Automatic diagnosis method for unknown fever causes based on PET image

Abstract

The invention discloses a PET-image-based automatic diagnosis method for unknown fever causes, which comprises the following steps of: simulating a film reading habit of a doctor of an imaging department, extracting maximum value mapping graphs from different angles, obtaining relation characteristics among the maximum value mapping graphs from various angles by utilizing a Bayesian convolutional neural network, forming characteristic vector representation of a 3D PET image, and finally obtaining a 3D PET image. A common machine learning model is utilized to construct a classification method of unknown cause fever causes, and an automatic diagnosis model of the unknown cause fever causes is formed. The method can effectively assist clinicians in definite diagnosis of causes of fever with unknown reasons, and can quickly and early screen malignant causes, so that the treatment efficiency of possible causes of patients is improved.

Description

Technical Field

[0001] The present invention relates to clinical automatic diagnosis technology, and in particular to an automatic diagnosis method for the cause of fever of unknown origin based on PET images. Background Art

[0002] Fever of unknown origin (FUO), also known as fever to be investigated, was proposed by Petersdorf and Beeson in 1961. Its definition is: generally referring to a patient with a fever lasting for more than three weeks, and their oral temperature rising above 38.3°C multiple times, and still unable to clearly diagnose the cause after completing various examinations within one week of admission. Usually, if the fever persists for more than three weeks and above, and the cause cannot be clearly identified through a detailed medical history, physical examination, and routine laboratory tests, such cases can be classified as fever of unknown origin. For such cases, doctors need to accurately judge whether it is caused by infection, rheumatism and immunity, tumors, or other factors based on their empirical knowledge. The undiagnosed rate of the cause of fever of unknown origin reaches 53%, which has also become one of the more difficult diseases to diagnose in current clinical practice. Therefore, if the cause is accurately identified, the right medicine can be prescribed to save lives; but if the diagnosis is incorrect, very serious consequences will occur.

[0003] In recent years, with the rapid development of positron emission tomography-computed tomography (PET-CT) technology, it is playing a very important role in the etiological diagnosis of fever of unknown origin, disease staging, and evaluating the prognosis of treatment effects. Currently, the commonly used positron radionuclides mainly include 18F, 11C, 13N, etc. Among them, 18F-labeled FDG (18F-FDG) is called a "molecule of the century" and is currently an important positron drug, which has been widely used in clinical imaging diagnosis, and its usage accounts for more than 95% of all PET imaging. The tracer 18F-FDG used in PET is a glucose analog. When the body has an unknown cause of fever due to tumors, infections, and inflammations, tumor cells and inflammatory cells often highly express glucose transporters, resulting in a large accumulation of 18F-FDG in the lesion, and thus being accurately located by PET at the functional and metabolic level, providing a powerful direction and basis for further auxiliary examinations and diagnoses. 18F-FDG PET can simultaneously display the morphological structure and functional and metabolic characteristics of tissues, and plays an important role in finding the cause of fever of unknown origin.

[0004] In addition, PET can scan the head and body parts in one go, covering all internal organs. It can identify the nature, location, and cause of the fever in patients with unknown fever causes, promptly detect hidden "bombs" in the body, comprehensively evaluate the systemic metabolic changes of the patients, and determine the location of the lesions causing the fever. It can also provide an accurate puncture target area for pathological examination. If the fever is caused by a tumor, a comprehensive evaluation of the tumor disease can be carried out.

[0005] Currently, the PET medical image reading habits of clinical radiologists are as follows: First, rotate and read the PET images in different axial directions to roughly judge the possible lesion areas, and then draw targeted lesions in the suspected lesion areas. Then, judge the location of the lesions in the human body and the appearance of the lesions to determine the cause of the unknown fever. Currently, the main causes of unknown fever that need to be diagnosed through PET images are infection, rheumatism and immunity, and tumors. If the cause is suspected to be a tumor or other reasons, the patient needs to be further arranged for a puncture biopsy to rule out the possibility of a tumor pathologically. This process is both cumbersome and time-consuming, and sometimes the best treatment time will be missed.

[0006] After consulting relevant materials, most of them are about the value of PET images in the diagnosis of the cause of unknown fever, and there is no automatic diagnosis method for the cause of unknown fever based on PET images from the perspective of image analysis. Summary of the Invention

[0007] In view of the problems existing in the prior art, the present invention provides an automatic diagnosis method for the cause of unknown fever based on PET images, which can effectively assist clinicians in diagnosing the cause of unknown fever, quickly screen for malignant causes, and improve the treatment efficiency of possible causes of patients.

