Naso-laryngoscope image auxiliary diagnosis system based on deep learning

Abstract

The invention provides a nasal laryngoscope image auxiliary diagnosis system based on deep learning. A model unit creates an image recognition model; the database unit creates an organ image library containing standard images of the nasal cavity and the laryngeal cavity; the acquisition unit acquires a to-be-diagnosed nasal laryngoscope image; the image processing unit matches the to-be-diagnosed nasolaryngoscope image with the organ image library, and identifies a nasal cavity area and a laryngeal cavity area of the to-be-diagnosed nasolaryngoscope image; segmenting the to-be-diagnosed nasal laryngoscope image according to the nasal cavity region and the laryngeal cavity region to obtain segmented pictures; the image recognition unit inputs the segmented pictures into an image recognition model; acquiring a disease identification result of each segmented picture output by the image identification model; the evaluation unit receives an artificial evaluation score for the disease identification result; and determining the diagnosed disease of the to-be-diagnosed nasal laryngoscope image according to the manual evaluation score and the disease recognition result. According to the system, the difficulty of interpretation of a nasal laryngoscope image by a doctor is reduced, and the objectivity of a diagnosis result is improved.

Description

Technical Field

[0001] The present invention belongs to the field of smart medical technology, and specifically relates to a nasal laryngoscope image-assisted diagnosis system based on deep learning. Background Art

[0002] In the process of digitalization and intelligent transformation of the medical and health industry, smart medical care has become a backbone force. Video laryngoscopy is an important auxiliary tool for modern otolaryngologists. Due to the location of the nasopharynx and the deep cavity, and its relatively complex anatomy, it cannot be directly observed with the naked eye, so some special examination methods are needed to diagnose otolaryngological diseases, and laryngoscopes are commonly used. However, nasal and laryngoscope images based on laryngoscopes require high naked eye recognition capabilities of doctors, and there is uncertainty in the interpretation of nasal and laryngoscope images by different doctors, which makes the diagnosis results and treatment plans more subjective. Summary of the invention

[0003] In response to the defects in the prior art, the present invention provides a nasolaryngoscope image-assisted diagnosis system based on deep learning, which reduces the difficulty of doctors in interpreting nasolaryngoscope images and improves the objectivity of the diagnosis results.

[0004] A nasolaryngoscope image-assisted diagnosis system based on deep learning, comprising:

[0005] Model unit: used to create an image recognition model; the image recognition model is obtained by training a deep learning architecture using image samples; the image samples include different nasal and laryngoscope images and annotated disease information;

[0006] Database unit: used to create an organ image library containing standard images of the nasal cavity and laryngeal cavity;

[0007] Acquisition unit: used to obtain nasal and laryngoscope images to be diagnosed;

[0008] Image processing unit: used to match the nasal and laryngoscope image to be diagnosed with the organ image library, identify the nasal cavity area and laryngeal cavity area of ​​the nasal and laryngeal cavity image to be diagnosed; segment the nasal and laryngeal cavity image to be diagnosed according to the nasal cavity area and laryngeal cavity area to obtain a segmented image;

[0009] Image recognition unit: used to input the segmented image into the image recognition model; obtain the disease recognition result of each segmented image output by the image recognition model;

[0010] Evaluation unit: used to receive manual evaluation scores for disease identification results; determine the disease to be diagnosed that has been confirmed by the nasal and laryngoscope image based on the manual evaluation scores and disease identification results.

[0011] Furthermore, the deep learning architecture includes: a combination of a dual combined deep residual network and a transfer learning network.

[0012] Furthermore, the disease information of the image sample includes diagnosed or undiagnosed diseases.

[0013] Furthermore, the collection unit is specifically used for:

[0014] Receive nasal and laryngoscope videos to be diagnosed;

[0015] Extract key frames from the nasal laryngoscope video to be diagnosed to obtain the nasal laryngoscope image to be diagnosed.

