Bacterial species identification device, bacterial species identification method, and bacterial species identification program
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2026-02-03
- Publication Date
- 2026-08-13
Smart Images
Figure JP2026003783_13082026_PF_FP_ABST
Abstract
Description
Bacterial species discrimination apparatus, bacterial species discrimination method, and bacterial species discrimination program
[0001] The present invention relates to a bacterial species discrimination apparatus, a bacterial species discrimination method, and a bacterial species discrimination program.
[0002] In order to take appropriate measures against infectious diseases, it is necessary to quickly identify the causative bacteria. Currently, as a method for quickly obtaining information on the causative bacteria of infectious diseases, it is generally practiced to observe a cultured specimen subjected to Gram staining treatment with an optical microscope using human eyes.
[0003] On the other hand, among the rapidly developing artificial intelligence in recent years, image recognition using deep learning plays a pioneering role and its application to medicine is being promoted.
[0004] For example, Non-Patent Document 1 describes a technique for discriminating whether a Gram-negative bacillus, a Gram-positive coccus (staphylococcus), or a Gram-positive coccus (diplococcus or streptococcus) is present by image recognition using deep learning.
[0005] Smith KP, Kang AD, Kirby JE. Automated Interpretation of Blood Culture Gram Stains by Use of a Deep Convolutional Neural Network. J Clin Microbiol. 2018 Feb 22;56(3):e01521-17
[0006] However, even bacteria with similar appearances, such as Staphylococcus aureus and coagulase-negative staphylococci, Pseudomonas aeruginosa and Escherichia coli, Streptococcus pneumoniae and Enterococcus faecalis, may have significantly different clinical treatments. Therefore, there is a need for a technique that can accurately discriminate bacterial species even for bacteria with similar appearances.
[0007] One aspect of the present invention aims to realize a bacterial species discrimination apparatus, a bacterial species discrimination method, and a bacterial species discrimination program that can accurately discriminate bacterial species even for bacteria with similar appearances.
[0008] To solve the above problems, a bacterial species discrimination device according to one aspect of the present invention includes a discrimination means for discriminating the bacterial species of a target bacterial by inputting an image of the target bacterial to a trained model corresponding to the culture conditions of the target bacterial from among a plurality of trained models corresponding to different culture conditions, wherein each of the plurality of trained models is a trained model that has been machine-learned using an image of a bacterial cultured under the culture conditions corresponding to the trained model.
[0009] A bacterial species discrimination method according to one aspect of the present invention includes a discrimination step in which one or more computers input images of the target bacteria to a trained model corresponding to the culture conditions of the target bacteria from among a plurality of trained models corresponding to different culture conditions, thereby discriminating the bacterial species of the target bacteria, wherein each of the plurality of trained models is a trained model that has been machine-learned using images of bacteria cultured under the culture conditions corresponding to the trained model.
[0010] Each aspect of the present invention may be implemented by a computer, in which case the bacterial species identification device that implements the bacterial species identification device by a computer by operating the computer as each part (software element) of the bacterial species identification device, and the computer-readable recording medium on which the program is recorded also fall within the scope of the present invention.
[0011] According to one aspect of the present invention, even if the bacteria have similar appearances, it is possible to distinguish between different bacterial species with high accuracy.
[0012] This is a functional block diagram showing an example of the configuration of a bacterial species discrimination device according to an embodiment of the present invention. This is a flowchart showing an example of the operation of a bacterial species discrimination device according to an embodiment of the present invention. This is a flowchart showing an example of creating a trained model used by a bacterial species discrimination device according to an embodiment of the present invention. This is an explanatory diagram of six-fold cross-validation. This is a graph showing the change in the value of the loss function in classification by the object detection model when creating a trained model due to repeated training. This is a figure showing coagulase-negative staphylococci correctly detected by an object detection model created from anaerobic culture data. This is a figure showing Staphylococcus aureus correctly detected by an object detection model created from aerobic culture data. This is a figure showing Staphylococcus aureus that is misdetected as coagulase-negative staphylococci by an object detection model created from data that does not consider culture conditions. This is a graph showing the change in the recognition accuracy in classification by the image classification model when creating a trained model due to repeated training.
