Differential CT image enhancement method based on noise level

Abstract

The invention belongs to the field of medical image processing, and provides a differential CT image enhancement method based on noise level, which comprises six steps of data preprocessing, construction and training of a denoising network, construction and training of a super-resolution network, construction and training of a noise level classification network, CT image enhancement and data post-processing. According to the method, the noise level of the CT image is judged, different deep learning algorithms, namely a denoising or super-resolution model, are used for optimizing the quality of the CT image, the generalization ability is higher, and more complex and variable input data can be dealt with; data with noise are processed by using various traditional filtering methods and super-resolution pre-training models to obtain gold standard images required by training, so that time and energy are not required to be spent on collecting high-quality gold standard images in advance; the CT value of the output image is calibrated through post-processing operation, it is ensured that the CT value difference before and after enhancement does not affect the accuracy of clinical diagnosis, and compared with other methods without CT value calibration, the method has more practical application value.

Description

Technical Field

[0001] The invention belongs to the field of medical image processing, in particular to a differential CT image enhancement method based on noise level. Background Art

[0002] Computed tomography (CT) is a medical imaging technique that uses X-rays and computer processing to generate two-dimensional or three-dimensional images representing the internal structure of the human body. CT is widely used in medical diagnosis, treatment planning, and disease monitoring. Due to a variety of factors such as CT equipment, scattering of human tissue, reduced scanning dose, and statistical quantum noise, CT scans will introduce a certain degree of noise, which will reduce image quality, reduce contrast, and may interfere with doctors' interpretation of images.

[0003] Improving the quality of low-quality CT images is a current research hotspot. The usual CT quality improvement includes two parts: denoising and super-resolution reconstruction. First, the CT image is denoised, and then the denoised CT image is reconstructed with super-resolution. In the case of image pairs of high-quality CT images and low-quality CT images, relying on the Convolutional Neural Network (CNN) to learn the nonlinear mapping between the source image and the target image end-to-end, a good CT quality improvement can be achieved. However, when there are no high-quality CT images, only CT images of average quality and noise, it is difficult to train an effective deep learning model for CT quality improvement with traditional supervised learning, and its quality can only be improved by using various unsupervised methods or pre-trained weights. In addition, unlike high-level semantic processing such as CT image segmentation, since each clinician has different preferences for reading films, it is impossible to unify the quantitative standard of "good", and the manual intervention required is far greater than the restoration and reconstruction of natural images.

[0004] The reconstruction quality of CT images is affected by many factors, among which the dose size is a key factor. High doses usually lead to high noise, while low doses lead to blurred images and loss of key details. Therefore, developing CT image enhancement methods that can cope with various noise levels can not only significantly improve image quality, but also greatly improve doctors' reading efficiency.

[0005] To this end, those skilled in the art have proposed a differential CT image enhancement method based on noise level. Summary of the invention

[0006] In order to solve the above technical problems, the present invention provides a noise level-based differentiated CT image enhancement method to solve the problems mentioned in the background technology.

[0007] The noise level-based differential CT image enhancement method comprises the following steps:

[0008] S1. Construct training dataset: obtain CT images in dicom format with various noise levels; divide the dataset into high noise level sub-dataset and low noise level sub-dataset according to the clinician's suggestion;

[0009] S2, data preprocessing: normalize the CT image and adjust the data range to between 0 and 1;

[0010] S3. Build and train the denoising network: Select SCUNet as the denoising network; use a variety of traditional unsupervised denoising methods to manually adjust the parameters of the CT images in the high-noise level sub-dataset for denoising. After the doctor confirms that it meets the clinical requirements, it is used as the gold standard image to obtain a paired training data set; train the network and obtain the final denoising model after convergence;

[0011] S4. Build and train a super-resolution network: Select the LIFF network as the super-resolution network; use the pre-trained model of the SwinIR image super-resolution network to perform super-resolution on the CT images in the low-noise level sub-dataset, and use them as gold standard images after doctors confirm that they meet clinical requirements; train the network and obtain the final super-resolution model after convergence;

[0012] S5. Construct and train the noise level classification network: Design a multi-layer perceptron as the classification network; read the attributes representing the CT dose in the DICOM in the training data set, match the attribute values ​​with the corresponding noise levels determined by the doctor, and make a classification network training set; train the network, and obtain the final noise level classification model after convergence;

[0013] S6, CT image enhancement: given a CT image in DICOM format, normalize it using the data preprocessing method described in S1; read the attribute representing the CT dose in DICOM, and use the noise level classification network described in S5 to determine its noise level; if the noise level is high, use the denoising model for reasoning, otherwise use the super-resolution model for reasoning to obtain the output CT image;

[0014] S7, data post-processing: For the output CT image, first adjust its size to be consistent with the input; then crop its data range to between 0 and 1, and denormalize it to the data range of the input image; then use the MeanShift algorithm to adjust the local CT value so that the difference between it and the CT value of the corresponding area on the input image is within the clinically acceptable range, and obtain the enhanced CT image; save the final result in dicom format, the calculation method of the MeanShift algorithm is:

[0015] output=MeanShift(input-output')+output'

[0016] MeanShift is a local mean convolution kernel with a kernel size of 51 and a step size of 1. Input, output', and output represent the input image, the output CT image after the denormalization step, and the enhanced CT image, respectively.

