Cell micronucleus detection method based on prior knowledge and frequency domain perception

Abstract

The invention discloses a priori knowledge and frequency domain perception-based cell micronucleus detection method, and relates to the technical field of computers. The method comprises the following steps: extracting cell micronucleus features of a to-be-detected cell image in a multi-level and multi-angle manner based on priori knowledge of cell micronucleus; the cell micronucleus characteristics comprise the volume of the cell micronucleus, the form of the cell micronucleus and the distribution area of the cell micronucleus; performing continuous channel compression on the cell micronucleus features, and extracting frequency domain information of the cell micronucleus features after channel compression through discrete wavelet transform; the frequency domain information comprises high-frequency information and low-frequency information; convolution processing is carried out on the high-frequency information and the low-frequency information, and convolution processing results are fused to obtain frequency domain sensing fusion features; analyzing the frequency domain perception fusion feature to obtain a plurality of bounding boxes in the to-be-detected image; the bounding box is the predicted position of the cell micronucleus. The method can improve the detection precision of the cell micronucleus.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method for detecting cell micronuclei based on prior knowledge and frequency domain perception. Background Technology

[0002] Micronuclei (MNs) are chromatin structures in the cytoplasm that resemble the cell nucleus in morphology but are smaller in size. Micronucleus formation reflects genomic instability and is of great significance for understanding cellular abnormalities, the mutagenic effects of environmental pollutants, and disease mechanisms.

[0003] Traditional micronucleus detection methods primarily rely on manual counting under a microscope. This method is not only time-consuming and labor-intensive but also easily affected by the operator's subjective judgment. In recent years, the rapid development of computer vision and artificial intelligence technologies has provided new possibilities for automated micronucleus detection. However, due to the high similarity between the morphology of cellular micronuclei and the main nucleus, the wide range of micronucleus sizes, and the high throughput requirements of cellular micronuclei, the detection accuracy of general target detection models is relatively low when applied to micronucleus identification. Summary of the Invention

[0004] Therefore, it is necessary to provide a cell micronucleus detection method based on prior knowledge and frequency domain sensing to address the aforementioned technical problems. This method can improve the detection accuracy of cell micronuclei.

[0005] The present invention adopts the following technical solution: This invention provides a method for detecting cell micronuclei based on prior knowledge and frequency domain awareness, comprising: Acquire an image of the cells to be detected; the image of the cells to be detected includes multiple cells; Based on prior knowledge of cell micronuclei, cell micronucleus features are extracted from the cell images to be detected from multiple levels and angles; cell micronucleus features include cell micronucleus volume, cell micronucleus morphology, and cell micronucleus distribution area; Continuous channel compression is performed on the micronucleus features of cells; frequency domain information of the channel-compressed micronucleus features is extracted by discrete wavelet transform; the frequency domain information includes high-frequency information and low-frequency information. High-frequency and low-frequency information are convolved separately, and the results of the convolution are fused to obtain frequency domain-aware fusion features; The frequency domain sensing fusion features are analyzed to obtain multiple bounding boxes in the image to be detected; the bounding boxes are the predicted locations of cell micronuclei.

[0006] Preferably, multiple bounding boxes in the image to be detected are extracted using a cell micronucleus detection model. The cell micronucleus detection model is an improvement upon the YOLOv8 model. The YOLOv8 model includes a backbone network, a feature fusion pyramid module, and a detection head. The improvement involves adding a micronucleus prior knowledge-enhanced feature extraction module after each Haar wavelet-based downsampling module in the backbone network, and introducing a wavelet transform-based feature enhancement module in the feature fusion pyramid module. The micronucleus prior knowledge-enhanced feature extraction module includes a micronucleus bottleneck block and a micronucleus attention module. The micronucleus bottleneck block includes two parallel main paths and a micronucleus path. The main paths include... Convolution; microkernel path includes Convolution; the microkernel attention module includes a shape feature extraction module and a distance awareness module; the shape feature extraction module includes sequentially connected... Convolution, batch normalization, corrected linear unit and Convolution; the distance-aware module includes dilated convolution, batch normalization, rectified linear units, and... convolution.

[0007] Preferably, the micronucleus features of the cells in the image to be detected are extracted at multiple levels and from multiple angles, specifically including: The prior information of the cell image to be detected is input into the backbone network, and then processed through the backbone network... Convolution performs channel mapping and separation to obtain the first and second features; For each of the multiple microkernel bottleneck blocks connected in series, through the main path Convolution extracts intermediate features from the first feature; through the micro-kernel path Convolution is used to extract the micro-kernel features of the first feature; the micro-kernel features and intermediate features are then weighted and fused to obtain the output features of the micro-kernel bottleneck block; The circular shape feature is extracted from the output features of the last micro-core bottleneck block using the shape feature extraction module. The spatial relationship features between the micronucleus and the main core are extracted from the output features of the last micronucleus bottleneck block using the distance sensing module. The enhanced features are obtained by weighted fusion of circular shape features and spatial relationship features; The second feature and the enhancement feature are combined Convolution is used to integrate the data and obtain the cell micronucleus features.

[0008] Preferably, the frequency domain sensing fusion features are analyzed to obtain multiple bounding boxes in the image to be detected, specifically including: Predict the geometric parameters of the bounding box; the geometric parameters include the offset of the cell micronucleus center point coordinates relative to the grid, the width of the bounding box, and the height of the bounding box; The detection head senses each grid point of the fused features in the frequency domain, generating multiple candidate detection boxes; Based on multiple candidate detection boxes, multiple bounding boxes are determined by combining the offset, the width of the bounding box, and the height of the bounding box.

[0009] Preferably, the training process of the cell micronucleus detection model specifically includes: A cell micronucleus dataset was obtained as the training set, and an initial cell micronucleus detection model was constructed. The cell micronucleus dataset includes multiple cell images; each cell image contains multiple cells. The initial cell micronucleus detection model was trained using the training set, and then the AdamW optimizer was used to adjust the initial cell micronucleus detection model to minimize its loss function, thus obtaining the cell micronucleus detection model.

[0010] Preferably, the loss function is: ; in, , , yes IoU Position penalty items, The weighting coefficient for the penalty term. The distance between the center points of the two bounding boxes. It is the length of the diagonal of the smallest rectangle enclosing the two boxes. This represents the distance between the current microkernel size ratio and the optimal range. The penalty coefficient is the microkernel size. and These represent the width and height of the micronucleus, respectively. The shape penalty scaling factor. The position penalty term weighting coefficient, The weight coefficient for the prior knowledge penalty term of the microkernel. To measure the intersection-union ratio of the detection bounding box and the target bounding box, It is a correction factor. for A perceptual adaptive scaling function, This deviates from the ideal aspect ratio.