[0008] The object of the present invention is achieved through the following technical solutions.

[0009] An automatic diagnosis method for the cause of unknown fever based on PET images, the steps include:

[0010] 1) Obtain 3D PET medical images;

[0011] 2) Generate maximum value mapping diagrams at different angles;

[0012] 3) Construct a Bayesian convolutional neural network;

[0013] 4) Use the Bayesian convolutional neural network to obtain the relationship features between the maximum value mapping diagrams at each angle;

[0014] 5) Construct a feature vector representation set of 3D PET images;

[0015] 6) Construct a cause classification model for unknown fever;

[0016] 7) Analyze the prediction results of the etiology of fever of unknown origin.

[0017] In step 2), the obtained PET medical images are rotated one week along the x-axis, y-axis, and z-axis at a speed of 10° / time, and the maximum value mapping diagram at different angles is calculated during the rotation.

[0018] The calculation process of the maximum value mapping diagram includes:

[0019] Rotate the 3D PET medical image along the x-axis at a speed of 10° / time, take the xOz plane as the reference plane, and construct an empty matrix XOz of the maximum value mapping diagram x×x , extract the gray value I of each column perpendicular to the xOz plane y (x, z), and calculate the maximum value in I y (x, z) Place at xOz(x, z), and construct 36 maximum value mapping diagrams along the x-axis in this way;

[0020] Rotate the 3D PET medical image along the y-axis at a speed of 10° / time, take the xOy plane as the reference plane, and construct an empty matrix xOy of the maximum value mapping diagram x×y , extract the gray value I of each column perpendicular to the xOy plane z (x, y), and calculate the maximum value in I z (x, y) Place at xOy(x, y), and construct 36 maximum value mapping diagrams along the y-axis in this way;

[0021] Rotate the 3D PET medical image along the z-axis at a speed of 10° / time, take the yOz plane as the reference plane, and construct an empty matrix yOz of the maximum value mapping diagram y×z , extract the gray value I of each column perpendicular to the yOz plane x (y, z), and calculate the maximum value in I x (y, z) Place at yOz(y, z), and construct 36 maximum value mapping diagrams along the z-axis in this way.

[0022] In step 5), the total 108 maximum value mapping diagrams generated in step 2 are scaled to a unified resolution size to construct a set of feature image representations of the 3D PET image.

[0023] Specifically, step 3) is:

[0024] The data labels for training the Bayesian convolutional neural network are the doctor's diagnosis results \(L\in\{0, 1, 2\}\), where \(0\) represents tumor, \(1\) represents rheumatology and immunology, and \(2\) represents infection;

[0025] The input maximum value mapping graph dataset is denoted as \(X\);

[0026] Select a common classification convolutional neural network as the backbone network;

[0027] Adopt Monte Carlo sampling for approximate integration and calculate the prediction mean:

[0028]

[0029] where is the new sample data with the prediction result, \(\mu\) p is the true data mean, is the mean of the prediction results of the Bayesian convolutional neural network, \(N\) is the total number of Monte Carlo samplings, represents the state of different discarded neurons in the \(n\)th sampling, \(\theta(\cdot)\) represents the neural network parameter weights, and \(p(\cdot)\) is the output of the Bayesian convolutional neural network;

[0030] Measure the uncertainty obtained by the prediction of the Bayesian convolutional neural network with the standard deviation:

[0031]

[0032] The specific content of step 4) is as follows:

[0033] Use the trained Bayesian convolutional neural network to calculate the uncertainty \(P\) of each maximum value mapping graph, so as to describe the relationship characteristics between the maximum value mapping graphs of the same 3D PET medical image at different angles, and generate a relationship feature representation vector with a length of 108 for each 3D PET medical image.

[0034] The specific method for step 6) to construct a classification model for the cause of fever of unknown origin by using the relationship features of the maximum value mapping graphs at different angles is as follows:

[0035] The relationship features of the maximum value mapping graphs at different angles are denoted as \(F\), and the label is \(L\in\{0, 1, 2\}\), where \(0\) represents tumor, \(1\) represents rheumatology and immunology, and \(2\) represents infection;

[0036] Use a machine learning model or a neural network model for feature fitting;

[0037] According to the selected model, perform model training and calculate the result of the loss function

[0038]

[0039] where \(C\) is the total number of categories;

[0040] Use the trained model to predict the category of the cause of fever of unknown origin.

