A fluorescent image recognition system for detecting fungi based on a deep learning algorithm

Abstract

The application provides a fluorescent image recognition system for detecting fungi based on a deep learning algorithm, comprising an image acquisition module, a single spore and mycelium recognition module and a mycelium cluster recognition module; the image acquisition module is used for obtaining a target fluorescent image; the single spore and mycelium recognition module and the mycelium cluster recognition module are used for inputting the target fluorescent image into a pre-established deep learning algorithm model, recognizing single spores or mycelia and mycelium clusters through the pre-established deep learning algorithm model, and outputting fungal negative or positive results according to the recognition results. Compared with the prior art, after the image acquisition module obtains the target fluorescent image, the target fluorescent image is input into the pre-established deep learning algorithm model for automatic analysis, and the analysis of spores / mycelia and mycelium clusters that are different in morphology is divided into two modules for processing, so that the spores, mycelia and mycelium clusters can be more accurately recognized and distinguished.

Description

Technical Field

[0001] This invention relates to the field of medical testing technology, and in particular to a fluorescence image recognition system for detecting fungi based on a deep learning algorithm. Background Technology

[0002] Fungi are widely distributed in nature. When the human body has a weakened immune system or unhealthy lifestyle habits, it is susceptible to fungal invasion and infection. Clinically, pathogenic fungi are classified into superficial and deep fungal infections based on the location of infection. Superficial fungal infections of the skin have a higher incidence in summer, including tinea corporis, tinea manuum, and tinea pedis. Their clinical symptoms are often atypical, and the condition is prone to recurrence. Deep fungal infections, on the other hand, have seen an increasing incidence and mortality rate globally over the past two decades. Despite this, rapid and efficient diagnostic techniques are still lacking.

[0003] In laboratory testing for fungal infections, traditional microscopic examination remains fundamental. This method relies on technicians operating a microscope for observation and is widely used in clinical diagnosis. For human samples from sterile sites, a positive microscopic examination confirms a fungal infection. However, this method requires manual analysis of the target fluorescence image by laboratory personnel, demanding a high level of experience. Furthermore, the interpretation of results is subjective, with potential discrepancies between different personnel, leading to lower sensitivity and positive rates. Additionally, manual interpretation is time-consuming and inefficient, requiring constant adjustments to the field of view under the microscope to ensure all areas are examined, consuming significant manpower and time.

[0004] Experienced laboratory technicians can identify fungi such as Candida, Aspergillus, Malassezia, Cryptococcus, and Mucor through direct microscopic examination. However, in reality, there is a relative shortage of technicians proficient in fungal microscopy, and the heavy workload inevitably affects the accuracy of test results. Especially in primary healthcare institutions, due to the lack of experience among technicians, there is a risk of missed detections due to failure to identify the fungi, or false positive results due to misinterpretation, all of which can impact the formulation of subsequent treatment plans.

[0005] In recent years, methods based on deep neural networks, especially deep convolutional neural networks, have been widely applied to many computer vision and pattern recognition tasks, achieving results surpassing human performance in the classification of natural images. Previous studies have mostly used the same artificial intelligence algorithm to process fungal spores, hyphae, and hyphal clusters. For example, CN115345828A provides a method for identifying lesions in fungal immunofluorescence staining images based on an artificial intelligence algorithm. This method identifies these three types of fungal morphology in images by setting different area thresholds for hyphae, hyphal clusters, and spores using an artificial intelligence algorithm. However, clusters of hyphae and / or spores have complex structures, resulting in significant differences in area in fungal staining images compared to isolated hyphae or spores. Furthermore, the morphology of spores and / or hyphae within hyphal clusters in fungal staining images may exhibit considerable changes compared to isolated spores or hyphae. Therefore, using a single artificial intelligence algorithm may struggle to effectively identify these three different fungal morphologies simultaneously, leading to decreased identification accuracy. Therefore, developing a new, automated fungal detection technology to improve the accuracy and efficiency of fungal detection is an urgent problem to be solved in the field of medical testing. Summary of the Invention

[0006] This invention provides a fluorescence image recognition system for detecting fungi based on a deep learning algorithm, which can effectively solve the problems of low efficiency and accuracy of fungal microscopic examination in the prior art.

[0007] Therefore, in a first aspect, the present invention provides a fluorescence image recognition system for detecting fungi based on a deep learning algorithm, comprising: an image acquisition module, a single spore and hyphae recognition module, and a hyphal cluster recognition module;

[0008] The image acquisition module is used to obtain the target fluorescence image;

[0009] The single spore and hyphae identification module and the hyphae cluster identification module are used to input the target fluorescence image into a pre-established deep learning algorithm model, identify single spores or hyphae and hyphae clusters through the pre-established deep learning algorithm model, and output fungal negative or positive results based on the identification results.

[0010] The pre-established deep learning algorithm model includes a deep learning-based target detection algorithm and an image classification algorithm. The single spore and hyphae identification module uses the deep learning-based target detection algorithm, and the hyphae cluster identification module uses the deep learning-based image classification algorithm.

