A generative intelligent driven ultrafast antimicrobial susceptibility testing method and apparatus

Abstract

The application discloses a kind of generatable intelligent driving's super-fast antibacterial drug sensitivity test method and equipment, belong to clinical microbiological examination and medical detection equipment technical field.The method includes: continuous acquisition agar plate's delay digital image;Image processing is carried out using interconnected target detection model and time series prediction model, target detection model identifies and quantifies evolution in bacteriostatic circle profile output time sequence measurement sequence, time series prediction model deduces growth dynamics trend by time series regression, reconstructs complete growth curve and calculates end point equivalent bacteriostatic circle diameter;End point equivalent bacteriostatic circle diameter is compared with pre-set clinical sensitivity standard and report is output.The application is accurately identified and end point is predicted to early weak bacteriostatic circle by deep learning joint algorithm framework, solve the problem that traditional drug sensitivity test period is long, artificial reading error is big and existing equipment recognition precision is low, realize super-fast, low-cost and high-precision drug sensitivity detection.

Description

Technical Field

[0001] This invention relates to a generative, intelligently driven, ultra-rapid method and equipment for testing the susceptibility of antimicrobial drugs, belonging to the technical field of clinical microbiology testing and medical detection equipment. Background Technology

[0002] Antimicrobial resistance has become a global public health crisis. The abuse and overuse of antibiotics have further accelerated the emergence of multidrug-resistant pathogens, leading to extreme dependence on empirical broad-spectrum antibiotics. Therefore, rapid and accurate antimicrobial susceptibility testing is crucial for guiding early targeted drug use in clinical practice and slowing the spread of resistance.

[0003] Among existing phenotypic drug susceptibility testing methods, the classic paper diffusion method is widely used globally due to its simplicity and extremely low cost. However, traditional paper diffusion phenotypic drug susceptibility testing suffers from a significant drawback: it relies heavily on the visual maturity endpoint outline formed after bacteria have fully proliferated overnight. Diameter interpretation typically requires 18 to 24 hours of incubation, a daily turnaround time that severely delays the golden treatment window for critically ill patients. Furthermore, manual measurement relying on the naked eye or measuring rulers introduces a high degree of subjectivity, resulting in extremely poor repeatability and standardization of results. Moreover, commercially available large-scale fully automated drug susceptibility testing systems are extremely expensive in terms of both construction and operating consumables, making them difficult to popularize in primary healthcare, small and medium-sized laboratories, or point-of-care testing scenarios.

[0004] In recent years, the development of machine vision and time-series deep learning algorithms has brought opportunities to digital microbiology. Studies have shown that the growth dynamics of the early inhibition zone during the first 5 to 9 hours of incubation are strongly correlated with the final endpoint morphology after 24 hours. If endpoint prediction can be achieved at this stage, same-day drug sensitivity reporting can be realized. However, the implementation of this concept faces unavoidable underlying technical challenges. On the one hand, during the first 5 to 9 hours of incubation, the colonies are not yet fully formed, the local illumination on the agar surface is uneven, and the background noise is strong, resulting in the inhibition zone being in an extremely weak, low-contrast, and continuously evolving dynamic morphology. Conventional machine vision detection models are prone to missed detections or edge misjudgments when faced with such visually blurred images, making it impossible to achieve continuous and stable time-series measurements. At the same time, the weak boundary of the early inhibition zone is not a single gray-level abrupt change; it often exhibits a complex signal with the superposition of dark ring response, texture uniformity changes, local gradient peaks, and radial gray-level gradual changes. If only fixed thresholds, single edge operators, or single-frame static detection are used, paper edges, culture medium textures, bubble scratches, or inner dark peaks are easily misidentified as inhibition zone boundaries. If relying entirely on manual annotation, the early low-contrast boundary annotation workload is large and inconsistent. On the other hand, the growth kinetics of microorganisms, colony optical characteristics, culture medium variability, and the physical diffusion rates of different antibiotic molecules vary greatly. This causes existing static single-frame AI drug sensitivity models to produce serious prediction reading biases and performance degradation when faced with unseen pathogen and drug combinations, lacking an effective mechanism for rapid local adaptive adaptation.

[0005] Furthermore, current automated flat-panel imaging equipment suffers from functional limitations and a severe lack of system integration. These devices mostly employ a simple timed static imaging mode, lacking a complete data loop and intelligent collaboration between hardware modules. They also lack an edge computing control core capable of deeply integrating continuous optical imaging with time-series generative AI trajectory prediction algorithms, and an intuitive graphical user interface. This results in their inability to perform high-precision real-time quantitative tracking of subtle diameter changes in the early stages, their inability to provide compatibility with clinical and laboratory standards association guidelines, and their inability to quickly generate quantitative drug sensitivity reports. This severely restricts the iterative upgrade of phenotypic drug sensitivity technology towards ultra-fast, low-cost, and highly generalizable applications. Summary of the Invention

[0006] To address the problems of long testing cycles and large human interpretation errors in paper diffusion method for drug susceptibility testing, as well as the low accuracy, poor generalization ability, and insufficient system integration of existing automated equipment when dealing with early low-contrast inhibition zones, this invention provides a generative, intelligently driven, ultra-rapid antimicrobial drug susceptibility testing method and equipment. The technical solution is as follows: In a first aspect, the present invention provides a method for testing antimicrobial susceptibility, applied to the parallel incubation process of MH agar plates inoculated with bacterial suspension and containing antimicrobial drug discs, comprising the following steps: Step 1: During the incubation period, time-lapse digital images of the MH agar plates are collected periodically and continuously; Step 2: The time-lapse digital image is input into the intelligent prediction module for processing. The intelligent prediction module has a built-in deep learning joint algorithm architecture, which includes an interconnected target detection model and a time series prediction model. The target detection model performs dynamic feature extraction on the time-lapse digital image, identifies and quantifies the evolving inhibition zone outline, and outputs a time series measurement sequence of the inhibition zone diameter. The time series prediction model receives the time series measurement sequence of the inhibition zone diameter, infers the growth dynamics trend through time series regression, reconstructs the complete incubation cycle growth curve, and calculates and outputs the corresponding endpoint equivalent inhibition zone diameter. Step 3: Compare the diameter of the equivalent inhibition zone at the endpoint with the preset clinical sensitivity standard, and output an antimicrobial susceptibility test report.

[0007] Optionally, the target detection model is a target detection model optimized based on the YOLO architecture, including: a heuristic backbone network, a hybrid attention mechanism module, and a geometric factor simplified detection head; The heuristic backbone network replaces the C2f module in the native YOLO architecture with a simplified convolutional block and a hierarchical channel reduction strategy, and constructs a feature pyramid by combining shallow residual connections. The hybrid attention mechanism module includes a cascaded adaptive channel attention module and a spatial attention module; The geometric factor simplified detection head uses a multi-point distance intersection-union loss function.

[0008] Optionally, the time series prediction model is a long short-term memory network enhanced with a time attention mechanism, comprising: Long short-term memory network units are stacked vertically, with discard layers configured between them; A time attention mechanism module integrated into the input end is used to calculate and assign nonlinear time step weights to the input sequence in order to focus on the abrupt change points of dynamic features; A mapping layer is used to map the output of the time attention mechanism module to the predicted value of the diameter of the equivalent inhibition zone at the endpoint.

[0009] Optionally, before the target detection model processes the data, the intelligent prediction module also performs a machine vision early weak boundary pre-identification and annotation assistance step, including: Flat field illumination correction and local contrast enhancement are performed on the input time-delayed digital image; Dark ring response, local texture response, gradient response and grayscale change response are extracted and fused to generate a two-dimensional antibacterial zone response map; Based on the two-dimensional inhibition zone response map, a conservative initial inhibition zone radius is obtained, and a one-dimensional radial compensation curve is constructed by polar coordinate radial sampling to determine the boundary compensation position. The compensated inhibition zone boundary labeling information is then output to assist in the training or inference of the target detection model.

