Detection Method for Fungal Cell Populations in High-Density Fluorescence Microscopic Images for Diffusion Models

Abstract

This application relates to the technical field of computer vision and biomedical image processing, and discloses a method for detecting fungal cell populations in high-density fluorescence microscopy images for diffusion models, including: performing multi-scale illumination estimation and adaptive compensation on the input fluorescence microscopy image to correct the illumination non-uniformity in the image; performing local density distribution adjustment on the image to achieve density equilibrium through minimizing the energy functional; performing object detection based on an improved DiffusionDet framework, where the improved DiffusionDet framework optimizes the conditional diffusion process through a dynamic noise scheduling strategy and combines a multi-scale feature fusion network to extract cell morphological features; calculating the spatial relationship and feature similarity between cells, and constructing an affinity matrix to optimize the detection results. This application improves the detection accuracy of high-density overlapping cells.

Description

Technical Field

[0001] This application relates to the technical fields of computer vision and biomedical image processing. Specifically, it is a method for detecting fungal cell populations in high-density fluorescence microscopy images for diffusion models, which is particularly suitable for detecting scenarios of high-density, overlapping cells and low signal-to-noise ratio fluorescence images. Background Art

[0002] The existing cell detection technologies mainly have the following problems:

[0003] 1. Traditional object detection algorithms perform poorly in dealing with high-density and overlapping cells: Region proposal-based methods are prone to generating duplicate detections in overlapping regions; Anchor box-based methods are difficult to adapt to cells with irregular shapes; Single-stage detectors are prone to missing detections in high-density regions. Specifically, traditional object detection algorithms show significant limitations in dealing with high-density and overlapping fungal cell populations: (1) Two-stage detection methods: Algorithms represented by Faster R-CNN have a misdetection rate as high as 35% when dealing with overlapping regions. Especially in regions where the cell density exceeds 200 cells / mm², the problem of duplicate detections is more prominent. (2) Anchor box-based detection methods: When dealing with irregular shapes, the detection accuracy generally drops by 20%-30%. For structures such as hyphae with an aspect ratio exceeding 5:1, the detection recall rate is as low as 40%. (3) Single-stage detectors: Although the inference speed is increased by 3-4 times, the missing detection rate in high-density regions (>300 cells / mm²) reaches 25%, and the recognition accuracy of cell growth states is only 65%.

[0004] 2. The inherent imaging characteristics of fluorescence microscopy images pose challenges: low signal-to-noise ratio, blurred cell boundaries; uneven illumination results in inconsistent image contrast; background fluorescence interference seriously affects detection accuracy. Specifically, fluorescence microscopy technology faces the following specific problems in fungal observation: The signal-to-noise ratio (SNR) of typical fluorescence images is only 8-12 dB, far lower than 20-25 dB of conventional optical images; The uneven illumination causes the brightness difference between the central and edge regions of the image to reach 40%; The autofluorescence intensity of the fungal cell wall can reach 15-20% of the target signal; During long-term observation, the signal intensity attenuation caused by photobleaching effect per hour is about 8-10%.

[0005] 3. Existing methods are difficult to effectively model the spatial relationships between cells: lack of modeling of the distribution characteristics of cell populations; do not fully utilize local structural information; the detection results are vulnerable to uneven density distributions. Specifically, existing algorithms have obvious defects in dealing with the distribution characteristics of fungal populations: for a typical exponentially growing fungal community, the density in the central region can reach 3 - 5 times that of the edge; when existing spatial modeling methods handle such density gradients, the detection accuracy in the edge region is 25% lower than that in the central region; in multi-scale feature fusion, the computational complexity increases exponentially with the number of feature layers, and it takes 2 - 3 seconds to process a 1024×1024 resolution image.

[0006] In recent years, the DiffusionDet algorithm based on diffusion models has shown unique advantages in the field of object detection. By introducing a diffusion process, this algorithm can gradually optimize the position and size of the detection box, effectively solving the performance bottleneck of traditional detectors when dealing with dense objects. The main advantages of DiffusionDet are as follows: it breaks through the dependence on prior boxes of traditional detectors and can adaptively learn the spatial distribution of objects; it improves the detection accuracy through an iterative refinement process, especially performing well when dealing with overlapping objects; it has an end-to-end training paradigm, simplifying the model optimization process.