[0011] Preferably, the target detection algorithm is a target detection algorithm based on convolutional neural networks, and the image classification algorithm is a lightweight image classification algorithm based on convolutional neural networks.

[0012] Preferably, the target detection algorithm is a YOLO series algorithm, and more preferably a YOLOX algorithm.

[0013] Preferably, the image classification algorithm is the MobileNet algorithm, and more preferably the MobileNet V2 algorithm.

[0014] Preferably, the mycelial cluster identification module further includes segmenting the target fluorescence image into blocks before inputting the target fluorescence image into the image classification algorithm model in the deep learning algorithm model, inputting the segmented blocks into the image classification algorithm model, and outputting a result that merges all the blocks of the target fluorescence image.

[0015] Preferably, the fluorescence image recognition system further includes integrating the output results of the single spore and hyphae recognition module with the output results of the hyphae cluster recognition module, and making a final judgment based on the integrated result to output a fungal negative or positive result.

[0016] Preferably, the target detection algorithm identifies the target and assigns a confidence level corresponding to the target, and the image classification algorithm classifies each small patch and assigns a confidence level indicating the presence of mycelial clusters. Furthermore, the integration is performed using a linear equation and an activation function, wherein the linear equation is expressed as follows:

[0017]

[0018] Where z is the result of a linear combination of the output of the single spore and hyphae identification module and the output of the hyphae cluster identification module; P 1. P 2. P 3 represents the probability that there are no spores, hyphae, or hyphal clusters in the target fluorescence image. ,in X i The nth spore or hyphae target with the highest confidence level identified by the target detection algorithm i The confidence level for a single spore or hyphae target is preferably n = 3-7. , where Z is the maximum confidence score identified by the image classification algorithm; β0 is the intercept term of the model, and β1, β2, and β3 correspond to the weights of spores, hyphae, and hyphal clusters, respectively;

[0019] The activation function is a non-linear activation function, preferably a sigmoid function, which is used to convert the linear combination result z into a probability value. P p When the probability value P p When the probability value is greater than or equal to a set threshold C, a fungal positive result is output; when the probability value is greater than or equal to a set threshold C, a fungal positive result is output.P p If the result is less than the set threshold C, output a negative result for fungi.

[0020] Preferably, the training process of the pre-established deep learning algorithm model specifically includes:

[0021] Obtain sample images of labeled fungal spores, hyphae, or hyphal clusters;

[0022] A deep learning algorithm model is constructed based on the object detection algorithm and image classification algorithm, and the deep learning model is iteratively trained using the sample images until the deep learning algorithm model converges.

[0023] Preferably, the image acquisition module includes a scanning module for scanning to generate a fluorescence image.

[0024] Preferably, the image acquisition module further includes an image quality optimization module, which is used to preprocess the fluorescent pattern to obtain a clearer fluorescent image.

[0025] Preferably, the fluorescence image recognition system also presents the analysis progress of the individual spore and hyphae recognition module and the hyphal cluster recognition module and / or the detection results of spore, fungus and hyphal cluster morphology in the form of visual annotations.

[0026] In a second aspect, the present invention relates to a fluorescence image recognition method for detecting fungi based on a deep learning algorithm, which uses the fluorescence image recognition system for detecting fungi of the first aspect for analysis.

[0027] In a third aspect, the present invention relates to a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to implement the fluorescent image recognition method for detecting fungi, as described in the second aspect.

[0028] The following benefits can be obtained by implementing the present invention:

[0029] This invention provides a fluorescence image recognition system for detecting fungi based on a deep learning algorithm, comprising: an image acquisition module, a single spore and hyphae recognition module, and a hyphal cluster recognition module; the image acquisition module is used to acquire a target fluorescence image; the single spore and hyphae recognition module and the hyphal cluster recognition module are used to input the target fluorescence image into a pre-established deep learning algorithm model, and to identify single spores, hyphae, and hyphal clusters through the pre-established deep learning algorithm model, and output a fungal negative or positive result based on the recognition result; wherein the pre-established deep learning algorithm model includes a deep learning-based target detection algorithm and an image classification algorithm, and the single spore and hyphae recognition module adopts the deep learning-based target detection algorithm, and the hyphal cluster recognition module adopts the deep learning-based image classification algorithm. Compared to existing technologies, this invention, after obtaining the target fluorescence image through the image acquisition module, inputs the fluorescence image into a pre-established deep learning algorithm model for automated analysis. The analysis of spores / hyphae and hyphal clusters, which differ significantly in morphology, is handled in two separate modules. This allows for more accurate identification and differentiation of spores, hyphae, and hyphal clusters, indicating the fungal morphology in the sample and providing quantity and visual target selection results (for spores and hyphae). This can serve as a valid basis for physician review. In this way, the cost of traditional manual fluorescence image interpretation can be significantly reduced, while improving the efficiency and accuracy of detection. Attached Figure Description

[0030] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention as described above or otherwise will become clearer.

[0031] Figure 1 This is a schematic diagram of the process for fungal identification using the fluorescent image recognition system for detecting fungi based on deep learning algorithms according to the present invention.