[0011] This invention provides a cell micronucleus detection device based on prior knowledge and frequency domain sensing, comprising: The acquisition module is used to acquire images of cells to be detected; the images of cells to be detected include multiple cells; The first extraction module is used to extract the micronucleus features of the cell image to be detected from multiple levels and angles based on prior knowledge of the cell micronucleus. The micronucleus features include the volume of the cell micronucleus, the morphology of the cell micronucleus, and the distribution area of ​​the cell micronucleus. The second extraction module is used to continuously compress the cell micronucleus features through channels and extract the frequency domain information of the compressed cell micronucleus features through discrete wavelet transform; the frequency domain information includes high-frequency information and low-frequency information. The fusion module is used to perform convolution processing on high-frequency information and low-frequency information respectively, and then fuse the results of the convolution processing to obtain frequency domain-aware fusion features; The parsing module is used to analyze the frequency domain sensing fusion features to obtain multiple bounding boxes in the image to be detected; the bounding boxes are the predicted locations of cell micronuclei.

[0012] The present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for detecting cell micronuclei based on prior knowledge and frequency domain awareness.

[0013] The present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-described method for detecting cell micronuclei based on prior knowledge and frequency domain awareness.

[0014] The above-mentioned at least one technical solution adopted in this invention can achieve the following beneficial effects: Based on prior knowledge of cell micronuclei, this method extracts cell micronucleus features from the cell image to be detected from multiple levels and angles, fusing prior knowledge features such as the volume, morphology, and distribution area of ​​cell micronuclei to significantly improve the completeness of feature representation. Continuous channel compression is applied to the cell micronucleus features to reduce redundant computation. Frequency domain information of the channel-compressed cell micronucleus features is extracted through discrete wavelet transform, which decomposes spatial features into high-frequency and low-frequency components, enhancing the model's adaptability to complex scenarios such as blurred micronucleus edges and small sizes. High-frequency and low-frequency information are convolved separately, and the results are fused to obtain frequency-domain perceptual fusion features. The frequency-domain perceptual fusion features are then analyzed to obtain multiple bounding boxes in the image to be detected. This method can improve the detection accuracy of cell micronuclei. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0016] Figure 1 A schematic diagram of a cell micronucleus detection method based on prior knowledge and frequency domain sensing provided by the present invention; Figure 2 This is a structural diagram of the YOLO-MN model provided by the present invention; Figure 3This is a simplified structural diagram of the cell micronucleus detection model provided by the present invention; Figure 4 This is a structural diagram of the HDC provided by the present invention; Figure 5 This is a schematic diagram of the microkernel attention module (MAM) provided by the present invention; Figure 6 The structural diagram of the HWD module provided by this invention; Figure 7 A schematic diagram of the wavelet feature enhancement module (WFE) provided by the present invention; Figure 8 A typical diagram of several impurities in the microkernel dataset provided by this invention; Figure 9 This is a schematic diagram illustrating the training variation process of mAP@50 for different models on the CMD dataset provided by this invention. Figure 10 This is a schematic diagram illustrating the training variation process of mAP@50-95 for different models on the CMD dataset provided by this invention. Figure 11 This is a schematic diagram illustrating the training process of recall for different models under the CMD dataset provided by this invention. Figure 12 The tenth-layer heatmap using the C2f module provided by this invention; Figure 13 The tenth-layer heatmap using the C2f_MAM module provided for this invention; Figure 14 This is a schematic diagram of the PR curve provided by the present invention; Figure 15 A schematic diagram of a cell micronucleus detection device based on prior knowledge and frequency domain sensing provided by the present invention; Figure 16 This is a schematic diagram of a computer device for implementing a cell micronucleus detection method based on prior knowledge and frequency domain perception, provided by the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0018] Devices such as desktop computers, servers, and laptops are capable of executing the solutions of this invention. For ease of explanation, the following description will focus on servers as the executing entity.

[0019] In the modern world, humans are exposed to various genotoxic substances present in contaminated environments. Therefore, testing is necessary to determine exposure levels and health risks. While many detection methods are classified as "in vivo biomonitoring," micronucleus detection is one of the best and most popular.

[0020] Cell micronucleus detection methods are mainly divided into two categories: traditional methods and deep learning-based methods. Traditional methods primarily rely on microscopic observation and manual counting, with accuracy improved through various image processing techniques. Researchers compared manual and automated scoring systems to provide a basis for equipment optimization. Researchers developed a computer-based micronucleus automatic scoring system (CBMN) to effectively distinguish between micronuclei and micronucleated cells. Several researchers improved micronucleus detection technology from the perspectives of DNA damage identification, peripheral blood smear detection, and RGB channel separation. Researchers used two mathematical algorithms from ImageJ software—fractal dimension (FD) and gray-level co-occurrence matrix (GLCM)—to analyze MN dimension, MN irregularity, and texture. Despite these advancements, traditional methods still struggle to meet the demands of large-scale screening and high-throughput, prompting research into more automated and efficient detection methods.

[0021] Deep learning technology has brought revolutionary breakthroughs to cell micronucleus detection. Researchers combined Convolutional Neural Networks (CNNs) with visual attention, using AlexNet as the backbone network, achieving an AP value of 93.2% with relatively low computational cost. They also fused CBMNs and CNNs to develop automated detection software capable of efficiently processing large numbers of images. Researchers innovatively combined YOLO with traditional image processing methods, first identifying normal binucleated cells and then performing micronucleus detection. They enhanced the detection capabilities of YOLOv5 through a Bidirectional Feature Pyramid Network (BiFPN) structure and micro-detection layers. Researchers designed an advanced screening feature fusion pyramid to address intercellular scale differences, and developed a lightweight blood cell detection model, significantly improving detection speed. Researchers have proposed a novel method for automatic instance segmentation of cell nuclei in digitized tissue samples using a deep learning architecture based on the Cell VisionTransformer (CellViT). They have developed STD-YOLOv7, which optimizes micronucleus feature extraction using a coordinate attention (CA) mechanism and a residual self-attention and convolution integration module (Res-ACmix). Despite these improvements, the challenge of addressing the small size of cell micronuclei and their significant scale differences from other cells remains.

[0022] To address the scale issue in micronucleus detection, multi-scale feature fusion techniques are widely used. These techniques improve detection performance by fusing strong semantic information from deep networks with detailed geometric information from shallow networks. Mainstream methods fall into two categories: parallel multi-branch networks and serial skip connection structures. Google Network's (GoogLeNet) Inception module extracts features at different scales through four parallel branches, while Spatial Pyramid Pooling Convolutional Networks (SPPNet) employs a sub-block branching strategy. Feature Pyramid Networks (FPNs) representatively realize the flow of semantic information from deep to shallow layers, while Adaptively Spatial Feature Fusion (ASFF) solves the problem of multi-scale feature inconsistency through dynamic selection and fusion strategies, providing effective technical support for the detection of small targets such as cell micronuclei. BiFPN is a weighted bidirectional pyramid network that can easily and quickly perform multi-scale feature fusion. Frequency-aware Feature Fusion (FreqFusion) integrates an Adaptive Low-pass Filter Generator (ALPF), an offset generator, and an Adaptive High-pass Filter Generator (AHPF) to enhance the high-frequency detail edge information lost during feature fusion.