[0007] However, directly applying DiffusionDet to the detection of high-density fungal fluorescence microscopy images still faces several key challenges: First, the algorithm is sensitive to image quality and signal-to-noise ratio, and its performance is unstable when dealing with low-quality fluorescence images, making it difficult to cope with the imaging characteristics of fungal samples; second, the algorithm is not optimized for the special morphology and growth characteristics of fungal cells and lacks the ability to model the dynamic changes of cell populations; finally, the iterative calculation of the diffusion process brings a large computational overhead, resulting in low efficiency when dealing with large-scale fungal samples. Summary of the Invention

[0008] The purpose of this application is to provide a method for detecting fungal cell populations in high-density fluorescence microscopy images oriented to diffusion models to solve the technical problems raised in the above background art.

[0009] To achieve the above purpose, this application discloses the following technical solutions: A method for detecting fungal cell populations in high-density fluorescence microscopy images oriented to diffusion models, the method comprising the following steps:

[0010] Step 1: Perform multi-scale illumination estimation and adaptive compensation on the input fluorescence microscopy image through a fluorescence signal enhancement module to correct the illumination non-uniformity in the image;

[0011] Step 2: Use a density balance optimization system to adjust the local density distribution of the image and achieve density balance by minimizing the energy functional.

[0012] Step 3: Perform object detection based on the improved DiffusionDet framework. The improved DiffusionDet framework optimizes the conditional diffusion process through a dynamic noise scheduling strategy and combines a multi-scale feature fusion network to extract cell morphological features;

[0013] Step 4: Calculate the spatial relationship and feature similarity between cells through a local affinity module, and construct an affinity matrix to optimize the detection results.

[0014] Preferably, in the said Step 1, the fluorescence signal enhancement module is specifically configured as:

[0015] A. Use four Gaussian kernels with different scales for multi-scale illumination estimation to generate illumination components with different receptive fields;

[0016] B. Based on the adaptive weight fusion of local variance, the weight calculation satisfies the following constraints:

[0017] Constraint 1: The sum of weights at each scale is 1;

[0018] Constraint 2: The range of a single weight is ;

[0019] Constraint 3: The difference between adjacent scale weights is less than or equal to 0.3;

[0020] C. Dynamically calculate the compensation coefficient according to the local illumination intensity , where:

[0021] When , ;

[0022] When , ;

[0023] When , .

[0024] Preferably, in the fluorescence signal enhancement module, the multi-scale feature maps of the input image are generated by four Gaussian kernels with standard deviations of being 3, 7, 15, and 31, and the corresponding receptive fields are 9×9, 21×21, 45×45, and 93×93 pixels respectively; the weight calculation is obtained through the following formula:

[0025]

[0026] where is the local variance of the scale feature map at scale , and is the weight adjustment parameter.

[0027] Preferably, in the step 2, the energy functional adopted by the density balance optimization system is specifically:

[0028]

[0029] where is the regularization parameter, is the target density level, is the image region at a certain point in it is the local cell density,

[0030] Preferably, in the energy functional, the regularization parameter is , and the target density level is set to , and the maximum number of iterations is 100.

[0031] Preferably, in the step 3, the improved DiffusionDet framework specifically includes:

[0032] a. Dynamic noise scheduling strategy: The noise intensity increases linearly with the time step , and the calculation formula is: , where is the total number of diffusion steps;

[0033] b. Multi-scale feature fusion network: Different-level feature fusion is performed through the formula , and the feature weights are optimized by combining the channel attention mechanism, where is the fused feature map, is the high-level feature map, represents the upsampling operation on the high-level features, represents the lateral connection feature map of the current layer, represents the convolution operation.

[0034] Preferably, the backbone network of the multi-scale feature fusion network is ResNet50, and a feature pyramid network is integrated for multi-scale feature extraction.

