Novel interstitial lung disease lesion detection method based on deep learning

Abstract

The invention discloses a novel interstitial lung disease lesion detection method based on deep learning, and relates to the technical field of intelligent medical treatment, and the method comprises the following steps: obtaining an HRCT image of a patient, and carrying out the format conversion of the HRCT image to obtain a CT image; inputting the CT image into a U-Net, and carrying out pathological pulmonary parenchyma segmentation on the CT image to obtain a segmented image; determining a lesion candidate region of the segmented image; and evaluating the lesion candidate region to determine whether the lung disease lesion occurs in the lesion candidate region. According to the method, the lesion areas of different CT images are accurately positioned, and the lesion is correctly identified and false positive areas are removed by extracting the statistical features, the shape features and the fractal features of the lesion candidate areas under the condition of interference of blood vessels and tracheas, so that the detection accuracy of novel interstitial lung disease lesion detection is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent medicine, and particularly relates to a novel method for detecting interstitial lung disease lesions based on deep learning. Background Art

[0002] Interstitial lung disease is a group of diseases that affect the lung interstitium. Idiopathic pulmonary fibrosis is the most common type of interstitial lung disease in adults, and the prognosis of patients is poor. Early diagnosis of idiopathic pulmonary fibrosis can guide anti-fibrotic treatment, thereby prolonging the survival period of patients and reducing acute exacerbation. The diagnosis and monitoring of interstitial lung disease usually rely on imaging examinations, especially HRCT. Existing manual diagnosis methods require doctors to have high professional knowledge and rich clinical experience, and reading a large amount of data is a great challenge to the doctor's attention and energy. However, traditional image analysis methods are time-consuming and laborious, and are greatly affected by the experience of the reading doctor. With the development of artificial intelligence technology, deep learning has become a powerful tool for medical image analysis. However, when current artificial intelligence detects whether there is interstitial lung disease in a patient's lungs, due to the interference of pulmonary blood vessels and trachea, misdetection and missed detection often occur, so improvement is urgently needed. Summary of the Invention

[0003] Aiming at the deficiencies of the existing technology, the purpose of the present invention is to provide a novel method for detecting interstitial lung disease lesions based on deep learning, aiming to solve the above technical problems.

[0004] To achieve the above purpose, the present invention adopts the following technical solutions:

[0005] A novel method for detecting interstitial lung disease lesions based on deep learning, comprising the following steps:

[0006] Obtain the HRCT image of the patient, and perform format conversion on the HRCT image to obtain a CT image;

[0007] Input the CT image into U-Net to perform pathological lung parenchyma segmentation on the CT image to obtain a segmentation image;

[0008] Determine the lesion candidate region of the segmentation image;

[0009] Evaluate the lesion candidate region to determine whether there is a lung disease lesion in the lesion candidate region.

[0010] Preferably, obtaining the HRCT image of the patient and performing format conversion on the HRCT image to obtain a CT image includes the following steps:

[0011] Step 1: Read in the HRCT images in DICOM format. The stored data values are obtained after encoding the CT values. The CT values are restored through the following formula:

[0012] ct value = va * sl + in

[0013] where ct value represents the CT value, va represents the encoded data value, sl represents the slope, and in represents the intercept. The slope and intercept are stored in the tags of the DICOM data;

[0014] Step 2: Convert the corresponding CT values to the range of 0 - 255 according to the window width and window level, which is achieved through the following formula:

[0015]

[0016] where wl represents the window level, w represents the window width, Lv represents the minimum CT value within the local window, Hv represents the maximum CT value within the local window, I represents the CT image after format conversion, and others represent the CT values greater than the minimum CT value and less than the maximum CT value within the local window.

[0017] Preferably, the CT image format conversion module for obtaining the HRCT images of the patient and performing format conversion on the HRCT images to obtain the CT image includes a preprocessing unit, a DICOM parsing unit, and a standardized output unit;

[0018] The preprocessing unit receives the HRCT raw data from the Picture Archiving and Communication System (PACS) and unifies the scan sequences with inconsistent slice thicknesses to 1mm equally spaced slices through an interpolation algorithm;

[0019] The DICOM parsing unit extracts the image pixel data and metadata tags, including the patient ID, scan parameters, and reconstruction matrix;

[0020] The standardized output unit uses the window width and window level adjustment technology to convert the raw data into an 8-bit grayscale JPEG format and normalizes it to a resolution of 256×256 through bicubic interpolation. The selected window width w is 1600Hu, and the window level wl is -400Hu.