[0016] Furthermore, the image processing unit is specifically used for:

[0017] A sliding window is set on the nasal laryngoscope image to be diagnosed to perform a sliding search, and the matching degree between the standard image in the organ image library and the nasal laryngoscope image to be diagnosed in the sliding window is calculated;

[0018] The area within the sliding window on the nasolaryngoscope image to be diagnosed with the highest matching degree is defined as the nasal cavity area or the laryngeal cavity area.

[0019] Furthermore, the image processing unit is specifically used for:

[0020] Adjust the size of the sliding window, and calculate the matching degree between the standard image and the nasal laryngoscope images to be diagnosed in sliding windows of different sizes;

[0021] Rotate the standard image and calculate the matching degree between the standard image at different angles and the nasal laryngoscope image to be diagnosed in the sliding window;

[0022] The edge information or shape descriptors of the standard image and the nasal laryngoscope image to be diagnosed in the sliding window are extracted, and the matching degree of the edge information or shape descriptors between the standard image and the nasal laryngoscope image to be diagnosed in the sliding window is calculated.

[0023] Furthermore, the disease recognition results of the segmented image include:

[0024] Disease information, confidence and evaluation value of the segmented image.

[0025] Furthermore, the evaluation unit is specifically used for:

[0026] Receive evaluation scores from multiple users for each evaluation value;

[0027] Set weights for each evaluation value;

[0028] The weights are used to perform a weighted average on all evaluation scores to obtain the manual evaluation score.

[0029] Furthermore, the model unit is also used to:

[0030] The nasal and laryngoscope images to be diagnosed, confirmed diseases, and manual evaluation scores are added to the image samples to train and optimize the image recognition model.

[0031] Furthermore, the model unit is also used to:

[0032] Reverse deduction is performed on the confirmed nasal and laryngoscope images to obtain the image features that characterize the disease;

[0033] The confirmed nasal and laryngoscope images and image features are added to the image samples to train and optimize the image recognition model.

[0034] It can be seen from the above technical scheme that the deep learning-based nasolaryngoscope image-assisted diagnosis system provided by the present invention applies the deep learning model to the medical image processing of the nasolaryngoscope, and quickly identifies the symptoms of the nasolaryngoscope image through the deep learning model, thereby realizing the early identification and warning of the disease, reducing the difficulty of doctors' interpretation of nasolaryngoscope images, and improving the objectivity of the diagnosis results. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for the specific embodiments or the description of the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual scale.

[0036] Figure 1 A module block diagram of a nasolaryngoscope image-assisted diagnosis system provided in an embodiment.

[0037] Figure 2 Schematic diagram of standard images of the laryngeal cavity under different lesions provided in the embodiment. DETAILED DESCRIPTION

[0038] The embodiments of the technical solution of the present invention will be described in detail below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and are therefore only used as examples, and the protection scope of the present invention cannot be limited thereto. It should be noted that, unless otherwise specified, the technical terms or scientific terms used in this application should be the common meanings understood by those skilled in the art to which the present invention belongs.

[0039] It should be understood that when used in this specification and the appended claims, the terms "include" and "comprises" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0040] It should also be understood that the terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0041] As used in this specification and the appended claims, the term "if" may be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrases "if it is determined" or "if [described condition or event] is detected" may be interpreted as meaning "upon determination" or "in response to determining" or "upon detection of [described condition or event]" or "in response to detecting [described condition or event]," depending on the context.

[0042] Example:

[0043] A deep learning-based nasal and laryngoscope image-assisted diagnosis system, see Figure 1 ,include:

[0044] Model unit: used to create an image recognition model; the image recognition model is obtained by training a deep learning architecture using image samples; the image samples include different nasal and laryngoscope images and annotated disease information;

[0045] Database unit: used to create an organ image library containing standard images of the nasal cavity and laryngeal cavity;

[0046] Acquisition unit: used to obtain nasal and laryngoscope images to be diagnosed;

[0047] Image processing unit: used to match the nasal and laryngoscope image to be diagnosed with the organ image library, identify the nasal cavity area and laryngeal cavity area of ​​the nasal and laryngeal cavity image to be diagnosed; segment the nasal and laryngeal cavity image to be diagnosed according to the nasal cavity area and laryngeal cavity area to obtain a segmented image;

[0048] Image recognition unit: used to input the segmented image into the image recognition model; obtain the disease recognition result of each segmented image output by the image recognition model;

[0049] Evaluation unit: used to receive the manual evaluation score for the disease identification result; determine the disease to be diagnosed in the nasolaryngoscope image according to the manual evaluation score and the disease identification result.

