations by 25%, preserving key features. Tensor decomposition compresses feature dimensions by 60%, increasing feature computation and transmission speeds by 40%. The overall module maintains 98% analysis accuracy while increasing computational efficiency by 50%, significantly reducing resource consumption on edge devices and providing strong support for real-time medical image analysis. Table 4 compares the performance of the various submodules and the entire computational acceleration module before and after optimization from multiple perspectives. The convolution expansion module increases the receptive field coverage by nearly 1.8 times while maintaining the same number of parameters, reducing repeated computations by 30%. The channel correlation coefficient calculation module reduces the proportion of redundant channels by 37.14% and the amount of computation by over 30% by identifying and penalizing redundant channels. The implicit feature conversion module significantly compresses feature dimensions, reducing computational time by nearly 46%. Overall, the entire computational acceleration module shortens model inference time by 43.75% while maintaining computational accuracy at 98%, effectively achieving the goal of significantly improving computational efficiency while maintaining high accuracy, validating the effectiveness and superiority of this module's technical solution. Summary of Technological Innovation Breakthrough in positioning accuracy: Through semantic segmentation + morphological filtering, the traditional threshold method solves the problem of missing low-contrast cells, and the positioning accuracy rate reaches 98.5%; Multi-scale adaptation: Dynamic window division based on the actual cell size increases the complete cell segmentation rate to 99% and reduces edge distortion by 60%; Intelligent enhancement strategy: Adaptive noise-morphology dual-dimensional parameters achieve a 30%-50% contrast improvement while improving noise suppression by 20%-30% compared to traditional methods; Engineering advantages: Processing delay ≤ 8ms / image (1024×1024), adapting to real-time clinical testing needs, and power consumption is reduced by 40% under hardware acceleration.

[0027] Through hardware-algorithm collaborative innovation, the above modules form a complete white blood cell image preprocessing technology solution, significantly improving the feature quality of subsequent convolutional neural networks, providing a new preprocessing paradigm for medical image computing, and possessing outstanding creativity and practicality.

[0028] Although specific embodiments of the present invention have been described above, those skilled in the art will appreciate that these specific embodiments are merely illustrative, and that those skilled in the art may omit, substitute, and modify the details of the methods and systems described above without departing from the principles and spirit of the present invention. For example, combining the above method steps to perform substantially the same functions in substantially the same manner to achieve substantially the same results falls within the scope of the present invention. Accordingly, the scope of the present invention is limited solely by the appended claims.

Claims

1. An improved convolutional neural network model for white blood cell counting, characterized by: include: Image preprocessing module; Acquiring a medical image containing leukocytes, the image preprocessing module comprising a leukocyte positioning module, an image segmentation module, a window size adaptation module, an overlapping area compensation module, and an adaptive image enhancement module; The image feature extraction module generates a multi-scale representation of the image through a Gaussian pyramid and extracts edge Canny operator, texture, and gray-level co-occurrence matrix features at different resolutions; Secondly, the preprocessed sub-images are normalized and input into the bottom layer of the improved convolutional neural network model module. The basic visual features of the cell membrane contour or cell nucleus texture are extracted and spliced along the channel dimension to form a composite feature vector containing semantic information and spatial structure. Improved convolutional neural network model module, including: Multi-scale convolution module, including parallel heterogeneous convolution kernel group, computation acceleration module, adaptive adjustment module, global average pooling module, global maximum pooling module and spatial attention mask module; A dual-layer attention mechanism module, including a channel attention mechanism and a spatial attention mechanism. The dual-layer attention mechanism module extracts contextual information in the channel dimension through global average pooling and global maximum pooling, and generates a channel attention weight of 1×1×C through a two-layer fully connected network to suppress the feature response of red blood cell-irrelevant areas. Secondly, a spatial dimension convolution operation (7×7 convolution) is performed to generate a spatial attention mask H×W×1, focusing on the white blood cell outline and cell nucleus area. Finally, the two attention weights are superimposed on the feature map through element-wise multiplication to enhance the key features of white blood cells, reduce background noise interference, and improve the discriminability of feature expression. A convolution layer calculation module, comprising a calculation acceleration module and a convolution expansion module connected to the calculation acceleration module, a channel correlation coefficient calculation module, and an implicit feature conversion module; A visualization terminal, used to display the data information output by the improved convolutional neural network model module; The output end of the image preprocessing module is connected to the input end of the image feature extraction module, the output end of the image feature extraction module is connected to the input end of the improved convolutional neural network model module, and the output end of the improved convolutional neural network model module is connected to the input end of the visualization terminal.