[0035] Preferably, in the step 4, the calculation method of the affinity matrix of the local affinity module is specifically:

[0036]

[0037] where is the feature vector of cell , is the feature vector of cell ​ Represents the calculation of cells and cells The characteristic similarity between them is and The dimensions of are the spatial position coordinates of cells are the spatial position coordinates of cells is the spatial attenuation coefficient

[0038] Advantageous effects: The method for detecting fungal cell populations in high-density fluorescence microscopic images for diffusion models in this application improves the detection accuracy of high-density overlapping cells, effectively overcomes the imaging defects of fluorescence microscopic images, makes full use of the spatial distribution information of cell populations, and enhances the robustness of the detection model under uneven density distributions. Specifically, by estimating the illumination component with a multi-scale Gaussian kernel, combining the adaptive weight fusion and segmented compensation strategies, the problem of uneven illumination is effectively corrected; based on the energy functional, the local density distribution is dynamically adjusted to solve the problems of missed detection in high-density regions and false detection in low-density regions; by improving the DiffusionDet framework, the detection recall rate of mycelium is increased, and the detection rate of tiny cells is improved, enhancing the detection efficiency; through the design of the local affinity module, the false detection rate is reduced, and the detection accuracy in the colony edge region is improved Brief Description of the Drawings

[0039] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following described drawings are only some embodiments of this application. For those skilled in the art, without creative efforts, other drawings can be obtained based on these drawings

[0040] Figure 1 is the flowchart of the method for detecting fungal cell populations in high-density fluorescence microscopic images for diffusion models provided by the embodiments of this application

[0041] Figure 2 is the schematic diagram of the operation of the system framework corresponding to the method for detecting fungal cell populations in high-density fluorescence microscopic images for diffusion models provided by the embodiments of this application

[0042] Figure 3 is the working principle diagram of the fluorescence signal enhancement module provided by the embodiments of this application

[0043] Figure 4 is the schematic diagram of the detection result obtained by the prior art

[0044] Figure 5 ​​Schematic diagram of the detection result obtained by using the method for detecting fungal cell populations in high-density fluorescence microscopy images for diffusion models provided in this embodiment. Detailed implementation manners

[0045] The technical solutions in the embodiments of the present application will be clearly and completely described below. Apparently, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative efforts shall fall within the protection scope of the present application.

[0046] In this article, the term "including" is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such a process, method, article or device. Without further limitation, the elements defined by the statement "including..." do not exclude the existence of other identical elements in the process, method, article or device including the said elements.

[0047] This embodiment discloses a method for detecting fungal cell populations in high-density fluorescence microscopy images for diffusion models as Figure 1 shown, and in combination with Figures 2 - 5 , a specific introduction to this method is given. Among them, Figure 2 shows four core modules of the entire system: Fluorescence Signal Enhancement Module (FSE), Density Adaptive Enhancement Module (DAE), improved DiffusionDet detection module, and Local Affinity Module (LAM). These modules process the input fluorescence microscopy image in sequence and finally output the cell detection result. Figure 3 shows that the Fluorescence Signal Enhancement Module adopts a multi-scale illumination estimation strategy, estimates the illumination component through four different Gaussian kernels, integrates the multi-scale results through an adaptive weight fusion method, and finally obtains the enhanced image through illumination compensation.

[0048] Specifically, the method includes the following steps:

[0049] Step 1: Perform multi-scale illumination estimation and adaptive compensation on the input fluorescence microscopy image through the Fluorescence Signal Enhancement Module to correct the illumination non-uniformity in the image;

[0050] Step 2: Use the density equalization optimization system to adjust the local density distribution of the image and achieve density equalization by minimizing the energy functional;

[0051] Step 3: Perform object detection based on the improved DiffusionDet framework. The improved DiffusionDet framework optimizes the conditional diffusion process through a dynamic noise scheduling strategy and combines a multi-scale feature fusion network to extract cell morphological features;

[0052] Step 4: Calculate the spatial relationship and feature similarity between cells through a local affinity module, and construct an affinity matrix to optimize the detection results.

[0053] In the actual fluorescence microscopy imaging process, due to the influence of various factors such as light source distribution, sample thickness variation, and the characteristics of the optical system itself, the acquired images often exhibit obvious illumination non-uniformity. This non-uniformity not only affects the visual quality of the images but also interferes with the subsequent image analysis and processing processes. Based on this, this embodiment adopts a collaborative working method of multi-scale illumination estimation, adaptive weight fusion, and compensation coefficient calculation. That is, in the above-mentioned Step 1, the fluorescence signal enhancement module is specifically configured as:

[0054] A - Multi-scale illumination estimation: Use four Gaussian kernels with different scales for multi-scale illumination estimation to generate illumination components with different receptive fields. Specifically, this step includes:

[0055] Set four reference scales for illumination estimation, and the standard deviation of each Gaussian kernel is 3, 7, 15, and 31 respectively, that is

[0056] The multi-scale feature maps of the input image are generated by Gaussian kernels with different scales. The specific expression for generating the Gaussian kernels is:

[0057]

[0058] The formula for calculating the kernel size is:

[0059] The receptive fields corresponding to Gaussian kernels with different scales are 9×9, 21×21, 45×45, and 93×93 pixels respectively.

