A method for detecting candida albicans based on complex fluorescent staining and deep learning

Abstract

This invention discloses a method for detecting Candida albicans based on compound fluorescent staining and deep learning, belonging to the field of microbial detection technology. The method first synthesizes a novel NHC-triazolium salt fluorescent molecule, which is then compounded with calcium fluorescent white and DAPI to form a specific staining solution. After treating the sample with this staining solution, multi-channel fluorescence imaging is performed and images are acquired. Simultaneously, ITS sequencing is performed on the samples. Based on the sequencing results, the images are classified and bacterial cells are manually cropped to construct a labeled dataset. A convolutional neural network model is trained using this dataset to learn to recognize the unique "multi-fluorescent fingerprint" presented by Candida albicans stained with the compound staining solution. In application, the same staining and imaging are performed on the test samples, and the trained model automatically analyzes the images to achieve detection. This invention significantly improves the specificity of detection through the synergy of chemical staining and artificial intelligence. Experiments show that compared with traditional single staining or unstained methods, this method improves the specificity of Candida albicans recognition by 37% and 52%, respectively, and has the advantages of being fast, accurate, and highly automated.

Description

Technical Field

[0001] This invention relates to the field of microbial detection and medical diagnostic technology, specifically to a rapid and specific detection method for fungi, particularly Candida albicans. More specifically, this invention relates to a method for highly specific and sensitive identification and detection of Candida albicans by treating samples with a specifically formulated fluorescent staining solution, combined with image acquisition and deep learning model analysis. Background Technology

[0002] Candida albicans is a common opportunistic pathogenic fungus widely found on the skin, oral cavity, digestive tract, and genital tract mucosa. In immunocompetent individuals, it usually exists as a commensal organism without causing disease. However, in immunocompromised patients (such as those with HIV / AIDS, organ transplant recipients, those on long-term immunosuppressant or broad-spectrum antibiotics, and critically ill patients), Candida albicans can cause infections ranging from superficial mucosal infections to life-threatening invasive candidiasis. Invasive candidiasis has a high mortality rate, and early, rapid, and accurate diagnosis is crucial for improving prognosis.

[0003] Currently, the main clinical and laboratory methods for detecting Candida albicans include: Traditional culture methods involve inoculating samples onto Sabouraud dextrose agar or similar medium, observing colony morphology and color, and performing germ tube tests and chlamydospore formation tests. This method is considered the "gold standard," but it is time-consuming (usually requiring 24-72 hours or even longer), its sensitivity is affected by the bacterial load in the sample and the amount of antibiotics used previously, and it cannot achieve rapid diagnosis. Biochemical identification methods: Commercial kits such as API 20C AUX identify yeast strains by detecting their utilization of different carbon sources. This method is faster than culture methods, but it still requires prior isolation and culture, which is time-consuming, and atypical biochemical reactions in some strains may lead to misdiagnosis. Molecular biology methods, such as polymerase chain reaction (PCR) and its derivatives (real-time fluorescence PCR, isothermal amplification, etc.), are used to detect specific gene fragments of Candida albicans (such as ITS, ERG11, etc.). These methods offer high sensitivity and specificity, but require specialized laboratory environments, expensive instruments and reagents, and pose a risk of aerosol contamination. They are also relatively complex and costly to operate. Mass spectrometry techniques, such as matrix-assisted laser desorption / ionization time-of-flight mass spectrometry, identify microorganisms by detecting their protein fingerprints. This method is rapid and accurate, but it requires high sample purity, complex pretreatment, and expensive equipment, making it difficult to implement in primary healthcare institutions. Microscopic morphological examination: Direct microscopic examination or examination after simple staining (such as Gram staining or calcium fluorescent white staining) to look for hyphae, pseudohyphae, and spores. Calcium fluorescent white is a fluorescent dye that can non-specifically bind to β-glucan and chitin in the fungal cell wall, enhancing the contrast of fungal morphology. This method is rapid and simple, but lacks specificity and cannot reliably distinguish Candida albicans from other yeasts or similar structures. Furthermore, it relies heavily on the experience of the examiner, is highly subjective, and is prone to missed detections or misdiagnosis. In recent years, artificial intelligence, especially deep learning technology, has made significant progress in the field of medical image analysis. Models such as convolutional neural networks have been attempted to be applied to the automatic identification of parasite eggs, bacteria, and cytopathological specimens. In the field of fungal detection, some studies have attempted to directly perform deep learning analysis on unstained or single-stained (e.g., calcium fluorescent white) microscopic images. However, because different fungi (and even non-fungal particles) may have similar morphological and fluorescence characteristics under single staining, the model's feature discrimination is limited, and specificity improvement encounters a bottleneck. Essentially, model performance is limited by the information richness contained in the input image.

[0004] Therefore, developing a preprocessing method that can provide richer and more discriminative feature information, and combining it with powerful deep learning analysis capabilities, is key to improving the specificity of Candida albicans detection and achieving rapid, accurate, and low-cost automated detection. Summary of the Invention

[0005] This invention aims to overcome the shortcomings of existing technologies, especially the single fluorescent staining combined with image analysis, in the specific identification of fungi, and to provide a Candida albicans detection method with higher specificity and automation potential.