[0050] In this embodiment, the image sample can be a confirmed or undiagnosed naso-laryngoscope image from the source hospital. The model unit trains an image recognition model, which is used to classify the input naso-laryngoscope images and identify the diseases of the naso-laryngoscope images. When training the deep learning architecture, the model unit uses naso-laryngoscope images annotated with disease information for training, which improves the recognition accuracy of the image recognition model. The organ image library stores standard images of the nasal cavity and laryngeal cavity, see Figure 2 , Figure 2 A is a normal laryngoscope image, B is a laryngoscope image of vocal cord nodules, C is a laryngoscope image of vocal cord white spots, D is a laryngoscope image of polyps, and E is a laryngoscope image of malignant tumors. For example, the system can manually mark typical areas of the nasal cavity or laryngeal cavity in known nasal laryngoscope images or medical image databases as standard images.

[0051] In this embodiment, after receiving the nasoscopic image to be diagnosed, the image processing unit matches the nasoscopic image to be diagnosed with the standard image, identifies the nasal area and laryngeal area of ​​the nasoscopic image to be diagnosed, and segments the nasoscopic image to be diagnosed according to the nasal area and laryngeal area to obtain a segmented image, so that independent nasal and laryngeal images can be obtained. The image recognition unit inputs the segmented image into the image recognition model, and the image recognition model outputs the disease recognition results of each segmented image. In order to further improve the accuracy of disease recognition, the system also adds a manual evaluation process to the prediction results of the model. The evaluation unit receives the manual evaluation score for the disease recognition result; determines the disease of the nasoscopic image to be diagnosed based on the manual evaluation score and the disease recognition result.

[0052] The system applies deep learning models to the medical image processing of nasal and laryngoscopes. Through the deep learning model, the symptoms of nasal and laryngoscope images can be quickly identified, achieving early identification and warning of diseases, reducing the difficulty of doctors' interpretation of nasal and laryngoscope images, and improving the objectivity of diagnostic results.

[0053] Furthermore, the deep learning architecture includes: a combination of a dual combined deep residual network and a transfer learning network.

[0054] In this embodiment, the system is trained based on the architecture of a dual combined deep residual network + transfer learning network to obtain an image recognition model.

[0055] Furthermore, the disease information of the image sample includes diagnosed or undiagnosed diseases.

[0056] In this embodiment, the image samples include not only naso-laryngoscope images, but also the annotated diagnosed or undiagnosed diseases for each naso-laryngoscope image. The system can also establish a database to record daily training, recognition and operation. The database can also save each annotation, improvement or reduction of training effect and operator, which is convenient for data backtracking and reverse deduction.

[0057] Furthermore, the collection unit is specifically used for:

[0058] Receive nasal and laryngoscope videos to be diagnosed;

[0059] Extract key frames from the nasal laryngoscope video to be diagnosed to obtain the nasal laryngoscope image to be diagnosed.

[0060] In this embodiment, the acquisition unit can filter the nasal and laryngoscope video to be diagnosed (such as videos in AVI, MP4, MIDI, and MOV formats), for example, filtering at a speed greater than 30 frames per second, and filtering out the key frames in the video to obtain the nasal and laryngeal endoscopy image to be diagnosed, and quickly locate the nasal and laryngeal organs that need to be examined.

[0061] Furthermore, the image processing unit is specifically used for:

[0062] A sliding window is set on the nasal laryngoscope image to be diagnosed to perform a sliding search, and the matching degree between the standard image in the organ image library and the nasal laryngoscope image to be diagnosed in the sliding window is calculated;

[0063] The area within the sliding window on the nasolaryngoscope image to be diagnosed with the highest matching degree is defined as the nasal cavity area or the laryngeal cavity area.