[0013] [Embodiment 1] An embodiment of the present invention will be described in detail below.
[0014] (Regarding bacterial species identification) Although the risk of death or severe illness from infectious diseases has been greatly reduced by the use of antimicrobial agents, appropriate measures against infectious diseases are still essential, not only in areas with poor sanitation, but also when dealing with patients with impaired immune function, and even in medical institutions where advanced treatments such as organ transplants and hematopoietic stem cell transplants are performed.
[0015] While it is possible to identify the causative bacteria using biochemical enzymatic reactions, mass spectrometry, and genomic DNA sequencing, these methods are not only costly but also time-consuming, and the worsening of symptoms during this time could lead to fatal consequences. Rapid and accurate identification of the causative bacteria and treatment based on those results directly impact the patient's prognosis.
[0016] Treatment that involves successively switching between various antibiotics without identifying the causative bacteria is not recommended from the perspective of appropriate use of antimicrobial agents, especially given the significant global concerns about drug-resistant bacteria in recent years.
[0017] As a method for rapidly obtaining information about the causative agents of infectious diseases in the early stages of illness, it is common practice to observe Gram-stained cultured specimens with a light microscope using the human eye. By observing Gram-stained specimens, similar-looking bacterial species are distinguished based on morphological differences such as whether the bacteria are visible in three dimensions, the size of the bacterial clusters, and the size of the bacteria themselves.
[0018] While bacteria observed under a light microscope exhibit diverse shapes, most can be classified as either rods or cocci, and further categorized as Gram-positive or Gram-negative based on whether they stain with Gram. In other words, they can be classified as Gram-positive rods, Gram-negative rods, Gram-positive cocci, and Gram-negative cocci. Furthermore, streptococci, which form chains, and staphylococci, which form aggregates, can be easily distinguished. For example, Gram-negative cocci can be identified using inexpensive Gram staining, such as identifying Neisseria gonorrhoeae.
[0019] However, in reality, even experts often misidentify bacterial species based on visual observation alone, as they look very similar despite having vastly different clinical treatments, such as Staphylococcus aureus and coagulase-negative staphylococci, Pseudomonas aeruginosa and Escherichia coli, and Streptococcus pneumoniae and Enterococcus.
[0020] While it is possible to identify bacteria using biochemical methods such as enzymatic reactions, mass spectrometry, and genomic DNA sequencing, these methods require significant equipment and reagent costs, specialized procedures, and considerable time to obtain results. Furthermore, the worsening of symptoms during this time could be life-threatening.
[0021] Therefore, it is considered useful to utilize deep learning-based image recognition for bacterial species identification using Gram staining, which can be obtained early in the disease's onset, as a rapid and inexpensive method of identification.
[0022] Gram staining of bacteria and fungi is inexpensive and simple, and can be performed anywhere as long as staining reagents and an optical microscope are available. Furthermore, the results can be transmitted quickly, easily, and inexpensively to any location in the world via mobile communication networks and the internet.
[0023] Deep learning is a supervised learning method that uses large amounts of annotated data to perform machine learning. Building a model through training requires enormous computational resources and takes a long time, but once the model is complete, classification can be done in just a few seconds. Furthermore, if the program is provided as a web service from the cloud, classification can be requested and the results obtained from anywhere in the world via the internet.
[0024] This embodiment aims to provide information for determining treatment strategies for patients with bacteremia and sepsis quickly and inexpensively, to facilitate early recovery from infection through the selection of appropriate antimicrobial agents, and to prevent the emergence of drug-resistant bacteria, thereby contributing to future medical care.
[0025] (Bacterial Species Identification Device) The bacterial species identification device according to this embodiment utilizes multiple trained models corresponding to different culture conditions. Each of the multiple trained models is a trained model that has been machine-learned using images of bacteria cultured under the culture conditions corresponding to that trained model. The identification means of the bacterial species identification device identifies the bacterial species of the target bacteria by inputting the images of the target bacteria into the trained model corresponding to the culture conditions of the target bacteria from among the multiple trained models.