[0010] Optionally, the intelligent prediction module also includes a user-oriented adaptive learning workflow, comprising: Receive time-lapse image datasets of specific pathogen-antibiotic combinations imported by the user; By utilizing a modular fine-tuning mechanism, the weights of the time series prediction model are incrementally modified while maintaining the overall structure of the deep learning joint algorithm architecture, in order to adapt to the dynamic evolution trajectory of the local strain.

[0011] Optionally, the step of outputting the antimicrobial susceptibility test report may also include early typing identification for different specific resistance phenotypes, including at least one of the following: Single-disk reverse interpretation and genotyping: Identifying specific drug resistance mechanism phenotypes by monitoring the temporal evolution characteristics of the inhibition zone of a single antimicrobial drug disk; Multi-disk difference co-typing: Identifying broad-spectrum enzyme inhibitor resistance phenotypes by recording or predicting the difference in inhibition zone diameter between at least two antimicrobial drug disks; Temporal dynamics phenotyping: Identifying super-resistant phenotypes by extracting inhibition zone expansion rate and specific growth trajectory features.

[0012] In a second aspect, the present invention provides an antimicrobial susceptibility testing device for implementing the method described in any of the preceding claims, comprising: Miniature box-type basic structure; A multi-channel parallel incubation system is set within the micro-box basic structure, including a temperature and humidity control unit and a motor-driven turntable, on which multiple slots for mounting agar plates are arranged. A fully automated image acquisition system, suspended and fixed directly above the turntable, includes an industrial-grade digital camera and a light source; An edge computing control core is integrated into the micro-box basic structure, and the edge computing control core integrates an intelligent prediction module. The intelligent prediction module has a built-in deep learning joint algorithm architecture, which includes an interconnected target detection model and a time series prediction model. The target detection model is configured to process the time-series images input from the fully automated image acquisition system, identify the boundaries of the antibacterial drug paper discs, quantify the radius of the evolving inhibition zone, and output a time-series measurement sequence of the inhibition zone diameter. The time-series prediction model is configured to receive the time-series measurement sequence of the inhibition zone diameter, deduce the growth dynamics trend through time-series regression, reconstruct the complete incubation cycle growth curve, and calculate and output the corresponding endpoint equivalent inhibition zone diameter. The graphical user interface adopts a three-tier software architecture, including a view layer, a business layer, and a model and data layer, and is divided into two collaborative monitoring system interfaces: a detection system interface and a prediction system interface.

[0013] Optional, The detection system interface is used to dynamically visualize and present time-lapse digital images, and overlay bounding boxes, antibiotic category labels, confidence scores, and real-time qualitative drug sensitivity status generated by the target detection model. The prediction system interface is used to display a summary table of prediction results and simultaneously plot the prediction effect curve and the actual prediction curve to show the dynamic evolution overlap between the time-series prediction trajectory and the actual trajectory.

[0014] Optionally, the fully automatic image acquisition system further includes a multi-dimensional camera adjustment mechanism, including a fixed bracket, a slider, an adjustment bracket, and a lens mount. The industrial-grade digital camera is suspended and fixed within the lens mount, and the light source is a ring-shaped shadowless light source that provides uniform diffuse illumination.

[0015] Optionally, the temperature and humidity control unit includes a fan, heating tube, humidification module and thermocouple installed in the inner cavity, and the edge computing control core maintains the constant temperature and humidity environment in the inner cavity of the box through proportional-integral-derivative closed-loop feedback regulation.

[0016] The beneficial effects of this invention are: This invention utilizes an intelligent prediction module with a built-in deep learning joint algorithm architecture to process delayed digital images collected during the incubation period. By employing interconnected target detection models and time series prediction models, it achieves accurate identification, quantification, and prediction of the equivalent diameter of the early weak inhibition zone contour. This solves the technical problems of long testing cycles and large subjective errors in manual interpretation of traditional paper diffusion methods for drug sensitivity testing, as well as the low identification accuracy and difficulty in achieving ultra-fast detection of early low-contrast inhibition zones by existing automated equipment. Experiments have demonstrated that this invention can accurately predict the 24-hour endpoint results as early as the incubation stage (e.g., 5-9 hours), significantly advancing the issuance time of drug sensitivity reports by 15-19 hours, thus significantly shortening the detection cycle and securing a valuable golden treatment window for early targeted drug therapy in clinical practice. This invention utilizes a deep learning model to replace manual visual interpretation, eliminating subjective errors and achieving pixel-level precise positioning and continuous stable tracking of low-contrast, weak-boundary inhibition zones, improving detection accuracy and objectivity, and ensuring the repeatability and standardization of test results. Furthermore, this invention seamlessly integrates image acquisition, intelligent analysis, and result interpretation, achieving fully automated closed-loop control from sample incubation to report generation, greatly improving detection efficiency and throughput while reducing labor costs.

[0017] This invention addresses the technical challenges of existing automated drug sensitivity testing systems, such as low system integration, lack of end-to-end data closure and intelligent collaboration, and inability to achieve early high-precision real-time quantitative tracking and rapid quantitative report generation, through an integrated equipment architecture comprising a miniature enclosure, a multi-channel parallel incubation system, a fully automated image acquisition system, an edge computing control core, and a graphical user interface. First, by deeply integrating multi-channel parallel incubation with the fully automated image acquisition system, and relying on the deep learning joint algorithm architecture built into the edge computing control core, it achieves fully automated closed-loop control from sample incubation and time-series image acquisition to intelligent predictive analysis, completely eliminating subjective errors and time delays introduced by manual operation, and significantly improving detection efficiency and throughput. Second, the deployment of the edge computing core enables real-time local processing of time-series images, avoiding data transmission bottlenecks and ensuring accurate identification of early weak inhibition zone contours and rapid prediction of the endpoint equivalent diameter, thus securing a valuable golden treatment window for clinical practice. Finally, a three-layer software architecture graphical user interface, through collaborative monitoring of the detection and prediction dual-system interfaces, not only achieves dynamic visualization of the detection status and intuitive presentation of the prediction trajectory, but also ensures compatibility with clinical and laboratory standards association guidelines. This greatly enhances the intelligence level and clinical applicability of the equipment, effectively promoting the iterative upgrade of phenotypic drug sensitivity technology towards ultra-fast, low-cost, and highly generalizable directions. The closed-loop temperature and humidity control and in-situ short-wave ultraviolet disinfection components inside the chamber further ensure a highly stable incubation environment, eliminating the risk of biosafety exposure and cross-contamination for operators. Overall, this invention combines high precision, extremely short processing time, ultra-low cost, and strong versatility and scalability, providing an advanced ultra-fast diagnostic solution for fields such as clinical precision anti-infective medicine and hospital epidemiological drug resistance monitoring, and has broad application prospects.

[0018] Furthermore, this invention fully retains the extremely high clinical accessibility and ultra-low consumable cost of the classic paper diffusion method. It is directly compatible with standard commercially available antibiotic paper discs and agar plates widely used globally, bypassing the heavy reliance of existing commercial automated drug sensitivity systems on expensive and closed-loop micro-liquid-based dilution card consumables. This significantly reduces the cost of single-test equipment and operating consumables in primary healthcare and resource-limited environments.

[0019] Furthermore, this invention employs an improved YOLO architecture as the target detection model, effectively solving the technical challenge of blurred edge recognition caused by low contrast and strong background noise interference in early inhibition zone images. By constructing a heuristic backbone network using simplified convolutional blocks and a hierarchical channel reduction strategy, the computational complexity and number of parameters of the model are significantly reduced while retaining the ability to extract multi-scale morphological features from low-contrast images. Combined with a hybrid attention mechanism module, adaptive weighting of feature map channels and space is achieved, effectively suppressing background noise such as culture medium texture. In addition, the geometric factor of the detection head is simplified using a multi-point distance intersection-union loss function, which strengthens the localization constraint of inhibition zone edges and significantly improves the accuracy and robustness of the model in recognizing weak inhibition zone contours in complex incubation environments.