[0060] B - Adaptive weight fusion: Adaptive weight fusion based on local variance. Specifically, the weight calculation is obtained through the following formula:

[0061]

[0062] where is the local variance of the scale feature map at scale and is the weight adjustment parameter.

[0063] And the weight calculation satisfies the following constraints:

[0064] Constraint 1: The sum of weights at each scale is 1;

[0065] Constraint 2: The range of a single weight is ;

[0066] Constraint 3: The difference between adjacent scale weights is less than or equal to 0.3.

[0067] The core idea of multi-scale illumination estimation is to comprehensively capture the illumination change characteristics through image analysis at different scales. Through Gaussian kernels of different sizes designed for four benchmark scales , it covers illumination changes at various scales from the internal structure of cells to large-scale tissue regions. And by using the formula for weight calculation, it ensures a smooth transition in the spatial domain and avoids artifacts caused by mutations. In the specific implementation, the design of the kernel size ensures that each Gaussian kernel can cover 99.7% of the effective information region while maintaining computational efficiency. Through the settings of these four scales, receptive fields of 9×9, 21×21, 45×45, and 93×93 pixels are respectively formed, which can correspond to illumination change characteristics at different spatial scales. Specifically: The first scale (9×9 pixels) is mainly responsible for capturing the most subtle illumination changes, which is particularly important for retaining the fine structure inside cells; the second scale (21×21 pixels) focuses on illumination changes within the size range of a single complete cell and can effectively smooth local illumination fluctuations; the third scale (45×45 pixels) mainly deals with illumination differences within the range of multiple cell groups; the fourth scale (93×93 pixels) is responsible for capturing the gradual illumination pattern in a larger range to ensure accurate estimation of the overall illumination trend.

[0068] Calculation of C compensation coefficient: The compensation coefficient calculation system adopts a piecewise function strategy and dynamically calculates the compensation coefficient according to the local illumination intensity . Among them,

[0069] Piecewise calculation of the basic compensation coefficient :

[0070]

[0071] Adaptive bandwidth calculation:

[0072]

[0073] Among them, the bandwidth constraint is: Minimum bandwidth: h min = 15, Maximum bandwidth: h max = 60.

[0074] The piecewise strategy adopted is:

[0075] ​Dark region processing (L < 0.2): That is, when , ; Use the maximum compensation coefficient, aiming to significantly enhance the signal strength in the dark region and avoid information loss in the dark region;

[0076] Middle region processing (0.2 ≤ L ≤ 0.8): Adopt an inverse proportional compensation strategy. It is feasible that when ,, In this way, the compensation intensity is dynamically adjusted with the illumination intensity to ensure smooth transition and avoid sudden changes;

[0077] Bright region processing (L > 0.8): That is, when ,, Use the minimum compensation coefficient, which can prevent signal saturation caused by excessive enhancement and maintain the original signal characteristics.

[0078] In this embodiment, the energy functional adopted by the density equalization optimization system in step 2 is specifically:

[0079]

[0080] For the energy functional, its parameter settings are specifically: the regularization parameter is defaulted to 0.1, the target density level is set to 0.5, is the local cell density at a certain point in the image region , is the density gradient, and the maximum number of iterations is 100.

[0081] Furthermore, in the energy functional, the local cell density is normalized to [0, 1], where 0 represents no cells and 1 represents the highest density. Its calculation method is obtained by counting the cell detection results, such as the number of cells detected per unit area. The target density level is used to force the optimized density distribution to tend to be globally uniform and avoid detection errors caused by local over-density or over- sparseness. The reason for setting its value to 0.5 is: Assuming that under ideal conditions, the fungal cell population is evenly distributed in space, so the intermediate value is set to balance the high-density and low-density regions. The regularization parameter is used to balance the weights of the two terms in the energy function. The first term : drives the density to approach the target value; the second term : constrains the smoothness of the density change. The experimental optimization results show that the best balance between density equalization and smoothness can be achieved by cross-validation when λ = 0.1. The larger the λ value, the stronger the influence of the density gradient, and the smoother the density distribution, but local details may be sacrificed. The density gradient Represents the rate of change of density in space, which can be calculated by discrete difference approximation or Sobel operator in the prior art. Its physical meaning is that the greater the gradient, the more drastic the density mutation. Minimizing the gradient square term can suppress density mutation and ensure a smooth transition of the optimized density distribution. For the maximum number of iterations, it is set to 100 to ensure that the function converges within a finite number of steps and avoid infinite loops. Experiments have shown that in most cases, the energy function has tended to be stable after 100 iterations.