[0026] From another perspective, the bacterial species identification method according to this embodiment utilizes multiple trained models corresponding to different culture conditions. Each of the multiple trained models is a trained model that has been machine-learned using images of bacteria cultured under the culture conditions corresponding to that trained model. One or more computers then input the images of the target bacteria into the trained model corresponding to the culture conditions of the target bacteria, thereby identifying the bacterial species of the target bacteria.
[0027] The following describes a configuration in which multiple pre-trained models include a pre-trained model that corresponds to anaerobic culture and a pre-trained model that corresponds to aerobic culture. In the following configuration, the target bacteria are cultured using anaerobic and aerobic methods, respectively, and the discrimination means identifies the species of the target bacteria by inputting images of the anaerobically cultured target bacteria into the pre-trained model that corresponds to anaerobic culture, and by inputting images of the aerobically cultured target bacteria into the pre-trained model that corresponds to aerobic culture.
[0028] However, this embodiment is not limited thereto, and any pre-trained models that correspond to multiple culture conditions with varying culture conditions, such as the composition of the culture medium, the pH of the culture medium, the drugs added to the culture medium, the culture temperature, the culture time, the pressure of the culture environment, the lighting during culture, the composition of the gas in the culture environment (partial pressure of each component), and the shaking speed, are acceptable. Furthermore, the number of pre-trained models may be multiple, and may be three or more.
[0029] Figure 1 is a functional block diagram showing an example configuration of the bacterial species identification device 100 according to this embodiment. As shown in Figure 1, the bacterial species identification device 100 includes an input unit 10, an output unit 20, and a discrimination unit (discrimination means) 30.
[0030] (Input Unit) The input unit 10 receives input of images of the target bacteria to be identified. The images of the target bacteria are distinguished according to the culture conditions of the target bacteria. In the example shown in Figure 1, images of the target bacteria cultured anaerobically (anaerobic culture data) and images of the target bacteria cultured aerobically (aerobic culture data) are distinguished and input.
[0031] The bacteria to be identified are those derived from the subject, for example, bacteria derived from the subject's blood. The bacteria to be identified may be cultured in a liquid medium or in a solid medium. If the bacteria to be identified are derived from the subject's blood, it is preferable to culture the blood using a liquid medium.
[0032] The target bacteria for discrimination may consist only of those cultured under a single culture condition, or they may consist of those cultured under multiple conditions. In the example shown in Figure 1, only those cultured under either anaerobic or aerobic conditions may be used, or both the anaerobic and aerobic cultured target bacteria for discrimination may be used.
[0033] Anaerobic culture data and aerobic culture data may be images captured by a camera installed on a slide scanner or microscope, or images captured by a tablet or smartphone camera of an image displayed on the slide scanner or microscope. The data format of the captured images is not particularly limited and may be TIFF, JPEG, HEIC, etc.
[0034] (Output Unit) The output unit 20 outputs the determination result of the determination unit 30. The output unit 20 may be, for example, a display device such as a display or touch panel, or it may be a communication unit that transmits data to a terminal used by the user via a network.
[0035] (Discrimination Unit) The discrimination unit 30 includes an object detection unit 31, an image extraction unit 32, an image classification unit 33, a trained model 40, and a trained model 50. The trained model 40 is a trained model that corresponds to anaerobic culture. The trained model 50 is a trained model that corresponds to aerobic culture.
[0036] Each trained model consists of an object detection model that detects the location and species of bacteria in the captured image, and an image classification model that classifies the objects located at the locations detected by the object detection model. Trained model 40 consists of an object detection model 41 and an image classification model 42. Trained model 50 consists of an object detection model 51 and an image classification model 52.
[0037] While not limited to these, object detection models that utilize deep learning, such as YOLO (You Only Look Once), R-CNN (Region-Convolution Neural Network), and SSD (Single Shot Multibox Detector), can be used.
[0038] While not limited to these, deep learning image classification models such as CNN (Convolutional Neural Network), VisionTransformer, and ResNet (Residual Neural Networks) can be used as image classification models.
