High-precision identification method of culture medium drug sensitive paper based on mixed model

Abstract

The invention discloses a high-precision identification method for a culture medium drug sensitive paper sheet based on a mixed model, and the method comprises the steps: S1, image collection and preprocessing: carrying out the standardization processing of a collected culture medium image, including noise removal, illumination correction and contrast enhancement; s2, semantic segmentation model processing: carrying out pixel-level classification on the preprocessed image by adopting a deep learning network architecture, accurately positioning the positions and boundaries of all drug sensitive paper sheets in a culture medium, and generating a segmentation mask; and S3, classification model processing: extracting a paper sheet area based on a result of the step S2, and identifying a drug code and concentration on each paper sheet by using a CNN classification model. And S4, hybrid model fusion: intelligently fusing the data in the steps S2 and S3, and outputting a final drug sensitive paper identification result through confidence weighting and result verification. According to the method, semantic segmentation and classification models are fused, the difficulties of low recognition precision, poor environmental adaptability and the like of a traditional drug sensitive test are overcome, and high-precision automatic recognition of the drug sensitive paper sheets is realized.

Description

Technical Field

[0001] The present invention relates to the field of medical detection technology, and in particular to a method for high-precision identification of culture medium drug-sensitive paper sheets using a hybrid model combining a semantic segmentation model and a classification model. Background Art

[0002] Antimicrobial susceptibility testing is a core technique in clinical microbiology. It determines bacterial susceptibility to antimicrobial drugs to guide rational clinical medication use. In the traditional Kirby-Bauer disc diffusion method, susceptibility strips containing different antimicrobial drugs are placed on the surface of a culture medium inoculated with the test bacteria. After incubation, inhibition zones are formed, and the diameter of these zones is measured to determine bacterial susceptibility to the drug.

[0003] The core workflow of the automated zone of inhibition measurement system includes image acquisition, identification and positioning of drug-sensitive paper discs, detection of the zone of inhibition boundary, diameter measurement and calculation, and result interpretation. Accurate identification and positioning of the drug-sensitive paper discs are crucial prerequisites for the entire measurement process, directly impacting the accuracy of subsequent zone of inhibition boundary detection and the reliability of measurement results.

[0004] The following analyzes the problems and shortcomings of the existing technology: 1. Limitations of traditional manual measurement methods: Highly subjective: Manual visual measurement of the inhibition zone diameter has obvious subjective judgment differences, and the measurement results of different inspectors may differ by 2-4mm; Inefficiency: A single culture dish usually contains 6-8 antibiotic susceptibility paper discs, and manual measurement takes 5-10 minutes; Poor reproducibility: The results of the same sample measured by the same inspector at different times will also vary; High labor intensity: Large laboratories need to process hundreds of culture dishes every day, and the workload of manual measurement is huge.

[0005] 2. Deficiencies of existing automated recognition technology: Single model accuracy limitations: Existing methods often use a single image processing algorithm or machine learning model, with recognition accuracy rates of only around 85% in complex backgrounds. Poor environmental adaptability: Sensitive to interference factors such as contamination, scratches, bubbles on the surface of the culture medium, and the recognition accuracy is significantly reduced; Lighting condition dependence: performance is unstable under different lighting conditions, and the shooting environment needs to be strictly controlled; Difficulty in distinguishing drug types: It is impossible to accurately distinguish the types of paper strips of different antimicrobial drugs, which affects the accuracy of automated interpretation.

[0006] 3. Special challenges of drug susceptibility strip identification: Morphological similarity: The paper discs of different drugs are highly similar in shape and size, and the main difference lies in the drug code printed on the surface; Background complexity: The surface of the culture medium may contain complex background interference such as bacterial growth, culture medium texture, and surface reflection; Small size: The diameter of drug-sensitive paper is usually only 6mm, and the printed characters are even smaller, which requires extremely high recognition accuracy; Batch variability: There are slight differences in color, font, etc. between different batches and manufacturers of drug-susceptibility papers. Summary of the Invention

[0007] In response to the above technical problems, the present invention provides a hybrid model drug-susceptibility paper recognition method based on the combination of a semantic segmentation model and a classification model, aiming to solve key problems in the existing technology such as low recognition accuracy, poor environmental adaptability, and difficulty in distinguishing drug types, and to achieve high-precision automatic recognition of drug-susceptibility paper strips in culture media.