[0020] Furthermore, this invention employs an improved Long Short-Term Memory (LSTM) network as the time series prediction model, effectively addressing the technical challenges of large differences in microbial growth kinetics and uneven diffusion rates of different antibiotics leading to significant deviations in early prediction readings. By integrating a time attention mechanism module, the model can automatically calculate and assign nonlinear time step weights to the input sequence, thereby accurately focusing on the kinetic mutation points in the inhibition zone growth process and effectively capturing key temporal information. Combining vertically stacked LSM network units with interlayer dropout layers enhances the model's ability to extract nonlinear features from long sequence data while effectively preventing overfitting. Finally, a mapping layer accurately transforms the temporal features into the equivalent inhibition zone diameter at the endpoint, significantly improving the model's generalization ability and prediction accuracy when facing unknown pathogens and drug combinations.

[0021] Furthermore, this invention introduces a machine vision-based early weak boundary pre-identification and annotation auxiliary unit, which has the advantages of interpretable parameters, fast operation speed, and no need for prior deep learning training. It can provide stable initial radius, radial compensation radius, boundary annotation, and diameter time series even when there is insufficient training data for the target detection model, manual annotation is difficult, or the early boundary is extremely weak. By combining two-dimensional conservative detection with one-dimensional global peak extrapolation, this unit avoids the problem of small boundaries in simple thresholding methods. It also suppresses excessive extrapolation and single-frame false detections through hard extrapolation upper limits, drug role constraints, and time series monotonic constraints. This improves the annotation efficiency, consistency, and target detection model training quality of weak inhibition zone boundaries in early drug sensitivity images (5-9 hours), and can further shorten drug sensitivity testing time.

[0022] Furthermore, the user-oriented adaptive learning workflow and multidimensional phenotyping mechanism introduced in this invention enable the system to perform incremental trajectory calibration on specific pathogens and novel drug combinations not seen locally using very few samples, without the need for global retraining or network architecture reorganization that requires a large amount of manual annotation. This perfectly overcomes the generalization bottleneck caused by microbial specific dynamics and culture medium contrast variation, and can efficiently and accurately identify a variety of complex high-risk drug resistance phenotypes, including methicillin resistance, extended-spectrum β-lactamase production, ceftazidime resistance, and vancomycin resistance. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a three-dimensional schematic diagram of the overall structure of an ultra-fast antibacterial drug sensitivity testing device according to one embodiment of the present invention.

[0025] Figure 2 This is a side view of the overall structure of an ultra-fast antibacterial drug sensitivity testing device according to one embodiment of the present invention.

[0026] Figure 3 This is a schematic diagram showing the position and layered assembly of the multi-channel parallel incubation system and the fully automatic image acquisition system in one embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram illustrating the paper diffusion principle and inhibition zone formation mechanism of phenotypic drug sensitivity testing in one embodiment of the present invention.

[0028] Figure 5 This is a schematic diagram illustrating the process of automatically locating, classifying, and labeling different antibiotic paper discs using a target detection network in one embodiment of the present invention.

[0029] Figure 6 This is a schematic diagram illustrating the process of extracting dynamic features of the inhibition zone diameter and performing real-time quantitative measurement during the early stage of incubation in one embodiment of the present invention.

[0030] Figure 7 This is a schematic diagram of the workflow of a fully automated dynamic digital image acquisition, image analysis and target detection system for intelligent drug sensitivity testing in one embodiment of the present invention.

[0031] Figure 8 This is a flowchart illustrating the machine vision early weak boundary pre-identification and annotation auxiliary unit in one embodiment of the present invention.

[0032] Figure 9 This is a bar chart showing the absolute error distribution between the measured diameter of the inhibition zone output by the target detection model in one embodiment of the present invention and the diameter measured manually in the traditional way.

[0033] Figure 10 This is a schematic diagram of the dual-modal deep learning algorithm architecture and workflow for target detection and time series prediction in one embodiment of the present invention.

[0034] Figure 11 This is a diagram of the core cascaded architecture of a prediction network based on the interconnection of a temporal attention mechanism and a long short-term memory network in one embodiment of the present invention.

[0035] Figure 12 This is a scatter plot showing the correlation between the very early direct reading diameter without algorithmic prediction and the final physical diameter over 24 hours, according to one embodiment of the present invention.

[0036] Figure 13 This is a scatter plot showing the correlation between the equivalent diameter output by the time-series prediction model of this invention and the final physical diameter over 24 hours in one embodiment of the invention.

[0037] Figure 14 This is a comparison chart of the mean absolute error, root mean square error, and coefficient of determination of the time series prediction model under different incubation time windows in one embodiment of the present invention.

[0038] Figure 15 This is a histogram showing the error distribution between the diameter predicted by the time-series prediction model and the actual diameter at the 24-hour endpoint in one embodiment of the present invention.

[0039] Figure 16 This is a confusion matrix diagram for determining the sensitive / intermediate / resistant (S / I / R) status using direct readings in the early immature stage according to one embodiment of the present invention.

[0040] Figure 17 This is a confusion matrix diagram for determining the sensitive / intermediate / resistant (S / I / R) status after prediction by the model of this invention in one embodiment of the present invention.

[0041] Figure 18 This is a block diagram of the modular fine-tuning mechanism for a user-oriented adaptive learning workflow in one embodiment of the present invention.

[0042] Figure 19 This is a time-series kinetics expansion trajectory curve of a novel, unknown clinical strain after adaptive fine-tuning, according to one embodiment of the present invention.

[0043] Figure 20 This is a scatter plot comparing the early reading deviation of a novel unknown strain before and after adaptive fine-tuning with the ideal correlation baseline in one embodiment of the present invention.

[0044] Figure 21 This is a schematic diagram of the end-to-end closed-loop analysis algorithm pipeline for the graphical user interface in one embodiment of the present invention.

[0045] Figure 22 This is a schematic diagram of a real-time image monitoring and target positioning interface of a graphical user interface detection system according to one embodiment of the present invention.

[0046] Figure 23 This is a schematic diagram of the graphical user interface for predicting the time-series growth trajectory analysis and endpoint report generation of a prediction system according to one embodiment of the present invention.

[0047] In the diagram, 1. Hollow square steel frame; 2. Support plate; 3. First heat dissipation plate; 4. Inner plate; 5. Second heat dissipation plate; 6. Right side plate; 7. Side cover plate; 8. Front cover plate; 9. Lower side plate; 10. Thermocouple; 11. Upper pivot; 12. Lower pivot; 13. Outer door panel; 14. Inner door panel; 15. Upper cover plate; 16. Rear side plate; 17. Left side plate; 18. Lower bottom plate; 19. Ventilation plate; 20. Rotary motor; 21. Turntable fixing bracket; 22. Side cover plate; 23. Temperature controller; 24. Controller; 25. Fixing bracket; 26. Slider; 27. Adjusting bracket; 28. Lens fixing bracket; 29. ​​Lens; 30. Light source; 31. Turntable; 32. Heating tube; 33. Bottom support plate; 34. Fan. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0049] Example 1: This embodiment provides a method for testing antimicrobial susceptibility. This method is based on generative intelligent driving and utilizes fully automated image acquisition and deep learning algorithms to accurately predict the 24-hour endpoint result in the early stages of incubation (5-9 hours). Specifically, it includes the following steps: Step 1: Perform multi-channel parallel incubation to prepare samples.

[0050] An MH agar plate inoculated with 0.5 μL of McFarland standard bacterial suspension and covered with antimicrobial drug paper discs is placed on the motor-driven rotary slot of the ultra-rapid antimicrobial susceptibility testing equipment.

[0051] The multi-channel parallel incubation system is started. Through the temperature and humidity closed-loop control unit (including fan, heating tube, humidification module and thermocouple), under the PID closed-loop feedback regulation of the edge computing control core, a constant temperature and humidity microenvironment for microbial incubation is maintained in the chamber at 35°C and 60%~70% relative humidity. The turntable is rotated in a controlled stepwise manner to achieve parallel incubation of multiple samples.