[0032] Figure 2 This is a schematic diagram of the software architecture of the fluorescence image recognition system for detecting fungi based on deep learning algorithms according to the present invention.

[0033] Figure 3 This is a schematic diagram showing the result of the single spore and hyphae recognition module of the fluorescence image recognition system for detecting fungi based on deep learning algorithm of the present invention recognizing the target fluorescence image.

[0034] Figure 4 This is a schematic diagram showing the result of the mycelial cluster recognition module of the fluorescent image recognition system for detecting fungi based on deep learning algorithm of the present invention recognizing the target fluorescent image.

[0035] Figure 5 This is a flowchart of Embodiment 1 of the present invention.

[0036] Figure 6 The receiver operating characteristic (ROC) curve for threshold C is plotted based on the dataset used in Example 1.

[0037] Figure 7 An exemplary fluorescent image of the target to be detected is shown in Example 2.

[0038] Figure 8 The visualization results show the target identification results after the fluorescence image of the target to be detected is analyzed by the YOLOX model, with spores marked in yellow boxes and hyphae marked in purple boxes. Detailed Implementation

[0039] Some specific embodiments of the invention are described here for illustrative purposes and not for limitation. It will be apparent to those skilled in the art that various modifications and variations can be made to the invention without departing from its scope or spirit. For example, features described or illustrated as part of one embodiment can be used in another embodiment to produce further embodiments.

[0040] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for descriptive purposes only and is not intended to limit the scope of the invention.

[0041] As used throughout, a range is used as a shorthand to describe each and all values ​​within that range. Any value within the range, such as an integer, can be used as the endpoint of the range. For example, the range 20%-50% is used to describe all values ​​within that range, such as 20%, 21%, 22%, 23%, 24%, 25%, 30%...50%, and includes all subranges, such as 20%-40%, 25%-35%, 30%-40%, etc.

[0042] I. Definition

[0043] As used in this invention, "fungus" refers to a large class of eukaryotic microorganisms with highly differentiated nuclei containing a nuclear membrane and nucleolus, complete organelles in the cytoplasm, cell walls composed of chitin or cellulose, and lacking chlorophyll. They do not differentiate into roots, stems, or leaves, and a few are unicellular, while most are multicellular structures.

[0044] As used in this invention, "spore" refers to a round or oval structure produced by reproductive hyphae. It is the reproductive structure of fungi and a major basis for fungal identification and classification.

[0045] As used in this invention, "hypha" refers to a tubular, single, filamentous structure in fungi, which is the structural unit of most fungi. Under suitable conditions, spores sprout young shoots called germ tubes, which gradually elongate into hyphae.

[0046] As used in this invention, "mycelium" refers to a vegetative body composed of many hyphae and / or spores linked together.

[0047] As used in this invention, "fluorescence" refers to a photoluminescence phenomenon, where a substance at room temperature is irradiated by incident light of a certain wavelength, absorbs light energy, enters an excited state, and immediately de-excites and emits outgoing light with a wavelength longer than the incident light. Outgoing light with this property is called fluorescence.

[0048] As used in this invention, "deep learning (DL)" is a branch of machine learning. It refers to a method of machine learning that uses multi-layered neural networks to learn the inherent patterns and representation levels of sample data. The ultimate goal is to enable machines to have analytical and learning capabilities like humans, and to be able to recognize data such as text, images, and sounds.

[0049] The "object detection algorithm" used in this invention is a type of computer vision algorithm that can identify multiple objects in an image and return the position (usually in the form of a bounding box) and category of each object.

[0050] The "YOLO series algorithm (You Only Look Once)" used in this invention is a target detection algorithm based on convolutional neural network (CNN). It is used to identify and locate multiple objects in images or videos in real time. Its main features are high speed and high accuracy, and it can achieve fast target detection in real-time scenarios.

[0051] The "image classification algorithm" used in this invention is a type of computer vision algorithm whose main task is to assign an input image to one or more predefined category labels. The input of the algorithm is an image, and the output is the probability distribution of the category to which the image belongs or the specific category label.

[0052] The "MobileNet algorithm" used in this invention is a lightweight image classification algorithm based on convolutional neural networks, designed specifically for mobile devices and embedded vision applications. Its core idea is to decompose the traditional convolution operation into two independent operations: depthwise convolution and pointwise convolution.

[0053] As used in this invention, "accuracy" refers to the proportion of correctly classified samples out of the total number of samples.

[0054] As used in this invention, "recall" refers to the proportion of samples that are actually positive that are correctly predicted as positive.

[0055] The "F1 score" used in this invention is a statistical indicator used to measure the precision of a binary classification model. It takes into account both the precision and recall of the classification model and can be regarded as a harmonic average of the model's precision and recall. The maximum value is 1 and the minimum value is 0.

[0056] As used in this invention, "microscope" is short for microscopic examination, which includes taking samples of the specimen to be examined, preparing slides, and observing, analyzing, and judging them under a microscope. The sample objects include fungi, exfoliated cells, or human tissues, animal tissues, plant cells, etc.