[0041] Specifically, step 7) is as follows:

[0042] If it is a tumor, the system issues a warning to remind the radiologist to give medical advice to the patient and recommend further biopsy and puncture for pathological examination.

[0043] If the pathological examination result confirms a tumor, implement the tumor treatment plan.

[0044] If the pathological examination result is not a tumor, rule out the possibility of a tumor and conduct routine symptomatic treatment, but long-term follow-up is also required.

[0045] If it is not a tumor, it is reconfirmed by a senior radiologist.

[0046] If the senior radiologist confirms it is a tumor, conduct a pathological examination for the final tumor diagnosis and exclusion.

[0047] If the senior radiologist confirms it is not a tumor, inform the patient, recommend specialist treatment for symptomatic treatment, and adhere to follow-up.

[0048] Compared with the prior art, the advantages of the present invention are as follows: The present invention does not require manual feature design or feature selection. It automatically extracts the maximum value mapping diagram of the PET image from different angles, simulates the process of a radiologist reading images, and represents the sparse representation of a patient's PET image in this way. It uses a Bayesian convolutional neural network to predict the uncertainty (probable probability value) of the maximum value mapping diagrams extracted from different angles, and uses the obtained uncertainty to construct the feature representation vector of each patient's PET image, thereby constructing a classification model for the cause of the disease, solving the problem that the process of determining the cause of fever of unknown origin is cumbersome and time-consuming at present, assisting the radiologist in diagnosing the cause of fever of unknown origin, facilitating the clinical auxiliary diagnosis of the cause of fever of unknown origin, shortening the diagnosis time of some causes, and improving the treatment efficiency of patients. The present invention uses the collected PET image data, simulates the reading habits of radiologists, constructs a multi-angle maximum value mapping diagram, uses a Bayesian convolutional neural network to obtain the relationship features between images at different angles, constructs a classification model for the cause of fever of unknown origin, and realizes a completely automatic diagnostic model for the cause of fever of unknown origin, which is conducive to assisting clinicians in making accurate diagnoses quickly. Description of the Drawings

[0049] Figure 1 It is a flow chart of the present invention.

[0050] Figure 2 It is a schematic diagram for generating the maximum value mapping diagrams of 3D PET images at different angles.

[0051] Figure 3 It is a flow chart for predicting the cause of fever of unknown origin. Specific implementation mode

[0052] The present invention will be described in detail below in conjunction with the accompanying drawings of the specification and specific embodiments.

[0053] Embodiment

[0054] The present invention simulates the film reading habits of radiologists, extracts the maximum value mapping diagrams from different angles, uses a Bayesian convolutional neural network to obtain the relationship features between the maximum value mapping diagrams at each angle, constructs a feature vector representation of the 3D PET image, uses common machine learning models to construct a classification method for the cause of fever of unknown origin, and forms an automatic diagnosis model for the cause of fever of unknown origin.

[0055] As Figure 1 shown in the flow chart of the present invention:

[0056] 1. Obtain PET medical images;

[0057] 2. Schematic diagram for generating maximum value mapping diagrams of 3D PET images from multiple angles, Figure 2 shown as:

[0058] Rotate the obtained PET medical images one week along the x-axis, y-axis, and z-axis in turn at a speed of 10° / time;

[0059] Calculate the maximum value mapping diagrams at different angles during the rotation process;

[0060] The calculation process of the maximum value mapping diagram is as follows:

[0061] The 3D PET medical image rotates along the x-axis at a speed of 10° / time. Taking the xOz plane as the reference plane, construct an empty matrix xOz of the maximum value mapping diagram x×z , extract the gray value I of each column perpendicular to the xOz plane y (x, z), calculate the maximum value in I y (x, z) Place at xOz(x, z) to construct 36 maximum value mapping diagrams along the x-axis;

[0062] The 3D PET medical image rotates along the y-axis at a speed of 10° / time. Taking the xOy plane as the reference plane, construct an empty matrix xOy of the maximum value mapping diagram x×y , extract the gray value I of each column perpendicular to the xOy plane z (x, y), calculate the maximum value in I z (x, y) Place Place it in xOy (x, y) to construct 36 maximum value mapping diagrams along the y-axis;

[0063] The 3D PET medical image rotates along the z-axis at a speed of 10° per time. Taking the yOz plane as the reference plane, construct an empty matrix yOz of the maximum value mapping diagram y×z , extract the gray value I of each column perpendicular to the yOz plane x (y, z), calculate I x the maximum value in (y, z)