[0023] Unlike existing methods, the YOLO-MN model proposed in this invention systematically addresses the aforementioned limitations: it explicitly encodes the morphological prior knowledge of the micronucleus through a micronucleus multiple attention module (MAM); it preserves and enhances the edge and texture features of the micronucleus in the frequency domain using a wavelet feature enhancement module (WFE); and it guides the model to more accurately locate and classify cell micronuclei through an innovative MNIoU loss function. This method, which deeply integrates domain knowledge with deep learning technology, provides a more accurate and efficient solution for cell micronucleus detection.

[0024] However, directly applying general object detection models to microkernel recognition still faces many challenges.

[0025] (a) Similarity between micronuclei and the main nucleus: Cellular micronuclei and the main nucleus are highly similar in staining characteristics, both being round or oval in shape, which poses a significant challenge to accurate staining.

[0026] (b) Large variation in micronucleus size: Depending on the formation mechanism of the micronucleus, its diameter is usually between 1 / 16 and 1 / 3 of the diameter of the main nucleus. This wide range of sizes means that the model must be able to detect small targets and distinguish targets of similar size at the same time.

[0027] (c) Loss of detail information: In the process of multi-scale feature fusion, downsampling and channel compression are required to align the scale and number of channels of the feature map. These operations achieve the fusion of high-level semantic information and low-level detail information, but result in the loss of edge and texture information.

[0028] (d) High-throughput requirement: In order to meet the needs of large-scale screening, cell micronucleus detection needs to have high-throughput capability, which can quickly process and analyze a large number of cell images.

[0029] To address these challenges, this invention proposes a high-throughput micronucleus detection model (You Only Look Once - Micronuclei, YOLO-MN) based on prior knowledge and frequency domain awareness. The YOLO-MN model is designed with reference to prior knowledge of micronuclei, significantly improving the accuracy and robustness of micronucleus detection. Specifically, this invention designs a Micronucleus Attention Module (MAM), integrating multiple mechanisms such as channel attention, scale awareness, shape features, and distance relationships. Starting from the morphological characteristics of the micronucleus, the model can focus on the discriminative features of the micronucleus, using the distance awareness module to capture the spatial relationship information between the micronucleus and the main nucleus. Simultaneously, this invention proposes a Wavelet Feature Enhancement (WFE) module, which decomposes the feature map into low-frequency and high-frequency components, processing the overall chromatin distribution and detailed texture respectively. This method effectively extracts multi-scale texture information and enhances the feature representation of edges.

[0030] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0031] Figure 1 This is a schematic diagram of a cell micronucleus detection method based on prior knowledge and frequency domain sensing in this invention, which specifically includes the following steps: S101: Acquire an image of the cell to be detected; the image of the cell to be detected includes multiple cells.

[0032] S102: Based on prior knowledge of cell micronuclei, extract cell micronucleus features from the cell image to be detected from multiple levels and angles; cell micronucleus features include cell micronucleus volume, cell micronucleus morphology, and cell micronucleus distribution area.

[0033] In an exemplary embodiment, multiple bounding boxes in the image to be detected are extracted using a cell micronucleus detection model. The cell micronucleus detection model is an improvement upon the YOLOv8 model. The YOLOv8 model includes a backbone network, a feature fusion pyramid module, and a detection head. The improvement involves adding a micronucleus prior knowledge-enhanced feature extraction module after each Haar wavelet-based downsampling module in the backbone network, and introducing a wavelet transform-based feature enhancement module in the feature fusion pyramid module. The micronucleus prior knowledge-enhanced feature extraction module includes a micronucleus bottleneck block and a micronucleus attention module. The micronucleus bottleneck block includes two parallel main paths and a micronucleus path. The main paths include... Convolution; microkernel path includes Convolution; the microkernel attention module includes a shape feature extraction module and a distance awareness module; the shape feature extraction module includes sequentially connected... Convolution, batch normalization, corrected linear unit and Convolution; the distance-aware module includes dilated convolution, batch normalization, rectified linear units, and... convolution.

[0034] Specifically, Figure 2 This is a structural diagram of the YOLO-MN model provided by the present invention. The YOLO-MN model is the cell micronucleus detection model proposed in this invention. Figure 2Therefore, this invention adds a micronuclei-based prior knowledge-enhanced feature extraction module (C2f with Micronuclei Attention Module, C2f_MAM) to the YOLOv8 backbone network to enhance the micronuclei feature extraction capability of the backbone network. WFE is added before FreqFusion (frequency-aware feature fusion), and finally, a precise detection head is used to predict the location and category. The YOLO-MN model is based on the YOLOv8 backbone and undergoes deep optimization and customized improvements. It mainly consists of three core components: a backbone network integrating micronuclei multi-attention modules, a frequency-aware feature fusion pyramid (FFPN) integrating wavelet feature enhancement, and a precise detection head. In the backbone network design, this invention innovatively encodes the unique biological characteristics of the micronuclei directly into the network architecture, developing a dedicated micronuclei attention module (MAM). Through deep integration and improvement with the C2f structure, this invention constructs a C2f_MAM specifically optimized for cell micronucleus detection, effectively enhancing the ability to perceive micronucleus structures. The FFPN, as a frequency-domain perceptual feature fusion pyramid, focuses on solving the common problem of blurred micronucleus boundaries in microscopic imaging and effectively addresses the challenge of significant scale differences between micronuclei and cells in cell micronucleus images. During feature fusion, to preserve the rich feature information of the image to the greatest extent, this invention introduces efficient downsampling techniques (HWD) to ensure that key features are fully preserved during multi-scale fusion. Finally, the feature map, after multi-level feature extraction and fusion, is fed into the detection head to achieve accurate localization and classification of micronuclei. This multi-module collaborative design significantly improves the model's detection performance and robustness for cell micronuclei.

[0035] Figure 2 In the backbone network, Conv stands for Convolutional Network, Conv2d stands for 2D Convolutional Layer, BN stands for Batch Normalization, SiLU stands for Sigmoid Linear Unit, SPFF stands for Spatial Pyramid Pooling Fast, Maxpool2d stands for 2D Max Pooling Layer, Concat represents a connection, Bboxloss is the bounding box loss, and Cls loss is the classification loss.

[0036] Specifically, the present invention provides as follows Figure 3 The diagram shown is a simplified version of the cell micronucleus detection model. Figure 3 The meaning of the module and Figure 2 Consistent.

[0037] In one exemplary embodiment, the multi-level and multi-angle extraction of cell micronucleus features from the cell image to be detected specifically includes: inputting prior features of the cell image to be detected into the backbone network, and then using the features of the backbone network... Convolution performs channel mapping and separation to obtain the first and second features; for each of the multiple sequentially connected micro-kernel bottleneck blocks, the first and second features are obtained through the main path. Convolution extracts intermediate features from the first feature; through the micro-kernel path Convolution extracts the micro-kernel features of the first feature; the micro-kernel features and intermediate features are weighted and fused to obtain the output features of the micro-kernel bottleneck block; the circular shape features in the output features of the last micro-kernel bottleneck block are extracted using the shape feature extraction module; the spatial relationship features between the micro-kernel and the main kernel in the output features of the last micro-kernel bottleneck block are extracted using the distance perception module; the circular shape features and spatial relationship features are weighted and fused to obtain the enhanced features; the second feature and the enhanced features are then processed... Convolution is used to integrate the data and obtain the cell micronucleus features.