2. The improved convolutional neural network model for white blood cell counting according to claim 1, characterized in that: The leukocyte positioning module uses the lightweight semantic segmentation model U-NetLite to pre-train three types of semantic labels: leukocytes, erythrocytes, and background. Morphological opening operations are used to remove erythrocyte adhesion interference and connected domain analysis to screen cell nucleus features, thereby achieving accurate positioning of leukocyte ROIs. The grayscale histogram threshold OTSU and texture feature LBP operator are integrated with the semantic segmentation results. The image segmentation module constructs a variable window library based on the 5-20 μm white blood cell size distribution and the corresponding pixels 50-200 px @ 1000× magnification. It calculates the minimum circumscribed circle diameter of white blood cells based on the Canny operator edge detection algorithm and automatically matches 32×32 small cells, 64×64 medium cells, and 128×128 large cell windows. Mirror compensation of edge pixels is implemented for cells across windows, and the grayscale continuity of the cut area is restored through bilinear interpolation; The window size adaptation module uses an intelligent adjustment algorithm based on diameter-pixel mapping. It establishes a mapping relationship between the actual diameter of white blood cells and the pixel size through microscopic image calibration to achieve physical adaptation of the window size. When the cell density in the ROI area is greater than 50 cells / mm², the window step size is automatically reduced from 40px to 20px to avoid missing high-density cells. The overlapping area compensation module calculates the gradient amplitude distribution of the overlapping area based on the weighted gradient fusion algorithm. When the gradient is greater than the threshold, a weight of 0.3 is applied to the edge pixels, and when the gradient is less than or equal to the threshold, a weight of 0.7 is applied to the center pixels to suppress the generation of artifacts. The adaptive image enhancement module calculates texture complexity through local binary pattern (LBP). When the noise density detects that the salt and pepper noise is greater than 10% or the Gaussian noise σ is greater than 15, a dynamic selection enhancement algorithm based on bilateral filtering is used to automatically match elliptical or polygonal structural elements according to the morphological parameters of white blood cells for expansion and corrosion. The roundness of 0.8-1.2 is for lymphocytes, and 0.6-0.8 is for neutrophils.

3. The improved convolutional neural network model for white blood cell counting according to claim 1, characterized in that: The intelligent adjustment algorithm based on diameter-pixel mapping adjusts the feature information of the image through the weighted fusion coefficient in image preprocessing. The weighted fusion coefficient is calculated by the following formula: In formula (1), represents the edge strength of the i-th pixel in the high-frequency detail layer, It represents the grayscale change of the i-th pixel in the low-frequency background layer, and N is the total number of image pixels.

4. The improved convolutional neural network model for white blood cell counting according to claim 1, characterized in that: The adaptive image enhancement module is based on a multi-scale feature collaborative optimization architecture, and incorporates a coding module and a data classification module to fuse a generative adversarial network (cGAN), a spatial attention mechanism, and a feature pyramid network to achieve intelligent improvement in the clarity of white blood cell images. In high-frequency detail layer processing, a conditional generative adversarial network (cGAN) is introduced, using prior data on white blood cell morphology as conditional input to drive the generator to learn and generate ideal edge-enhanced images that conform to real cell morphology. The unsharp mask parameters are optimized jointly by the feature matching loss function and the adversarial loss function, so that the edge enhancement process can not only enhance the white blood cell contour details but also effectively suppress noise and artifacts. In processing the low-frequency background layer, a spatial attention-weighted histogram equalization strategy is adopted. Based on a convolutional neural network, feature maps of background subregions are extracted. A dual-channel attention mechanism is used to calculate subregion weights from both the channel and spatial dimensions. The histogram equalization strength is dynamically adjusted to accurately suppress over-enhancement of non-critical background areas and highlight the contrast of potential white blood cell areas. In the multi-scale fusion link, the multi-resolution processing framework based on the feature pyramid network FPN performs adaptive enhancement on the high-frequency detail layer and low-frequency background layer at different scales respectively. Through the top-down and horizontally connected feature fusion paths, the organic unity of local detail enhancement and global structure optimization is achieved, effectively balancing noise suppression and detail retention, and significantly improving the visual distinction between white blood cells and the background.