[0052] Step 2: Acquire time-series image data.

[0053] During the complete incubation cycle of 0 to 24 hours, the fully automatic image acquisition system is controlled to periodically and continuously acquire time-delayed digital images of MH agar plates in the early kinetic diffusion stage at a preset time resolution.

[0054] The acquired raw images are named according to the rule "strain-antibiotic-time.jpg" and stored in the model and data layers.

[0055] Step 3: Image preprocessing and enhancement.

[0056] The acquired raw time-series images are preprocessed, including lightweight calibration, background normalization, denoising, and image enhancement preprocessing. The image enhancement preprocessing further includes flat field illumination correction, contrast-limited adaptive histogram equalization, local standard deviation texture calculation, Sobel gradient magnitude calculation, and dark residual response extraction, in preparation for subsequent weak boundary recognition.

[0057] Step 4: The intelligent prediction module is used to perform dynamic feature extraction and temporal evolution analysis on the preprocessed time-delayed digital image. In the early stage when the visual inhibition zone is not mature, the complete growth dynamic trajectory of the inhibition zone is reconstructed, and the predicted diameter of the inhibition zone at the end of 24 hours is generated.

[0058] The intelligent prediction module incorporates a deep learning joint algorithm architecture, including an interconnected target detection model and a time series prediction model. The target detection model processes the time-series images input from the fully automated image acquisition system, identifying the boundaries of different antibacterial drug-containing paper discs within the images. It accurately locates and quantifies the radius of the evolving weak inhibition zone under early low-contrast conditions, outputting a continuous time-series measurement sequence of inhibition zone diameters. The time-series prediction model receives the time-series measurement sequence of inhibition zone diameters, uses time-series regression to deduce the growth kinetic trend, reconstructs a complete 24-hour growth curve, and directly calculates and outputs the corresponding 24-hour endpoint equivalent diameter.

[0059] The target detection model is a torsion-optimized target detection model, including: A heuristic backbone network replaces the convolutional modules in the native detection network, employs minimal convolutional blocks and hierarchical channel reduction strategies, and combines shallow residual connections to construct a feature pyramid, which is used to extract multi-scale morphological features of low-contrast images in the early immature incubation stage. The hybrid attention mechanism module integrates a dual attention mechanism consisting of an adaptive channel attention module and a spatial attention module connected in series in the feature fusion network. This mechanism is used to share the feature map and suppress background noise caused by uneven agar matrix and the growth of contaminants. The multi-dimensional geometric constraint detection head, in the bounding box regression component, adopts the multi-point distance intersection-union ratio loss function. By directly minimizing the Euclidean distance between the predicted bounding box and the corresponding diagonal vertices of the true bounding box, it simultaneously calculates the center point, corner points, diagonal length, and contour overlap area, achieving pixel-level accurate positioning of the edge of the early weak antibacterial zone.

[0060] The time series prediction model is a long short-term memory network enhanced by a time attention mechanism. It consists of two vertically stacked long short-term memory network units, each containing 64 hidden layer neurons. A dropout layer with a ratio of 0.2 is configured between the two long short-term memory network units to suppress overfitting. The network input integrates a time attention mechanism module, which normalizes the input sequence measured at 144 time points within 0 to 24 hours, calculates and assigns nonlinear time step weights, and adaptively focuses on the early transient dynamic feature mutation points that have the most predictive information for the 24-hour endpoint. The output of the time attention mechanism module is mapped through a fully connected layer to directly obtain the predicted value of the final 24-hour inhibition zone diameter for this specific pathway.

[0061] Furthermore, the intelligent prediction module in this embodiment also includes a machine vision early weak boundary pre-identification and annotation auxiliary unit set before the target detection model. This unit first performs flat-field illumination correction and local contrast enhancement on the flat image, and calculates the dark residual response map, local standard deviation texture map, gradient magnitude map, and grayscale change map respectively. Then, it weights and fuses the dark ring response, low texture response, and low gradient background suppression term to generate a two-dimensional inhibition zone response map. Within the annular region outside each antibacterial drug paper disc, connected component analysis is performed using multiple percentile thresholds to select candidate regions that are connected to the seed ring near the paper disc, have high angular coverage, and strong average response, thus obtaining a conservative initial inhibition zone radius. Based on this, polar coordinate radial sampling is performed with the paper disc center as the pole, and a one-dimensional radial compensation curve is constructed by integrating texture rise, gradient peak, grayscale change, and dark ring change. Within the limited external compensation search range, the global maximum peak is selected as the weak boundary external compensation position. The boundary contour, diameter annotation, and diameter time series output by the unit are used for automatic annotation, annotation verification, and early boundary learning of the target detection model training samples.

[0062] Furthermore, the intelligent prediction module also incorporates a user-oriented adaptive learning workflow to eliminate model generalization bottlenecks caused by differences in strain types, culture medium conditions, and local epidemiology. This includes: allowing local laboratory users to import minimum-latency image datasets of specific pathogens and antibiotic combinations that were not involved in the global retraining of the original model via a graphical user interface; and utilizing the modular fine-tuning mechanism of the adaptive learning workflow to incrementally correct the temporal evolution trajectory weights of the time series predictor using a very small number of samples, without requiring architecture reconstruction, global retraining, or large-scale manual annotation of the entire deep learning network. This is used to reconstruct and correct the inhibition zone dynamics evolution trajectory of novel strains or drug combinations in the early immature incubation stage, eliminate early prediction bias, and achieve generalization calibration and adaptation of the system to local multi-source clinically isolated strains and drug sensitivity phenotypes of novel antibiotic combinations.

[0063] Step 6: Phenotypic quantitative interpretation and report generation.

[0064] The system automatically compares the predicted diameter of the equivalent inhibition zone at the 24-hour endpoint with the preset clinical susceptibility disc diffusion breakpoint standard, outputs a quantitative drug susceptibility report for sensitive / intermediate / resistant (S / I / R), and achieves early and rapid typing identification of specific drug resistance phenotypes in the early stage of 5-9 hours.

[0065] This embodiment enables early and rapid typing and identification of different specific drug resistance phenotypes, including: Single-disk reverse interpretation typing, for specific Gram-positive or Gram-negative bacterial samples, by real-time monitoring and prediction of the time evolution characteristics of the inhibition zone of a single specific antimicrobial drug disk, directly distinguishes and identifies the corresponding specific drug resistance mechanism phenotype at an early stage; Multi-disk difference co-typing, for specific Gram-negative bacterial samples, directly confirms specific broad-spectrum enzyme inhibitor resistance phenotypes at an early stage by recording or predicting the real-time or predicted inhibition zone diameter difference between at least two antimicrobial drug disks and combining disk evolution mechanisms. Temporal kinetic characterization, targeting Gram-positive or Gram-negative clinically isolated strains with specific drug resistance evolution trajectories, directly identifies the corresponding super-resistant phenotype by extracting the inhibition zone expansion rate and specific growth trajectory characteristics in the early kinetic stages.

[0066] Example 2: This embodiment provides an antimicrobial susceptibility testing device, such as... Figures 1 to 3 As shown, it includes a micro-box basic structure, a multi-channel parallel incubation system, a fully automatic image acquisition system, and an edge computing control core with an integrated intelligent prediction module.

[0067] The basic structure of the miniature enclosure includes an overall frame constructed from a hollow square steel frame 1. The exterior and sides are covered and fixed with a right side plate 6, a lower side plate 9, a rear side plate 16, and a left side plate 17. The top and bottom ends are connected to an upper cover plate 15 and a lower bottom plate 18, respectively. The interior is enclosed by a support plate 2 and an inner plate 4, forming a highly insulated incubation cavity. A ventilation window for installing a vent plate 19 is partially opened on the top surface of the frame. The outer side front wall and lower front side of the main frame are respectively enclosed by a shell structure with a side cover plate 7, a front cover plate 8, and a side cover plate 22, and embedded with a temperature controller 23 and a controller 24, facilitating real-time monitoring and control of the testing environment and equipment parameters by operators from the outside. An outer door plate 13 and an inner door plate 14 for sealing the enclosure are mounted on the front of the main body. The outer door plate 13 and the inner door plate 14 are hinged to one side of the frame via an upper pivot 11 and a lower pivot 12, allowing them to flip outwards when the enclosure is opened.