[0008] The technical solution adopted by the present invention to solve its technical problem is: A high-precision identification method for culture medium drug-sensitive paper strips based on a hybrid model comprises the following steps: Step S1. Image acquisition and preprocessing: The collected culture medium images are standardized, including noise removal, illumination correction, and contrast enhancement operations, to provide high-quality input images for subsequent model processing and eliminate the impact of environmental factors on recognition accuracy.

[0009] Furthermore, the illumination correction is to unify the image brightness through histogram equalization and gamma correction technology to eliminate the influence of uneven illumination in different shooting environments.

[0010] Furthermore, the contrast enhancement is to improve the visibility of image details by using adaptive histogram equalization.

[0011] Step S2. Semantic segmentation model processing: Use a deep learning network architecture to perform pixel-level classification on the image preprocessed in step S1, accurately locate the position and boundaries of all drug-sensitive paper strips in the culture medium, generate a segmentation mask, and achieve paper strip positioning in a complex background.

[0012] Furthermore, the network architecture adopts DeepLabV3+ semantic segmentation network as the core architecture, ResNet-50 as the feature extraction backbone network, integrates void convolution to expand the receptive field, integrates void spatial pyramid pooling module to realize multi-scale feature fusion, and introduces decoder module for feature refinement to build an end-to-end drug-susceptibility paper accurate segmentation system.

[0013] Furthermore, the pixel-level classification adopts the DeepLabV3+ semantic segmentation network architecture, with ResNet-50 as the feature extraction backbone network, and realizes the end-to-end pixel-level binary classification task through the encoder-decoder structure, accurately marking each pixel in the input image as "background" or "drug-sensitive paper" category, and outputting a dense prediction mask with the same resolution as the original image, thereby realizing pixel-level precise positioning and boundary delineation of the drug-sensitive paper.

[0014] Step S3. Classification model processing: Based on the paper area extracted by the semantic segmentation result of step S2, the CNN classification model is used to identify the drug code and concentration on each paper, so as to achieve accurate classification and identification of different antimicrobial drug paper slips.

[0015] Furthermore, the drug code refers to the standardized antimicrobial drug abbreviation code printed on the surface of the drug-susceptibility paper, which uses a combination of 2-4 uppercase English letters and follows the unified naming conventions of the international antimicrobial drug sensitivity test standards as a unique identifier for drug identification.

[0016] Furthermore, the CNN classification model is an end-to-end classifier designed specifically for the task of drug code recognition on drug-susceptibility paper strips using a deep convolutional neural network architecture. It uses ResNet-50 as the backbone network and integrates multi-scale feature extraction, attention mechanism, and transfer learning strategy to achieve high-precision automatic recognition and classification of drug codes in pre-segmented paper strip areas, and outputs drug category probability distribution and confidence assessment.

[0017] Step S4. Hybrid model fusion: Intelligently fuse the location information of the semantic segmentation model in step S2 with the drug type information of the classification model in step S3. Through the strategy of confidence weighting and result verification, the final drug sensitivity paper recognition result is output to ensure the accuracy and reliability of the recognition.

[0018] Furthermore, the confidence weighting adopts a dual-reset confidence evaluation mechanism of semantic segmentation and classification models, calculates the segmentation confidence and classification confidence respectively through the softmax probability value and probability ratio, obtains the final confidence based on the adaptive weight fusion strategy, and sets multi-level thresholds to achieve differentiated post-processing. When the confidence is lower than 0.8, manual review is automatically triggered to ensure the high reliability of the recognition results.

[0019] Furthermore, the result verification is to establish a multi-dimensional cross-validation mechanism, ensure the rationality of paper distribution through spatial consistency verification, verify the geometric parameters of paper through morphological feature verification, and identify abnormal results through historical data comparison verification, and build a full range of automatic error correction system from spatial layout, physical characteristics to statistical laws to ensure the accuracy and consistency of recognition results.