[0039] In the example shown in Figure 1, the discrimination unit 30 has (stores) each learned model, but this embodiment is not limited to this, and each learned model may be stored in another device that can communicate with the fungal species discrimination device 100.
[0040] Figure 2 is a flowchart showing an example of the operation of the bacterial species identification device 100.
[0041] In step S1 (input processing), the input unit 10 receives input of images of the target bacteria for discrimination. The input unit 10 may receive only anaerobic culture data and aerobic culture data, or it may receive both anaerobic culture data and aerobic culture data of the same target bacteria. The anaerobic culture data and aerobic culture data may each contain multiple images.
[0042] In step S2 (object detection processing), the object detection unit 31 detects whether each candidate bacterial species is present in each captured image input to the input unit 10. Candidate bacterial species refer to candidate bacterial species that can be identified by the bacterial species identification device 100. Candidate bacterial species may be limited to, for example, bacterial species that are similar in appearance. The number of candidate bacterial species is not particularly limited, but the following description will explain the case where the candidate bacterial species are bacterial species X and bacterial species Y.
[0043] The object detection unit 31 inputs the captured image of the target bacteria input to the input unit 10 into an object detection model corresponding to the culture conditions of the target bacteria. That is, the object detection unit 31 inputs anaerobic culture data into the object detection model 41 and aerobic culture data into the object detection model 51. The object detection unit 31 obtains the detected bacterial species, its position in the captured image, and the probability from the object detection model.
[0044] In step S3 (image extraction process), the image extraction unit 32 calculates the center coordinates of the target bacteria shown in the captured image based on the position detected by the object detection unit 31, and cuts out an image fragment of a fixed number of pixels on all four sides as a crop. The number of pixels may correspond to a specific length such as 10 μm, for example.
[0045] In step S4 (image classification process), the image classification unit 33 inputs the crop cut out by the image extraction unit 32 into an image classification model corresponding to the culture conditions of the target bacteria. That is, the image classification unit 33 inputs the crop cut out from the anaerobic culture data into the image classification model 42 and the crop cut out from the aerobic culture data into the image classification model 52. The image classification unit 33 obtains the classified bacterial species, that is, whether it is either bacterium X or bacterium Y, and the probability from the image classification model.
[0046] In step S5 (discrimination process), the discrimination unit 30 discriminates the bacterial species. In one example, the discrimination unit 30 discriminates the bacterial species of the target bacteria based on the bacterial species and probability output by the object detection model and / or the bacterial species and probability output by the image classification model.
[0047] In step S6 (output process), the output unit 20 outputs the discrimination result of the bacterial species discriminated by the discrimination unit 30.
[0048] Here, bacteria and fungi exhibit slight morphological differences depending on whether they are cultured anaerobically or aerobically. While these morphological differences are generally difficult to distinguish, by learning these differences using deep learning, it is possible to obtain higher recognition accuracy than conventional models when using a pre-trained model for anaerobic culture data, and when using a pre-trained model for aerobic culture data.
[0049] For example, the technology described in Non-Patent Literature 1 performs learning without considering anaerobic and aerobic culture data. Furthermore, the technology described in Non-Patent Literature 1 only distinguishes between broad categories such as Gram-negative bacilli, Gram-positive cocci (Staphylococcus), and Gram-positive cocci. In contrast, according to this embodiment, even bacteria that look similar can be distinguished.
[0050] While this embodiment can distinguish between bacteria with similar appearances, it can naturally also be used to distinguish between bacteria with dissimilar appearances. When used to distinguish between bacteria with dissimilar appearances, it can achieve even higher accuracy.
[0051] As described above, this embodiment makes it possible to identify the type of bacteria or fungi causing bacteremia or sepsis without the need for experts or skilled personnel. It enables the identification of similar bacteria that are difficult even for experts to distinguish, providing information for determining the initial treatment plan for patients requiring urgent care. Furthermore, by selecting an appropriate antimicrobial agent, the emergence of drug-resistant bacteria can be suppressed, and the effectiveness of finite antimicrobial agents can be maintained for the future.