[0057] As used in this invention, "IoU (Intersection over Union)" refers to an evaluation metric that measures the accuracy of detecting corresponding objects in a specific dataset. It is also known as the intersection-over-union ratio, which compares the predicted bounding box with the ground truth bounding box (i.e., manually labeled box) to evaluate the accuracy of the model and measure the degree of overlap between the predicted bounding box and the ground truth bounding box.

[0058] As used in this invention, "confidence interval" refers to the probability that the population parameter value falls within a certain interval of the sample statistic, or the degree of certainty that the population parameter value falls within a certain interval of the sample statistic. It reflects the reliability of the sampling.

[0059] As used in this invention, "training set" refers to data samples used for model fitting and for training parameters in a neural network.

[0060] As used in this invention, the "validation set" refers to a set of samples reserved separately during model training. It can be used to adjust the model's hyperparameters and to conduct a preliminary evaluation of the model's capabilities. It is typically used during iterative model training to verify the current model's generalization ability (accuracy, recall, etc.) in order to decide whether to stop training.

[0061] As used in this invention, the "test set" refers to a sample set used to evaluate the generalization ability of the final model. However, it cannot be used as the basis for algorithm-related selections such as parameter tuning and feature selection.

[0062] The "block cutting" used in this invention refers to dividing an image into multiple small patches, which can solve problems such as limited resources and difficult data management, thereby improving image processing efficiency and accuracy.

[0063] II. The Method of the Invention

[0064] This invention provides a fluorescence image recognition system for detecting fungi based on a deep learning algorithm, comprising: an image acquisition module, a single spore and hyphae recognition module, and a hyphal cluster recognition module;

[0065] The image acquisition module is used to obtain the target fluorescence image;

[0066] The single spore and hyphae identification module and the hyphae cluster identification module are used to input the target fluorescence image into a pre-established deep learning algorithm model, identify single spores or hyphae and hyphae clusters through the pre-established deep learning algorithm model, and output fungal negative or positive results based on the identification results.

[0067] The pre-established deep learning algorithm model includes a deep learning-based target detection algorithm and an image classification algorithm. The single spore and hyphae identification module uses the deep learning-based target detection algorithm, and the hyphae cluster identification module uses the deep learning-based image classification algorithm.

[0068] In a preferred embodiment, the fluorescent image recognition system for detecting fungi also presents the analysis progress and / or detection results of spore, fungus, and hyphal cluster morphology in the form of visual annotations for individual spore and hyphal cluster identification modules and hyphal cluster identification modules. For hyphal cluster targets, only presence or absence is indicated; for hyphal and spore targets, in addition to presence or absence, information on quantity, location, and size is provided. Furthermore, the fluorescent image recognition system for detecting fungi can also display sample information.

[0069] In a preferred embodiment, the training process of the pre-established deep learning algorithm model specifically includes: acquiring sample images of labeled fungal spores, hyphae, or hyphal clusters; constructing a deep learning algorithm model based on the target detection algorithm and image classification algorithm; and iteratively training the deep learning model using the sample images until the deep learning algorithm model converges. In a more preferred embodiment, the target detection algorithm model in the deep learning model is trained and validated using sample images of labeled spores and / or hyphae, and the image classification algorithm model in the deep learning model is trained and validated using sample images of labeled hyphal clusters. In an even more preferred embodiment, the sample images are divided into a training set and a validation set according to a ratio (e.g., 6:4, 7:3, or 8:2, etc.), the deep learning model is iteratively trained using the sample images in the training set, and the deep learning algorithm model is confirmed to have converged using the sample images in the validation set. For example, when the accuracy of both the training set and the validation set tends to stabilize or reaches a certain preset high threshold, the model can be considered to have converged. In a preferred embodiment, the annotation of fungal spores, hyphae, or hyphal clusters in the sample image is performed using annotation software, such as APTime.

[0070] The deep learning algorithm model used in the fluorescence image recognition system for detecting fungi of the present invention includes a deep learning-based target detection algorithm and an image classification algorithm. For the deep learning-based target detection algorithm, the input is an image, and the output is the bounding box (e.g., fungal spores or hyphae) of each detected target object (in the embodiments of the present invention) in the labeled image. Figure 3 As shown), and each detected target object is assigned a category label; for deep learning-based image classification algorithms, the main task is to assign the input image to one or more predefined category labels (such as...). Figure 4 (As shown).

[0071] In a preferred embodiment, the object detection algorithm is a convolutional neural network-based object detection algorithm, including but not limited to the YOLO series algorithms, RetinaNet detection algorithm, SSD detection algorithm, Faster R-CNN detection algorithm, and Mask R-CNN detection algorithm; and the image classification algorithm is a lightweight convolutional neural network-based image classification algorithm, including but not limited to SqueezeNet, MobileNet, ShuffleNet, and Xception.