[0068] A multi-channel parallel incubation system is configured within the cavity, including a temperature and humidity control unit and a motor-driven turntable 31 controlled by an edge computing control core. Figure 3 The layered assembly structure shown features a rotary motor 20 vertically mounted above the support plate 2 at the bottom of the inner cavity via a turntable fixing frame 21. The output axis of the rotary motor 20 extends upwards and drives the horizontally connected turntable 31. The turntable 31 has multiple slots axially arranged for mounting and interlocking standard circular MH agar plates, enabling precise step-by-step rotation and parallel incubation of drug-sensitive samples. The temperature and humidity control unit includes a fan 34, a heating tube 32, a humidification module, and a thermocouple 10 installed within the inner cavity. The space below the turntable 31 further houses the heating tube 32, the bottom support plate 33, and the fan 34. The fan 34 drives airflow through the heating tube 32, providing efficient and uniform heat circulation to the inner cavity. The equipment features a staggered arrangement of a first heat dissipation plate 5 and a second heat dissipation plate 3 within its internal cavity frame and side structure, providing excellent heat convection and barrier functions. Through proportional-integral-derivative closed-loop feedback regulation of the edge computing control core, a constant temperature and humidity microenvironment for microbial incubation is maintained within the chamber at 35°C and 60%–70% relative humidity. Short-wave ultraviolet disinfection lamps are also fixed to the inner cavity sidewalls for in-situ sterilization of the incubation chamber after a single drug sensitivity test.

[0069] The fully automated image acquisition system is suspended and fixed directly above the turntable 31, including an industrial-grade digital camera and a light source 30 providing uniform diffused illumination. The light source 30 adopts a ring-shaped light-emitting diode shadowless light source structure. The multi-dimensional camera adjustment mechanism includes a vertically fixed bracket 25 fixedly mounted on the support plate 2. A slider 26 that can slide up and down is mounted on the fixed bracket 25. The slider 26 is connected to an adjustment bracket 27, and the front end of the adjustment bracket 27 is connected to a lens mounting bracket 28. The industrial-grade digital camera is suspended and fixed inside the lens mounting bracket 28. A lens 29 is threadedly connected to the lower end of the camera. The lens 29 is vertically fixed at the center hole of the light source 30, and its optical axis is vertically aligned with the current flat working position of the turntable 31, enabling high-definition full-field digital image capture of drug-sensitive samples when the turntable 31 rotates stepwise to switch working positions. A side cover plate 7 is also provided on the edge of the frame to protect the transmission components. Figure 3 As shown, a heating tube 32, a bottom support plate 33, and a fan 34 are further sealed and assembled below the multi-dimensional camera adjustment mechanism and at the transmission layer to prevent the penetration of warm and humid gases into the chassis electrical compartment.

[0070] During the time-series image acquisition process, the main control program controls an industrial-grade digital camera to acquire multi-channel images at a time resolution of 10 minutes per frame, continuously recording the complete full-cycle dynamic evolution of the inhibition zone within 0-24 hours. Its underlying basic physical diffusion and initial growth mechanisms are as follows: Figure 4 As shown, on MH agar plates covered with antibacterial drug discs, the growth state of microorganisms and the continuous diffusion of antibiotic molecules are antagonistic. In the early diffusion stage of incubation, the resulting low-contrast, faint outlines are extremely difficult to discern directly with the naked eye. The acquired raw images undergo lightweight calibration, background normalization, denoising, and image enhancement preprocessing within the edge computing control core to eliminate interference from equipment lighting fluctuations and the agar plate background. They are then stored in the model and data layers using a strain-antibiotic-time naming convention.

[0071] Furthermore, the intelligent prediction module incorporates a deep learning joint algorithm architecture that includes interconnected object detection and time series prediction models. During operation, this method cascades the object detection and time series prediction models to transform very early, low-confidence weak contour signals into high-fidelity 24-hour endpoint prediction diameters. First, the object detection model processes the time-series images input from the fully automated image acquisition system. It extracts multi-scale features using a convolutional neural network. The specific paper target localization and category labeling process is as follows: Figure 5 As shown, it can automatically identify and locate the boundaries of different antibacterial drug paper pieces in an image. Based on this, the target detection model accurately locates and quantifies the radius of the evolving weak inhibition zone under early low-contrast conditions. The dynamic feature extraction and real-time numerical quantization process of the weak edge contour of the inhibition zone is as follows: Figure 6As shown, this is used to output a continuous time-series measurement sequence of the inhibition zone diameter. To address the challenge of large-scale dataset labeling before model training, this system provides a method such as... Figure 7 The workflow shown is a fully automated dynamic digital image acquisition, image analysis, and target detection process. The intelligent prediction module first calls, for example... Figure 8 The machine vision early weak boundary pre-identification and annotation auxiliary unit shown performs early antibacterial zone boundary pre-identification and diameter sequence extraction on time-lapse digital images. Its output is used to generate initial annotations for the target detection model, assist in manual verification, expand early weak boundary training samples, and further form a continuous diameter sequence that can be input into the time series prediction model, significantly reducing the cost of manual labeling. The target detection model is based on an improved YOLOv8 target detection architecture, including a backbone network, a feature fusion network, and a multi-dimensional geometric constraint detection head. The absolute error distribution between its detected diameter and manually measured values ​​is shown below. Figure 9 As shown, it exhibits extremely strong positioning stability.

[0072] Furthermore, the dual-modal deep learning algorithm architecture and overall workflow built into the intelligent prediction module are as follows: Figure 10 As shown, a joint operation of the object detection model and the time series prediction model is employed. The time series prediction model is a long short-term memory network enhanced by a time attention mechanism, and its core cascaded architecture and transmission mechanism are as follows: Figure 11 As shown, the network comprises two vertically stacked Long Short-Term Memory (LSTM) network units, each containing 64 hidden layer neurons. A dropout layer with a ratio of 0.2 is positioned between the two LSM network units to suppress overfitting. The network's input integrates a time-attention mechanism module, which normalizes the input sequence measured at 144 time points within 0-24 hours, calculates and assigns nonlinear time-step weights, adaptively focusing on early transient dynamic feature mutation points that provide the most predictive information for the 24-hour endpoint. The output of the time-attention mechanism module is mapped through a fully connected layer to directly obtain the predicted 24-hour final inhibition zone diameter for this specific drug sensitivity pathway. Based on this prediction regression mechanism, the correlation scatter plot between the diameter of the very early direct reading (without algorithm prediction) and the 24-hour physical final diameter is shown below. Figure 12 As shown, the correlation scatter plot between the equivalent diameter output by the time-series prediction model of this invention and the final physical diameter over 24 hours is as follows: Figure 13 As shown. Meanwhile, a comparison of the mean absolute error, root mean square error, and coefficient of determination of the time-series prediction model under different incubation time windows is presented. Figure 14 As shown, the histogram of the error distribution between the predicted diameter and the actual endpoint diameter over 24 hours is as follows. Figure 15 As shown. To visually assess the accuracy of qualitative classification, the confusion matrix for determining the sensitive / intermediate / resistant (S / I / R) status using direct readings in the early immature stage is as follows. Figure 16As shown, the confusion matrix for determining the sensitive / intermediate / resistant (S / I / R) status after prediction by the model of this invention is as follows: Figure 17 As shown.