[0052] Furthermore, in the example shown in Figure 1, the bacterial species discrimination device 100 as a whole utilizes four models: object detection models 41 and 51 and image classification models 42 and 52. Based on the bacterial species and probability output from each model, more accurate discrimination can be performed. Specifically, the discrimination unit 30 may input anaerobic culture data into the trained model 40 to determine the bacterial species of the target bacteria based on the first bacterial species and probability of the first bacterial species detected by the object detection model 41, the second bacterial species and probability of the second bacterial species detected by the image classification model 42, and input aerobic culture data into the trained model 50 to determine the bacterial species of the target bacteria based on the third bacterial species and probability of the third bacterial species detected by the object detection model 51, and the fourth bacterial species and probability of the fourth bacterial species detected by the image classification model 52. For example, the discrimination unit 30 may determine the bacterial species by taking into account the probabilities of the first to fourth bacterial species and performing a majority vote among the first to fourth bacterial species. Specifically, the discrimination unit 30 may compare the sum of the probabilities of detecting bacterium X with the sum of the probabilities of detecting bacterium Y and determine that the bacterium is the one with the higher probability.
[0053] As a result, the bacterial species identification device 100 can perform identification with greater accuracy.
[0054] Furthermore, in the example shown in Figure 1, multiple anaerobic and aerobic culture data sets may be input. For example, if the recognition accuracy of each model is approximately 0.8, then incorrect judgments will be made in about one out of five images. However, by preparing about 10 different images for both anaerobic and aerobic culture data, it becomes possible to accurately distinguish between similar bacteria. For example, the discrimination unit 30 may determine the species of the target bacteria based on multiple discrimination results obtained by inputting multiple images into each trained model.
[0055] (Creation of a pre-trained model) Figure 3 is a flowchart showing an example of creating a pre-trained model used by the bacterial species identification device 100.
[0056] To create each trained model, first, training data for supervised learning is created. In step S11 (culture process), blood samples are collected from patients with bacteremia or sepsis, and anaerobic and aerobic cultures are performed in parallel. Since the optimal culture conditions for an unknown causative bacterium are unknown, culture is attempted under both conditions, and culture samples that show bacterial or fungal growth are evaluated by Gram staining at that point, and then subculture is performed to use the results for subsequent identification of the bacterial species. On the other hand, in this embodiment, it is preferable to use data obtained from both culture conditions. This is because even if the external appearance of the bacterial cells is similar, if they show different morphologies due to differences in culture conditions, this can serve as a basis for identification.
[0057] Since the subjects are similar-looking bacteria, we will assume here that we are distinguishing between bacterial species X and bacterial species Y.
[0058] In step S12 (annotation), the type of bacterium or fungus is determined (annotated) by biochemical analysis, mass spectrometry, or genomic DNA sequencing. This result becomes the correct label in deep learning. From this point onward, only bacterial species X and Y are targeted, and other bacterial species are not considered.
[0059] In step S13 (imaging process), Gram staining is performed for each culture condition, and photographs are taken with an optical microscope. The quantity of photographs is important for model construction, so multiple images are acquired while changing the location on the slide. There are situations where bacteria form clumps and appear three-dimensional, and this state is also a factor in the judgment, so multiple photographs with different focal points are acquired. When evaluating the model, if the image data of the same location is separated into training data and test data, the recognition accuracy may be calculated to be higher than it actually is, so this point should be taken into consideration.
[0060] In step S14 (position detection process), from each microscope image, a limited area representing the characteristics of the observed bacterial species X or Y is manually positioned as a rectangular bounding box, and the bacterial species is labeled. The bounding boxes are designed to capture the characteristics of each bacterial species, such as whether they exist individually, in chains, or in clumps. It is preferable to prepare several thousand to tens of thousands of bounding boxes. Furthermore, it is desirable to surround all observed bacteria with bounding boxes so that they can be clearly distinguished from other objects such as red blood cells and the background during object detection.