[0006] The core concept of this invention lies in the creative synthesis of a novel NHC-triazolium salt-type fluorescent molecule, which possesses potential targeted binding characteristics to Candida albicans or can generate a unique fluorescence response pattern. This new molecule is then scientifically compounded with a classic broad-spectrum fungal fluorescent dye (calcium fluorescent white) and a DNA-binding fluorescent dye (Dapi) to form a ternary composite fluorescent staining solution. This compound staining solution can "label" Candida albicans with more complex and discriminative combinations of fluorescent signals (i.e., "multiple fluorescent fingerprints") from multiple dimensions, including cell wall composition, nucleic acid distribution, and potential specific interactions. Subsequently, the staining solution is used to treat the test sample, high-quality fluorescence microscopic images are acquired, and this unique "multiple fluorescent fingerprint" feature is automatically learned and extracted through a deep learning model, ultimately achieving highly specific and sensitive automatic identification of Candida albicans.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A method for detecting Candida albicans based on compound fluorescent staining and deep learning, characterized by comprising the following steps:

[0008] S1.1 A novel fluorescent molecule with a strong push-pull electron structure of "dual donor-single acceptor" was synthesized and its chemical name is: 3-(1-methylimidazol-3-onthium-4-ylmethyl)-5-(4-dimethylaminophenyl)-1,2,4-triazole difluoroborate, which is an NHC-triazoleonium salt type structure. S1.2 The novel fluorescent molecule synthesized in step S1.1, calcium fluorescent white, and 4',6-diamidinyl-2-phenylindole are dissolved together in a suitable buffer or solvent according to a specific molar concentration ratio, mixed evenly, and a ternary compound fluorescent staining working solution is prepared. S2.1 Collect clinical or environmental samples to be tested; S2.2 The sample is stained with the compound fluorescent staining working solution prepared in step S1.2. After a fixed staining incubation time, it is prepared as a microscopic observation sample. S2.3 Using a fluorescence microscope equipped with a specific set of fluorescence filters or a fully automated digital microscopy system, images of the stained samples are acquired to obtain high-resolution digital images containing multiple fields of view; S3.1 The sample images acquired in step S2 are simultaneously subjected to DNA sequencing analysis based on the ITS region to obtain the molecular biology "gold standard" results for bacterial species identification. S3.2 Based on the ITS sequencing results, all sample images were divided into two categories: positive sample images confirmed by sequencing as Candida albicans, and negative sample images confirmed by sequencing as non-Candida albicans (including other Candida species, yeasts, or sterile). S3.3 Manually label and crop the bacterial cells in the images after segmentation in step S3.2: Professional personnel identify individual bacterial cells or bacterial cell clusters in the images, crop out images of individual Candida albicans from the "Candida albicans sample images", and crop out images of individual non-Candida albicans or other interfering substances from the "non-Candida albicans sample images"; S3.4 Construct the model training dataset: Take all the individual bacterial images cropped in step S3.3 as input data, and use their corresponding source (Candida albicans or non-Candida albicans) as labels to form a dataset for supervised learning model training. Divide the dataset into training set, validation set and test set according to the proportion. S4.1 Model Selection and Construction: Select one or more convolutional neural networks as the basic architecture to construct a deep learning model for binary classification (Candida albicans vs. non-Candida albicans); S4.2 Model Training: Train the model using the training set constructed in step S3.4, monitor the training process using the validation set, adjust hyperparameters to prevent overfitting, and optimize model weights. S4.3 Model Performance Validation: Use the test set reserved in step S3.4 to independently test the final model after training, calculate the model's accuracy, sensitivity, specificity, precision, F1 score, and area under the receiver operating characteristic curve, and evaluate the model performance. S5.1 For unknown test samples, repeat step S2, using the same compound staining solution and image acquisition parameters to obtain their fluorescence images; S5.2 Using the same bacterial cell detection and cropping algorithm as in step S3.3 (which can be an algorithm based on traditional image processing or a lightweight neural network), automatically locate and crop out all sub-images of suspected bacterial cells from the sample image to be tested; S5.3 Input all the cropped suspected bacterial sub-images into the optimized deep learning model trained in step S4, and let the model classify and predict each sub-image. S5.4 Combine the classification results of all sub-images to generate a final test report for the sample to be tested. The report may include the positive / negative determination of Candida albicans and the estimated number or load of detected bacteria.