[0038] Specifically, the innovative feature extraction module (C2f with MicronucleiAttention Module, C2f_MAM) designed in this invention addresses the unique characteristics of micronuclei, key biomarkers of cytotoxicity and gene damage, such as their small size, near-circular shape, and location typically around the periphery of the main nucleus. The C2f_MAM module cleverly encodes these biological priors into the network architecture, significantly improving the accuracy and sensitivity of micronucleus detection. The feature extraction module is systematically optimized based on the efficient Cross-Stage Partial (CSP) structure, greatly enhancing the ability to perceive specific morphological features of micronuclei while maintaining computational efficiency. C2f_MAM consists of three carefully designed core components: an improved CSP architecture, a micronucleus bottleneck block (MNBottleneck), and a micronucleus attention module (MAM). This multi-level, multi-angle feature extraction mechanism enables the model to accurately capture subtle features of micronuclei in complex cellular contexts.

[0039] First, the prior features are channel-mapped and separated using convolution, generating two features. The first feature is processed through multiple consecutive MNBottleneck modules, and the output of the last bottleneck block is further enhanced by MAM. All processed features are finally concatenated and integrated through convolution to form the cellular micronucleus feature.

[0040] The working process of the feature extraction module enhanced by microkernel prior knowledge is represented as shown in formula (1): (1) in, and These are the first feature and the second feature, respectively. For the first The output characteristics of each bottleneck block i =1,2,... n , This is a microkernel attention module.

[0041] The micro-core bottleneck block includes a main path and a micro-core path. The first feature is input into multiple sequentially connected micro-core bottleneck blocks for processing. Specifically, for each of the multiple sequentially connected micro-core bottleneck blocks, the intermediate features of the first feature are extracted through the main path; the micro-core features of the first feature are extracted through the micro-core path; and the micro-core features and intermediate features are weighted and fused to obtain the output features of the micro-core bottleneck block. Figure 4 for n A structural diagram of a series of micro-core bottleneck blocks connected in series.

[0042] Specifically, the microkernel bottleneck block MNBottleneck is a bottleneck structure designed specifically for microkernels. This structure contains two parallel paths: the main path extracts intermediate features through two convolutions; the microkernel path uses convolutions to specifically capture microkernel features. The outputs of the main path and the microkernel path are weighted and fused proportionally, and the weighted fusion formula is shown in formula (2): (2) in, Main path, For the microkernel path, and For weights.

[0043] Specifically, Figure 5 This is a schematic diagram of the Microkernel Attention Module (MAM). The MAM contains two parallel processing paths: the upper shape feature extraction module captures the circular shape features of the microkernel through 3×3 convolution; the lower distance perception module expands the receptive field using dilated convolution (dilation=2) to perceive the spatial relationship between the microkernel and the main kernel.

[0044] The Microkernel Attention Module (MAM) is the core innovative component of C2f_MAM, specifically designed to enhance the capture capability of microkernel-specific features, such as... Figure 5As shown, the micronucleus attention module (MAM) consists of a shape feature extraction module and a distance awareness module, forming a multi-dimensional feature enhancement system. In the field of cell micronucleus detection, micronuclei have three key characteristics: their size is significantly smaller than that of the main nucleus, their shape is nearly circular, and they are usually located near the main nucleus. Traditional object detection models often fail to fully utilize this prior knowledge, while MAM weightedly fuses the shape and distance features of the micronucleus.

[0045] Shape feature extraction is specifically designed for the characteristics of microkernel shapes, enhancing the detection sensitivity for circular features through a specific convolutional structure. Firstly, through... The convolutional layers extract edge and shape information, which is then subjected to batch normalization and ReLU activation. Then, it uses... Convolution converts the feature map into a single-channel output and obtains the shape attention map through the Sigmoid function. The corresponding formulas for shape feature extraction are shown in formulas (3) and (4):

[0046] (3) (4) in, The input is a feature map, and ReLU is the activation function. For the Sigmoid function, The result of ReLU processing. This is a shape attention map.

[0047] The distance sensing module leverages the spatial location characteristics of the microkernel, typically found within the main kernel, to enhance the perception of relative spatial location through dilated convolution. It uses a dilation rate (...) )of Convolution expands the receptive field to better capture the spatial relationship between the microkernel and the main kernel, and then through... Convolution and the Sgmoid function yield the distance attention map. The formulas corresponding to the distance-aware module are shown in formulas (5) and (6):

[0048] (5) (6) in, For spatial relationships, This is a distance attention map.

[0049] Ultimately, in the given The fusion of shape attention and distance perception attention is shown in formula (7): (7) in, It is the input feature map. These are the weight parameters for merging convolutions. It is the Sigmoid function.

[0050] S103: Perform continuous channel compression on the cell micronucleus features and extract the frequency domain information of the channel-compressed cell micronucleus features through discrete wavelet transform; the frequency domain information includes high-frequency information and low-frequency information.

[0051] S104: Perform convolution processing on high-frequency and low-frequency information respectively, and fuse the results of the convolution processing to obtain frequency domain sensing fusion features.

[0052] Specifically, YOLOv8's neck employs a Frequency-Aware Feature Fusion Pyramid (FFPN) structure for multi-scale feature fusion, effectively addressing the challenge of scale differences between micronuclei and cells. Deep features in the FFPN typically utilize the FreqFusion module and the Haar wavelet-based downsampling module HWD for cross-scale feature fusion. This combination effectively mitigates the loss of edge and texture details in deep features as the number of network layers increases in the detection of small target cell micronuclei. While the FreqFusion module fuses features at different depths, avoiding target boundary blurring and offset issues caused by standard upsampling, it requires channel compression using convolution in the early stages of fusion. This compression method cannot effectively distinguish and process information at different frequencies, easily losing high-frequency details during compression, thus affecting subsequent frequency-domain feature fusion.

[0053] The Haar wavelet-based downsampling module (HWD) consists of two carefully designed sub-modules: a lossless feature encoding block and a feature representation learning block. The lossless feature encoding block is primarily responsible for efficient feature transformation and spatial resolution reduction, with its core technology based on Haar. The HWD model structure diagram is shown below. Figure 6 As shown.