[0020] Beneficial effects of the present invention The present invention provides an innovative fusion semantic segmentation and classification model to overcome the defects of traditional drug sensitivity tests, such as low recognition accuracy, poor environmental adaptability, and difficulty in distinguishing drug types. The present invention uses the DeepLabV3+ network to achieve pixel-level precision positioning, combined with the CNN classification model to accurately identify drug codes, and through confidence weighting and multi-dimensional verification mechanisms, ensures that the recognition accuracy rate reaches 100%, and the average confidence exceeds 0.95, far exceeding the 85% accuracy of the existing technology. The present invention shows excellent adaptability to complex backgrounds, lighting changes, and tiny characters, covers a variety of antimicrobial drug types, and has stable and consistent repeatability verification, greatly improving the efficiency and reliability of automated drug sensitivity tests, and providing advanced technical support for clinical microbiology testing. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a flow chart of the overall technical implementation of the present invention.

[0022] Figure 2 It is a schematic diagram of the structural composition of the culture medium image acquisition device.

[0023] Figure 3 It is the structure diagram of the semantic segmentation network.

[0024] Figure 4 It is a CNN classification network structure diagram.

[0025] Figure 5 This is a comparison chart of the segmentation effect of drug-susceptibility paper strips.

[0026] Figure 6 This is the result diagram of identification of different drug types. DETAILED DESCRIPTION

[0027] The following is combined with Figure 1-6 And specific implementation method further details: like Figure 1 As shown in the flowchart, the complete process of high-precision identification of drug-susceptibility paper strips based on mixed model culture medium is described, which includes four main steps (S1-S4): S1 Image Acquisition and Preprocessing Objective: To standardize the collected culture medium images to eliminate environmental interference such as noise, uneven lighting, and insufficient contrast, providing high-quality input for subsequent models. Specifically, this includes: S1.1 Noise removal: A Gaussian blur filter (kernel size 3x3, standard deviation 1.0) was used to initially smooth the image and remove high-frequency noise; a median filter (window 5x5) was then used to remove noise.

[0028] S1.2 Lighting Correction: Histogram equalization is applied to enhance the overall brightness distribution. Gamma correction (γ = 1.2) is combined to adjust the nonlinear lighting gradient and unify the brightness under different shooting conditions.

[0029] S1.3 Contrast enhancement: Adaptive histogram equalization (CLAHE, cliplimit = 2.0, tilegridsize = 8x8) was used to enhance the visibility of image details, especially the texture of the culture medium and the edges of the drug-susceptibility paper.

[0030] S2 semantic segmentation model processing Objective: Use a deep learning network to perform pixel-level classification on preprocessed images and accurately locate the boundaries of all drug-susceptibility paper discs in culture medium.

[0031] S2.1 Network architecture initialization: DeepLabV3+ semantic segmentation network is used, and the pre-trained ResNet-50 is used as the feature extraction backbone network; the input image resolution is adjusted to 512x512 to ensure a balance between computational efficiency and accuracy.

[0032] S2.2 Feature extraction: Use atrous convolution (with dilation rates of 6, 12, and 18, respectively) to expand the receptive field; integrate the atrous spatial pyramid pooling (ASPP) module to fuse the multi-scale features of 1x1, 3x3, and 6x6 convolution kernels.

[0033] S2.3 Segmentation mask generation: The low-resolution feature map is restored to the original image resolution through the decoder module (upsampling factor 4x); the softmax activation function is used for pixel-level binary classification ("background" or "drug-sensitive paper"), and a dense prediction mask with the same resolution as the input image is output.

[0034] S3 classification model processing Objective: Based on the paper area extracted by segmentation results, use CNN to identify the drug code and concentration on each paper.

[0035] S3.1 Region extraction: Extract connected regions from the segmentation mask of S2 (using OpenCV's findContours), crop each drug-susceptibility paper region of interest (ROI), and resize it to 64x64 pixels.

[0036] S3.2 Feature analysis: Resnet50 is used as the backbone network and the ROI image is input; multi-scale convolutional layers (3x3, 5x5) are used to extract drug code features, and the SE (Squeeze-and-Excitation) attention mechanism is integrated to enhance the focus on key areas.