5. The improved convolutional neural network model for white blood cell counting according to claim 1, characterized in that: In the parallel heterogeneous convolution kernel group, the convolution layers with different convolution kernel sizes include 3×3, 5×5, and 7×7 convolution layers, and each convolution layer is connected to a batch normalization layer and an activation function layer; The computing acceleration module includes a tensor processing unit of the NPU and a convolution kernel dimensionality reduction processing unit connected to the tensor processing unit of the NPU, an 8-channel parallel computing unit, and a fixed-point conversion module, wherein the tensor processing unit of the NPU is a multi-scale convolution kernel processor based on the AI-SSC369G chip; The adaptive adjustment module includes a data training classifier, a weight dynamic monitoring module, a weight dynamic adjustment module and a computing power dynamic allocation module; wherein the output end of the data training classifier is connected to the input end of the weight dynamic monitoring module, the output end of the weight dynamic monitoring module is connected to the input end of the weight dynamic adjustment module, and the output end of the weight dynamic adjustment module is connected to the input end of the computing power dynamic allocation module; The global average pooling module calculates the global mean of each channel, retaining the average grayscale of the leukocyte cytoplasm and the average temperature of the device surface in the feature map to suppress local noise interference; The global maximum pooling module extracts the global maximum value of each channel, highlights the high grayscale value area of the cell nucleus or the abnormal hot spot of the device in the feature map, and strengthens the discriminative power of key features; The spatial attention mask module generates an attention mask H×W×1 through a convolution operation in the spatial dimension, and the focus target area contains at least: Edge enhancement: A 7×7 convolution with a dilationrate of 2 is used to expand the receptive field to capture the edges of white blood cell contours or the location of device terminals of the target structure. Saliency detection: Combining Sobel edge detection with grayscale threshold segmentation and Otsu algorithm, a binary saliency map is generated. After Gaussian blurring with a kernel size of 5×5, the map is converted into a continuous mask with values between 0 and 1. Mask fusion: The mask is multiplied element-by-element with the feature map, which increases the feature response of the target area by 1.5 times and attenuates the response of the background area to less than 0.6 times, thereby improving the recognition rate of abnormal areas by 22%.

6. The improved convolutional neural network model for white blood cell counting according to claim 5, characterized in that: The working method of the global average pooling module is: (S1) Perform average pooling and maximum pooling on the input feature map in the channel dimension to obtain two different two-dimensional feature maps; (S2) The two 2D feature maps are concatenated in the channel dimension and subjected to a convolutional layer for feature fusion. (S3) The fused feature map is passed through the Sigmoid activation function to obtain the spatial weight matrix.

7. The improved convolutional neural network model for white blood cell counting according to claim 5, characterized in that: The global average pooling module calculates the global mean of each channel, and the temperature coefficient adjustment factor is calculated by the following formula: In formula (2), is the preset adjustment parameter, is the complexity measure of the white blood cell image, which is obtained by calculating the degree of overlap and edge blur of different cells in the image.