[0054] Wavelet transform is an advanced method that can effectively reduce the resolution of feature maps while fully preserving information integrity. The subsequent feature representation learning block is composed of finely tuned standard convolutional layers, batch normalization layers, and ReLU activation functions, which are specifically used to extract highly discriminative features and filter redundant information to ensure efficient processing of downstream tasks. Haar wavelet transform, as a mathematical tool with wide applications in image coding, edge feature extraction, and binary logic design, plays a key role in our module. Its theoretical basis of first-order one-dimensional transform can be elegantly expressed through wavelet basis functions and scaling functions, as shown in Equations (8)-(11):

[0055] (8) (9) (10) (11) Formula (9) is The definition, in which, The scale and ordinate of the function are represented. Equation (11) shows that the first-order Haar transform can be represented by the second-order Haar basis functions. Figure 7 As shown, the Haar wavelet transform is applied to effectively decompose the original image into four components with half the spatial resolution, capturing low-frequency and high-frequency information respectively to retain more image information. This transformation reduces image resolution while increasing the number of channels in the feature map to maintain information integrity, thereby optimizing the subsequent multi-scale feature extraction process.

[0056] This is a schematic diagram of the Wavelet Feature Enhancement (WFE) module. It applies Discrete Wavelet Transform (DWT) to decompose the features into four sub-bands: high-frequency detail components (cH, cV, cD) and low-frequency approximation components (cA). Lossless embedding fills the upper left corner with the wavelet transform results, leaving the rest of the area as 0.

[0057] To better address the challenge of high-frequency detail loss caused by pre-fusion channel compression in the FreqFusion module, this invention proposes a wavelet transform-based feature enhancement module (WFE), such as... Figure 6 As shown, this module utilizes wavelet analysis to achieve channel compression while preserving key information, facilitating input into the subsequent Frequency Domain Aware Feature Fusion Pyramid (FFPN) for frequency domain feature fusion. Wavelet transform, as an effective signal processing tool, can decompose an image into components of different frequencies, thereby distinguishing between major structures and detailed information. In cell micronucleus detection, enhancing the edge and texture features of the micronucleus is particularly important. The module mainly includes three key steps: channel compression, wavelet transform processing, and feature fusion. The input features are first subjected to continuous channel compression, then wavelet transform is applied to extract frequency domain information, performing feature enhancement processing on low-frequency and high-frequency information respectively. Finally, the original compressed features are fused with the wavelet-processed features to generate output features rich in information and with a reduced number of channels.

[0058] The channel compression stage employs two consecutive convolutions to progressively reduce the number of channels in the feature map. Each compression layer consists of convolution, batch normalization, and a ReLU activation function, ensuring efficient information transfer. Wavelet transform processing is the core innovation of this module. It performs a two-dimensional discrete wavelet transform (DWT) on the compressed features, decomposing the features into low-frequency components (cA) and three high-frequency components (cH, cV, cD), capturing detailed information in the horizontal, vertical, and diagonal directions, respectively. First, DWT is performed on the feature maps of each batch and each channel. The low-frequency components preserve the semantic information of the image and are enhanced through convolution. The high-frequency components capture the edge and texture information of the image and are processed through detail-enhancing convolutions. Finally, the compressed features and the wavelet-processed features are concatenated along the channel dimension.

[0059] (12) (13) (14) (15) (16) (17) in, This represents a two-dimensional discrete wavelet transform operator. This indicates the selected wavelet basis function. For low-frequency processing operators, by It consists of convolution, batch normalization, and ReLU activation functions. For high-frequency processing operators, This indicates enhanced detail operations, including The process involves grouped convolutions and channel summation. In the feature fusion stage, the channel-compressed features are concatenated with the wavelet-processed features along the channel dimension, and then... The final output of the convolutional convolution is shown in formula (18):

[0060] (18) In one exemplary embodiment, the training process of the cell micronucleus detection model specifically includes: A cell micronucleus dataset was obtained as a training set, and an initial cell micronucleus detection model was constructed. The cell micronucleus dataset includes multiple cell images, and each cell image contains multiple cells. The initial cell micronucleus detection model was trained using the training set, and the AdamW optimizer was used to adjust the initial cell micronucleus detection model to minimize the loss function of the initial cell micronucleus detection model, thus obtaining the cell micronucleus detection model.

[0061] Specifically, to train the cell micronucleus detection model, this invention uses a cell micronucleus dataset as the training set. The cell micronucleus dataset (CMD) was provided by the radiology departments of several provincial disease control centers, representing the challenges of real-world clinical applications. The CMD dataset was precisely labeled by domain experts using the LabelImg tool and includes two key sample types: 726 images containing micronucleated cells and 10,419 control images without micronucleated cells, forming an ideal foundation for evaluating the model's detection capabilities.

[0062] High-resolution synthetic images (1170×1170 pixels) were used for training, with each image containing 100 single cells arranged in a 10×10 grid, including approximately 10-20 micronucleated cells. This design simulates the high-density samples in actual clinical screening while providing sufficient micronucleus morphological variation for the cell micronucleus detection model to learn. This invention expands the original dataset to 263 high-quality images. To ensure the objectivity and reliability of the evaluation, the data was randomly divided in an 8:2 ratio before the experiment, forming a standard testing environment with 211 training images and 52 validation images, providing a solid foundation for evaluating the performance of the cell micronucleus detection model.

[0063] The YOLO-MN model was implemented on Ubuntu 20.04, using Python version 3.8.0, PyTorch deep learning framework 2.0.0, and CUDA framework 11.8. The hardware configuration of the environment consisted of an NVIDIA GeForce RTX2080Ti and an Intel(R) Xeon(R) Platinum 8255C CPU @ 2.50GHz.

[0064] The model was trained in 300 iterations with a batch size of 8. The input image size in the network was [size missing]. The initial learning rate of the model was 0.01. The AdamW optimizer was used to tune the model, and other parameter configurations are shown in Table 1.

[0065] Table 1 In an exemplary embodiment, the loss function of the cell micronucleus detection model is shown in Equation (19): (19) in, , , yes IoU Position penalty items, The weighting coefficient for the penalty term. The distance between the center points of the two bounding boxes. It is the length of the diagonal of the smallest rectangle enclosing the two boxes. This represents the distance between the current microkernel size ratio and the optimal range. The penalty coefficient is the microkernel size. and These represent the width and height of the micronucleus, respectively. The shape penalty scaling factor. The position penalty term weighting coefficient, The weight coefficient for the prior knowledge penalty term of the microkernel. To measure the intersection-union ratio of the detection bounding box and the target bounding box, It is a correction factor. for A perceptual adaptive scaling function, This deviates from the ideal aspect ratio.

[0066] Specifically, the design of the loss function in object detection has a significant impact on model performance. Traditional intersection-to-union (IoU) loss functions and their variants, such as GIoU

[28] , DIoU

[29] , and CIoU

[30] , perform well in general object detection tasks. However, these loss functions do not fully utilize prior knowledge specific to the domain, resulting in failure to achieve optimal performance in specialized scenarios such as cell micronucleus detection. Cell micronucleus detection has its unique challenges, as micronuclei typically exhibit specific size ratios and morphological features relative to the cell nucleus. Figure 7 As shown, this paper illustrates typical occurrences of various impurities in the micronucleus dataset, with red boxes representing micronuclei and blue boxes representing impurities. The micronucleus dataset contains misleading impurities of diverse sizes and shapes. Based on this observation, this invention proposes a Micronucleus Optimized Intersection over Union (IoU) loss function (MNIoU). This function retains the advantages of standard IoU while incorporating domain knowledge of micronucleus morphology, thus optimizing the micronucleus detection task.