[0017] Preferably, when constructing the training data set, the data set contains a sufficient number of CT images scanned by machines from various manufacturers, with various resolutions and doses. Doctors will divide all CT images into high-noise level sub-datasets and low-noise level sub-datasets from a diagnostic perspective based on their clinical film reading experience.

[0018] Preferably, during data preprocessing, a fixed minimum-maximum normalization operation is applied to all CT images in the data set, i.e.

[0019] img=(img'-v min ) / (v max -v min )

[0020] Among them, img', img represent the images before and after normalization, v min , v max They are 0 and 4095 respectively, representing the maximum data range of CT images in dicom format.

[0021] Preferably, each CT image is optimized using traditional methods and confirmed by doctors to meet clinical requirements as the corresponding gold standard. A deep learning network SCUNet with strong generalization ability and short processing time is trained and set as a denoising network.

[0022] Preferably, a super-resolution network is used to improve the quality of CT images with low noise levels, and a SwinIR natural image super-resolution network is used to train the network using a pre-trained model to perform super-resolution on CT images in a low noise level sub-dataset as a gold standard.

[0023] Preferably, when training the noise level classification network, the attributes representing the CT dose in the DICOM are read, including but not limited to the exposure type, exposure time, tube voltage, and tube current, and these attributes are matched with the noise level classification results determined by the doctor, and used as the training set to train the MLP network to obtain the noise level classification model.

[0024] Preferably, the CT value reflects the degree of absorption of X-rays by different tissues. Different tissues and lesions, including but not limited to fat, muscle, blood, bone and tumor, have different CT values. The MeanShift algorithm is used to reduce the local CT value difference between the input and output images through convolution filtering to reach the clinically accepted difference range standard.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. The present invention optimizes the quality of CT images by determining the noise level of CT images and using different deep learning algorithms, namely denoising or super-resolution models. Compared with the existing single-model denoising or enhancement methods, the present invention has stronger generalization ability and can cope with more complex and changeable input data.

[0027] 2. The present invention uses a variety of traditional filtering methods and super-resolution pre-training models to process noisy data to obtain the gold standard image required for training. Therefore, there is no need to spend time and effort in advance to collect high-quality gold standard images, and there is no data registration problem.

[0028] 3. The present invention calibrates the CT value of the output image through post-processing operations to ensure that the difference in CT values ​​before and after enhancement does not affect the accuracy of clinical diagnosis. Compared with other methods without CT value calibration, the present invention has more practical application value.

[0029] 4. The noise of the CT image processed by the method of the present invention is significantly reduced, the edges are clear, and the quality is close to or even exceeds that of a real high-quality CT image. In practical applications, only low-dose CT images are collected, and the effect of a full-dose CT image can be obtained after enhancement, which can reduce the radiation damage to patients during the scanning process and is conducive to improving the efficiency of doctors' reading of films. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a flow chart of the noise level-based differentiated CT image enhancement method of the present invention;

[0031] Figure 2 It is a CT image enhancement module and data post-processing diagram of the present invention;

[0032] Figure 3 The following are comparison diagrams of CT images before and after enhancement using the present invention, wherein Figure A.1 is the input original CT image with high noise level, Figure A.2 is the CT image enhanced by the denoising model, Figure B.1 is the input CT image with low noise level, and Figure B.2 is the CT image enhanced by the super-resolution model. DETAILED DESCRIPTION

[0033] The following embodiments of the present invention are described in further detail in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0034] As attached Figure 1 To Attachment Figure 3 As shown:

[0035] Embodiment 1: The present invention provides a method for differential CT image enhancement based on noise level, comprising the following steps: S1, constructing a training data set: obtaining CT images in dicom format with various noise levels; dividing the data set into a high noise level sub-data set and a low noise level sub-data set according to the clinician's suggestion. Taking a single-channel CT image with a size of 512×512 as an example, the obtained high and low noise images are as follows: Figure 3 .A.1,3.B.1;

[0036] S2. Data preprocessing: normalize the CT image and adjust the data range to between 0 and 1.

[0037] S3. Build and train the denoising network: Select SCUNet (Swin-Conv-UNet) as the denoising network; use a variety of traditional unsupervised denoising methods to manually adjust the parameters of the CT images in the high-noise level sub-dataset for denoising. After the doctor confirms that it meets the clinical requirements, it is used as the gold standard image to obtain a paired training data set; train the network, and obtain the final denoising model after convergence.