[0061] In step S15 (object detection model creation process), separate object detection models are created for anaerobic data from data obtained from anaerobic culture and for aerobic data from data obtained from aerobic culture. These are further divided into training data and test data, and deep learning is used to create object detection models corresponding to anaerobic culture and aerobic culture. The object detection models can output which bacterial species is present and where it is located in the photograph, along with a probability of accuracy.
[0062] In step S16 (image classification model creation process), an image classification model is created to perform a two-classification to distinguish between bacterial species X and bacterial species Y. A different neural network is used than the one used for object detection. This model also distinguishes between data obtained from anaerobic culture and aerobic culture, and creates separate image classification models for each. The center coordinates are calculated from the two-dimensional position information of the bounding boxes labeled in step S14, and image fragments of a fixed number of pixels on each side are cropped to correspond to a specific length, such as 10 μm, depending on the size and spread of the bacterial species. These are divided into training data and test data, and an image classification model is created using deep learning.
[0063] Steps S15 and S16 may be performed in parallel or individually in any order. By combining the obtained object detection model with the image classification model, a trained model corresponding to each culture condition can be created.
[0064] (Effects of this embodiment) According to this embodiment, causative bacterial species of bacteremia and sepsis can be identified with high accuracy without the need for experts or skilled personnel, and furthermore, it becomes possible to distinguish similar bacteria, which was difficult even for experts. Not only can medical expenses be reduced by avoiding biochemical methods, mass spectrometry, and base sequencing, but even if these procedures are necessary, identification results can be obtained quickly. The determination of early initial treatment strategies, such as drug selection and removal of medical devices, is extremely important for patients with bacteremia and sepsis, as it can be a matter of life and death, and the role played by this invention is significant. Regarding the selection of antimicrobial agents, it also provides one measure against the problem of drug-resistant bacteria, which is a global concern. This system can be built and provided as a web service on the cloud, and can be used even in medically underserved areas and disaster areas, as long as image data can be uploaded. Although legal issues may remain, if drug information corresponding to the bacterial species can be provided, it will be a system that enables early response to patients even when a doctor is not present.
[0065] [Example of implementation using software] The function of the fungal species identification device 100 (hereinafter referred to as "the device") is a program that causes a computer to function as the device, and can be realized by a fungal species identification program that causes a computer to function as each control block of the device (especially each part included in the identification unit 30).
[0066] In this case, the device includes a computer having at least one control device (e.g., a processor) and at least one storage device (e.g., memory) as hardware for executing the program. By executing the program using this control device and storage device, the functions described in each of the embodiments are realized.
[0067] The above program may be recorded on one or more computer-readable recording media, not temporary ones. These recording media may or may not be provided by the above device. In the latter case, the program may be supplied to the above device via any wired or wireless transmission medium.
[0068] Furthermore, some or all of the functions of each of the above control blocks can also be realized by logic circuits. For example, an integrated circuit in which logic circuits functioning as each of the above control blocks are formed is also included in the scope of the present invention. In addition, it is also possible to realize the functions of each of the above control blocks by, for example, a quantum computer.
[0069] [Examples] An embodiment of this invention will be described below with reference to examples.
[0070] In this example, coagulase-negative staphylococci (CNS) and Staphylococcus aureus (SAU) were distinguished. Both CNS and SAU are Gram-positive cocci that form clusters resembling bunches of grapes, making them difficult to distinguish visually. However, as evidenced by the conventional practice of identifying bacteria based on the presence or absence of coagulase, which causes plasma coagulation, CNS and SAU differ significantly in their clinical treatment, including pathogenicity, drug resistance, tissue invasiveness, and association with artificial materials, leading to different treatment strategies for patients with bacteremia. Therefore, techniques that can distinguish between these visually similar bacteria are extremely useful.
[0071] Blood samples were collected from numerous patients with bacteremia caused by either CNS or SAU, and anaerobic and aerobic cultures were performed. Gram staining was applied to the samples cultured under each condition, and photographs were taken using a light microscope. 4,965 and 3,858 microscopic images were obtained for the anaerobic and aerobic cultured samples, respectively. Specifically, 2,498 images were of anaerobic CNS, 2,467 images were of aerobic CNS, 2,046 images were of anaerobic SAU, and 1,812 images were of aerobic SAU. 1,800 images were randomly selected from each group, and 1,500 were used as training validation data and 300 as test data to create an object detection model.