[0064] In this embodiment, the image processing unit is used to find an area similar to the standard image in the naso-laryngoscope image, and locate the nasal cavity area and the laryngeal cavity area in the naso-laryngoscope image to be diagnosed. Specifically, the standard image is searched in a sliding manner in the naso-laryngoscope image to be diagnosed, that is, it is moved pixel by pixel at each position of the naso-laryngoscope image to be diagnosed, and the matching degree between the standard image and the current sliding window area is calculated, and the area in the sliding window on the naso-laryngoscope image to be diagnosed with the highest matching degree is defined as the nasal cavity area or the laryngeal cavity area.

[0065] In this embodiment, the matching degree can be calculated by a variety of methods, including:

[0066] 1) Mutual correlation coefficient: Calculate the correlation coefficient between the standard image and the sliding window endonasal laryngoscope image. The higher the correlation coefficient, the higher the matching degree.

[0067] 2) Mean square error: Calculate the mean square error between the standard image and the nasal laryngoscope image in the sliding window. The smaller the mean square error, the higher the matching degree.

[0068] 3) Normalized cross-correlation: The obtained correlation coefficient is normalized to reduce the influence of factors such as illumination and contrast. The higher the normalized cross-correlation, the higher the matching degree.

[0069] 4) Feature matching: extract feature points of standard images in the nasal cavity and laryngeal cavity, use SIFT algorithm to match the feature points of the nasal and laryngeal endoscope images to be diagnosed, and obtain the matching degree. For example, feature points of nasal and laryngeal endoscope images under various lesions can be extracted for matching.

[0070] 5) Deep learning method: The convolutional neural network (CNN) deep learning model is used to classify and identify the nasal and laryngeal cavities of the nasal and laryngeal endoscopic images to be diagnosed. The convolutional neural network deep learning model classifies and identifies the input nasal and laryngeal endoscopic images to be diagnosed based on pre-trained feature points.

[0071] The method for matching the nasolaryngoscope images to be diagnosed is intuitive, simple, and easy to implement, and can accurately locate the nasal cavity and laryngeal cavity in the nasolaryngoscope images.

[0072] Furthermore, the image processing unit is specifically used for:

[0073] Adjust the size of the sliding window, and calculate the matching degree between the standard image and the nasal laryngoscope images to be diagnosed in sliding windows of different sizes;

[0074] Rotate the standard image and calculate the matching degree between the standard image at different angles and the nasal laryngoscope image to be diagnosed in the sliding window;

[0075] The edge information or shape descriptors of the standard image and the nasal laryngoscope image to be diagnosed in the sliding window are extracted, and the matching degree of the edge information or shape descriptors between the standard image and the nasal laryngoscope image to be diagnosed in the sliding window is calculated.

[0076] In this embodiment, since the operation of sliding windows will result in a large amount of calculation for large-sized naso-laryngoscope images and standard images, the image processing unit can match the standard image and the naso-laryngoscope image to be diagnosed at different scales to adapt to the matching of images of different sizes. Since the matching effect will be significantly reduced if the shape, size or direction of the nasal cavity or laryngeal cavity in the naso-laryngoscope image to be diagnosed is significantly different from that of the standard image, the image processing unit allows the standard image to be matched at different rotation angles to adapt to the rotation transformation. Since noise and illumination changes in the naso-laryngoscope image to be diagnosed may affect the accuracy of the matching degree, the image processing unit uses edge detection or shape descriptors to enhance the robustness of the standard image matching.

[0077] Furthermore, the disease recognition results of the segmented image include:

[0078] Disease information, confidence and evaluation value of the segmented image.

[0079] In this embodiment, in addition to outputting the disease information to which the segmented image belongs, the image recognition model also outputs the confidence and evaluation value of the disease information, which are used to assist in confirming the disease of the naso-laryngoscope image to be diagnosed. The model unit can also use the confidence to adjust the classification threshold of the model to optimize a certain evaluation value of the model. For example, the model unit can balance the precision and recall rate by selecting different confidence thresholds to achieve the best model performance.