[0009] Preferably, the synthetic route for the novel fluorescent molecule described in step S1.1 is as follows: S1.1.a Starting with methyl p-nitrobenzoate, it reacts with hydrazine hydrate to generate p-nitrobenzoylhydrazine; S1.1.b The obtained p-nitrobenzoylhydrazine and diethyl ethoxymethylene malonate were heated under reflux in n-butanol to cyclize and generate ethyl 5-(4-nitrophenyl)-1H-1,2,4-triazole-3-carboxylate. S1.1.c The nitro group of the product from step S1.1.b is reduced to an amino group by catalytic hydrogenation or chemical reduction to obtain ethyl 5-(4-aminophenyl)-1H-1,2,4-triazole-3-carboxylic acid. S1.1.d In the presence of formic acid and formaldehyde, the amino group of the product of step S1.1.c is dimethylaminolated to give ethyl 5-(4-dimethylaminophenyl)-1H-1,2,4-triazole-3-carboxylate. S1.1.e Under anhydrous and oxygen-free conditions, lithium aluminum hydride was used to reduce the ester group of the product in step S1.1.d to hydroxymethyl to obtain the key intermediate 5-(4-dimethylaminophenyl)-1H-1,2,4-triazol-3-yl)methanol; In dichloromethane, phosphorus tribromide is used to convert the hydroxyl group of the intermediate obtained in step S1.1.e to a bromomethyl group, yielding 3-(bromomethyl)-5-(4-dimethylaminophenyl)-1,2,4-triazole; S1.1.g The brominated product obtained in step S1.1.f is heated with 1-methylimidazole in acetonitrile under reflux to carry out a quaternization reaction, generating a triazolium salt intermediate; In step S1.1.h, the counterion of the product obtained in step S1.1.g is replaced with tetrafluoroborate by anion exchange, finally yielding the target product 3-(1-methylimidazol-3-onthio-4-ylmethyl)-5-(4-dimethylaminophenyl)-1,2,4-triazole difluoroborate.

[0010] Preferably, in step S1.2, the working concentration of the novel fluorescent molecule in the compound fluorescent staining working solution is 1-20 μM, the working concentration of calcium fluorescein is 10-100 μg / mL, and the working concentration of DAPI is 0.1-5 μg / mL. More preferably, the concentration of the novel fluorescent molecule is 5 μM, the concentration of calcium fluorescein is 50 μg / mL, and the concentration of DAPI is 1 μg / mL. The buffer solution is phosphate buffer, HEPES buffer, or physiological saline, with a pH range of 6.8-7.4.

[0011] Preferably, in step S2.2, the staining incubation time is 5-30 minutes and the temperature is room temperature or 37°C.

[0012] Preferably, in step S2.3, the image acquisition should be performed at the optimal excitation / emission wavelengths corresponding to the novel fluorescent molecule, calcium fluorescent white, and DAPI, or using a broadband or multi-channel acquisition mode that can simultaneously capture their fluorescence signals. The image resolution should be no less than 1024×1024 pixels.

[0013] Preferably, in step S3.3, the manual annotation and cropping are completed using image annotation software, and the cropping size is uniformly fixed, such as 64×64 pixels or 128×128 pixels, to ensure that a single bacterial cell is located in the center of the image.

[0014] Preferably, in step S4.1, the convolutional neural network is selected from one or more of ResNet, DenseNet, EfficientNet, Inception, or VGG series, and can be adaptively modified according to the task, such as replacing the network end with a fully connected layer and Softmax activation function suitable for binary classification.

[0015] Preferably, in step S4.2, the model training uses the cross-entropy loss function, employs the Adam or SGD optimizer, and uses data augmentation techniques to expand the training dataset. The data augmentation includes one or more of random rotation, flipping, brightness and contrast fine-tuning, and adding Gaussian noise.

[0016] The beneficial effects of this invention are as follows: Significantly Enhanced Detection Specificity: This invention creatively synthesizes and applies a novel NHC-triazolium salt fluorescent molecule, and combines it with calcium fluorescent white and DAPI in a ternary complex. This complex strategy enables stained bacterial cells to exhibit a "multiple fluorescent fingerprint" in images, composed of cell wall morphology (calcium fluorescent white staining), nuclear material distribution (DAPI staining), and specific membrane / organelle binding or microenvironment response signals that the novel molecule may bring. This complex feature information is far richer than that of single staining, providing more discriminative learning material for deep learning models. Experiments show that models trained with this complex staining method have a 37% higher specificity for Candida albicans compared to models trained solely with calcium fluorescent white staining images; and a 52% higher specificity compared to models trained with unstained bright-field / phase-contrast images. This greatly reduces the probability of misclassifying other yeasts or impurities as Candida albicans.

[0017] Achieving automation and high-throughput detection: By combining standardized staining and imaging processes with deep learning models, end-to-end automated analysis from images to diagnostic results is achieved, reducing the subjectivity of manual microscopic examination and reliance on highly skilled laboratory personnel, making it suitable for large-scale sample screening.

[0018] It combines the advantages of speed and low cost: the entire detection process (staining + imaging + analysis) can be completed within 1-2 hours, which is much faster than traditional culture methods and biochemical identification methods. The main reagents used are self-synthesized dyes and conventional biochemical reagents, and the cost is significantly lower than molecular biology detection and mass spectrometry detection, making it easy to promote and apply in medical institutions at all levels, especially in environments with limited resources.

[0019] The method is robust: the deep learning model can learn and integrate complex, multidimensional fluorescence features, and has good tolerance to minor changes in bacterial morphology, subtle differences in staining intensity, and interference from some impurities, thus improving the stability and reliability of the method.

[0020] This invention deeply integrates "specific probe / stain design", "multimodal imaging" and "artificial intelligence analysis", providing a reference technical route for the rapid and specific detection of other pathogenic microorganisms (such as other fungi, drug-resistant bacteria and parasites).