[0072] In a preferred embodiment, the target detection algorithm is a YOLO series algorithm, such as YOLOv1-v10, YOLOX, etc. In a more preferred embodiment, the target detection algorithm model in the deep learning algorithm model used in the fluorescence image recognition system of the present invention is built based on the YOLOX algorithm. YOLOX is a latest high-performance target detection model, whose performance surpasses previous YOLO series models. In other words, the single spore and hyphae identification module uses the YOLOX algorithm to identify single spores or hyphae. For each detected target object in the image, YOLOX outputs a bounding box, which is typically represented by four values: the x and y coordinates of the object's center point (ratios relative to the image width and height), the object's width and height, and each detected object is assigned a category label. Based on these bounding boxes and category labels, the quantity of each type of fungal morphology can be counted, and a qualitative three-classification result can be obtained: hyphae present, spores present, and no fungi present.

[0073] In a preferred embodiment, the image classification algorithm is the MobileNet algorithm, such as MobileNetV1, MobileNetV2, MobileNetV3, etc. In a preferred embodiment, the mycelial cluster identification module using the MobileNet algorithm further includes segmenting the target fluorescence image into blocks before inputting it into the MobileNet algorithm model, inputting the segmented blocks into the MobileNet algorithm model, and outputting the result of merging all the blocks of the target fluorescence image. Introducing block segmentation can transform the target fluorescence image into fine image blocks, improving recognition accuracy. Its core idea is to process the image into blocks, classify them separately, and then integrate them to obtain the classification judgment of the whole image, which is especially suitable for areas that are difficult to identify by the YOLOX algorithm. In addition, the key points of block segmentation are: 1. When processing high-resolution images, the total image data volume is huge, and direct processing can easily lead to insufficient memory. 2. By segmenting the image into small blocks, the amount of data processed at one time can be reduced, thereby adapting to memory limitations. 3. Since the fungal target is small, block segmentation can make the model more focused on the local features of the image, improve the robustness of the model to small targets, and also reduce redundant calculations caused by the overlap of adjacent pixels in traditional convolutional neural networks. In a preferred embodiment, the image is segmented into 4-64 blocks (e.g., 4, 8, 16, 32, or 64 blocks, depending on the image size and model requirements; for example, a 2048×2048 pixel image can be segmented into 8 512×512 pixel images or 16 256×256 pixel images). In a preferred embodiment, the image segmentation uses an overlapping segmentation method, with the overlapping pixels adjusted according to the image size and model requirements, for example, 20%-50% of the image block pixels. In a more preferred embodiment, the image segmentation uses a 160-pixel overlapping segmentation method, segmenting the image into 8 512*512 pixel images, which are then input into the MobileNet model.

[0074] In a more preferred embodiment, the image classification algorithm model in the deep learning algorithm model used in the fluorescence image recognition system of the present invention is constructed based on the MobileNet V2 algorithm. In other words, the mycelial cluster recognition module uses the MobileNet V2 algorithm to classify whether there are mycelial clusters. The MobileNet V2 algorithm is only used to classify whether there are mycelial clusters. As a convolutional neural network architecture, it can reduce the amount of computation and improve the classification accuracy, accurately identifying complex structures such as mycelial clusters. For each small block, it outputs two results: no mycelial clusters or mycelial clusters. Finally, it merges all the results of the small blocks to give the conclusion of the whole image.

[0075] In a preferred embodiment, the fluorescence image recognition system further includes integrating the output results of the single spore and hyphae recognition module with the output results of the hyphal cluster recognition module, and making a final judgment based on the integrated result to output a fungal negative or positive result.

[0076] In a more preferred embodiment, the target detection algorithm identifies targets and assigns a confidence level corresponding to the targets, and the image classification algorithm classifies each patch and assigns a confidence level for mycelial clusters, and the integration is performed using a linear equation and an activation function, wherein the linear equation is expressed as follows:

[0077]

[0078] Where z is the result of a linear combination of the output of the single spore and hyphae identification module and the output of the hyphae cluster identification module; P 1. P 2. P 3 represents the probability that there are no spores, hyphae, or hyphal clusters in the target fluorescence image. ,in The nth spore or hyphae target with the highest confidence level identified by a target detection algorithm (such as the YOLOX algorithm) i The confidence level of each spore or hyphae target is determined (the n spores with the highest confidence levels are selected to prevent false positives caused by many objects similar to spores being predicted as having only a slight resemblance to spores). Preferably, n is 3-7, for example, n can be 3, 4, 5, 6, or 7. In a most preferred embodiment, n is 5. ; , where Z is the maximum confidence score identified by the image classification algorithm (e.g., MobileNet V2 algorithm); β0 is the intercept term of the model, and β1, β2, and β3 correspond to the weights of spores, hyphae, and hyphal clusters, respectively.

[0079] The above linear expression equation can be derived through the following process: the probability that there are no arbitrary targets (including spores, hyphae, and hyphal clusters) on the target fluorescence image. This formula can be transformed by taking the logarithm of both sides of the equation. , log P 1. log P 2. log P Using 3 as input, the linear expression equation can be obtained by modeling with logistic regression. In the preferred implementation, the logarithm of both sides of the equation is taken as the natural logarithm.