[0021] Preferably, the CT image is input into the U-Net for pathological lung parenchyma segmentation of the CT image to obtain a segmented image. The loss function used by the U-Net is as follows:

[0022] L = L iou + L bce

[0023] where L iouis the IOU loss function, L bce is the binary cross-entropy loss function.

[0024] Preferably, the IOU loss function L iou has the following calculation formula:

[0025]

[0026] where N represents the total number of pixels in the CT image, y i is the true label of pixel i, and when y i is -1, it means that the area where pixel i is located is the background, and when y i is 1, it means that the area where pixel i is located is the lung parenchyma area, represents the predicted value of pixel i.

[0027] Preferably, the binary cross-entropy loss function L bce has the following calculation formula:

[0028]

[0029] where N represents the total number of pixels in the CT image, y i is the true label of pixel i, and when y i is -1, it means that the area where pixel i is located is the background, and when y i is 1, it means that the area where pixel i is located is the lung parenchyma area, represents the predicted value of pixel i.

[0030] Preferably, determining the lesion candidate region of the segmented image includes the following steps:

[0031] Step 1, input the CT image I;

[0032] Step 2, use Laplacian sharpening to enhance the edges and details of the CT image to obtain the image I l ;

[0033] Step 3, adopt Harris corner detection to detect corners in the image I l to obtain the corner distribution image I J ;

[0034] Step 4, count the number of corners within the 32×32 neighborhood of each pixel point and use this number of corners as the gray value of the pixel to obtain the corner distribution intensity map I Q ;

[0035] Step 5, multiply the corner distribution intensity map I Q pixel by pixel with the lung mask image of the same size, and then obtain the binary image I E ;

[0036] Step 6, perform a filling operation on the binary image I E to obtain a filled image I T ;

[0037] Step 7, perform a connected component analysis on the filled image I T to remove connected components with a total number of pixels less than 7, thereby obtaining a lesion candidate region I B .

[0038] Preferably, the dynamic threshold segmentation is implemented using the Otsu algorithm.

[0039] Preferably, evaluate the lesion candidate region to determine whether the lesion candidate region has a pulmonary disease lesion, including the following steps:

[0040] Step 1, extract the statistical features, shape features, and fractal features of the lesion candidate region I B to obtain a feature vector;

[0041] Step 2, input the feature vector into a detection model to determine the detection result.

[0042] Preferably, the detection model is constructed as follows:

[0043] Step 1, obtain a number of HRCT images of new interstitial lung disease lesions and HRCT images of healthy lungs;

[0044] Step 2, label the lung regions and new interstitial lung disease lesion regions in the HRCT images of new interstitial lung disease lesions and HRCT images of healthy lungs;

[0045] Step 3, preprocess the HRCT images of new interstitial lung disease lesions and HRCT images of healthy lungs after labeling to obtain a CT image set;

[0046] Step 4, divide the CT image set into a training set and a validation set;

[0047] Step 5, use the training set to train an SVM model, validate it with the validation set, and adjust the model parameters of the SVM model according to the performance of the validation set to finally obtain a detection model.

[0048] In summary, due to the adoption of the above technical solutions, the beneficial effects of the present invention are:

[0049] The present invention accurately locates the lesion areas in different CT images. In the case of interference from blood vessels and trachea, by extracting the statistical features, shape features, and fractal features of the lesion candidate areas, the lesions are correctly identified, the false positive areas are removed, and thus the detection accuracy of the detection of novel interstitial lung disease lesions is greatly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following described drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0051] Figure 1 The flowchart of a novel interstitial lung disease lesion detection method based on deep learning is shown.

[0052] Figure 2 The flowchart of evaluating the lesion candidate areas is shown.

[0053] Figure 3 The flowchart of constructing the detection model is shown. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0054] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0055] Refer to Figures 1 to 3 A further description is made of an embodiment of a novel interstitial lung disease lesion detection method based on deep learning of the present invention.