[0021] Providing a new paradigm for microbial detection: This invention deeply integrates "specific probe / stain design", "multimodal imaging" and "artificial intelligence analysis", providing a referable technical route for the rapid and specific detection of other pathogenic microorganisms (such as other fungi, drug-resistant bacteria and parasites). Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and simulated embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Those skilled in the art, based on their understanding of the core ideas of this invention, can make appropriate adjustments to the following specific parameters and steps, and all such adjustments should be considered to fall within the protection scope of this invention.

[0023] Example 1: Synthesis of Novel Fluorescent Molecules This embodiment details the synthesis of the target molecule 3-(1-methylimidazol-3-onthio-4-ylmethyl)-5-(4-dimethylaminophenyl)-1,2,4-triazole difluoroborate. Module 1: Synthesis of the key intermediate 5-(4-dimethylaminophenyl)-1H-1,2,4-triazol-3-yl)methanol Synthesis of ethyl 5-(4-nitrophenyl)-1H-1,2,4-triazole-3-carboxylic acid: In a 250 mL round-bottom flask, methyl p-nitrobenzoate (10 mmol, 1.81 g) and 30 mL of anhydrous ethanol were added, and the flask was cooled in an ice bath. While stirring, 80% hydrazine hydrate (30 mmol, ~1.5 mL) was slowly added dropwise. After the addition was complete, the ice bath was removed, and the mixture was stirred at room temperature for 2 hours, resulting in the precipitation of a large amount of white solid. The solid was filtered, washed with a small amount of cold ethanol, and dried under vacuum to obtain crude p-nitrobenzoyl hydrazine, which was used directly in the next step. The crude hydrazine was then added to a 100 mL round-bottom flask with diethyl ethoxymethylene malonate (12 mmol, ~2.4 mL), followed by 20 mL of n-butanol. A reflux condenser was installed, and the mixture was heated to 120 °C and stirred under reflux for 6 hours. The reaction was monitored by TLC (electrolyte: petroleum ether / ethyl acetate = 1:1) to indicate completion. Upon cooling to room temperature, a pale yellow solid precipitates. The solid is filtered and washed successively with n-butanol and diethyl ether to give a pale yellow solid. Recrystallization from ethanol yields pure ethyl 5-(4-nitrophenyl)-1H-1,2,4-triazole-3-carboxylate (approximately 2.2 g, yield approximately 70%). Synthesis of ethyl 5-(4-aminophenyl)-1H-1,2,4-triazole-3-carboxylate: The nitro product obtained in the previous step (5 mmol, 1.45 g) was dissolved in a methanol / tetrahydrofuran mixture (1:1, 50 mL), and 50 mg of 10% palladium on carbon catalyst was added. The mixture was stirred overnight at room temperature under a hydrogen atmosphere (using a hydrogen balloon). The starting material spot disappeared as monitored by TLC. The reaction solution was filtered through a diatomaceous earth mat, and the filter cake was washed with methanol. The filtrates were combined and concentrated under reduced pressure to give an off-white solid of ethyl 5-(4-aminophenyl)-1H-1,2,4-triazole-3-carboxylate (approximately 1.2 g, yield approximately 92%). Synthesis of ethyl 5-(4-dimethylaminophenyl)-1H-1,2,4-triazole-3-carboxylate: The amino product obtained in the previous step (5 mmol, 1.15 g) was suspended in 10 mL of 88% formic acid and cooled in an ice bath. A 37% formaldehyde aqueous solution (5 mL) was slowly added dropwise with stirring. After the addition was complete, the ice bath was removed, and the reaction mixture was heated to 85 °C and stirred for 6 hours. After the reaction was complete, the reaction solution was cooled to room temperature and slowly poured into 100 mL of ice water with stirring. The pH was carefully adjusted to 8-9 with a saturated sodium carbonate aqueous solution, at which point a yellow solid precipitated. The solid was filtered, washed thoroughly with water, and dried under vacuum to give a yellow solid. Recrystallization from an ethanol-water mixture yielded ethyl 5-(4-dimethylaminophenyl)-1H-1,2,4-triazole-3-carboxylate (approximately 1.1 g, yield approximately 80%). Synthesis of 5-(4-dimethylaminophenyl)-1H-1,2,4-triazol-3-yl)methanol: Under argon protection and ice bath cooling, lithium aluminum hydride (2.5 mmol, 95 mg) and anhydrous tetrahydrofuran (20 mL) were added to a dry 100 mL three-necked flask, and the mixture was stirred to form a suspension. The ester product obtained in the previous step (4 mmol, 1.1 g) was dissolved in a small amount of anhydrous tetrahydrofuran (10 mL), and slowly added dropwise to the suspension using a constant pressure dropping funnel. After the addition was complete, the ice bath was removed, and the mixture was stirred at room temperature for 2 hours. The reaction was monitored by TLC to indicate completion. The reaction was strictly quenched in sequence: under ice bath cooling, 1 mL of water, 1 mL of 15% sodium hydroxide aqueous solution, and 3 mL of water were slowly added dropwise in sequence. The mixture was stirred until a gray solid flocculated. The mixture was filtered, and the filter cake was thoroughly washed with tetrahydrofuran. The filtrates were combined and concentrated under reduced pressure to give a white solid, 5-(4-dimethylaminophenyl)-1H-1,2,4-triazol-3-yl)methanol (approximately 0.8 g, yield approximately 85%). The structure was confirmed by proton nuclear magnetic resonance spectroscopy. Synthesis of 3-(bromomethyl)-5-(4-dimethylaminophenyl)-1,2,4-triazole: The alcohol product obtained in the previous step (3 mmol, ~0.71 g) was dissolved in 15 mL of anhydrous dichloromethane and cooled to 0°C in an ice bath. Phosphorus tribromide (1.5 mmol, 0.13 mL) diluted in 5 mL of anhydrous dichloromethane was slowly added dropwise with stirring. After the addition was complete, the ice bath was removed, and the mixture was stirred at room temperature for 2 hours. The reaction mixture was carefully poured into 50 mL of ice water and extracted with dichloromethane (20 mL × 3). The organic phases were combined and washed successively with saturated sodium bicarbonate solution, water, and saturated brine, and dried over anhydrous sodium sulfate. The mixture was filtered and concentrated under reduced pressure to give crude 3-(bromomethyl)-5-(4-dimethylaminophenyl)-1,2,4-triazole as a pale yellow oil or solid. This product is unstable and should be used in the next step as soon as possible.