[0073] Furthermore, the intelligent prediction module also has built-in features such as... Figure 18 The user-oriented adaptive learning workflow mechanism framework is shown. This workflow allows local laboratory users to import minimum-latency image datasets of specific pathogens and antibiotic combinations that were not involved in the global retraining of the original model via a graphical user interface. Utilizing a modular fine-tuning mechanism, incremental retraining and correction of the temporal evolution trajectory weights of the time-series predictor are performed using a very small number of samples, without requiring architecture reconstruction, global retraining, or large-scale manual annotation of the entire deep learning network. This reconstructs and corrects the inhibition zone kinetics evolution trajectory of novel strains or drug combinations in the early, immature incubation stage. The adaptively fine-tuned temporal kinetics expansion trajectory curve of the drug sensitivity of novel, unknown clinical strains is shown in the figure. Figure 19 As shown in the figure. This mechanism eliminates early prediction bias caused by low image contrast and visual blur, making the fine-tuned predicted diameter output closely match the equivalent diameter of the actual endpoint at 24 hours with the ideal correlation baseline. The scatter plot comparing the early reading bias of the novel unknown strain before and after adaptive fine-tuning with the ideal correlation baseline is shown in the figure. Figure 20 As shown, the system has achieved generalized calibration and adaptation for local multi-source clinical isolation strains and novel antibiotic combination susceptibility phenotypes.

[0074] Furthermore, the graphical user interface adopts a three-layer software architecture consisting of a view layer, a business layer, and a model and data layer. Its end-to-end closed-loop analysis algorithm pipeline for the graphical user interface is as follows: Figure 21 As shown, the software is divided into two collaborative monitoring interfaces: the graphical user interface for real-time image monitoring and target localization is illustrated in the diagram below. Figure 22 As shown, the system is used to dynamically visualize time-lapse digital images, overlaying bounding boxes with antibiotic category labels and confidence scores generated by the target detection model onto the images in real time. It displays a real-time status indicator including total incubation time and earliest detection time, and the absolute coordinates, calibration diameter, inference time, and current S / I / R qualitative drug sensitivity status of each paper piece are updated in a scrolling detection log table at the bottom of the interface. The graphical user interface for the prediction system's time-series growth trajectory analysis and endpoint report generation is illustrated below. Figure 23 As shown, the interface features a prediction cutoff time control axis, which displays a 24-hour prediction result summary table to compare the predicted diameter, the actual final diameter, and the corresponding drug sensitivity qualitative results. Simultaneously, the interface plots the prediction effect curve and the actual prediction curve on a central coordinate system, showcasing the dynamic evolution and overlap between the time-series prediction trajectory and the actual trajectory. This interface supports batch export of test results in standard CSV format, enabling seamless data interconnection with the laboratory information management system.

[0075] Furthermore, the phenotypic quantitative interpretation step enables early and rapid typing and identification of different specific drug resistance phenotypes, including: single-disk reverse interpretation typing, which, for specific Gram-positive or Gram-negative bacterial samples, directly distinguishes and identifies the corresponding specific drug resistance mechanism phenotype at an early stage by real-time monitoring and prediction of the temporal evolution characteristics of the inhibition zone of a single specific antimicrobial drug disk; multi-disk difference collaborative typing, which, for specific Gram-negative bacterial samples, directly confirms the specific broad-spectrum enzyme inhibitor resistance phenotype at an early stage by recording or predicting the real-time or predicted inhibition zone diameter difference between at least two antimicrobial drug disks through combined disk evolution mechanisms; and temporal kinetic characteristic typing, which, for Gram-positive or Gram-negative clinical isolation strains with specific drug resistance evolution trajectories, directly identifies the corresponding super-drug-resistant phenotype by extracting the inhibition zone expansion rate and specific growth trajectory characteristics in the early kinetic stage.

[0076] In one specific implementation, as described above Figure 8 As shown, the machine vision early weak boundary pre-identification and annotation assistance unit operates according to the following process: (1) Read the flat plate images at each time point, perform large-scale Gaussian background estimation on the grayscale image and perform flat field illumination correction, and then use contrast-limited adaptive histogram equalization to enhance local contrast. (2) Calculate the local standard deviation map to characterize the uniformity of the bacterial moss texture, calculate the Sobel gradient magnitude map to characterize the edge intensity, and obtain the dark residual map by subtracting the large-scale background map from the enhanced grayscale map. (3) The dark residual normalization result, the reverse local standard deviation result and the reverse gradient result are fused into a two-dimensional antibacterial zone response map according to the weights, and a conservative initial boundary is obtained by multi-threshold connected domain screening in the pre-set annular area outside the paper.

[0077] This process is set up before the object detection model. Its output is not a replacement for the final deep learning detection result of the object detection model, but rather serves as a preliminary step for early weak boundary labeling assistance, training label generation, and diameter sequence extraction.

[0078] Subsequently, the system maps the enhanced grayscale image, local standard deviation image, gradient magnitude image, and dark residual image to polar coordinate space, using the center of the antibacterial drug paper as the pole. It then performs truncated mean calculation and smoothing on the angular sampled values ​​at each radius, forming radial grayscale, radial texture, radial gradient, and radial dark ring curves. The system further calculates the texture increase, grayscale change, and dark ring change, and robustly normalizes and combines them with the radial gradient peak to obtain a one-dimensional radial compensation curve. Within the restricted search band outside the two-dimensional initial radius, the system selects the global maximum peak as a candidate position for weak boundary compensation, and determines the final compensation diameter through the maximum compensation pixel threshold, compensation diameter ratio threshold, and petri dish edge constraints. The final compensation radius and its corresponding contour can be converted into bounding boxes, circular contours, masks, or diameter annotation files required by the target detection model, and correspond to the strain-antibiotic-time image file to construct an early inhibition zone weak boundary training dataset.

[0079] To verify the technical effectiveness of the present invention, the following experiments were conducted: Experiment Example 1: Performance Evaluation of Optimized Target Detection Model in Early Low-Contrast Antibacterial Zone Recognition This experiment uses a dataset of 1500 independent standard clinical pathogen-antibiotic combinations to train and evaluate the performance of an optimized YOLOv8 object detection model on an edge computing control core equipped with an advanced graphics processing unit. Input images were uniformly scaled to 640×640 pixels, the training batch size was set to 16, and the number of iterations was 300. To simulate the complex variations in clinical flat-panel illumination, data augmentation strategies such as multi-angle rotation (±45°), affine transformation, brightness / contrast perturbation, Gaussian noise, and random occlusion were introduced during the training phase. The test results are as described above. Figure 9 As shown, during the early, immature incubation stage (5-9 hours), the bacterial flora is not yet fully formed, and the agar surface exhibits significant low contrast and blurred edge noise. The optimized YOLOv8 detection model of this invention, through a feature pyramid architecture and a hybrid dual attention mechanism, significantly enhances the extraction capability of small target paper boundaries and extremely weak evolutionary inhibition zone edge morphological features. Compared to traditional edge detection algorithms and unoptimized conventional target detection models, the bounding box regression loss of this invention's model decreases more smoothly, demonstrating superior performance in convergence speed and stability in capturing weak contours. In the full-index evaluation, the model strictly controls the absolute error in recognizing the diameter of the very early low-contrast inhibition zone within ±1 mm, significantly reducing measurement variability caused by visual interpretation or conventional algorithms. The output detection trajectory exhibits high smoothness and strong discriminative power, providing a high-quality data foundation for subsequent time-series prediction modules.

[0080] Experiment Example 2: Validation of the weak boundary detection branch of machine vision in early low-contrast inhibition zones with annotation assistance.