[0082] To fully consider the particularity of fluorescence microscopy images and significantly improve the performance of the model in cell detection tasks. In this embodiment, in-depth improvements and optimizations are made to the traditional DiffusionDet object detection framework. Specifically, the improvement process is as follows:

[0083] (1)Conditional diffusion process

[0084] In the conditional diffusion process, a diffusion mechanism based on Gaussian distribution is adopted, and the diffusion process is controlled through a designed probability distribution function In this distribution function, The term controls the degree of retention of the original signal, while The term determines the intensity of the injected noise. This design not only ensures the controllability of the diffusion process but also maintains sufficient randomness, enabling the model to fully explore the possible target state space. In practical applications, this balance is crucial for accurately detecting cells with different brightness and morphologies.

[0085] To more precisely control the diffusion process, a dynamic noise scheduling strategy is designed, where the noise intensity increases linearly with the time step , and the calculation formula is:

[0086] , where is the total number of diffusion steps.

[0087] This strategy uses a smaller noise intensity at the initial stage of diffusion and gradually increases the noise level as time progresses. This progressive noise addition method is particularly suitable for the characteristics of fluorescence microscopy images. At the initial stage of diffusion, the smaller noise helps to retain the fine structure and morphological features of cells; while in the later stage, the larger noise provides sufficient randomness, enabling the model to better handle different cell morphologies and distributions. By setting a base noise level of 0.1 and a maximum increment of 0.9, it is ensured that the noise intensity remains within a reasonable range throughout the diffusion process.

[0088] (2)Feature extraction network

[0089] It is feasible that the backbone network of the multi-scale feature fusion network is ResNet50, and a feature pyramid network is integrated for multi-scale feature extraction. The formula of the multi-scale feature fusion network is:

[0090]

[0091] The multi-scale feature fusion network is used for feature fusion at different levels, and the channel attention mechanism is combined to optimize the feature weights. Among them, is the fused feature map, is the high-level feature map, coming from the deeper layer of the backbone network (such as ResNet50), represents the upsampling operation on the high-level features, for example, bilinear interpolation (BilinearUpsampling) is used for upsampling, represents the lateral connection feature map of the current layer, coming from the shallower layer of the backbone network, represents the convolution operation, which is used to fuse the upsampled high-level features and the lateral connected low-level features. In terms of the feature extraction network, the fusion mechanism of the multi-scale feature fusion network includes an upsampling operation and lateral connection two key components. The upsampling operation is responsible for transmitting the high-level semantic features to the current layer, while the lateral connection retains the local detailed information of the current layer. Finally, these features are adaptively fused through the convolution operation Conv() to generate a richer and more effective feature representation.

[0092] Calculation example:

[0093] Suppose the input is a 1024×1024 image and the backbone network is ResNet50:

[0094] : The feature map from Stage5 (32×32 resolution, 2048 channels);

[0095] : Bilinearly upsampled to 64×64 resolution, and the number of channels remains 2048;

[0096] : The feature map from Stage3 (64×64 resolution, 512 channels), adjusted to 2048 channels through 1×1 convolution;

[0097] Fusion and convolution:

[0098] Add the upsampled feature map (64×64×2048) and the lateral feature (64×64×2048) element by element;

[0099] Generate the final feature map through 3×3 convolution (output channels 256) (64×64×256).

[0100] The design of this feature extraction network is particularly suitable for processing cell detection tasks in fluorescence microscopy images. In actual fluorescence microscopy images, cells often exhibit diverse morphological features: some cells appear as regular circles or ellipses, while others have irregular shapes; some cells have clear boundaries, while others are more blurred; at the same time, there are significant differences in the size of cells, and from tiny organelles to large cell clusters may appear in the same image simultaneously. Through the designed multi-scale feature fusion mechanism, the network can capture these features of different scales and morphologies simultaneously, thus providing more accurate detection results. In practical applications, the detection recall rate of irregular cells (such as mycelia) has been increased from 40% to 75%; the detection rate of tiny cells (diameter <5μm) has been increased from 55% to 80%, the multi-scale fusion module has increased the feature extraction speed by 50%, and the processing time of a single image has been shortened to 0.8 seconds (1024×1024 resolution).