[0064] Place in yOz (y, z) to construct 36 maximum value mapping diagrams along the z-axis;

[0065] 3. Construction of the Bayesian convolutional neural network:

[0066] The data label for training the Bayesian convolutional neural network is the doctor's diagnosis result L ∈ {0, 1, 2}, where 0 represents tumor, 1 represents rheumatic immunity, and 2 represents infection;

[0067] The input maximum value mapping diagram data set is denoted as X;

[0068] Select common classification convolutional neural networks as the backbone network, such as ResNet50, DenseNet121, etc.;

[0069] Adopt Monte Carlo sampling to approximate integration and calculate the prediction mean:

[0070]

[0071] Among them, is the new sample data with the prediction result, μ p is the true data mean, is the prediction result mean of the Bayesian convolutional neural network. N is the total number of Monte Carlo samplings, represents the state of different discarded neurons in the nth sampling. θ(·) represents the neural network parameter weights, and p(·) is the output of the Bayesian convolutional neural network;

[0072] Measure the uncertainty obtained by the prediction of the Bayesian convolutional neural network with the standard deviation:

[0073]

[0074] 4. Use the trained Bayesian convolutional neural network to calculate the uncertainty (possible probability value) P of each maximum value mapping diagram, so as to describe the relationship characteristics between the maximum value mapping diagrams at different angles of the same 3D PET medical image. In this way, a relationship feature representation vector with a length of 108 can be generated for each 3D PET medical image;

[0075] 5. Scale the total of 108 maximum value mapping images generated in step 2 to a unified resolution size (such as 224×224) to construct a set of feature image representations of the 3D PET image;

[0076] 6. Use the relationship features of the maximum value mapping images obtained from different angles to construct an etiological classification model for fever of unknown origin, Figure 3 as shown;

[0077] The relationship features of the maximum value mapping images from different angles are denoted as F, and the labels are the same as those in step 3, L∈{0, 1, 2}, where 0 represents tumor, 1 represents rheumatic immunity, and 2 represents infection;

[0078] Use a machine learning model for feature fitting. Traditional machine learning models such as support vector machine (SVM) and gradient boosting decision tree (GBDT) can be selected, or it can also be a neural network model such as convolutional neural network (CNN) and graph neural network (GCN), etc.;

[0079] According to the selected model, perform model training and calculate the result of the loss function

[0080]

[0081] where c is the total number of categories.

[0082] Use the trained model to predict the category of the cause of fever of unknown origin;

[0083] 7. Analyze the prediction results of the cause of fever of unknown origin;

[0084] If it is a tumor, the system issues a warning to remind the radiologist to give medical advice to the patient and recommend further biopsy and puncture for pathological examination;

[0085] If the pathological examination result confirms a tumor, implement the tumor treatment plan;

[0086] If the pathological examination result is not a tumor, exclude the possibility of a tumor and conduct routine symptomatic treatment, but long-term follow-up is also required;

[0087] If it is not a tumor, it is reconfirmed by a senior radiologist;

[0088] If the senior radiologist confirms it is a tumor, conduct a pathological examination for final tumor diagnosis and exclusion;

[0089] If the senior radiologist confirms it is not a tumor, inform the patient, recommend specialist treatment for symptomatic treatment, and adhere to follow-up.

[0090] The main advantages of the present invention are as follows: it simulates the habits of radiologists in reading PET medical images, generates maximum value mapping diagrams from different angles, trains and generates a Bayesian convolutional neural network model with a common classification convolutional neural network as the backbone network, obtains the relationship feature representation vectors between the maximum value mapping diagrams of the same 3D PET medical image from different angles, constructs and generates a classification prediction model for the causes of fever of unknown origin that can assist radiologists, helps in the rapid diagnosis of the causes of fever of unknown origin clinically, and improves the efficiency of symptomatic treatment for patients.

Claims

1. An automatic diagnosis method for the cause of fever of unknown origin based on PET images, characterized by the steps Including: 1) Obtain 3D PET medical images; 2) Generate maximum intensity projection (MIP) maps at different angles; 3) Construct a Bayesian convolutional neural network; 4) Use the Bayesian convolutional neural network to obtain the relationship features between the MIP maps at different angles; 5) Form a set of feature vector representations of the 3D PET image; 6) Construct an etiological classification model for fever of unknown origin; 7) Analyze the prediction results of the etiology of fever of unknown origin.