[0056] A novel interstitial lung disease lesion detection method based on deep learning includes the following steps:

[0057] Step 1: Obtain the HRCT images of the patient and convert the format of the HRCT images to obtain CT images;

[0058] By converting the format of the HRCT images to obtain CT images, subsequent processing is facilitated.

[0059] Step 2: Input the CT images into U-Net to perform pathological lung parenchyma segmentation on the CT images to obtain segmented images;

[0060] Step 3: Determine the lesion candidate areas of the segmented images;

[0061] Step 4, evaluate the lesion candidate region to determine whether a pulmonary lesion has occurred in the lesion candidate region.

[0062] Obtain the patient's HRCT image and perform format conversion on the HRCT image to obtain a CT image, including the following steps:

[0063] Step 11, read in the HRCT image in DICOM format, where the stored data value is obtained after encoding the CT value. Restore the CT value through the following formula:

[0064] ct value = va * sl + in

[0065] where ct value represents the CT value, va represents the encoded data value, sl represents the slope, and in represents the intercept. The slope and intercept are stored in the tags of the DICOM data;

[0066] It should be noted that the Digital Imaging and Communications in Medicine (DICOM) standard has been widely accepted in the medical field, unifying the medical image storage and transmission formats of different manufacturers and regions. The maximum gray value of DICOM format images is usually not less than 12 bits. The data in this technical solution uses 16-bit encoding, and ordinary computers can only display an 8-bit gray range. When doctors view HRCT images, they use the windowing technique to display the images, opening a local window, which is jointly determined by the window width and window level. The CT values within this local window are converted to between 0 and 255, and then the HRCT image is displayed.

[0067] Step 12, convert the corresponding CT values to between 0 and 255 according to the window width and window level, and implement it through the following formula:

[0068]

[0069] where wl represents the window level, w represents the window width, Lv represents the minimum CT value within the local window, Hv represents the maximum CT value within the local window, I represents the CT image after format conversion, and others represent CT values greater than the minimum CT value within the local window and less than the maximum CT value within the local window.

[0070] The CT image format conversion module used for obtaining the patient's HRCT image and performing format conversion on the HRCT image to obtain a CT image includes a preprocessing unit, a DICOM parsing unit, and a standardized output unit;

[0071] The preprocessing unit receives the original HRCT data from the Picture Archiving and Communication System (PACS), and unifies the scanning sequences with inconsistent slice thicknesses to 1-mm equally-spaced slices through an interpolation algorithm.

[0072] The DICOM parsing unit extracts the image pixel data and metadata tags, including the patient ID, scanning parameters, and reconstruction matrix.

[0073] The standardized output unit uses the window width and window level adjustment technology to convert the original data into an 8-bit grayscale JPEG format, and normalizes it to a resolution of 256×256 through bicubic interpolation. The selected window width w is 1600 Hu, and the window level wl is -400 Hu.

[0074] It should be noted that in this technical solution, the window widths and window levels commonly used by doctors in a provincial people's hospital are statistically analyzed. Therefore, the selected window width w is 1600 Hu, and the window level wl is -400 Hu.

[0075] For example, in the application scenario of the radiology department of a tertiary hospital, a technician imports the HRCT sequence (slice thickness 0.6 mm, interval 0.4 mm) collected by a Siemens SOMATOM Force CT device into the system. The preprocessing unit first detects the interlayer registration error and uses an elastic registration algorithm based on B-spline to control the displacement field within 0.3 mm. The DICOM parsing unit automatically identifies the lung window parameters (window width 1600 HU, window level -500 HU) and corrects the geometric distortion caused by the tilting of the scanning bed. The standardized output unit generates an image file that conforms to the DICOM Part10 standard, while retaining the linear relationship of the original CT values to ensure the quantitative accuracy of subsequent analysis.

[0076] For example, for the compressed HRCT data (slice thickness 2 mm, JPEG-LS lossy compression) transferred from a primary hospital, the system starts a redundant data processing protocol: the preprocessing unit restores the compressed high-frequency information through local histogram matching and uses an artifact suppression model based on U-Net (the training set contains 500 labeled data) to eliminate block artifacts. The standardized output unit enhances the texture details of the low-resolution image through a generative adversarial network (the generator loss function is L GAN =E[logD(x)]+E[log(1-D(G(z) ))] ) and finally outputs an image sequence that meets the analysis requirements.