[0081] This experimental example uses early antimicrobial susceptibility plate images at consecutive time points as input, with image files named in chronological order of incubation time. The system first positions the petri dish and antimicrobial drug discs, and automatically assigns corresponding drug roles such as FOX, LZD, and LEV based on the relative spatial positions of the discs. Then, it calls the aforementioned... Figure 8 The annotation-assisted process shown performs flat-field illumination correction, contrast-limited adaptive histogram equalization enhancement, local standard deviation calculation, gradient magnitude calculation, and dark residual extraction on each frame of the image to construct a two-dimensional antibacterial zone response map. Within the annular region outside the paper, the system uses multiple response percentile thresholds to filter connected components, obtaining a conservative but reliable initial antibacterial region. For potential weak boundaries outside this initial region, the system performs radial sampling with the paper center as the pole, integrating texture uplift, gradient peak, grayscale changes, and dark ring changes to form a one-dimensional radial compensation curve, and selects the global maximum peak as the boundary compensation point within the limited outer compensation range. These results serve as early weak boundary pre-identification results before training the target detection model, used for semi-automatic generation of antibacterial zone bounding boxes and circular mask annotation files. Experimental results show that this unit can first obtain stable inner candidate boundaries in the early two-dimensional feature map, and then appropriately expand the smaller boundaries outward through one-dimensional radial global peak compensation. For cases where strong peaks are obvious near the LZD paper and weak outer boundaries are easily missed, the drug role constraint can skip the strong inner peaks and enhance the detection of weak outer boundaries. After median filtering and monotonicity constraints, the abrupt changes and noise in the inhibition zone diameter sequence were significantly reduced in the continuous time-point results. This stable sequence was successfully applied to the pre-labeling and automated verification of large-scale training data for target detection models, laying a robust dataset foundation for subsequent deep learning networks to efficiently learn very early weak boundary features.

[0082] Experiment Example 3: Validation of Endpoint Prediction and Correction of Time Series Forecasting Model in Early Time Windows This experiment verifies the ability of a long short-term memory network enhanced with a time attention mechanism to retrospectively extrapolate the early drug sensitivity growth kinetics trajectory and predict the endpoint. The input data consisted of a time-series sequence of inhibition zone diameters recorded every 10 minutes (144 time points from 0 to 24 hours), extracted by the detection module, and processed using max-min normalization. The network employed a two-layer stacked long short-term memory unit structure, including a time attention mechanism layer, and used mean squared error as the loss function, iteratively trained 200 times using an elastic gradient optimizer. The experimental evaluation results are as described above. Figures 10 to 17 As shown. Figure 10 , Figure 12 and Figure 14As shown, if traditional manual measurement or direct reading methods are used to interpret the drug susceptibility of immature inhibition zones in the early stage (5-9 hours), the bacteria are in the early logarithmic growth phase and the drug molecules have not yet reached diffusion equilibrium, leading to extremely serious judgment bias. This results in as many as 148 major errors of misreporting resistant bacteria as susceptible bacteria and 297 major errors of misreporting susceptible bacteria as resistant bacteria. The scatter distribution shows severe dispersion and weak correlation. However, as... Figure 11 , Figure 13 and Figure 15 As shown, the time series prediction model of this invention, through a time attention mechanism, dynamically updates the hidden state and assigns high weights to key transient feature mutation points within the early 5-9 hours, successfully reconstructing a complete 24-hour inhibition zone growth curve with high morphological fidelity and temporal extensibility. Scatter regression analysis and error distribution histograms show that within the early time window of 5-9 hours, the predicted output equivalent diameter of this invention exhibits a perfect linear correlation with the final true value of the 24-hour physical test, with the coefficient of determination being comprehensively dominant, and the prediction error tightly constrained within ±1 mm. The final classification accuracy reached 100%, completely eliminating extremely significant and significant errors. This indicates that the characteristic information contained in the early kinetic signals has been fully extracted by the generative time series model, significantly advancing the output time of quantitative phenotypic drug sensitivity reports.

[0083] Experiment Example 4: Extended Validation of User-Oriented Adaptive Learning Workflow for Multi-Source, Multi-Phenotype Superdrug-Resistant Bacteria This experimental example aims to verify the system's generalization performance and adaptive fine-tuning mechanism when faced with novel pathogen-drug combinations (unseen combinations) outside the original core training set.

[0084] The experiment introduced six clinically significant super-drug-resistant strains, including *Klebsiella pneumoniae* (standard strain and extended-spectrum beta-lactamase-producing strain), *Pseudomonas aeruginosa* (standard strain and ceftazidime-resistant strain), and *Enterococcus* (standard strain and vancomycin-resistant strain). Each strain was matched with disks containing three commonly used clinical antibiotics (e.g., ceftazidime, imipenem, amoxicillin / clavulanic acid, vancomycin, linezolid, etc.), constructing an independent validation set of 1080 independent data instances. During the experiment, local laboratory users imported a small number of early time-lapse digital images of these new strains via the graphical user interface. The system triggered the aforementioned... Figure 16 The modular fine-tuning mechanism shown performs incremental retraining adjustments to the weights of the time series predictor using only a very small number of samples, without changing the underlying network architecture or performing global retraining. The verification results are as described above. Figure 17 and Figure 18As shown. Before implementing adaptive learning, due to biological variations in the optical opacity, background contrast, and specific expansion rate of the bacterial flora on agar plates, the system's early predicted trajectories exhibited certain reading deviations. However, after activating the adaptive learning workflow, the system successfully corrected the trajectory deviations in the early, immature stage through lightweight local weight fine-tuning. The determination coefficients between the fine-tuned prediction results and the actual measurements at the 24-hour endpoint all exceeded 0.96, and the mean absolute percentage error remained strictly at an extremely low level of 0.51%~0.70%. Meanwhile, as... Figure 17 The dynamic tracking shown demonstrates that the system automatically activates specific availability activation threshold times based on the tissue-dependent dynamics of different microorganisms: reliable S / I / R susceptibility reports are generated for standard Klebsiella pneumoniae at 530 minutes and for extended-spectrum beta-lactamase-producing Klebsiella pneumoniae at 340 minutes; accurate predictions are achieved for standard Pseudomonas aeruginosa and ceftazidime-resistant strains at 500 minutes and 350 minutes, respectively; and reliable reporting endpoints are locked for Enterococci and vancomycin-resistant Enterococci at 420 minutes and 470 minutes, respectively. The prediction errors exhibit a compact clustered distribution. This embodiment confirms that the present invention possesses excellent cross-laboratory generalization capabilities, effectively addressing the generalization limitations of conventional diagnostic artificial intelligence systems.

[0085] Experimental Example 5: Fully Automated Closed-Loop Antimicrobial Susceptibility Testing of Clinically Isolated Strains and In Vivo Mouse Skin Infection Model Samples This experimental example comprehensively verifies the closed-loop synergy of hardware and software in the ultra-rapid phenotypic drug susceptibility testing equipment and its practical translational medical application value by simulating real-world clinical medical scenarios and biologically complex samples. On one hand, the experiment prospectively introduced 155 clinical isolates from real patients' blood, urine, sputum, and wound exudates (including 60 Staphylococcus aureus and 95 Escherichia coli isolates, covering a large number of extended-spectrum beta-lactamase-producing and methicillin-resistant high-risk phenotypes). On the other hand, a full-thickness skin wound infection model was constructed in mice (using 6-8 week old female infected mice, a 1.2 cm wound, and an inoculation dose of 10,000,000 colony-forming units per milliliter). Wound exudate samples with high biological complexity and noise from blood and bodily secretions were collected 48 hours after infection. The experimental procedure is as follows: Figures 1 to 3The ultra-fast phenotypic drug susceptibility testing equipment shown features fully automated closed-loop operation. The inoculated agar plates are inserted into the turntable 31 driven by the rotary motor 20. The outer door panel 13 and inner door panel 14 are opened, and the temperature and humidity closed-loop control unit is activated. Fan 34 drives air convection through the heating tube 32, working in conjunction with the humidification module to precisely maintain the environment enclosed by the inner cavity framework and inner panel 4 at a constant temperature and humidity of 35℃ and 60%~70% relative humidity for microbial incubation. The main control program drives the rotary motor 20 to move the turntable 31 through multi-channel parallel displacement switching. An industrial-grade digital camera, combined with a ring-shaped uniform light source 30, performs full-field automatic focusing and dynamic image acquisition on each working plate every 10 minutes. The edge computing control core performs online background normalization preprocessing on the time-series digital images and calls the tomographically optimized target detection model and time series prediction model. During this process, the software... Figure 22 and Figure 23 The graphical user interface monitoring and analysis replay shown presents users with real-time bounding boxes of drug sensitivity test pieces with confidence labels, continuous feature tracking tables, and longitudinal trajectory prediction alignment curves.