[0080] In a preferred embodiment, the weighting coefficients β1, β2, and β3 are optimized from the training data using maximum likelihood estimation to quantify the contribution of spores, hyphae, and hyphal clusters to the overall positive probability; and β0 represents the baseline probability when there are no spores, hyphae, or hyphal clusters. For example, in one embodiment, β0, β1, β2, and β3 are -2.80871019, -1.33605743, -1.52315955, and -0.40059451, respectively.

[0081] The activation function is a non-linear activation function, preferably the Sigmoid function, which is used to convert the linear combination result z into a probability value. P p The Sigmoid function is known to have the following form:

[0082]

[0083] Substituting the linear combination result z into the Sigmoid function yields the hypothesis function for logistic regression, thus transforming the linear combination result z into a probability value. P p The expression is:

[0084]

[0085] When the probability value P p When the probability value is greater than or equal to a set threshold C, a fungal positive result is output; when the probability value is greater than or equal to a set threshold C, a fungal positive result is output. P p When the value is below a set threshold C, a negative fungal result is output. The selection of threshold C is based on ROC curve analysis to balance the sensitivity and specificity of the model, ensuring its diagnostic efficacy in practical applications. An ROC curve is a coordinate graph analysis tool where the horizontal axis represents the false positive rate (FPR) and the vertical axis represents the true positive rate (TPR). When selecting the threshold, the point where sensitivity and specificity achieve the optimal balance is chosen as the value of threshold C based on the desired diagnostic efficacy. For example, in a preferred embodiment, threshold C can be set to 0.28.

[0086] In a preferred embodiment, the image acquisition module includes a scanning module for scanning to generate fluorescence images. In a more preferred embodiment, the scanning module consists of multiple scanners connected to a computer, which can work in parallel and in a fully automated manner. More specifically, the scanners have dimensions of 130mm*110mm*130mm, weigh 2.3kg, and have a slice size of 26mm*76mm.

[0087] In a preferred embodiment, the image acquisition module further includes an image quality optimization module, which preprocesses the fluorescence pattern to obtain a clearer fluorescence image. The image quality optimization module can identify images with poor image quality, remove noise signals, and enhance image quality and target sharpness.

[0088] This invention also provides a fluorescence image recognition method for detecting fungi based on a deep learning algorithm, which uses the fluorescence image recognition system of this invention for analysis. In one embodiment, the recognition method is as follows: Figure 1 As shown.

[0089] According to another aspect of the present invention, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, causes the processor to implement the fluorescence image recognition method of the present invention.

[0090] The term "computer-readable medium" as used herein includes various types of computer storage media, and can be any available medium accessible to a general-purpose or special-purpose computer. For example, a computer-readable medium may include RAM, ROM, EPROM, E2PROM, registers, hard disks, removable disks, CD-ROMs or other optical disc storage devices, magnetic disk storage devices or other magnetic storage devices, or any other temporary or non-temporary medium capable of carrying or storing desired units of program code in the form of instructions or data structures and accessible by a general-purpose or special-purpose computer, or a general-purpose or special-purpose processor. As used herein, disks and discs include compact discs (CDs), laser discs, optical discs, digital multipurpose discs (DVDs), floppy disks, and Blu-ray discs, wherein disks typically magnetically copy data, while discs optically copy data using lasers. Combinations of the above should also be included within the scope of protection of computer-readable media. An exemplary storage medium is coupled to a processor so that the processor can read and write information from / to the storage medium. In an alternative, the storage medium may be integrated into the processor. The processor and storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In the alternative, the processor and storage media can reside as discrete components in the user terminal.

[0091] Example

[0092] The following will further illustrate the concept, specific structure, and technical effects of the present invention with reference to embodiments, so that those skilled in the art can fully understand the purpose, features, and effects of the present invention. Those skilled in the art will understand that the embodiments described herein are merely illustrative, and the scope of the present invention is not limited thereto.

[0093] Example 1. Establishing a deep learning algorithm model for a fluorescence image recognition system for fungal detection and evaluating model performance.

[0094] Figure 5 The flowchart illustrates the process of building a deep learning algorithm model for a fluorescence image recognition system used to detect fungi and evaluating the model's performance. The specific steps are as follows:

[0095] 1. Sampling and slide preparation: Moisten the lesion area with a sterile cotton swab soaked in 75% ethanol, and select skin debris from the edge of the active lesion. Add fluorescent staining solution to the specimen, cover with a coverslip and wait for 1 minute. Mix the staining solution and specimen thoroughly, stain for a few seconds, and remove any excess solution.

[0096] 2. Fluorescence scanning: The prepared sample is sent into a fluorescence scanner, the appropriate magnification is adjusted and the whole slide is scanned;

[0097] 3. Image annotation: APTime was used to select all fungal targets and assign category labels: spores, hyphae, or hyphal clusters. A total of 3894 spore targets and 3331 hyphae targets were annotated, totaling 1223 images used for training and validation of the YOLOX algorithm model. For hyphal clusters, a total of 659 images were annotated for training and validation of the MobileNet V2 algorithm model. The two datasets were divided into training and validation sets in a 7:3 ratio for modeling.