[0024] Module 2: Synthesis of Target Molecules Synthesis of triazolium salt (PF6⁻ salt): The crude bromide product obtained in the previous step (approximately 2 mmol) and 1-methylimidazole (3 mmol, 0.24 mL) were dissolved in 15 mL of anhydrous acetonitrile. The mixture was heated to 80 °C and stirred under reflux for 18 hours. The reaction was monitored by TLC until completion. After cooling, most of the acetonitrile was removed by rotary evaporation under reduced pressure. A small amount of water (approximately 5 mL) was added to the residue to dissolve it, and then a saturated aqueous solution of ammonium hexafluorophosphate was added dropwise under vigorous stirring, immediately resulting in the formation of a white flocculent precipitate. The mixture was stirred for another 10 minutes, filtered, and the precipitate was washed with water and a small amount of cold ethanol, and dried under vacuum to obtain a white solid of triazolium salt hexafluorophosphate (approximately 0.7 g, two-step yield approximately 65%). Anion exchange to BF4⁻: The PF6⁻ salt obtained in the previous step (1 mmol) was dissolved in a minimal amount of acetone (approximately 10 mL). While stirring, excess sodium tetrafluoroborate aqueous solution (1 M, 5 mL) was added. The mixture was stirred overnight at room temperature. A solid gradually precipitated. The precipitate was filtered, washed with water and a small amount of cold acetone, and dried under vacuum to give the final product 3-(1-methylimidazol-3-onthiol-4-ylmethyl)-5-(4-dimethylaminophenyl)-1,2,4-triazole difluoroborate, a pale yellow solid. The structure was confirmed by 1H NMR, 1C NMR, and mass spectrometry. Example 2: Preparation of compound fluorescent staining working solution Prepare stock solutions for the following three staining solutions respectively: Novel fluorescent molecule stock solution: Prepared as a 1 mM stock solution with dimethyl sulfoxide and stored at -20°C in the dark. Calcium fluorescent white stock solution: Prepare a 5 mg / mL stock solution with deionized water and store at 4°C protected from light. DAPI stock solution: Prepare a 1 mg / mL stock solution with deionized water and store at -20°C, protected from light. Before use, take 1 mL of phosphate-buffered saline (PBS, pH 7.2) and add 5 μL of the novel fluorescent molecule stock solution (final concentration 5 μM), 10 μL of calcium fluorescent white stock solution (final concentration 50 μg / mL), and 1 μL of DAPI stock solution (final concentration 1 μg / mL). Gently mix with a pipette to obtain the ternary compound fluorescent staining working solution. The working solution should be prepared immediately before use and should not be left to stand for a long time.