[0086] Experimental data show that the system exhibits extremely high fidelity in quantitative prediction for both multi-source clinical patient isolates and mouse wound exudate samples with complex matrix background noise. Through a single-disc reverse interpretation typing mechanism and a multi-disc difference collaborative typing mechanism, the system successfully locks in the 24-hour endpoint drug susceptibility diameter within an early time window of 300 to 480 minutes (approximately 5-9 hours) after the start of incubation, enabling the output of ultra-rapid quantitative drug susceptibility reports on the same day, significantly shortening the detection cycle of the traditional disk diffusion method by 15 to 19 hours. Within 5-9 hours, the system achieves 100% accuracy in typing and identifying methicillin-sensitive and resistant phenotypes, standard Escherichia coli and extended-spectrum beta-lactamase-producing phenotypes, and ceftazidime-resistant and vancomycin-resistant phenotypes. After a single drug susceptibility test, the edge computing control core automatically triggers the short-wave ultraviolet disinfection lamps on the top of the chamber for in-situ ultraviolet sterilization. This embodiment confirms that the present invention achieves ultra-fast, high-throughput, standardized, and highly biosafety-based precise diagnosis of clinical drug sensitivity phenotypes through integrated intelligent collaboration of software and hardware, without changing the classic physical format and ultra-low consumable costs.

[0087] Some steps in the embodiments of the present invention can be implemented using software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk.

[0088] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for testing antimicrobial susceptibility, applied to the parallel incubation process of MH agar plates inoculated with bacterial suspension and containing antimicrobial drug discs, characterized in that, Includes the following steps: Step 1: During the incubation period, time-lapse digital images of the MH agar plates are collected periodically and continuously; Step 2: The time-lapse digital image is input into the intelligent prediction module for processing. The intelligent prediction module has a built-in deep learning joint algorithm architecture, which includes an interconnected target detection model and a time series prediction model. The target detection model performs dynamic feature extraction on the time-lapse digital image, identifies and quantifies the evolving inhibition zone outline, and outputs a time series measurement sequence of the inhibition zone diameter. The time series prediction model receives the time series measurement sequence of the inhibition zone diameter, infers the growth dynamics trend through time series regression, reconstructs the complete incubation cycle growth curve, and calculates and outputs the corresponding endpoint equivalent inhibition zone diameter. Step 3: Compare the diameter of the equivalent inhibition zone at the endpoint with the preset clinical sensitivity standard, and output an antimicrobial susceptibility test report.

2. The method for testing bacterial drug susceptibility according to claim 1, characterized in that, The target detection model is an optimized target detection model based on the YOLO architecture, including: a heuristic backbone network, a hybrid attention mechanism module, and a geometric factor simplified detection head; The heuristic backbone network replaces the C2f module in the native YOLO architecture with a simplified convolutional block and a hierarchical channel reduction strategy, and constructs a feature pyramid by combining shallow residual connections. The hybrid attention mechanism module includes a cascaded adaptive channel attention module and a spatial attention module; The geometric factor simplified detection head uses a multi-point distance intersection-union loss function.

3. The method for testing bacterial drug susceptibility according to claim 1, characterized in that, The time series prediction model is a long short-term memory network enhanced with a time attention mechanism, including: Long short-term memory network units are stacked vertically, with discard layers configured between them; A time attention mechanism module integrated into the input end is used to calculate and assign nonlinear time step weights to the input sequence in order to focus on the abrupt change points of dynamic features; A mapping layer is used to map the output of the time attention mechanism module to the predicted value of the diameter of the equivalent inhibition zone at the endpoint.

4. The method for testing bacterial drug susceptibility according to claim 1, characterized in that, Before the target detection model processes the data, the intelligent prediction module also performs a machine vision early weak boundary pre-identification and annotation assistance step, including: Flat field illumination correction and local contrast enhancement are performed on the input time-delayed digital image; Dark ring response, local texture response, gradient response and grayscale change response are extracted and fused to generate a two-dimensional antibacterial zone response map; Based on the two-dimensional inhibition zone response map, a conservative initial inhibition zone radius is obtained, and a one-dimensional radial compensation curve is constructed by polar coordinate radial sampling to determine the boundary compensation position. The compensated inhibition zone boundary labeling information is then output to assist in the training or inference of the target detection model.

5. The method for testing bacterial drug susceptibility according to claim 1, characterized in that, The intelligent prediction module also includes a user-oriented adaptive learning workflow, comprising: Receive time-lapse image datasets of specific pathogen-antibiotic combinations imported by the user; By utilizing a modular fine-tuning mechanism, the weights of the time series prediction model are incrementally modified while maintaining the overall structure of the deep learning joint algorithm architecture, in order to adapt to the dynamic evolution trajectory of the local strain.

6. The method for testing bacterial drug susceptibility according to claim 1, characterized in that, The steps for generating antimicrobial susceptibility test reports also include early typing identification for different specific resistance phenotypes, including at least one of the following: Single-disk reverse interpretation and genotyping: Identifying specific drug resistance mechanism phenotypes by monitoring the temporal evolution characteristics of the inhibition zone of a single antimicrobial drug disk; Multi-disk difference co-typing: Identifying broad-spectrum enzyme inhibitor resistance phenotypes by recording or predicting the difference in inhibition zone diameter between at least two antimicrobial drug disks; Temporal dynamics phenotyping: Identifying super-resistant phenotypes by extracting inhibition zone expansion rate and specific growth trajectory features.

7. An antimicrobial susceptibility testing device, used to implement the method as described in any one of claims 1 to 6, characterized in that, include: Miniature box-type basic structure; A multi-channel parallel incubation system is set within the micro-box basic structure, including a temperature and humidity control unit and a motor-driven turntable, on which multiple slots for mounting agar plates are arranged. A fully automated image acquisition system, suspended and fixed directly above the turntable, includes an industrial-grade digital camera and a light source; An edge computing control core is integrated into the micro-box basic structure, and the edge computing control core integrates an intelligent prediction module. The intelligent prediction module has a built-in deep learning joint algorithm architecture, which includes an interconnected target detection model and a time series prediction model. The target detection model is configured to process the time-series images input from the fully automated image acquisition system, identify the boundaries of the antibacterial drug paper discs, quantify the radius of the evolving inhibition zone, and output a time-series measurement sequence of the inhibition zone diameter. The time-series prediction model is configured to receive the time-series measurement sequence of the inhibition zone diameter, deduce the growth dynamics trend through time-series regression, reconstruct the complete incubation cycle growth curve, and calculate and output the corresponding endpoint equivalent inhibition zone diameter. The graphical user interface adopts a three-tier software architecture, including a view layer, a business layer, and a model and data layer, and is divided into two collaborative monitoring system interfaces: a detection system interface and a prediction system interface.

8. The antimicrobial susceptibility testing equipment according to claim 7, characterized in that: The detection system interface is used to dynamically visualize and present time-lapse digital images, and overlay bounding boxes, antibiotic category labels, confidence scores, and real-time qualitative drug sensitivity status generated by the target detection model. The prediction system interface is used to display a summary table of prediction results and simultaneously plot the prediction effect curve and the actual prediction curve to show the dynamic evolution overlap between the time-series prediction trajectory and the actual trajectory.

9. The antimicrobial susceptibility testing equipment according to claim 7, characterized in that, The fully automatic image acquisition system also includes a multi-dimensional camera adjustment mechanism, including a fixed bracket, a slider, an adjustment bracket, and a lens mount. The industrial-grade digital camera is suspended and fixed in the lens mount, and the light source is a ring-shaped shadowless light source that provides uniform diffuse illumination.

10. The antimicrobial susceptibility testing equipment according to claim 7, characterized in that, The temperature and humidity control unit includes a fan, heating element, humidification module and thermocouple installed in the inner cavity. The edge computing control core maintains the constant temperature and humidity environment in the inner cavity of the box through proportional-integral-derivative closed-loop feedback regulation.

Images