[0098] 4. Model Establishment

[0099] A deep learning algorithm model for a fluorescence image recognition system for detecting fungi is established using the two datasets mentioned above. In this embodiment, the YOLOX algorithm is selected for target detection, and the MobileNet V2 algorithm is selected for image classification. Before inputting the MobileNet V2 algorithm model, the image is cut into 512*512 pixel blocks using a 160-pixel overlap cutting method.

[0100] In this embodiment, the five highest confidence values ​​(i.e., n=5) of the YOLOX algorithm for spores / hyphae are selected. (To prevent false positives caused by many objects that resemble spores being predicted as having some resemblance to spores), calculate the probability that there are no spores / hyphae in the image. The MobileNet V2 algorithm assigns a confidence score to each segment, takes the maximum score among all segments as Z, and calculates the probability that there are no mycelial clusters in the image. Furthermore, the natural logarithm is used when calculating the logarithm of the probability values. Optimization using maximum likelihood estimation on the above dataset yields β0, β1, β2, and β3 of -2.80871019, -1.33605743, -1.52315955, and -0.40059451, respectively. The ROC curve plotted based on this dataset is shown below. Figure 6As shown, in this embodiment, the threshold C is set to 0.28, at which point the sensitivity is 0.95 and the specificity is 0.86.

[0101] 5. Performance Evaluation:

[0102] Performance evaluation was used to assess the consistency between the YOLOX algorithm and the MobileNet V2 algorithm in clinical decision-making, and was divided into three categories: 1) Evaluation of the consistency between the YOLOX algorithm and physician diagnosis; 2) Evaluation of the consistency between the MobileNet V2 algorithm and physician diagnosis; 3) Evaluation of the consistency between the overall judgment results and physician diagnosis. All evaluations used precision, recall, and F1 score as metrics.

[0103] The YOLOX algorithm model performance was used to identify individual spores and hyphae. This evaluation used 129 images. For each target object, YOLOX outputs a confidence score, representing the model's certainty about its detection results. In the evaluation phase, the IoU value was used to determine whether the target bounding boxes labeled by the doctor were consistent with those labeled by the AI ​​(i.e., the deep learning algorithm model of this invention). The IoU value was set at 0.3. The results showed high consistency with the doctor's results, with recall rates all above 85% and F1 scores above 80%; see Table 1 for the results.

[0104] The small-block classification model was used to identify hyphal clusters. This evaluation used 119 images, with a recall rate of 100% and an F1 score of 93%, showing high consistency with the doctor's results.

[0105] Finally, a full-film consistency assessment was conducted, and the results showed a precision of 93.5%, a recall of 98.6%, and an F1 score of 96%.

[0106] Table 1. Performance Evaluation Results of YOLOX Algorithm and Small Patch Classification Model

[0107]

[0108] The results above demonstrate that the fluorescence image recognition system of this invention combines the YOLOX algorithm with the MobileNet V2 algorithm. Through a pre-established deep learning algorithm model, it automates the analysis of fluorescence images, achieving good recognition results for fungal spores, hyphae, and hyphal clusters, with F1 scores exceeding 80%. Furthermore, the overall image judgment result integrating the outputs of both algorithms is highly consistent with the doctor's diagnosis (F1 score reaches 96%). This indicates that the fluorescence image recognition system of this invention improves detection efficiency while ensuring accuracy, possessing high clinical application reference value and contributing to the diagnosis and treatment of fungal diseases.

[0109] Example 2. Exemplary fluorescence image recognition using the deep learning algorithm model established above.

[0110] Using the deep learning algorithm model established above, Figure 7 The exemplary target fluorescence image shown is used for fluorescence image recognition. The visualization result of the target identified after analysis using the YOLOX model is as follows: Figure 8 As shown, spores are marked with yellow boxes and mycelia with purple boxes. For all identified spore and mycelial targets, the five targets with the highest confidence scores were selected: spores [0.789452850818634, 0.7716000080108643, 0.7684229016304016, 0.7675924897193909, 0.7639334201812744], and mycelia [0.779513418674469, 0.7745692729949951, 0.7687157988548279, 0.7590088248252869, 0.7552815079689026]. Meanwhile, MobileNet V2 analysis showed that the probability Z of the presence of mycelial clusters in the image was 0.09601539.

[0111] Based on the above data, the probabilities of images being free of spores, hyphae, and hyphal clusters were calculated. P 1. P 2. P 3 are 0.00061098, 0.00067797, and 0.90398461 respectively. Substituting these values ​​into the formula of the integrated data model (where the logarithm is the natural logarithm) yields... P p The value is 0.99999998, which is above the threshold C (0.28), and is therefore judged as positive.

[0112] The above results further demonstrate that the fluorescence image recognition system of the present invention can accurately identify and distinguish spores, hyphae, and hyphal clusters, and indicate the morphology of fungi in the sample, providing the quantity and visual target selection results (for spores and hyphae). This can serve as an effective basis for doctors' review and further ensure the accuracy of the detection.