[0077] The present invention realizes the adaptive parsing of heterogeneous DICOM data by setting up a multi-layer preprocessing architecture, solving the compatibility problem of devices from different manufacturers; adopting a standardized process that combines deep learning and traditional image processing, while retaining the lesion features, controlling the image processing time within 3 seconds per case (Intel Xeon Gold 6248R processor). Verified clinically (sample size n = 120), there is no significant difference in the detection rate of ground-glass opacities between the images converted by this system and the original data (p>0.05, Wilcoxon test), meeting the requirements for quantitative analysis of interstitial lung diseases.

[0078] Input the CT image into U-Net to perform pathological lung parenchyma segmentation on the CT image to obtain a segmented image, where the loss function used by U-Net is as follows:

[0079] L = L iou + L bce

[0080] where L iou is the IOU loss function, and L bce is the binary cross-entropy loss function.

[0081] The calculation formula of the IOU loss function L iou is as follows:

[0082]

[0083] where N represents the total number of pixels in the CT image, y i is the true label of pixel i, and when y i is -1, it means that the area where pixel i is located is the background, and when y i is 1, it means that the area where pixel i is located is the lung parenchyma area, represents the predicted value of pixel i.

[0084] The calculation formula of the binary cross-entropy loss function L bce is as follows:

[0085]

[0086] where N represents the total number of pixels in the CT image, y i is the true label of pixel i, and when y i is -1, it means that the area where pixel i is located is the background, and when y i is 1, it means that the area where pixel i is located is the lung parenchyma area, represents the predicted value of pixel i.

[0087] It should be noted that by using the U-Net to perform pathological lung parenchyma segmentation on the CT image, the method has a good segmentation effect, finely preserves the lung parenchyma in complex regions such as the hilum of the lung and the apex of the lung, removes large tracheas, and improves the subsequent detection accuracy to a certain extent.

[0088] Determine the lesion candidate region of the segmented image, including the following steps:

[0089] Step 31, input the CT image I;

[0090] Step 32, use Laplacian sharpening to enhance the edges and details of the CT image to obtain the image I l ;

[0091] By operating like this, the purpose is to make the lesion area in the CT image I more obvious, thereby generating more corner points and reducing the probability of missed detection.

[0092] Step 33, use Harris corner detection to detect corner points in the image I l to obtain the corner point distribution image I J ;

[0093] Step 34, count the number of corner points in the 32×32 neighborhood range of each pixel point, and use this number of corner points as the gray value of the pixel to obtain the corner point distribution intensity map I Q ;

[0094] Step 35, multiply the corner point distribution intensity map I Q pixel by pixel with the lung mask image of the same size, and then obtain the binary image I through dynamic threshold segmentation E ;

[0095] Step 36, perform a filling operation on the binary image I E to obtain the filled image I T ;

[0096] By operating like this, it is to avoid the appearance of small holes in the connected region, which may affect the subsequent detection.

[0097] Step 37, perform connected region analysis on the filled image I T to remove the connected regions with the total number of pixels less than 7, and then obtain the lesion candidate region I B .

[0098] The dynamic threshold segmentation in Step 35 is implemented using the Otsu algorithm. The Otsu algorithm is a common dynamic threshold segmentation method, which can automatically find the optimal segmentation threshold to maximize the between-class variance between the foreground and the background. Therefore, it will not be elaborated here.

[0099] By determining the lesion candidate regions of the segmented image, the method can better retain the edges of the possible lesion regions in various cases, preliminarily determine the location of the possible lesion, without missing any detections, thereby improving the accuracy of subsequent detections.

[0100] Evaluate the lesion candidate regions to determine whether there is a pulmonary disease lesion in the lesion candidate regions, including the following steps:

[0101] Step 41, extract the statistical features, shape features, and fractal features of the lesion candidate region I B to obtain a feature vector;

[0102] It should be noted that the statistical features include histogram features and gist features; the shape features include Zernike distance and morphological features. Among them, Zernike distance is commonly used in the classification and retrieval of medical images, with rotational invariance and orthogonality, and can globally describe the shape features of the candidate region with minimal redundancy. By extracting the mean and standard deviation of the black and white top-hat transformation as shape features, after the black and white top-hat transformation of the linear structure of the lesion region, the mean and variance change little; while after the black and white top-hat transformation of the linear structure of the healthy region, the linear structures such as blood vessels are strengthened or suppressed, and the mean and variance change greatly, thus avoiding misdetection of regions with complex trachea and blood vessels; the fractal features include fractal dimension.