2. The automatic diagnosis method for the cause of fever of unknown origin based on PET images according to claim 1, characterized in that In step 2), the obtained PET medical images are rotated one week along the x-axis, y-axis, and z-axis in turn at a speed of 10° / time, and the MIP maps at different angles are calculated during the rotation.

3. The automatic diagnosis method for the cause of fever of unknown origin based on PET images according to claim 2, characterized in that The calculation process of the MIP map includes: Rotate the 3D PET medical image along the x-axis at a speed of 10° per time, and construct the maximum value mapping map space matrix xOz with the xOz plane as the reference plane x×z , extract the gray value I of each column perpendicular to the xOz plane y (x, z), and calculate I y the maximum value in (x, z) Place at xOz(x, z), and thus construct 36 maximum value mapping maps along the x-axis; Rotate the 3D PET medical image along the y-axis at a speed of 10° per time, and use the xOy plane as the reference plane to construct the maximum mapping map space matrix xOy x×y , extract the gray value I of each column perpendicular to the xOy plane z (x, y), calculate I z The maximum value in (x, y) Place at xOy(x, y), and thus construct 36 maximum mapping maps along the y-axis; Rotate the 3D PET medical image along the z-axis at a speed of 10° per time, and construct the maximum value mapping map empty matrix yOz with the yOz plane as the reference plane. y×z , Extract the gray value I of each column perpendicular to the yOz plane. x (y, z), Calculate I x The maximum value in (y, z). Place at yOz(y, z), and thus construct 36 maximum value mapping maps along the z-axis.

4. The automatic diagnosis method for the cause of fever of unknown origin based on PET images according to claim 3, characterized in that In step 5), the total of 108 MIP maps generated in step 2 are scaled to the same resolution size to construct a set of feature image representations of the 3D PET image.

5. The automatic diagnosis method for the cause of fever of unknown origin based on PET images according to claim 1, characterized in that Step 3) is specifically: The data label for training the Bayesian convolutional neural network is the doctor's diagnosis result L∈{0, 1, 2}, where 0 represents tumor, 1 represents rheumatic immunity, and 2 represents infection; The input MIP map dataset is denoted as X; Select a commonly used classification convolutional neural network as the backbone network; Adopt Monte Carlo sampling for approximate integration and calculate the prediction mean: Among them, is the new sample data with prediction results, and μ p is the true data mean, is the mean of the prediction results of the Bayesian convolutional neural network, N is the total number of Monte Carlo samplings, represents the state of different discarded neurons in the nth sampling, θ(·) represents the neural network parameter weights, and p(·) is the output of the Bayesian convolutional neural network; Use the standard deviation to measure the uncertainty obtained by the prediction of the Bayesian convolutional neural network:

6. The automatic diagnosis method for the cause of fever of unknown origin based on PET images according to claim 1, characterized in that Step 4) is specifically: Use the trained Bayesian convolutional neural network to calculate the uncertainty P of each MIP map, so as to describe the relationship features between the MIP maps at different angles of the same 3D PET medical image, and generate a relationship feature representation vector with a length of 108 for each 3D PET medical image.

7. The automatic diagnosis method for the cause of fever of unknown origin based on PET images according to claim 1, characterized in that In step 6), use the relationship features of the MIP maps at different angles extracted to construct an etiological classification model for fever of unknown origin, specifically as follows: The relationship features of the MIP maps at different angles are denoted as F, and the label is L∈{0, 1, 2}, where 0 represents tumor, 1 represents rheumatic immunity, and 2 represents infection; Use a machine learning model or a neural network model for feature fitting; According to the selected model, conduct model training and calculate the loss function result where C is the total number of categories; Use the trained model to predict the category of the etiology of fever of unknown origin.

8. The automatic diagnosis method for the cause of fever of unknown origin based on PET images according to claim 1, characterized in that Step 7) is specifically: If it is a tumor, the system issues a warning to remind the radiologist to give medical advice to the patient and suggest further biopsy and puncture for pathological examination; If the pathological examination result confirms a tumor, implement the tumor treatment plan; If the pathological examination result is not a tumor, exclude the possibility of tumor and conduct routine symptomatic treatment, but long-term follow-up is also required; If it is not a tumor, it is reconfirmed by a senior radiologist; If the senior radiologist confirms a tumor, conduct a pathological examination for the final tumor diagnosis and exclusion; If the senior radiologist confirms that it is not a tumor, inform the patient, suggest specialist treatment and symptomatic treatment, and adhere to follow-up.

Images