[0037] S3.3 Code Recognition Input the features into the fully connected layer and output the probability distribution of 50 common drug codes (such as AMP, GEN, CIP); use the 2-4 digit uppercase letter code defined by the CLSI / EUCAST standard. Add Dropout (p=0.5) to prevent overfitting and output the highest probability category. S3.4 Confidence Assessment: Calculate the softmax output probability as the classification confidence C_cls, with the threshold set to 0.9; when it is lower than 0.7, it is marked as low confidence, triggering subsequent manual review.

[0038] S4 Hybrid Model Fusion Goal: Intelligently fuse semantic segmentation and classification results to output the final recognition result and ensure reliability.

[0039] S4.1 Confidence Weighting: Calculate the segmentation confidence C_seg (based on the Dice coefficient, target > 0.95) and classification confidence C_cls; use the adaptive weight fusion formula C_final=w1 C_seg+w2 C_cls (w1+w2=1, w1 initialized to 0.6), weights are optimized by gradient descent. Multi-level thresholds are set (>0.9 high confidence, 0.7-0.9 medium confidence, <0.7 low confidence), and a value below 0.8 triggers manual review.

[0040]

[0041] S4.2 Result Verification: Spatial consistency check: Verify that the paper disc distribution conforms to 6-8 pieces per culture dish, and the center spacing is >10mm.

[0042] Drug type rationality verification: exclude drug combinations that do not meet CLSI / EUCAST standards (such as the simultaneous presence of conflicting antibiotics).

[0043] Morphological feature verification: the paper disc diameter (6 ± 0.5 mm) and circularity (> 0.85) were measured, and an ellipse fitting algorithm was used.

[0044] like Figure 2 As shown in the figure, this is a schematic diagram of the structure of the culture medium image acquisition device, and the specific structure includes the following four levels: 2.1 Data Collection Layer ( Figure 2 (Structures 2 and 3 shown in the figure) Industrial camera (2): located above the system, responsible for high-precision image acquisition, capturing detailed image information of the drug-sensitive paper in the culture dish; Shadowless Light Source (3): Provides uniform and stable lighting conditions, eliminates shadow interference, and ensures consistent image quality.

[0045] 2.2 Mechanical control layer ( Figure 2 (Shown in Structure 1) Mechanical motion controller (1): controls the automated operation of the entire system, including the positioning of the culture dish, movement of the camera, and other mechanical actions; Industrial camera bracket system: provides precise positioning and stable support for cameras.

[0046] 2.3 Sample processing layer ( Figure 2 (Shown in Structure 5) Culture dish carrying platform (5): Place the culture dish to be tested. The circular culture dish and the distribution of drug-sensitive paper sheets inside it can be seen in the figure.

[0047] 2.4 Data Processing Layer ( Figure 2 (Shown in structures 4 and 6) Industrial host (4): Integrates a hybrid model algorithm, including a semantic segmentation model and a classification model, responsible for core processing tasks such as image preprocessing, drug-susceptibility paper identification and positioning, and drug type classification; Display terminal (6): provides a human-computer interaction interface and displays recognition results and system status.

[0048] like Figure 3 As shown in the figure, based on the DeepLabV3+ semantic segmentation network architecture diagram, the following describes the network workflow step by step: 3.1 Input Layer Input image: An image of a petri dish containing drug-sensitive paper strips. Multiple white drug-sensitive paper strips can be seen distributed within the circular petri dish. 3.2 Encoder DCNN backbone network: uses deep convolutional neural network as feature extractor; AtrousConv (Atrous Convolution): Expand the receptive field through atrous convolution with different expansion rates and maintain the feature map resolution.

[0049] Multi-scale feature extraction: 1×1Conv: performs dimension transformation; 3×3Convrate6: dilated convolution with a dilation rate of 6; 3×3Convrate12: dilated convolution with a dilation rate of 12; 3×3Convrate18: dilated convolution with a dilation rate of 18; ImagePooling: Global average pooling obtains global context information.

[0050] 3.3 Feature Fusion ASPP module: Concatenates features of different scales; 1×1Conv: Perform dimensionality reduction on the fused features.