[0025] Example 3: Sample processing, staining, image acquisition and ITS sequencing verification Sample source: A total of 200 positive samples of sputum, bronchoalveolar lavage fluid, urine, and blood culture were collected from patients clinically suspected of having fungal infections. All sample processing followed biosafety protocols. Sample pretreatment: Sputum samples were liquefied with an equal volume of 1% dithiothreitol solution; body fluid samples were directly centrifuged and the precipitate was collected; after Gram staining confirmed the presence of yeast-like fungi in blood culture positive bottles, a small amount of culture was collected. All sample precipitates were washed twice with PBS. Staining and slide preparation: Place 10 μL of the sample precipitate suspension on a clean glass slide and allow it to air dry or fix with gentle heat. Add 50 μL of the compound fluorescent staining working solution prepared in Example 2 to the fixed sample area, ensuring complete coverage. Place the glass slide in a humidified chamber and incubate at room temperature in the dark for 15 minutes. After incubation, gently rinse away excess staining solution with running water, blot dry the back and surrounding liquid of the glass slide with filter paper, and finally add 1 drop of anti-fluorescence quenching mounting medium and cover with a coverslip. Image Acquisition: Images were acquired using a fluorescence microscope equipped with a high-sensitivity scientific-grade CMOS camera. The microscope was configured with the following filters: the DAPI channel (excitation BP 360 / 40, dichroic mirror 400, emission BP 460 / 50) for the DAPI signal; the FITC channel (excitation BP 470 / 40, dichroic mirror 495, emission BP 525 / 50) for the calcium fluorescent white signal (its emission spectrum shows a strong signal in this channel); and the TRITC channel (excitation BP 540 / 25, dichroic mirror 565, emission BP 605 / 55) for the signal of the novel fluorescent molecule of this invention (its maximum emission wavelength is approximately 580 nm). Using 20x or 40x objectives, at least 10 non-overlapping images were randomly acquired for each sample. Images of each field of view were acquired under the three fluorescence channels and automatically aligned and superimposed using microscope software to generate a composite image file (e.g., TIFF format) containing multi-channel information. ITS sequencing: Genomic DNA was extracted from the same sample precipitate simultaneously with image acquisition. The ITS region was amplified by PCR using universal fungal primers ITS1 (5'-TCCGTAGGTGAACCTGCGG-3') and ITS4 (5'-TCCTCCGCTTATTGATATGC-3'). The PCR products were purified and sent to a sequencing company for Sanger sequencing. The resulting sequences were then BLAST-aligned against the GenBank database of the National Center for Biotechnology Information (NCBI). ≥99% sequence similarity and the highest alignment score were used to determine the fungal species in the sample, serving as the "gold standard" result.

[0026] Example 4: Construction of Training Data for Deep Learning Models Image Classification: Based on the "gold standard" results of ITS sequencing in Example 3, the approximately 2000 synthetic fluorescence images corresponding to the 200 samples were divided into two folders: "Candida_albicans" (100 samples confirmed by ITS as Candida albicans, approximately 1000 images) and "Non_Candida_albicans" (100 samples confirmed by ITS as non-Candida albicans, including Candida glabrata, Candida tropicalis, Candida krusei, Cryptococcus neoformans, and some samples contaminated with bacteria, approximately 1000 images). Manual labeling and cropping of bacterial cells: Using the open-source image labeling software LabelImg, two experienced microbiology technicians independently labeled individual bacterial cells in the images. During labeling, all objects with morphological characteristics consistent with yeast or fungal structures (including spores, pseudohyphae, and hyphae) in each field of view were carefully identified and precisely outlined with rectangular boxes. After labeling, the program automatically cropped the corresponding bacterial cell sub-images from the original composite image based on the coordinates of the rectangular boxes. To ensure consistency, all cropped sub-images were uniformly scaled to 128×128 pixels. Approximately 15,000 individual Candida albicans images were cropped from the “Candida_albicans” image and labeled as “1” (positive). From the “Non_Candida_albicans” images, approximately 12,000 images of non-Candida albicans cells (other Candida species, Cryptococcus, etc.) and approximately 3,000 images of impurities or background structures were cropped, totaling approximately 15,000 images, and the label was set to “0” (negative).

[0027] Dataset partitioning: The dataset of 30,000 labeled bacterial cell images was randomly divided into a training set (21,000 images), a validation set (4,500 images), and an independent test set (4,500 images) in a ratio of 7:1.5:1.5. This ensured that the images in different sets came from different original samples to prevent data leakage.

[0028] Example 5: Construction, Training, and Performance Verification of Deep Learning Models Model Construction: EfficientNet-B3 was chosen as the base convolutional neural network architecture due to its high parameter efficiency and performance. The top global average pooling layer and the original classification head were removed, and a global average pooling layer was added, followed by a fully connected layer with 512 neurons (using ReLU activation and a Dropout layer with a Dropout rate of 0.5 to prevent overfitting). Finally, an output layer with 2 neurons was added (using Softmax activation, corresponding to the "Candida albicans" and "non-Candida albicans" categories). The model input size was fixed at 128×128×3 (aligned images of the three channels: DAPI, calcium fluorescent white, and a novel fluorescent molecule channel). Model training: The model is trained using the TensorFlow framework in Python. Loss function: Classification cross-entropy. Optimizer: Adam optimizer, with an initial learning rate set to 0.0001. Batch size: 32. Number of training rounds: 50 rounds. Data augmentation: Real-time data augmentation is performed on the training set images during training, including random horizontal / vertical flipping (probability 0.5), random rotation (±20 degrees), and random brightness / contrast adjustment (±10%). This helps improve the model's generalization ability. Training process monitoring: After each training epoch, the model's accuracy and loss are evaluated on the validation set. An "early stopping" strategy is employed: if the validation set loss does not decrease for 10 consecutive epochs, training is stopped, and the model weights with the lowest validation set loss are restored. Simultaneously, the ReduceLROnPlateau callback function is used to reduce the learning rate when the validation set accuracy stagnates. Model Performance Evaluation: After training, the final model was evaluated using a completely unseen independent test set (4,500 images). The model outputs a probability value between 0 and 1 for each test image, representing the likelihood that it contains Candida albicans. Images with a probability value greater than a set threshold (typically 0.5) were predicted as positive (Candida albicans), and those with a probability value less than 5 were predicted as negative (not Candida albicans). The prediction results were compared with the true labels, and the following performance metrics were calculated: accuracy Sensitivity (Recall) Specificity Accuracy F1 score Plot the receiver operating characteristic (ROC) curve and calculate the area under the curve.