[0113] The foregoing descriptions are merely exemplary embodiments or examples of the present invention and are not intended to limit the invention. Those skilled in the art will recognize that the present invention can be modified and varied in many ways. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention are included within the scope of the claims of this application.

Claims

1. A fluorescence image recognition system for detecting fungi based on a deep learning algorithm, characterized in that, include: Image acquisition module, single spore and hyphae identification module, and hyphae cluster identification module; The image acquisition module is used to obtain the target fluorescence image; The single spore and hyphae identification module and the hyphae cluster identification module are used to input the target fluorescence image into a pre-established deep learning algorithm model, wherein the pre-established deep learning algorithm model includes a deep learning-based target detection algorithm and an image classification algorithm. The single spore and hyphae identification module is used to input the target fluorescence image into a pre-trained deep learning-based target detection algorithm, detect single spores and / or hyphae in the target fluorescence image, and output the target identification result and the confidence level corresponding to the target. The mycelium cluster identification module is used to input the target fluorescence image into a pre-trained deep learning-based image classification algorithm, classify whether the target fluorescence image contains mycelium clusters, and output the classification result and the confidence level of the presence of mycelium clusters. The fluorescence image recognition system further includes integrating the output results of the single spore and hyphae recognition module with the output results of the hyphae cluster recognition module, the integration including the following steps: Based on the confidence level, determine the probability that there are no spores, hyphae, or hyphal clusters on the target fluorescence image. P 1. P 2. P 3, of which, P 1 or P 2. The confidence scores of spore or hyphae targets output by the target detection model are inverted and multiplied together for calculation. P 3. The confidence value of the image classification model outputting the mycelial clusters is inverted for calculation. The probability is calculated using a linear equation. P 1. P 2. P 3. Perform a linear combination, and the expression of the linear equation is as follows: Where β0 represents the baseline probability when there are no spores, hyphae, or hyphal clusters, and β1, β2, and β3 correspond to the weights of spores, hyphae, and hyphal clusters, respectively. The linear combination result z is transformed into a probability value through a nonlinear activation function. P p ; and the probability value P p The probability value is compared with a set threshold C. P p When the probability value is greater than or equal to a set threshold C, a fungal positive result is output. P p If the result is less than the set threshold C, output a negative result for fungi.

2. The fluorescence image recognition system according to claim 1, characterized in that, P 1 or P 2. Calculated using the following formula: ， in X i The nth spore or hyphae target with the highest confidence level identified by the target detection algorithm i The confidence level of a single spore or hyphae target, n is 3-7.

3. The fluorescence image recognition system according to claim 1, characterized in that, The nonlinear activation function is the Sigmoid function.

4. The fluorescence image recognition system according to claim 1, characterized in that, The target detection algorithm is a target detection algorithm based on convolutional neural networks, and the image classification algorithm is a lightweight image classification algorithm based on convolutional neural networks.

5. The fluorescence image recognition system according to any one of claims 1-4, characterized in that, The target detection algorithm is a YOLO series algorithm.

6. The fluorescence image recognition system according to claim 5, characterized in that, The target detection algorithm is the YOLOX algorithm.

7. The fluorescence image recognition system according to any one of claims 1-4, characterized in that, The image classification algorithm is the MobileNet algorithm.

8. The fluorescence image recognition system according to claim 7, characterized in that, The image classification algorithm is the MobileNet V2 algorithm.

9. The fluorescence image recognition system according to claim 7, characterized in that, The mycelial cluster identification module further includes segmenting the target fluorescence image into blocks before inputting it into the image classification algorithm model in the deep learning algorithm model, inputting the segmented blocks into the image classification algorithm model, and outputting a result that merges all the blocks of the target fluorescence image.

10. The fluorescence image recognition system according to claim 1, characterized in that, The training process of the pre-established deep learning algorithm model specifically includes: Obtain sample images of labeled fungal spores, hyphae, or hyphal clusters; A deep learning algorithm model is constructed based on the object detection algorithm and image classification algorithm, and the deep learning model is iteratively trained using the sample images until the deep learning algorithm model converges.

11. The fluorescence image recognition system according to claim 1, characterized in that, The image acquisition module includes a scanning module, which is used to scan and generate fluorescence images.

12. The fluorescence image recognition system according to claim 1, characterized in that, The image acquisition module further includes an image quality optimization module, which is used to preprocess the fluorescence pattern to obtain a clearer fluorescence image.

13. The fluorescence image recognition system according to claim 1, characterized in that, The fluorescent image recognition system also presents the analysis progress and / or detection results of spore, fungus and mycelium morphology in the form of visual annotations for the individual spore and hyphae recognition module and the hyphal cluster recognition module.

14. A fluorescence image recognition method for detecting fungi based on a deep learning algorithm, characterized in that, The method uses the fluorescence image recognition system according to any one of claims 1-13 for analysis.

15. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, causes the processor to implement the fluorescence image recognition method of claim 14.

Images