[0103] Step 42, input the feature vector into the detection model to determine the detection result.

[0104] The construction steps of the detection model are as follows:

[0105] Step 421, obtain a number of HRCT images of new interstitial lung disease lesions and HRCT images of healthy lungs;

[0106] Step 422, label the lung regions and new interstitial lung disease lesion regions in the HRCT images of new interstitial lung disease lesions and HRCT images of healthy lungs;

[0107] It should be noted that the labeling of the lung regions and new interstitial lung disease lesion regions in the HRCT images of new interstitial lung disease lesions and HRCT images of healthy lungs is completed manually by professional radiologists.

[0108] Step 423, preprocess the HRCT images of new interstitial lung disease lesions and HRCT images of healthy lungs after labeling to obtain a CT image set;

[0109] Step 424, divide the CT image set into a training set and a validation set;

[0110] Step 425: Train the SVM model using the training set, verify it with the validation set, and adjust the model parameters of the SVM model according to the performance of the validation set to finally obtain the detection model.

[0111] The present invention accurately locates the lesion regions of different CT images. In the case of interference from blood vessels and trachea, by extracting the statistical features, shape features, and fractal features of lesion candidate region I B correctly identifies the lesions, removes false positive regions, and thus greatly improves the detection accuracy of the detection of novel interstitial lung disease lesions.

[0112] During use, first obtain the chest CT scan image of the patient, and input the CT image data in DICOM format into the pre-trained U-Net neural network model. The U-Net model adopts an encoder-decoder structure. The encoder part contains 4 downsampling modules, and each module consists of two 3×3 convolutional layers, a ReLU activation function, and a 2×2 max pooling layer; the decoder part contains 4 upsampling modules, and each module consists of a 2×2 transposed convolutional layer, a feature concatenation operation, and two 3×3 convolutional layers. The last layer of the model uses a 1×1 convolution combined with a Sigmoid activation function to output the segmentation result. The multi-scale features of the encoder are fused with the corresponding layer features of the decoder through skip connections to achieve precise segmentation of the pathological lung parenchyma region and output a binary segmentation image. The segmentation result eliminates discrete noise points through morphological post-processing, and finally obtains the lung parenchyma region containing interstitial lesion features such as fibrotic lesions and ground-glass opacities.

[0113] By adopting the U-Net deep learning architecture, combining the symmetric structure of the encoder-decoder and skip connections, the present invention effectively solves the technical problems of variable target morphology and blurred boundaries in medical image segmentation; through the multi-scale feature fusion mechanism, it realizes the precise capture of the feature regions of interstitial lung disease from local details to global structures; adopting an end-to-end training method enables the model to automatically learn lesion features and avoids the limitations of manually designed features in traditional methods. The present invention has high segmentation accuracy and strong generalization ability, and can provide a reliable quantitative analysis basis for clinical diagnosis.

[0114] The above description of the embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A novel method for detecting interstitial lung disease lesions based on deep learning, characterized in that, It includes the following steps: Obtain the HRCT image of the patient, and perform format conversion on the HRCT image to obtain a CT image; Input the CT image into U-Net to perform pathological lung parenchyma segmentation on the CT image to obtain a segmented image; Determine the lesion candidate region of the segmented image; Evaluate the lesion candidate region to determine whether the lesion candidate region has a lung disease lesion; Determine the lesion candidate region of the segmented image, including the following steps: Step 1, input CT image I; Step 2: Use Laplacian sharpening to enhance the edges and details of the CT image to obtain image I l ; Step 3, use Harris corner detection on the image I l to detect corner points and obtain the corner point distribution image I J ; Step 4, count the number of corner points within the 32×32 neighborhood of each pixel point, and use this number of corner points as the gray value of the pixel to obtain the corner point distribution intensity map I Q ; Step 5, multiply the corner point distribution intensity map I Q by the lung mask image of the same size pixel by pixel, and then obtain the binary image I E ; Step 6, perform a filling operation on the binary image I E to obtain a filled image I T ; Step 7, perform connected component analysis on the filled image I T to remove connected components with a total number of pixels less than 7, thereby obtaining the lesion candidate region I B .