[0080] Confidence describes the degree of certainty of the image recognition model about the disease information, and the value range is usually 0 to 1. The confidence of the image recognition model can be calculated in the following ways: 1) Softmax confidence: For multi-classification problems, the Softmax function converts the output of the image recognition model into a probability distribution, and the confidence is the category corresponding to the highest probability. 2) Confidence in binary classification problems: For binary classification problems, the confidence can be directly obtained from the probability output by the image recognition model.

[0081] Evaluation values ​​are used to measure the performance of image recognition models. Evaluation values ​​can include accuracy, precision, recall, F1 score, etc. Accuracy measures the proportion of correct predictions made by the image recognition model. Precision measures the proportion of true positive samples among samples predicted as positive. Recall measures the proportion of all true positive samples that are correctly predicted as positive. F1 score is the harmonic mean of precision and recall, and is used to evaluate the performance of image recognition models in unbalanced data sets. The area under the ROC curve (AUC-ROC) is an important indicator for evaluating the performance of a binary classification model, which represents the area under the ROC curve. The ROC curve depicts the relationship between the false positive rate (FPR) and the true positive rate (TPR) of an image recognition model.

[0082] Furthermore, the evaluation unit is specifically used for:

[0083] Receive the evaluation scores given by users for each evaluation value;

[0084] Set weights for each evaluation value;

[0085] All evaluation scores are weighted averaged using weights to obtain the manual evaluation score.

[0086] In this embodiment, when the evaluation unit receives the evaluation scores of the user for each evaluation value, it performs weighted summation on the evaluation scores to obtain a comprehensive manual evaluation score. The evaluation unit allows the user to assign different weights to each evaluation value according to specific needs or preferences. The user who performs the manual evaluation can be a professional medical professional team. For example, assuming there are n evaluation values, each evaluation value has an evaluation score of Si, and each user has a weight of Wi. The manual evaluation score is:

[0087]

[0088] in: or i is the variable and n is the total number of evaluation values.

[0089] Suppose that when evaluating an image recognition model, three indicators are considered: accuracy, precision, and recall, and their weights are 0.5, 0.3, and 0.2, respectively. When the accuracy, precision, and recall are evaluated as 0.85, 0.80, and 0.75, the Weighted Scor e =0.5*0.85+0.3*0.8+0.2*0.75=0.815.

[0090] The evaluation unit can use weighted scoring to comprehensively consider the performance of multiple evaluation values. For example, in the process of tuning the image recognition model, weights can be assigned according to business needs to balance precision and recall. In multi-attribute decision analysis, weights can be set by experts or decision makers based on actual conditions to evaluate the pros and cons of different solutions. The weights can be set directly based on experience or domain knowledge, determined by expert scoring, or determined by principal component analysis (PCA) or entropy method.

[0091] Furthermore, the model unit is also used to:

[0092] The nasal and laryngoscope images to be diagnosed, confirmed diseases, and manual evaluation scores are added to the image samples to train and optimize the image recognition model.

[0093] In this embodiment, after the medical professional team completes the manual evaluation, the model unit is also used to add the nasoscopic images to be diagnosed, the confirmed diseases, and the manual evaluation scores to the image samples, train and optimize the image recognition model, and improve the accuracy of the image recognition model.

[0094] Furthermore, the model unit is also used to:

[0095] Reverse deduction is performed on the confirmed nasal and laryngoscope images to obtain the image features that characterize the disease;

[0096] The confirmed nasal and laryngoscope images and image features are added to the image samples to train and optimize the image recognition model.

[0097] In this embodiment, for diseases that cannot be identified, such as nasal and laryngoscope images of some special diseases, the nasal and laryngoscope images of the special diseases can be enhanced first, and then input into the image recognition model trained with image features to identify the disease of the nasal and laryngoscope images of the special diseases. The model unit reversely derives the confirmed nasal and laryngoscope images to obtain image features, adds the confirmed nasal and laryngoscope images and image features to the image samples, and trains and optimizes the image recognition model.

[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and specification of the present invention.