[0051] 3.4 Decoder Low-LevelFeatures: Extract low-level features from the encoder; Feature processing: perform 1×1 convolution on low-level features; Upsampleby4: Upsample high-level features by a factor of 4; Concat: Connect the upsampled high-level features with the processed low-level features; 3×3Conv: Convolution processing is performed on the connected features; Upsampleby4: Upsamples the image by a factor of 4 again to restore it to the original size.

[0052] 3.5 Output Layer Segmented image: Outputs pixel-level segmentation results. The red area in the image represents the location of the identified drug-sensitive paper, achieving accurate positioning and segmentation of the drug-sensitive paper in the culture dish. This network architecture achieves high-precision semantic segmentation of drug-susceptibility paper strips through an encoder-decoder structure combined with dilated convolution and multi-scale feature fusion.

[0053] like Figure 4 As shown, based on this CNN classification model architecture diagram, the technical implementation process of the network is described step by step using patent terms: 4.1 Input Data Preprocessing Input image: Receive the drug-susceptibility paper area image extracted by the semantic segmentation model as input data for the classification network; Image normalization: The input drug-susceptibility paper image is preprocessed by size normalization and pixel value normalization.

[0054] 4.2 First Convolutional Feature Extraction Layer Convolution kernel operation: Use multiple convolution kernels to perform convolution operations on the input image to extract low-level features; Feature map generation: Multi-channel feature maps (Featuredmaps) are generated through convolution operations to capture basic features such as the edges and textures of the drug-susceptible paper.

[0055] 4.3 First Pooling Dimensionality Reduction Layer Pooling operation: Perform maximum pooling or average pooling operations on the convolutional feature map to achieve feature dimensionality reduction and spatial invariance; Pooled feature maps: Generate pooled feature maps (PooledFeaturedmaps) with reduced spatial resolution to reduce computational complexity.

[0056] 4.4 Second Convolutional Feature Extraction Layer Deep feature extraction: further extract high-level abstract features through the second layer of convolution operation; Feature map fusion: The outputs of multiple convolution kernels are fused to form a richer feature representation.

[0057] 4.5 Second Pooling Dimensionality Reduction Layer Feature compression: Pooling operation is performed on the second-layer convolutional feature map to further compress the feature dimension; Pooled feature optimization: Generate the final pooled feature maps (PooledFeaturedmaps) to retain key drug recognition features.

[0058] 4.6 Fully Connected Classification Layer Feature flattening: Flatten the two-dimensional pooling feature map into a one-dimensional feature vector (Flattenlayer); Fully connected mapping: Mapping feature vectors to classification space through a fully connected neural network; Softmax activation: The softmax activation function is used to calculate the probability distribution of each drug category.

[0059] 4.7 Output classification results Drug category identification: Output the specific drug type corresponding to the drug sensitivity paper, such as KZ30 (clindamycin 30μg), TE30 (tetracycline 30μg), MH30 (minocycline 30μg), etc. Confidence evaluation: Simultaneously output the probability value of each category (0.1, 0.1, 0.98, etc.) for subsequent confidence weighted fusion processing; The CNN classification model achieves high-precision automatic recognition and classification of drug codes on drug-susceptibility paper strips through a hierarchical feature learning mechanism of multi-layer convolution-pooling-full connection.

[0060] like Figure 5 As shown in the figure, based on this semantic segmentation model processing effect comparison chart, the following step-by-step description of the technical implementation process: 5.1 Figure 5 Step 1: Original image input analysis Figure 1 Original input: An image of a culture medium containing multiple circular white drug-sensitive paper discs. The discs are irregularly distributed, and the background contains complex interference factors such as the culture medium surface texture, uneven lighting, and traces of bacterial growth. Figure 3Original input: Another petri dish sample image with different number and spatial layout of drug-susceptibility paper strips, and also with environmental interference such as background noise and lighting changes.

[0061] 5.2 Figure 5 Step 2 semantic segmentation network processing Pixel-level feature extraction: The DeepLabV3+ network analyzes the input image pixel by pixel and extracts the spatial features and boundary information of the drug-susceptibility paper strips through multi-scale convolution and dilated convolution mechanisms; Background-foreground binary classification: The network accurately classifies each pixel into two categories: "background" (culture medium surface) or "foreground" (drug-sensitive paper).