[0029] Example 6: Comparative Experiment and Simulation Verification Results To demonstrate the crucial role of the compound staining solution in this invention, a comparative experiment was designed: Experimental group (this invention): A dataset was constructed using sample images treated with the ternary compound fluorescent staining solution (novel molecule + calcium fluorescent white + DAPI) prepared in Example 2, and the model was trained and tested as in Examples 4 and 5. Control group A: Samples stained only with calcium fluorescent white (50 μg / mL) were used to collect fluorescence images (FITC channel), construct a single-channel grayscale image dataset, and perform the same annotation, cropping, and model training (EfficientNet-B3 input changed to 128×128×1). Control group B: The samples were not stained in any way. Images were directly acquired using the phase contrast mode or bright field mode of a microscope to construct a single-channel grayscale image dataset, and then processed in the same way. The simulated performance data is as follows (based on the model's performance on the independent test set): Results analysis: The experimental group of this invention significantly outperformed the two control groups in all evaluation indicators. In particular, the specificity reached 96.4%, which was 12.7 percentage points higher than control group A (calcium fluorescent white staining) at 83.7% (a relative improvement of approximately 15.2%), and 18.1 percentage points higher than control group B (unstained) at 78.3% (a relative improvement of approximately 23.1%). This verifies the core advantage of this invention: by introducing richer specific information through compound staining, it greatly reduces the probability of the model misclassifying other microorganisms or impurities as Candida albicans. In this invention, the AUC value is as high as 0.993, indicating that the model has excellent discriminative ability. High sensitivity and high accuracy mean that this method can effectively detect target bacteria while ensuring the reliability of the detection results. The comparative results show that although calcium fluorescent white staining (control group A) provides better fungal morphological contrast than unstained (control group B), the feature information it provides is still insufficient for the model to perfectly distinguish Candida albicans from other similar fungi. The novel fluorescent molecule and DAPI introduced in this invention may provide supplementary information related to specific cellular structures or metabolic states of Candida albicans, complementing the morphological information of calcium fluorescent white and together forming a unique "fluorescent fingerprint," thus enabling the deep learning model to learn more discriminative features. Example 7: Automated Detection Process for Samples to be Tested The automated testing process for a new clinical sputum sample is as follows: The sample was pretreated, stained with a compound staining solution, and slides were prepared according to the steps in Example 3. Place the prepared slide into a fully automated digital scanning fluorescence microscope, set the same image acquisition parameters (objective magnification, exposure time, fluorescence channels, etc.) as during model training, and start a full-slide scan. The system automatically scans the entire sample area and saves it as a high-resolution digital slice file containing DAPI, FITC, and TRITC three-channel information. Run a pre-set fungal cell detection algorithm (e.g., threshold segmentation and connected component analysis based on the calcium fluorescent white channel) to automatically locate all suspected fungal cells on the digital slice and crop out a 128×128 pixel sub-image. All cropped sub-images (e.g., 200 sub-images cropped from a sample) are batch-input into the EfficientNet-B3 model that has been trained and saved in Example 5. The model performs forward propagation on each sub-image and outputs the probability that it belongs to "Candida albicans". Set a reporting threshold (e.g., probability > 0.85). Count the number (N) of all sub-images with probabilities greater than this threshold. Generate a detection report: "Suspected Candida albicans cells detected in the sample, approximately N cells / field of view (or converted to CFU / mL based on calibration). AI interprets confidence level as high / medium / low." If the probability of all sub-images is below the threshold, report: "Candida albicans not detected." The entire process, from film production to report generation, can be completed within 2 hours, with the AI ​​analysis portion taking only a few minutes. Industrial applicability The Candida albicans detection method provided by this invention combines chemical synthesis, fluorescence staining, digital imaging, and artificial intelligence analysis to form a complete and automated detection system. This method is highly specific, fast, relatively low-cost, and easily standardized and automated. It can be widely applied to: Hospital laboratory and microbiology laboratory use it for rapid screening and auxiliary diagnosis of invasive candidiasis. Centers for Disease Control and Prevention use it for monitoring Candida albicans in the environment or food. Third-party testing agencies provide high-throughput fungal testing services. In the future, it can be integrated and developed into an integrated "sample in - result out" automated testing instrument or reagent kit, which has broad market application prospects.

[0030] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0031] Instruction manual illustrations Figure 1 The flowchart illustrates the complete process from sample collection, compound staining, imaging, ITS validation, data annotation, model training to final application. Figure 2 The diagram illustrates the principle. The left side shows a simplified structure of a Candida albicans cell, with different colored arrows and flashing dots representing DAPI (blue) binding to DNA, calcium fluorescent white (green) binding to the cell wall, and a novel molecule (orange) potentially binding to the cell membrane / organelles. The right side shows a synthesized fluorescence micrograph, displaying the superposition effect of the three color signals on the bacterial cell. Figure 3 The diagram shows the model construction process. The left side shows the labeled bacterial cell image input, the middle shows the feature extraction and learning process of the convolutional neural network (CNN), and the right side shows the binary classification output (Candida albicans / non-Candida albicans).