[0101] In addition, the problem of uneven illumination commonly exists in fluorescence microscopy images, which results in cells with significantly different brightness levels may coexist in the same image. To address this challenge, an attention mechanism is introduced into the feature extraction network. This mechanism can adaptively adjust the feature weights of different spatial positions and channels, enabling the network to better focus on those image regions that are of great significance for cell detection, even if the brightness of these regions varies significantly.

[0102] Through these carefully designed improvements, the improved DiffusionDet framework demonstrates excellent performance, especially when dealing with complex fluorescence microscopy images. Experimental results show that the improved framework can not only accurately detect cells of different sizes and morphologies, but also exhibits good robustness to problems such as noise and uneven illumination in the images. These improvements provide a more reliable and efficient tool for the field of biomedical image analysis.

[0103] In this embodiment, the local affinity module involves affinity feature extraction and affinity matrix calculation.

[0104] Specifically, the affinity feature extraction is achieved through non-local feature calculation, and the calculation formula is:

[0105] The calculation of the affinity matrix is obtained through an improved affinity metric, and the calculation formula of the corresponding improved affinity metric is:

[0106]

[0107] Among them, is the cell The eigenvector is extracted from the input feature map through a convolutional layer. is for the cell The eigenvector is extracted from the same feature map through another convolutional layer. represents calculating the and the between cells is and dimensions, with a default , the higher the similarity, the greater the affinity. is for the cell spatial position coordinates, is for the cell spatial position coordinates, is to calculate the squared Euclidean distance between cells. The closer the distance, the stronger the spatial correlation. is the spatial decay coefficient, which controls the sensitivity of position correlation. The default value is 5. The larger it is, the weaker the spatial decay, allowing cells at greater distances to be correlated. The smaller it is, only the correlation of neighboring cells (close distance) is significant.

[0108] Calculation example:

[0109] Assume the features and positions of cells and are as follows:

[0110] Eigenvector: is [0.8, 0.2], is [0.7, 0.3]. For simplicity of calculation, take as a simplified example;

[0111] Position coordinates: is (10, 20), is (12, 22), with σ = 5.

[0112] Calculation steps:

[0113] Feature similarity score: ≈0.438

[0114] Spatial decay term: ≈0.852

[0115] Softmax normalization:

[0116] Assume the sum of similarity scores of other cell pairs is 1.5, then: ≈0.65

[0117] Final affinity: ≈0.554

[0118] Through the design of the local affinity module, it can be obtained through calculation and experiments that the continuous modeling of cell populations reduces the false detection rate by 15%, and the detection accuracy of the colony edge region is increased by 25%. Through sparse calculation (only processing adjacent cell pairs), the computational load of the local affinity module is reduced by 70%.

[0119] From the above, it can be seen that the method for detecting fungal cell populations in high-density fluorescence microscopic images for diffusion models provided in this embodiment has the following technical advantages:

[0120] Significantly improves the detection accuracy of high-density overlapping cells; effectively overcomes the imaging defects of fluorescence microscopic images; makes full use of the spatial distribution information of cell populations; and improves the robustness of the detection model under uneven density distributions.

[0121] In the embodiments provided in this application, it should be understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, code, or any appropriate combination thereof. For hardware implementation, the processor can be implemented in one or more of the following units: application-specific integrated circuit (ASIC), digital signal processor (DSP), digital signal processing device (DSPD), programmable logic device (PLD), field-programmable gate array (FPGA), processor, controller, microcontroller, microprocessor, or other electronic units designed to implement the functions described herein, or a combination thereof. For software implementation, part or all of the processes of the embodiments can be completed by a computer program instructing the relevant hardware. When implemented, the above program can be stored in a computer-readable storage medium or transmitted as one or more instructions or codes on a computer-readable storage medium. The computer-readable storage medium includes computer storage media and communication media, where the communication media includes any medium that facilitates the transfer of a computer program from one place to another. The storage media can be any available medium that can be accessed by a computer. The computer-readable storage medium can include, but is not limited to, RAM, ROM, EEPROM, CD-ROM, or other optical disc storage, magnetic disk storage media, or other magnetic storage devices, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer.