Claims

1. A nasal laryngoscope image-assisted diagnosis system based on deep learning, characterized in that: include: Model unit: used to create an image recognition model; the image recognition model is obtained by training a deep learning architecture using image samples; the image samples include different nasal and laryngoscope images and annotated disease information; Database unit: used to create an organ image library containing standard images of the nasal cavity and laryngeal cavity; Acquisition unit: used to obtain nasal and laryngoscope images to be diagnosed; Image processing unit: used for matching the nasal and laryngoscope image to be diagnosed with the organ image library, identifying the nasal cavity area and laryngeal cavity area of ​​the nasal and laryngeal cavity image to be diagnosed; segmenting the nasal and laryngeal cavity image to be diagnosed according to the nasal cavity area and laryngeal cavity area to obtain a segmented image; Image recognition unit: used for inputting the segmented picture into the image recognition model; obtaining the disease recognition result of each segmented picture output by the image recognition model; Evaluation unit: used for receiving a manual evaluation score for the disease identification result; and determining the disease of the nasolaryngoscope image to be diagnosed according to the manual evaluation score and the disease identification result.

2. The deep learning-based nasolaryngoscope image-assisted diagnosis system according to claim 1, characterized in that: The deep learning architecture includes: a combination of a dual combined deep residual network and a transfer learning network.

3. The deep learning-based nasolaryngoscope image-assisted diagnosis system according to claim 1 is characterized in that: The disease information of the image sample includes diagnosed or undiagnosed diseases.

4. The deep learning-based nasolaryngoscope image-assisted diagnosis system according to claim 1, characterized in that: The acquisition unit is specifically used for: Receive nasal and laryngoscope videos to be diagnosed; Extract key frames from the nasal laryngoscope video to be diagnosed to obtain the nasal laryngoscope image to be diagnosed.

5. The deep learning-based nasolaryngoscope image-assisted diagnosis system according to claim 1, characterized in that: The image processing unit is specifically used for: Setting a sliding window on the nasal laryngoscope image to be diagnosed to perform sliding search, and calculating the matching degree between the standard image in the organ image library and the nasal laryngoscope image to be diagnosed in the sliding window; The area within the sliding window on the nasolaryngoscope image to be diagnosed with the highest matching degree is defined as the nasal cavity area or the laryngeal cavity area.

6. The deep learning-based nasolaryngoscope image-assisted diagnosis system according to claim 5, characterized in that: The image processing unit is specifically used for: Adjusting the size of the sliding window, and calculating the matching degree between the standard image and the nasal laryngoscope images to be diagnosed in sliding windows of different sizes; Rotating the standard image, and calculating the matching degree between the standard image at different angles and the nasal laryngoscope image to be diagnosed in the sliding window; The edge information or shape descriptors of the standard image and the nasal laryngoscope image to be diagnosed in the sliding window are extracted, and the matching degree of the edge information or shape descriptors between the standard image and the nasal laryngoscope image to be diagnosed in the sliding window is calculated.

7. The deep learning-based nasolaryngoscope image-assisted diagnosis system according to claim 1, characterized in that: The disease recognition results of the segmented image include: Disease information, confidence and evaluation value of the segmented image.

8. The deep learning-based nasolaryngoscope image-assisted diagnosis system according to claim 1, characterized in that: The evaluation unit is specifically used for: Receive evaluation scores from multiple users for each evaluation value; Set weights for each evaluation value; The weights are used to perform a weighted average on all evaluation scores to obtain the manual evaluation score.

9. The deep learning-based nasolaryngoscope image-assisted diagnosis system according to claim 8, characterized in that: The model unit is also used to: The nasal and laryngoscope images to be diagnosed, the diagnosed diseases, and the manual evaluation scores are added to the image samples to train and optimize the image recognition model.

10. The deep learning-based nasolaryngoscope image-assisted diagnosis system according to claim 1, characterized in that: The model unit is also used to: Reverse deduction is performed on the confirmed nasal and laryngoscope images to obtain the image features that characterize the disease; The confirmed nasal and laryngoscope images and image features are added to the image samples to train and optimize the image recognition model.

Images