2. The novel interstitial lung disease lesion detection method based on deep learning according to claim 1, wherein Obtain the HRCT image of the patient, and perform format conversion on the HRCT image to obtain a CT image, including the following steps: Step 1, read in the HRCT image in DICOM format, where the stored data value is obtained after encoding the CT value, and restore the CT value through the following formula: ct value = va * sl + in Among them, ct value represents the CT value, va represents the encoded data value, sl represents the slope, and in represents the intercept. The slope and intercept are stored in the tags of the DICOM data; Step 2, according to the window width and window level, convert the corresponding CT value to between 0 and 255, and achieve it through the following formula: Among them, wl represents the window level, w represents the window width, Lv represents the minimum CT value within the local window, Hv represents the maximum CT value within the local window, I represents the CT image after format conversion, and others represent values greater than the minimum CT value within the local window and less than the maximum CT value within the local window.

3. A novel interstitial lung disease lesion detection method based on deep learning according to claim 2, characterized in that, The CT image format conversion module used for obtaining the HRCT image of the patient and performing format conversion on the HRCT image to obtain a CT image includes a preprocessing unit, a DICOM parsing unit, and a standardized output unit; The preprocessing unit receives the HRCT raw data from the Picture Archiving and Communication System (PACS), and unifies the scan sequence with inconsistent slice thicknesses to 1mm equally spaced slices through an interpolation algorithm; The DICOM parsing unit extracts the image pixel data and metadata tags, including the patient ID, scan parameters, and reconstruction matrix; The standardized output unit uses the window width and window level adjustment technology to convert the raw data into an 8-bit grayscale JPEG format, and normalizes it to a resolution of 256×256 through bicubic interpolation. The selected window width w is 1600Hu, and the window level wl is -400Hu.

4. A novel interstitial lung disease lesion detection method based on deep learning according to claim 3, characterized in that, Input the CT image into U-Net to perform pathological lung parenchyma segmentation on the CT image to obtain a segmented image, where the loss function used by U-Net is as follows: L = L iou + L bce where L iou is the IOU loss function, and L bce is the binary cross-entropy loss function.

5. A novel interstitial lung disease lesion detection method based on deep learning according to claim 4, characterized in that IOU loss function L iou The calculation formula is as follows: Among them, N represents the total number of pixels in the CT image, y i is the true label of pixel i, and when y i is -1, it indicates that the area where pixel i is located is the background. When y i is 1, it indicates that the area where pixel i is located is the lung parenchyma area, represents the predicted value of pixel i.

6. A novel interstitial lung disease lesion detection method based on deep learning according to claim 5, characterized in that, Binary cross-entropy loss function L bce The calculation formula is as follows: Among them, N represents the total number of pixels in the CT image, and y i is the true label of pixel i, and when y i is -1, it means that the area where pixel i is located is the background. When y i is 1, it means that the area where pixel i is located is the lung parenchyma area, represents the predicted value of pixel i.

7. A novel interstitial lung disease lesion detection method based on deep learning according to claim 6, characterized in that, Dynamic threshold segmentation is implemented using the Otsu algorithm.

8. A novel interstitial lung disease lesion detection method based on deep learning according to claim 7, characterized in that, Evaluate the lesion candidate region to determine whether the lesion candidate region has a lung disease lesion, including the following steps: Step 1, extract the statistical features, shape features, and fractal features of the lesion candidate region I B to obtain a feature vector; Step 2, input the feature vector into the detection model to determine the detection result.

9. A novel interstitial lung disease lesion detection method based on deep learning according to claim 8, characterized in that, The construction steps of the detection model are as follows: Step 1, obtain several HRCT images of patients with novel interstitial lung disease lesions and HRCT images of healthy lungs; Step 2, label the lung regions and novel interstitial lung disease lesion regions in the HRCT images of patients with novel interstitial lung disease lesions and HRCT images of healthy lungs; Step 3, preprocess the HRCT images of patients with novel interstitial lung disease lesions and HRCT images of healthy lungs after annotation to obtain a CT image set; Step 4: Divide the CT image set into a training set and a validation set; Step 5: Use the training set to train the SVM model, and use the validation set for validation. Adjust the model parameters of the SVM model according to the performance of the validation set to finally obtain a detection model.

Images