[0072] The annotation of the photographs used for model creation was done manually using Microsoft VoTT 2.2.0. The bounding box position information and labels (CNS or SAU) were saved as a JSON file.
[0073] From the location information, a 171-pixel square image corresponding to a 10 μm square of the micrograph used in the example was extracted. 11,480 crops were obtained from the anaerobic culture data of CNS, 27,550 from the aerobic culture data of CNS, 12,473 from the anaerobic culture data of SAU, and 7,141 from the aerobic culture data of SAU. 6,000 crops were randomly selected from each, and 5,000 crops were divided into training and validation data and 1,000 crops into test data. An image classification model (two-classification) was created using six-fold cross-validation as shown in Figure 4.
[0074] Deep learning was performed using a computer equipped with a Tesla V100-SXM2-16GB and with CUDA 12.6 installed. The operating system was Linux 6.1.109, the distribution was Amazon Linux® 2023, the scripting language was Python 3.12.6, and the deep learning framework used was PyTorch 2.4.0.
[0075] The object detection model was created using YOLO v8 8.3.23 over 40 epochs. Three object detection models were created: one trained using anaerobic culture data, one trained using aerobic culture data, and, for comparison, one trained using data without considering culture conditions. For the data without considering culture conditions, 900 images each were randomly selected from both the anaerobic and aerobic culture datasets.
[0076] In evaluating object detection models, both location and classification are evaluated. However, since the ultimate goal is to identify bacterial species, we investigated the change in the value of the loss function during repeated training cycles, using 1200 training images and 400 validation images, for the evaluation of classification. Figure 5 is a graph showing the change in the value of the loss function (Loss) in the classification of the object detection model with respect to repeated training cycles (Epoch). As shown in Figure 5, within equivalent training, the loss function values were generally kept low in training using anaerobic culture data and training using aerobic culture data, indicating that the object detection model was created more efficiently than when using data that did not consider culture conditions.
[0077] The created object detection models were tested. As described earlier, 300 test data points each were prepared for anaerobic culture data of CNS, aerobic culture data of CNS, anaerobic culture data of SAU, and aerobic culture data of SAU, which were not used for training. A total of 1200 data points were input according to the culture conditions, and the number of false detections in the object detection models was measured. The results are shown in Table 1. As shown in Table 1, it was found that the models created from anaerobic culture data and the models created from aerobic culture data had higher recognition accuracy than the models created from data that did not consider the culture conditions.
[0078]
[0079] Figure 6 shows coagulase-negative staphylococci correctly detected by an object detection model created from anaerobic culture data. Although not shown in the figure, Staphylococcus aureus was also correctly detected by the object detection model created from anaerobic culture data.
[0080] Figure 7 shows Staphylococcus aureus correctly detected by an object detection model created from aerobic culture data. Although not shown in the figure, coagulase-negative staphylococci were also correctly detected by the object detection model created from aerobic culture data.
[0081] Figure 8 shows Staphylococcus aureus that was misidentified as coagulase-negative staphylococcus in an object detection model created from data that did not take culture conditions into account. As can be seen by comparing Figures 6 and 7, coagulase-negative staphylococcus and Staphylococcus aureus are similar in appearance, and object detection models created from data that did not take culture conditions into account were more prone to misidentification than object detection models created from data that did take culture conditions into account.
[0082] The image classification model was created using transfer learning with Inception V3 on ImageNet. In addition to models using anaerobic and aerobic culture data, an object detection model was created for comparison using data without considering culture conditions, by randomly selecting 3000 crops from each of the two datasets.
[0083] Image data augmentation was performed using the following PyTorch specifications: RandomHorizontalFlip(p = 0.45) RandomVerticalFlip(p = 0.45) RandomAffine(degrees = (-9, 9), scale = (0.9, 1.1)) RandomErasing(p = 0.2, scale = (0.01, 0.05), ratio = (0.8, 1.2)) RandomPerspective(p = 0.2) ColorJitter(brightness = (0.2, 1.7), contrast = 0.2, saturation = 0.1, hue = 0.1) RandomAutocontrast(p = 0.2) Stochastic gradient descent was used for parameter updates, with a learning rate of 0.03, momentum of 0.9, and weighted decay of 10. －４ The training was repeated for 20 epochs.