[0122] Finally, it should be noted that the above are only the preferred embodiments of this application and are not used to limit this application. Although this application has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of this application shall be included in the protection scope of this application.

Claims

1. A method for detecting fungal cell populations in high-density fluorescence microscopy images for diffusion models, characterized in that, The method includes the following steps: Step 1: Perform multi-scale illumination estimation and adaptive compensation on the input fluorescence microscopic image through a fluorescence signal enhancement module to correct the illumination non-uniformity in the image; Step 2: Use a density equalization optimization system to adjust the local density distribution of the image and achieve density equalization by minimizing the energy functional; the specific energy functional adopted by the density equalization optimization system is: where λ is the regularization parameter, f target is the target density level, f(z) is the local cell density at a point z in the image region Ω, is the density gradient; Step 3: Perform object detection based on an improved DiffusionDet framework. The improved DiffusionDet framework optimizes the conditional diffusion process through a dynamic noise scheduling strategy and combines a multi-scale feature fusion network to extract cell morphological features; the improved DiffusionDet framework specifically includes: a. Dynamic noise scheduling strategy: noise intensity β t increases linearly with the time step t, and the calculation formula is: where Total is the total number of diffusion steps; b. Multi-scale feature fusion network: Through the formula F l = Conv(U(F l+1 ) + Lateral l ), feature fusion at different levels is performed, and the channel attention mechanism is combined to optimize the feature weights. Among them, F l is the fused feature map, F l+1 is the high-level feature map, U(F l+1 ) represents the upsampling operation on the high-level features, Lateral l represents the lateral connection feature map of the current layer, and Conv(·) represents the convolution operation; Step 4: Calculate the spatial relationship and feature similarity between cells through a local affinity module, and construct an affinity matrix to optimize the detection results.

2. The method for detecting fungal cell populations in high-density fluorescence microscopy images for diffusion models according to claim 1, wherein In the above Step 1, the fluorescence signal enhancement module is specifically configured as: A. Perform multi-scale illumination estimation using four Gaussian kernels with different scales to generate illumination components with different receptive fields; B. Adaptive weight fusion based on local variance, and the weight calculation satisfies the following constraints: Constraint 1: The sum of weights at each scale is 1; Constraint 2: The range of a single weight is 0.1 ≤ w k ≤ 0.6; where k is the scale; Constraint 3: The difference between adjacent scale weights is less than or equal to 0.3; C. Calculate the local illumination intensity L(x, y) and the dynamic compensation coefficient α(x, y), where: When L(x, y) < 0.2, α(x, y) = 2.5; When 0.2 ≤ L(x, y) ≤ 0.8, When L(x, y) > 0.8, α(x, y) = 0.

4.

3. The method for detecting fungal cell populations in high-density fluorescence microscopic images for diffusion models according to claim 2, characterized in that, In the fluorescence signal enhancement module, the multi-scale feature maps of the input image are generated by four Gaussian kernels with standard deviations of σ k being 3, 7, 15, and 31, and the corresponding receptive fields are 9×9, 21×21, 45×45, and 93×93 pixels respectively; the weights are calculated by the following formula: Among them, V k is the local variance of the scale feature map at scale k, and σ w is the weight adjustment parameter.

4. The method for detecting fungal cell populations in high-density fluorescence microscopy images for diffusion models according to claim 1, wherein, In the energy functional, the regularization parameter is λ = 0.1, the target density level is set to f target = 0.5, and the maximum number of iterations is 100.

5. The method for detecting fungal cell populations in high-density fluorescence microscopic images for diffusion models according to claim 1, characterized in that, The backbone network of the multi-scale feature fusion network is ResNet50, and a feature pyramid network is integrated for multi-scale feature extraction.

6. The method for detecting fungal cell populations in high-density fluorescence microscopic images for diffusion models according to claim 1, wherein In the above Step 4, the calculation method of the affinity matrix of the local affinity module is specifically: Among them, Q i is the feature vector of cell i, K j is the feature vector of cell j, represents calculating the feature similarity between cell i and cell j, d k is the dimension of Q i and K j , p i is the spatial position coordinate of cell i, p j is the spatial position coordinate of cell j, and ξ is the spatial attenuation coefficient.

Images