[0067] S104: Analyze the frequency domain sensing fusion features to obtain multiple bounding boxes in the image to be detected; the bounding boxes are the predicted locations of cell micronuclei.

[0068] In an exemplary embodiment, the frequency domain sensing fusion features are parsed to obtain multiple bounding boxes in the image to be detected. Specifically, this includes: predicting the geometric parameters of the bounding boxes; the geometric parameters include the offset of the cell micronucleus center point coordinates relative to the grid, the width of the bounding box, and the height of the bounding box; generating multiple candidate detection boxes at each grid point of the frequency domain sensing fusion features using the detection head; and determining multiple bounding boxes based on the multiple candidate detection boxes, combined with the offset, the width of the bounding box, and the height of the bounding box.

[0069] This invention first constructs a set of prior knowledge indicators for micronuclei, which precisely quantifies the morphological characteristics of micronuclei, laying a solid theoretical foundation for the design of the MNIoU loss function. The morphological prior knowledge of micronuclei serves as a key basis for screening and scoring in micronucleus experiments, possessing significant clinical and research value. To systematically apply this prior knowledge, this invention focuses on two core morphological features of micronuclei: size ratio and shape characteristics. Based on literature review, this invention identifies two key features: first, the diameter of the micronucleus is typically 1 / 16 to 1 / 3 of the diameter of the main nucleus; second, micronuclei are mostly circular or elliptical, and in a few cases may be pyramidal, hemispherical, or cylindrical, with very few cases exhibiting irregular shapes. Based on these two features, this invention uses the area ratio of the micronucleus to the main nucleus and the aspect ratio of the micronucleus as quantitative indicators to precisely quantify the size and shape characteristics of the micronucleus, respectively. To adapt to the characteristics of the dataset in this study, this invention conducted statistical analysis on all micronucleus and master nucleus labels in the dataset. The results showed that the area ratio of micronuclei to master nuclei was mainly distributed between 0.05 and 0.20, and the aspect ratio of micronuclei was concentrated between 1.0 and 1.3. These precise quantitative results provide a scientific basis for constructing a highly specific micronucleus morphological screening standard in this invention, significantly improving the accuracy of detection.

[0070] Building upon the traditional CIoU loss function, this invention innovatively introduces a penalty term P based on microkernel prior knowledge, making the loss function more adaptable to the specific needs of microkernel detection. This penalty term consists of two key components: a size-ratio-based penalty term and a shape penalty term. In size-ratio evaluation, this invention designs an adaptive penalty mechanism, ensuring that the penalty intensity is within the optimal range for the size ratio of the detection box to the target box, aligning with the IoU value. , When the aspect ratio deviates from the optimal range, no penalty is applied; when the aspect ratio deviates from the optimal range, the penalty intensity is positively correlated with the degree of deviation, making the system more sensitive to size differences under high IoU conditions, thereby improving detection accuracy. In terms of shape evaluation, the loss function encourages the microkernel to maintain a near-ideal circular or elliptical shape, and penalizes deviations from the ideal aspect ratio. The detection bounding box is subject to a moderate penalty. Furthermore, this invention introduces an IoU-aware adaptive scaling mechanism that automatically reduces the penalty intensity at low IoU values, effectively avoiding the negative impact of excessive penalty on the initial detection stage and ensuring the stability and robustness of the detection.

[0071] Figure 8 This is a geometrical interpretation diagram of the two punishment mechanisms. Figure 8 The areas of the red and green boxes in Figure (a) are not within the optimal range. Figure 8 In Figure (b), the aspect ratio of the red and green boxes deviates from the ideal aspect ratio.

[0072] In one exemplary embodiment, to verify the effectiveness of the model, a comprehensive comparative experiment was conducted on the cell micronucleus dataset. YOLO-MN was systematically compared with current mainstream object detection models, including traditional detection architectures such as Faster-RCNN and SDD, as well as multiple generations of evolved versions of the YOLO series: YOLOv5, YOLOv7, YOLOv8, YOLOv9, YOLOv10, YOLOv11, and YOLOv12, the latest Transformer detector TR-DETR, and the YOLOv5-CEB method specifically for cell micronucleus detection. Furthermore, this invention also provided detailed evaluations of different model-scale variants of YOLOv5, YOLOv8, YOLOv9, YOLOv10, YOLOv11, and YOLOv12 to comprehensively demonstrate the performance advantages of the YOLO-MN model of this invention.

[0073] The results of the YOLO-MN model on the cell micronucleus dataset are shown in Table 2. The proposed YOLO-MN-s model achieved accuracies of 94.7% and 59.1% on the CDM dataset. These results demonstrate that the micronucleus attention module (MAM) and wavelet feature enhancement module (WFE) used in the model of this invention effectively improve the accuracy of cell micronucleus detection. Compared with the state-of-the-art YOLO series algorithm YOLOv12, the model of this invention improves P, R, and by 3.3%, 2.8%, 2.8%, and 5.6%, respectively. Compared with the end-to-end real-time Transformer model RT-DETR, the YOLO-MN model improves by 2.5%, 1.1%, 0.7%, and 2.0%, respectively. Notably, compared with the baseline model YOLOv8, all four metrics show significant improvements. Furthermore, compared with the current state-of-the-art cell micronucleus target detection method YOLOv5-CEB, YOLO-MN achieves improvements of 5.8% and 3.3% in R and , respectively. Figure 9 , Figure 10 and 11 The figures show the changes in mAP@50, mAP@50-95, and recall after training the model of the same scale on the CMD dataset for 300 epochs. These three figures show that YOLO-MN outperforms most other models while maintaining relatively stable performance throughout the training process.

[0074] Table 2 In an exemplary embodiment, to deeply evaluate the contribution of each improved module to the algorithm performance, this invention designed and implemented three sets of systematic ablation experiments. First, this invention compared and analyzed the performance differences between the proposed MNIoU loss function and existing loss functions, quantifying its advantages in the cell micronucleus detection task. Second, in the YOLOv8 Neck structure, this invention conducted comparative experiments with other mainstream feature fusion schemes using the proposed frequency-domain perceptual feature fusion module FFPN with wavelet feature enhancement capabilities, fully verifying the effectiveness of FFPN in small target detection scenarios such as cell micronuclei. Finally, through progressive overall ablation experiments, this invention gradually introduced each improved module, comprehensively evaluating the gain contribution of each component to the final detection performance, thus providing a scientific basis for model design selection.