[0084] Of the 6000 crops, 5000 were used as training data and 400 as validation data. The change in recognition accuracy per epoch using the validation data was then examined. Figure 9 is a graph showing the change in recognition accuracy of the image classification model with each training iteration. As shown in Figure 9, the models using anaerobic culture data and aerobic culture data showed higher recognition accuracy than the models created from data that did not consider culture conditions.
[0085] Of the four sets of crops, each set containing 6000 crops, 5000 were used as training data, 400 as validation data, and 600 as test data. Six-fold cross-validation was performed, and the results of the six tests, along with the average, are summarized in Table 2. The models using anaerobic culture data and aerobic culture data showed higher recognition accuracy than the models created from data that did not consider culture conditions.
[0086]
[0087] The present invention is not limited to the embodiments described above, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present invention.
[0088] 10...Input unit 20...Output unit 30...Discrimination unit 31...Object detection unit 32...Image extraction unit 33...Image classification unit 40...Trained model 41...Object detection model 42...Image classification model 50...Trained model 51...Object detection model 52...Image classification model 100...Bacteria species discrimination device
Claims
1. A bacterial species discrimination device comprising a discrimination means for discriminating the species of a target bacterium by inputting an image of the target bacterium into a trained model corresponding to the culture conditions of the target bacterium from among a plurality of trained models corresponding to different culture conditions, wherein each of the plurality of trained models is a trained model that has been machine-learned using an image of a bacterium cultured under the culture conditions corresponding to the trained model.
2. The bacterial species discrimination device according to claim 1, wherein the plurality of trained models include a trained model corresponding to anaerobic culture and a trained model corresponding to aerobic culture.
3. The bacterial species discrimination device according to claim 2, wherein the target bacteria are cultured by anaerobic culture and aerobic culture, and the discrimination means inputs an image of the target bacteria cultured by anaerobic culture into a trained model corresponding to the anaerobic culture, and inputs an image of the target bacteria cultured by aerobic culture into the trained model corresponding to the aerobic culture to determine the bacterial species of the target bacteria.
4. The fungal species discrimination device according to claim 3, wherein each of the plurality of trained models comprises an object detection model that detects the location and species of fungi in the captured image, and an image classification model that classifies the objects located at the locations detected by the object detection model.
5. The bacterial species discrimination device according to claim 4, wherein the discrimination means inputs an image of the target bacterial species cultured anaerobically into a trained model corresponding to the anaerobic culture, and determines the bacterial species of the target bacterial species based on the likelihood of a first bacterial species detected by the object detection model of the trained model, the likelihood of a second bacterial species detected by the image classification model of the trained model, and inputs an image of the target bacterial species cultured aerobically into a trained model corresponding to the aerobic culture, the likelihood of a third bacterial species detected by the object detection model of the trained model, and the likelihood of a fourth bacterial species detected by the image classification model of the trained model.
6. The bacterial species discrimination device according to any one of claims 1 to 5, wherein the image of the target bacterial species includes a plurality of images that are different from each other, and the discrimination means determines the bacterial species of the target bacterial species based on a plurality of discrimination results obtained by inputting each of the plurality of images into the trained model.
7. The bacterial species identification device according to claim 1, wherein the target bacteria for identification are cultured in a liquid medium or a solid medium.
8. A method for determining the species of a target bacterium, comprising a determination step in which one or more computers input images of the target bacterium into a trained model corresponding to the culture conditions of the target bacterium from among a plurality of trained models corresponding to different culture conditions, wherein each of the plurality of trained models is a trained model that has been machine-learned using images of bacteria cultured under the culture conditions corresponding to the trained model.
9. A bacterial species identification program for causing a computer to function as a bacterial species identification device according to claim 1, wherein the bacterial species identification program causes the computer to function as the identification means.