[0062] 5.3 Figure 5 Step 3: Segmentation mask generation Figure 2 Segmentation results: Outputs an accurate binary segmentation mask, in which the gray square area accurately identifies the location and boundary contour of the drug-susceptible paper disc, achieving pixel-level precision positioning of the circular disc in the original image. Figure 4 Segmentation results: corresponding Figure 3 The segmentation output also accurately marks the drug-susceptibility paper area with a gray square mask, demonstrating the algorithm's ability to generalize and recognize different samples. The comparison results show that the semantic segmentation model can achieve high-precision automatic positioning and boundary extraction of drug-susceptibility paper strips in a complex culture medium background, providing an accurate regional segmentation basis for subsequent classification and recognition.

[0063] like Figure 6 As shown in the figure, based on the hybrid model drug susceptibility paper identification result verification diagram, the technical verification process is described step by step below: 6.1 Figure 6 Step 1 Test sample set construction Diverse drug coverage: A standardized test dataset containing 23 different types of antimicrobial drugs was constructed, covering commonly used clinical drug sensitivity papers such as CIP5, FOX30, MH30, MEM10, TZP110, IPM10, KZ30, SAM20, ETD1, TGC15, PB30, PRL100, FOT200, CRO30, TGC15, AMP10, SAM20, IPM10, ATM30, MEM10, VA30, CZA50, P10, FOX30, and AK30; Repeatability validation design: Multiple repeated tests are performed on the same drug type (e.g., MH30-1, MH30-2) to verify the stability and reproducibility of the recognition algorithm.

[0064] 6.2 Figure 6 Step 2: Hybrid model prediction processing Semantic segmentation and positioning: The semantic segmentation model first accurately locates the drug-susceptibility paper area and extracts its boundaries in the input image; Classification model recognition: The CNN classification network recognizes the drug code and classifies the type of the segmented paper area; Confidence fusion calculation: adopt confidence weighted fusion strategy to output the final prediction result and the corresponding confidence score.

[0065] 6.3 Figure 6 Analysis of recognition result output in steps 3 and 4 Prediction accuracy verification: The prediction results (PredictedCombined) of all 23 test samples are completely consistent with the actual labels (Actual), achieving 100% recognition accuracy.

[0066] Confidence Assessment Analysis: High confidence identification: The confidence level of drugs such as CIP5, FOX30, MEM10, and TZP110 reached 1.00, indicating that the model has extremely high recognition certainty for these drug types; Medium confidence recognition: MH30 (0.94), KZ30 (0.93), IPM10 (0.92), etc., all have confidence levels exceeding the 0.9 threshold, meeting high-precision recognition requirements; The confidence distribution is reasonable: more than 99% of the samples have a confidence level exceeding 0.9, proving that the mixed model has excellent discriminative ability.

[0067] 6.4 Figure 6 Technical performance indicator evaluation in steps 3 and 4 Recognition accuracy: reached 100% (23 / 23), far exceeding the 85% recognition accuracy of existing technologies; Confidence stability: The average confidence level exceeds 0.95, meeting the high reliability requirements of clinical applications; Drug type coverage: Successfully identified discs covering multiple antimicrobial drug classes, including β-lactams, aminoglycosides, and quinolones; Reproducibility verification: The results of multiple tests on similar drugs are consistent, proving that the algorithm has good stability; The verification results show that the hybrid model-based drug-susceptibility paper identification method has extremely high recognition accuracy and reliability in practical applications, and can meet the precision requirements of clinical microbiology testing.