[0075] To verify the effectiveness of the MNIoU loss function, this invention was compared with traditional CIoU, DIoU, EIoU, GIoU, and ShapeIoU, as shown in Table 3. While maintaining a precision (P) comparable to CIoU (90.5%), MNIoU significantly improved the recall (R) to 92.3%, 2.7 percentage points higher than CIoU. This improvement fully demonstrates the effectiveness of the loss function incorporating prior knowledge of micronucleus morphology in reducing false negatives. In terms of overall detection performance metrics, MNIoU... It reached 93.9%, 0.5 percentage points higher than the best-performing DIoU. In a more stringent... Under the evaluation criteria, MNIoU reached 57.8%, demonstrating a significant advantage over other loss functions. These results indicate that by incorporating micronucleus size distribution and morphological features into the loss function design, MNIoU can more accurately guide the detector to learn the localization and classification of micronuclei. This is particularly true for small, uniquely shaped targets such as cellular micronuclei; loss functions incorporating domain knowledge can effectively improve detection accuracy. Table 3 compares MNIoU with other losses on the MCD.

[0076] Table 3 To address the challenges of small micronucleus targets and significant scale differences between micronuclei and cells in cell micronucleus images, as well as the easy loss of edge and texture information during multi-scale feature fusion, this invention designs an innovative wavelet feature-enhanced frequency domain perceptual feature fusion (FFPN) module. This module effectively integrates high-level semantic information with low-level detail features, while enhancing edge and texture features, thereby achieving more accurate micronucleus localization and classification.

[0077] To verify the effectiveness of the proposed multi-layer adaptive feature fusion module, this invention compared FFPN with several advanced multi-scale feature fusion methods, including AFPN, BiFPN, Gold-YOLO, RepGFPN, HS-FPN, and ASFF, under the same baseline model YOLOv8 and the same dataset CMD. As shown in Table 4, FFPN improved performance metrics by 3.7% and 3.3% compared to ASFF, respectively. Compared with other feature fusion methods, the FFPN proposed in this invention demonstrates significant superiority in the cell micronucleus detection task.

[0078] Experimental results strongly demonstrate that FFPN not only enhances the model's ability to detect small targets but also improves the consistency between different feature scales, thereby promoting the effective fusion of low-level edge features and high-level semantic information in cell micronucleus images and providing reliable technical support for high-precision micronucleus detection. Table 4 compares FFPN with other feature fusion methods on MCD.

[0079] Table 4 According to the improved method proposed in this invention, an ablation experiment of the overall algorithm was conducted. The experiment started with the baseline model YOLOv8s and gradually introduced the proposed improved components. First, after introducing the FreqF module for frequency domain-aware feature fusion into the baseline model, the precision improved by 0.8% and the recall improved by 2.0%, showing a slight improvement. Subsequently, after adding the wavelet feature enhancement (WFE) module, the precision was further improved to 94.0%, reaching 57.9%, confirming the effectiveness of wavelet transform in preserving texture and edge details. Integrating the frequency domain-aware feature fusion and wavelet feature enhancement to form the complete FFPN module significantly improved the model's performance at high IoU thresholds, especially the recall, which increased to 92.1%, indicating that FFPN effectively enhanced the detector's localization accuracy for small targets such as cell micronuclei. Further introducing the micronucleus attention feature extraction module C2f_MAM improved the recall to 93.3%, reaching 94.8%, highlighting the importance of the micronucleus attention mechanism (MAM) in extracting micronucleus features.

[0080] Ultimately, the addition of the loss function MNIoU, which incorporates prior knowledge of the micronucleus, optimizes the model performance, achieving a precision of 91.7% and a recall of 93.4%. The performance improvement is particularly significant at high IoU thresholds, reaching 59.1%, strongly demonstrating the crucial role of the loss function incorporating domain prior knowledge in enhancing detection accuracy. Comprehensive ablation experiments show that each proposed improvement module positively contributes to the model's performance, collectively forming a highly efficient cell micronucleus detection system. Table 5 compares the overall model improvement methods.

[0081] Table 5 pass Figure 12 , 13 The comparison shows that the C2f_MAM module significantly improves feature representation, focusing attention more intently on the micro-core region and clearly distinguishing between the main core and micro-cores. The thermal distribution better matches the morphological characteristics of the micro-core, exhibiting a more regular circular structure. This comparison intuitively demonstrates the effectiveness of the proposed micro-core multi-attention module (MAM) in the paper. By integrating shape feature extraction and distance awareness, the C2f_MAM module enables the model to more accurately identify micro-core features, reduces false activations of irrelevant regions, and improves sensitivity to the micro-core region.

[0082] YOLO-MN was tested on MCD and a PR curve was obtained, such as Figure 14 As shown, the closer the curve is to the upper right corner (1,1), the better the performance of the YOLO-MN model. Figure 14 As can be observed, the curve for cell categories (yellow line) maintains extremely high precision (0.988) almost across the entire recall range, indicating that the model can reliably identify cell structures. For the more challenging micronucleus category (green line), the model also maintains high precision in the lower recall region, but its performance declines somewhat in the high recall region (>0.8), reflecting the inherent difficulty in detecting small micronuclei.

[0083] This invention addresses the challenging problem of micronucleus detection in medical image analysis by proposing a novel deep learning-based detection model, YOLO-MN. Through systematic improvements, this invention successfully solves key problems in micronucleus detection, such as the difficulty of detecting small targets, significant scale differences, and insufficient preservation of edge features during feature fusion.

[0084] This invention designs three innovative modules: (1) a frequency-domain perceptual feature fusion pyramid FFPN based on wavelet feature enhancement WFE, which effectively enhances edge and texture features and improves the model's ability to detect small targets; (2) a micronucleus attention feature extraction module (C2f_MAM), which accurately captures the salient features of micronuclei and enhances the model's ability to recognize micronuclei; and (3) a loss function (MNIoU) combining micronucleus prior knowledge, which reasonably guides the model to learn the localization and classification of micronuclei. Extensive experiments show that YOLO-MN achieves high accuracy and efficiency in cell micronucleus detection, demonstrating significant advantages over existing methods.

[0085] Compared with existing general target detection models and cell micronucleus detection methods, YOLO-MN exhibits superior performance in detecting cell micronuclei, significantly improving overall detection accuracy. The main contributions of this invention are as follows:

[0086] (a) The present invention designs a microkernel multi-attention module (MAM) with shape features and distance awareness and an improved CSP feature extraction network (C2f_MAM) to capture the prior knowledge features of the microkernel in the backbone network.

[0087] (b) The Frequency Domain Sensing Feature Fusion Pyramid (FFPN) proposed in this invention. A wavelet transform-based feature enhancement module (WFE) is designed in the FFPN to address the problem of loss of detail information during feature fusion. The high-frequency subband is enhanced, enabling the YOLO-MN model to capture the subtle differences in chromatin structure between the micronucleus and the main nucleus.

[0088] (c) In order to improve the detection accuracy of the model and accelerate the convergence of the model, the present invention designs the MNIoU loss function by combining the shape characteristics of the microkernel and the positional relationship between the microkernel and the main kernel.

[0089] (d) The YOLO-MN model proposed in this invention performs well in the cell micronucleus image recognition task, achieving a mAP@50 of 94.7%, a recall of 93.4%, and a precision of 91.7%, which is higher than the existing micronucleus detection models and improves the accuracy by 3.3%.