8. The improved convolutional neural network model for white blood cell counting according to claim 5, characterized in that: (1) The working principle of the data training classifier is as follows: the 128-dimensional edge features obtained by OTSU image segmentation, the control features of the 32-dimensional fuzzy logic output, and the 64-dimensional original sensor data are mapped to a 256-dimensional unified feature space through a fully connected layer, and an improved cross-entropy loss function is used to enhance the learning of difficult samples; Loss function formula: In formula (3), is the category balance factor For the focus parameters, abnormal samples is the predicted probability, is the true label; (2) The working method of the weight dynamic monitoring module is: to monitor the distribution characteristics of the neural network weight matrix in real time through a multi-dimensional weight anomaly detection mechanism, and to identify weight anomalies caused by noise interference or data drift; the core includes: calculating the information entropy of the weight gradient of each layer , the function reflecting the degree of chaos of gradient distribution is: In formula (4), when hour, is the historical average, is the standard deviation, and the weight update is judged to be abnormal; is a gradient element; is the normalization term, and the absolute value of the gradient Normalized to probability distribution; L2 norm of monitoring weight matrix norm , when the dynamic threshold Initial norm, identifying weight explosion caused by overfitting; generating weight anomaly signals , triggering the weight dynamic adjustment module to perform parameter correction; (3) The working method of the weight dynamic adjustment module is: first, the multi-objective optimization objective function is constructed: a three-dimensional benefit function including feature extraction accuracy, computational efficiency, and overfitting suppression is defined to achieve the global optimal scheduling of weight updates: In formula (5), the characteristic benefit term : Based on significant differences in leukocyte nuclear and cytoplasmic characteristics; component benefit value , benefit term based on component analysis; regularized benefit term : benefit term based on regularization method; They are The weight coefficient of ; Design channel-specific benefit function: In formula (6), is the characteristic channel of the cell nucleus, It is a cytoplasmic characteristic channel. is the channel weight, The grayscale contrast of the nucleus / cytoplasm area is calculated in real time by the leukocyte positioning results of the preprocessing module and is set as Prioritize feature extraction to ensure that the weight of key features such as the edge of the cell nucleus and cytoplasm granularity is increased by 30%; computational efficiency : Suppress the weight update of redundant feature channels, definition: In formula (7), is the number of inactivated feature channels, The computing power threshold of the edge computing unit is ≥80%. When the computing power utilization rate is greater than 80%, the weight update step size of non-critical channels is automatically reduced by 50%. is the total number of all feature channels in the model, used as the denominator to normalize FreeChan to a ratio , represents the proportion of inactivated channels; The number of floating-point operations, the total number of floating-point operations required for the model to perform one forward / backward propagation; overfitting suppression term : Introducing weight norm constraints and feature correlation penalties: In formula (8), is the regularization coefficient, by calculating the Pearson correlation coefficient between channels , penalizing highly correlated redundant features, thus improving the independence of feature channels by 40%; designing a weight update rule with timing constraints: if the current layer is the cell nucleus contour extraction layer, and the cell nucleus position confidence output by the previous layer is greater than 0.9, then the priority promotion mechanism is triggered: In formula (9), is the gain coefficient, which increases the weight update speed of the cell nucleus feature channel by 50%, ensuring the rapid convergence of key edge features; As the baseline update amount, adjustments are made on this basis; is the adjusted weight update amount, which is used to update the actual parameters, increase the weight update amplitude of the cell nucleus feature channel, and accelerate convergence; The confidence score of channel C for the cell nucleus feature is usually [0, 1]. A higher value indicates a more significant response of the channel to the cell nucleus feature. Delay-sensitive weight correction: When a white blood cell classification delay of > 20ms is detected, the edge computing unit's computing power monitoring module provides real-time feedback and automatically activates the lightweight adjustment mode. In formula (10), only the gradient update of the key feature channel is retained, and the weight of the non-key channel is decayed by 5%. In the scenario of limited computing power, the classification accuracy loss is still kept less than 3%; is the weight of channel c, and different update rules are executed depending on whether the channel belongs to the key channel; is the learning rate, which controls the magnitude of the key channel weight update; is the gradient of channel c, which determines whether the key channel performs enhanced update; 0.95 is the attenuation coefficient, which attenuates the weight of non-key channels and weakens their influence; to avoid feature extraction failure caused by abnormal weights, three layers of resource capacity constraints are established: Channel capacity constraint: The weight update range of each feature channel does not exceed the medical prior threshold defined in the preprocessing stage: ; When the lower limit of the weight of the cell nucleus edge feature channel is set to 0.8; the upper limit of the background noise channel weight is set to 0.3; computing power capacity constraint: the floating-point operation amount of the single-layer weight update does not exceed the edge computing power threshold: In formula (11), The computing power utilization safety threshold is set. When the threshold is approached, it automatically switches to fixed-point calculation, replacing 32-bit with 16-bit calculation, which increases the calculation speed by 3 times and the accuracy loss is less than 1%. Measures the computational complexity of model training and the number of floating-point operations during model update. Timing constraints: weight updates for dependent features must be initiated after the previous features converge. In formula (12), To reduce feature dependency delays, timestamps are used to ensure orderly optimization of hierarchical features and avoid classification errors caused by feature misalignment. To mark the starting moment of cytoplasm feature training, the cytoplasm feature channel Timestamp when optimization started; To mark the completion time of cell nucleus feature training, End timestamp optimized for the cell nucleus feature channel; Microscope resolution adaptation: Dynamically adjust the weight update step size according to the objective lens magnification 40×-400×: In formula (13), To address the insensitivity of weight optimization caused by the reduction of feature size under high-magnification microscopes, the accuracy of cell nucleus segmentation under a 400× objective lens was improved by 9%. The step size for updating the dynamically adjusted weights is adaptively adjusted according to the objective lens magnification to ensure optimization efficiency at different resolutions; As the basic step length, as the benchmark value, through Zoom in or out; (4) The working method of the dynamic allocation module of computing power is as follows: The working principle of the dynamic allocation module of computing power is to build a multi-dimensional decision model based on task complexity and resource load, and integrate the n-dimensional task parameters of image resolution R, feature map size S, number of network layers L and CPU usage. , GPU memory occupancy The m-dimensional resource parameters are mapped to the resource allocation weight vector W through the linear weighting function F; the weight calculation adopts the dynamic adjustment formula: In formula (14), Assign a weight to the resource corresponding to the i-th parameter, is the parameter importance adjustment factor, is the normalized value of the i-th parameter, is the nonlinear adjustment coefficient; when it is detected that the task parameter exceeds the threshold When resource parameters reach the warning state, the dynamic adjustment mechanism is activated to adjust Strengthen the allocation of critical task resources to achieve dynamic balance and efficiency optimization of computing resources.