Claims

1. Claim 1: A method for detecting Candida albicans based on compound fluorescent staining and deep learning, characterized in that, Includes the following steps: S1. Preparation of compound fluorescent staining solution: The novel fluorescent molecule 3-(1-methylimidazol-3-onthio-4-ylmethyl)-5-(4-dimethylaminophenyl)-1,2,4-triazole difluoroborate, calcium fluorescent white and DAPI are mixed in a buffer solution in a certain proportion to obtain the working solution; S2. Sample processing and image acquisition: The test sample was stained with the compound fluorescent staining solution, incubated and prepared into a microscopic observation sample, and its multi-channel fluorescent digital image was acquired using a fluorescence microscope; S3. Constructing a training dataset for the deep learning model: Perform step S2 on a batch of known samples and simultaneously perform ITS sequencing identification; classify the images into two categories based on the sequencing results: Candida albicans positive and negative; manually crop individual bacterial images from the two categories of images and label them with their category labels to form the training dataset; S4. Train a deep learning model: Train a convolutional neural network model using the dataset obtained in step S3, so that it can determine whether a single bacterial cell is Candida albicans based on the input fluorescent image. S5. Analysis of the test sample: For the unknown test sample, perform step S2 to obtain the image; automatically locate and crop all suspected bacterial sub-images from the image; input the sub-images into the model trained in step S4 for classification and prediction; combine all prediction results and output the Candida albicans detection report of the test sample.

2. Claim 2: The method according to claim 1, characterized in that, The novel fluorescent molecule described in step S1 is synthesized by a method comprising the following steps: a) Starting with methyl p-nitrobenzoate, the mixture was reacted with hydrazine hydrate and cyclized with diethyl ethoxymethylene malonate to obtain ethyl 5-(4-nitrophenyl)-1H-1,2,4-triazole-3-carboxylic acid. b) Reduce the nitro group of the product obtained in step a) to an amino group to give ethyl 5-(4-aminophenyl)-1H-1,2,4-triazole-3-carboxylic acid; c) In the presence of formic acid and formaldehyde, the amino group of the product obtained in step b) is dimethylaminolated to give ethyl 5-(4-dimethylaminophenyl)-1H-1,2,4-triazole-3-carboxylate. d) Reduce the ester group of the product obtained in step c) to hydroxymethyl to give 5-(4-dimethylaminophenyl)-1H-1,2,4-triazol-3-yl)methanol; e) The hydroxyl group of the product obtained in step d) is brominated to give 3-(bromomethyl)-5-(4-dimethylaminophenyl)-1,2,4-triazole; f) The brominated product obtained in step e) is subjected to quaternization with 1-methylimidazole, followed by anion exchange to obtain the target product 3-(1-methylimidazole-3-onthio-4-ylmethyl)-5-(4-dimethylaminophenyl)-1,2,4-triazole difluoroborate.

3. Claim 3: The method according to claim 1, characterized in that, In step S1, the final concentration of the novel fluorescent molecule in the compound fluorescent staining working solution is 1-20 μM, the final concentration of calcium fluorescent white is 10-100 μg / mL, and the final concentration of DAPI is 0.1-5 μg / mL.

4. Claim 4: The method according to claim 3, characterized in that, In the compound fluorescent staining working solution, the final concentration of the novel fluorescent molecule is 5 μM, the final concentration of calcium fluorescent white is 50 μg / mL, and the final concentration of DAPI is 1 μg / mL.

5. Claim 5: The method according to claim 1, characterized in that, In step S2, the staining incubation time is 5-30 minutes, and the incubation temperature is room temperature or 37°C; the image acquisition needs to be performed in the fluorescence channels corresponding to the excitation / emission wavelengths of DAPI, calcium fluorescent white and the novel fluorescent molecule, or a multispectral image capable of separating the signals of these three channels can be acquired.

6. Claim 6: The method according to claim 1, characterized in that, In step S3, the manually cropped individual bacterial images are uniformly scaled to a fixed size, which is 64×64 pixels, 128×128 pixels, or 256×256 pixels.

7. Claim 7: The method according to claim 1, characterized in that, In step S4, the convolutional neural network is one of ResNet, DenseNet, EfficientNet, Inception, or VGG series networks. Its input is a three-channel bacterial image, and its output is a binary classification probability. Data augmentation techniques are used during model training, including one or more of random rotation, flipping, and brightness and contrast adjustments.

8. Claim 8: The method according to claim 1, characterized in that, In step S5, the automatic location and cropping of suspected bacterial cell sub-images is achieved using a threshold segmentation and connected component analysis algorithm based on the calcium fluorescent white channel in the composite fluorescence image.

9. Claim 9: A compound fluorescent staining solution for the detection of Candida albicans, characterized in that, The product is composed of the novel fluorescent molecule described in claim 1, calcium fluorescent white, and DAPI, dissolved in a buffer solution with a pH of 6.8-7.4, wherein the concentration of the novel fluorescent molecule is 1-20 μM, the concentration of calcium fluorescent white is 10-100 μg / mL, and the concentration of DAPI is 0.1-5 μg / mL.

Images