[0090] When applying the cell micronucleus detection method based on prior knowledge and frequency domain sensing provided by this invention, it is not necessary to rely on... Figure 1 The steps shown are executed in sequence. The specific execution order of each step can be determined as needed, and this invention does not impose any restrictions on it.

[0091] The above describes a cell micronucleus detection method based on prior knowledge and frequency domain awareness, provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding cell micronucleus detection device based on prior knowledge and frequency domain awareness, such as… Figure 15 As shown.

[0092] Figure 15 A schematic diagram of a cell micronucleus detection device based on prior knowledge and frequency domain sensing provided by the present invention includes: The acquisition module 1501 is used to acquire an image of the cell to be detected; the image of the cell to be detected includes multiple cells.

[0093] The first extraction module 1502 is used to extract the micronucleus features of the cell image to be detected from multiple levels and angles based on prior knowledge of the micronucleus. The micronucleus features include the volume of the micronucleus, the morphology of the micronucleus, and the distribution area of ​​the micronucleus.

[0094] The second extraction module 1503 is used to continuously compress the cell micronucleus features through channels and extract the frequency domain information of the compressed cell micronucleus features through discrete wavelet transform; the frequency domain information includes high-frequency information and low-frequency information.

[0095] The fusion module 1504 is used to perform convolution processing on high-frequency information and low-frequency information respectively, and fuse the results of the convolution processing to obtain frequency domain-aware fusion features.

[0096] The parsing module 1505 is used to parse the frequency domain sensing fusion features to obtain multiple bounding boxes in the image to be detected; the bounding boxes are the predicted locations of cell micronuclei.

[0097] Specific limitations regarding the cell micronucleus detection device based on prior knowledge and frequency domain awareness can be found in the limitations of the cell micronucleus detection method based on prior knowledge and frequency domain awareness mentioned above, and will not be repeated here. Each module in the aforementioned cell micronucleus detection device based on prior knowledge and frequency domain awareness can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0098] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 The proposed method for detecting cell micronuclei is based on prior knowledge and frequency domain awareness.

[0099] The present invention also provides Figure 16 The schematic diagram of the computer device shown is as follows: Figure 16 As shown, at the hardware level, this computer device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then executes it to achieve the above. Figure 1 The proposed method for detecting cell micronuclei is based on prior knowledge and frequency domain awareness.

[0100] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0101] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

Claims

1. A method for detecting cell micronuclei based on prior knowledge and frequency domain sensing, characterized in that, include: Acquire images of the cells to be detected; The image of the cells to be detected includes multiple cells; Based on prior knowledge of cell micronuclei, cell micronucleus features of the cell images to be detected are extracted at multiple levels and from multiple angles. The characteristics of the cell micronucleus include the volume of the cell micronucleus, the morphology of the cell micronucleus, and the distribution area of ​​the cell micronucleus; The cell micronucleus features are subjected to continuous channel compression, and the frequency domain information of the channel-compressed cell micronucleus features is extracted by discrete wavelet transform. The frequency domain information includes high-frequency information and low-frequency information; The high-frequency information and the low-frequency information are convolved separately, and the results of the convolution are fused to obtain frequency domain-aware fusion features; The frequency domain sensing fusion features are analyzed to obtain multiple bounding boxes in the image to be detected; the bounding boxes are the predicted locations of cell micronuclei.

2. The method as described in claim 1, characterized in that, Multiple bounding boxes are extracted from the image to be detected using a cell micronucleus detection model. The cell micronucleus detection model is an improvement on the YOLOv8 model. The YOLOv8 model includes a backbone network, a feature fusion pyramid module, and a detection head. The improvement method is to add a feature extraction module enhanced with micronucleus prior knowledge after each downsampling module based on Haar wavelet in the backbone network, and to introduce a feature enhancement module based on wavelet transform in the feature fusion pyramid module. The microkernel prior knowledge-enhanced feature extraction module includes a microkernel bottleneck block and a microkernel attention module; The microkernel bottleneck block includes two parallel main paths and a microkernel path; the main path includes... Convolution; the microkernel path includes Convolution; the microkernel attention module includes a shape feature extraction module and a distance awareness module; the shape feature extraction module includes sequentially connected... Convolution, batch normalization, corrected linear unit and Convolution; the distance-aware module includes dilated convolution, batch normalization, corrected linear units, and... convolution.

3. The method as described in claim 2, characterized in that, The multi-level, multi-angle extraction of cell micronucleus features from the cell image to be detected specifically includes: Prior features of the cell image to be detected are input into the backbone network, and then processed through the backbone network... Convolution performs channel mapping and separation to obtain the first and second features; For each of the multiple microkernel bottleneck blocks connected in series, through the main path Convolution extracts intermediate features from the first feature; through the microkernel path Convolution is used to extract the micro-kernel features of the first feature; the micro-kernel features and the intermediate features are weighted and fused to obtain the output features of the micro-kernel bottleneck block; The circular shape feature is extracted from the output features of the last microkernel bottleneck block by the shape feature extraction module. The spatial relationship features between the micronucleus and the main core are extracted from the output features of the last micronucleus bottleneck block using the distance sensing module. The circular shape feature and the spatial relationship feature are weighted and fused to obtain the enhanced feature; The second feature and the enhanced feature are connected via The convolution process is used to integrate the data, resulting in the cell micronucleus features.

4. The method as described in claim 2, characterized in that, The step of parsing the frequency domain sensing fusion features to obtain multiple bounding boxes in the image to be detected specifically includes: Predict the geometric parameters of the bounding box; the geometric parameters include the offset of the cell micronucleus center point coordinates relative to the grid, the width of the bounding box, and the height of the bounding box; The detection head senses each grid point of the fused features in the frequency domain, generating multiple candidate detection boxes; Based on multiple candidate detection boxes, and combined with the offset, the width of the bounding box, and the height of the bounding box, multiple bounding boxes are determined.

5. The method as described in claim 2, characterized in that, The training process of the cell micronucleus detection model specifically includes: A cell micronucleus dataset was obtained as a training set, and an initial cell micronucleus detection model was constructed; the cell micronucleus dataset includes multiple cell images; each cell image includes multiple cells; The initial cell micronucleus detection model is trained using the training set, and then the AdamW optimizer is used to adjust the initial cell micronucleus detection model to minimize the loss function of the initial cell micronucleus detection model, thereby obtaining the cell micronucleus detection model.

6. The method as described in claim 5, characterized in that, The loss function is: ； in, , , yes IoU Position penalty items, The weighting coefficient for the penalty term. The distance between the center points of the two bounding boxes. It is the length of the diagonal of the smallest rectangle enclosing the two boxes. This represents the distance between the current microkernel size ratio and the optimal range. The penalty coefficient is the microkernel size. and These represent the width and height of the micronucleus, respectively. The shape penalty scaling factor. The position penalty term weighting coefficient, The weight coefficient for the prior knowledge penalty term of the microkernel. To measure the intersection-union ratio of the detection bounding box and the target bounding box, It is a correction factor. for A perceptual adaptive scaling function, This deviates from the ideal aspect ratio.

Images