9. The improved convolutional neural network model for white blood cell counting according to claim 1, characterized in that: The computing acceleration module includes a convolution expansion module, a channel correlation coefficient calculation module and an implicit feature conversion module; The convolution expansion module adopts dilated convolution and dynamically adjusts the receptive field of the convolution kernel through the expansion rate d. The calculation formula is: In formula (15), is the convolution output feature map, is the input feature map, is the convolution kernel, n is the input feature map size, m is the convolution kernel size, s is the convolution step size, and d is the dilation rate. Through dynamic adjustment, the network's perception range of white blood cell morphological characteristics is expanded without increasing the number of parameters, reducing repeated calculations.

10. The channel correlation coefficient calculation module uses the Pearson correlation coefficient to evaluate the correlation between feature channels and accelerates the calculation by penalizing redundant channels; the calculation formula is: In formula (16), For channel The correlation coefficient between Channel The i-th eigenvalue, is the mean of the corresponding channel eigenvalues, and N is the number of eigenvalues. By calculating the correlation coefficients of all channel pairs, the weights of highly correlated channels are identified and reduced to reduce the amount of computation. The implicit feature conversion module decomposes the high-dimensional feature tensor T into a combination of low-dimensional tensors based on tensor decomposition. The formula is as follows: In formula (17), T is the original high-dimensional feature tensor, are the elements of the core tensor, is a factor matrix, is the rank of the corresponding dimension, and ∘ represents the tensor outer product operation. By reducing the tensor rank and the feature dimension, the feature calculation and transmission process is accelerated, thereby improving the overall computational efficiency of the model.

Images