Claims

1. A high-precision identification method for culture medium drug-sensitive paper based on a hybrid model, characterized in that: The steps include: Step S1. Image acquisition and preprocessing: Standardize the acquired culture medium images, including noise removal, illumination correction, and contrast enhancement, to provide high-quality input images for subsequent model processing and eliminate the impact of environmental factors on recognition accuracy; Step S2. Semantic segmentation model processing: A deep learning network architecture is used to perform pixel-level classification on the image pre-processed in step S1, accurately locate the position and boundaries of all drug-sensitive paper strips in the culture medium, generate a segmentation mask, and achieve paper strip location in complex backgrounds; Step S3. Classification model processing: Based on the paper area extracted by the semantic segmentation result in step S2, a CNN classification model is used to identify the drug code and concentration on each paper, thereby achieving accurate classification and identification of different antimicrobial drug paper slips; Step S4. Hybrid model fusion: Intelligently fuse the location information of the semantic segmentation model in step S2 with the drug type information of the classification model in step S3, and output the final drug-susceptibility paper recognition result through the strategy of confidence weighting and result verification.

2. The method for high-precision identification of culture medium drug-sensitive paper sheets based on a hybrid model according to claim 1, characterized in that: In step S1, the illumination correction is to unify the image brightness by using histogram equalization and gamma correction technology to eliminate the influence of uneven illumination in different shooting environments.

3. The method for high-precision identification of culture medium drug-sensitive paper sheets based on a hybrid model according to claim 1, characterized in that: In step S1 , the contrast enhancement is to improve the visibility of image details by using adaptive histogram equalization.

4. The method for high-precision identification of culture medium drug-sensitive paper sheets based on a hybrid model according to claim 1, characterized in that: In step S2, the network architecture adopts DeepLabV3+ semantic segmentation network as the core architecture, ResNet-50 as the feature extraction backbone network, integrates void convolution to expand the receptive field, integrates void space pyramid pooling module to realize multi-scale feature fusion, and introduces decoder module for feature refinement to build an end-to-end drug-sensitive paper accurate segmentation system.

5. The method for high-precision identification of culture medium drug-sensitive paper sheets based on a hybrid model according to claim 1, characterized in that: In step S2, the pixel-level classification adopts the DeepLabV3+ semantic segmentation network architecture, uses ResNet-50 as the feature extraction backbone network, and implements the end-to-end pixel-level binary classification task through the encoder-decoder structure. Each pixel in the input image is accurately marked as "background" or "drug-sensitive paper" category, and a dense prediction mask with the same resolution as the original image is output to achieve pixel-level precise positioning and boundary delineation of the drug-sensitive paper.

6. The method for high-precision identification of culture medium drug-sensitive paper sheets based on a hybrid model according to claim 1, characterized in that: In step S3, the drug code refers to the standardized antimicrobial drug abbreviation code printed on the surface of the drug-susceptibility paper, which is in the form of a combination of 2-4 uppercase English letters and follows the unified naming convention of the international antimicrobial drug sensitivity test standard, and serves as a unique identifier for drug identification.

7. The method for high-precision identification of culture medium drug-sensitive paper sheets based on a hybrid model according to claim 1, characterized in that: In step S3, the CNN classification model is an end-to-end classifier designed specifically for the task of drug code recognition on drug-susceptibility paper strips using a deep convolutional neural network architecture. It uses ResNet-50 as the backbone network and integrates multi-scale feature extraction, attention mechanism, and transfer learning strategy to achieve high-precision automatic recognition and classification of drug codes in pre-segmented paper strip areas, and outputs drug category probability distribution and confidence assessment.

8. The method for high-precision identification of culture medium drug-sensitive paper sheets based on a hybrid model according to claim 1, characterized in that: In step S4, the confidence weighting adopts a dual-reset confidence evaluation mechanism of semantic segmentation and classification models, calculates the segmentation confidence and classification confidence respectively by softmax probability value and probability ratio, obtains the final confidence based on the adaptive weight fusion strategy, and sets multi-level thresholds to achieve differentiated post-processing. When the confidence is lower than 0.8, manual review is automatically triggered.

9. The method for high-precision identification of culture medium drug-sensitive paper sheets based on a hybrid model according to claim 1, characterized in that: In step S4, the result verification is to establish a multi-dimensional cross-validation mechanism, ensure the rationality of paper distribution through spatial consistency verification, verify the geometric parameters of paper through morphological feature verification, and identify abnormal results through historical data comparison verification, and build a full-range automatic error correction system from spatial layout, physical characteristics to statistical laws.

Images