[0038] Among them, the unsupervised denoising method includes one or a combination of micro-BM3D filtering, discrete cosine transform, wavelet sharpening, Perona-Malik filtering and TV filtering.

[0039] S4. Build and train a super-resolution network: Select the LIFF (Local Implicit Image Function) network as the super-resolution network; use the pre-trained model of the SwinIR (Transformer for Image Restoration) image super-resolution network to perform super-resolution on the CT images in the low-noise level sub-dataset, and use them as gold standard images after doctors confirm that they meet clinical requirements; train the network and obtain the final super-resolution model after convergence.

[0040] Among them, the 2x or 4x pre-trained model of SwinIR is used; for the gold standard image obtained using the 4x pre-trained model, the Lancoz method is used to perform 2x image scaling to reduce the image size.

[0041] S5. Construct and train a noise level classification network: Design a multi-layer perceptron (MLP) as the classification network; read the attributes representing the CT dose in the dicom in the training data set, match the attribute values ​​with the corresponding noise levels determined by the doctor, and make a classification network training set; train the network, and obtain the final noise level classification model after convergence.

[0042] Among them, the attributes representing CT dose include exposure type, exposure time, tube voltage, tube current, etc.

[0043] S6, CT image enhancement: Given a CT image in DICOM format, normalize it using the data preprocessing method described in S1; read the attribute representing the CT dose in DICOM, and use the noise level classification network described in S5 to determine its noise level; if the noise level is high, use the denoising model for inference, otherwise use the super-resolution model for inference to obtain the output CT image.

[0044] S7, data post-processing: For the output CT image, first adjust its size to be consistent with the input; then crop its data range to between 0 and 1, and denormalize it to the data range of the input image; then use the MeanShift algorithm to adjust the local CT value so that the difference between it and the CT value of the corresponding area on the input image is within the clinically acceptable range, and obtain the enhanced CT image; save the final result in dicom format. The calculation method of the MeanShift algorithm is,

[0045] output=MeanShift(input-output')+output'where MeanShift is a local mean convolution kernel with a kernel size of 51 and a step size of 1. input, output', and output represent the input image, the output CT image after the denormalization step, and the enhanced CT image, respectively.

[0046] Embodiment 2: This embodiment is basically the same as the previous embodiment, except that:

[0047] In the data preprocessing step, the images are normalized to other ranges, such as between -1 and 1.

[0048] When building a super-resolution network and training it, you can choose Stable-Diffusion as the super-resolution network.

[0049] As can be seen from the above, the present invention can improve the quality of CT images without collecting high-quality gold standard images as training data in advance. The noise of the CT images after quality improvement is significantly reduced, the edges are clear, and the quality is close to or even exceeds that of real high-quality CT images.

[0050] High-quality CT can improve the efficiency of doctors' clinical diagnosis. However, in practical applications, high-quality CT is often not available due to factors such as lack of equipment, high price or radiation. The present invention optimizes low-quality CT images using an unsupervised method to obtain high-quality CT images, and uses them as the gold standard to train a deep learning network. The quality of low-dose CT images enhanced by the method of the present invention can reach or even exceed that of CT images scanned with standard doses, greatly reducing the radiation damage to patients during the scanning process.

[0051] As can be seen from the above, the present invention optimizes the quality of CT images by determining the noise level of CT images, using different deep learning algorithms, i.e., denoising or super-resolution models, with stronger generalization ability and capable of coping with more complex and variable input data; uses a variety of traditional filtering methods and super-resolution pre-training models to process noisy data to obtain the gold standard image required for training, so there is no need to spend time and energy in advance to collect high-quality gold standard images; calibrates the CT value of the output image through post-processing operations to ensure that the difference in CT values ​​before and after enhancement does not affect the accuracy of clinical diagnosis, and has more practical application value compared to other methods that do not calibrate CT values; the noise of the CT images processed by the method of the present invention is significantly reduced, the edges are clear, and the quality is close to or even exceeds that of real high-quality CT images. In practical applications, only low-dose CT images are collected, and the effect of full-dose CT images can be obtained after enhancement, which can reduce the radiation damage to patients during scanning and help improve the efficiency of doctors' film reading.

[0052] Additionally, in order to provide a concise description of exemplary embodiments, all features of an actual embodiment (ie, those features that are not relevant to the best mode presently contemplated for carrying out the invention or those that are not relevant to implementing the invention) may not be described.

[0053] It will be appreciated that in the development of any actual implementation, as in any engineering or design project, numerous implementation-specific decisions may be made. Such a development effort may be complex and time-consuming, but will be a routine task of design, fabrication, and production for those of ordinary skill having the benefit of this disclosure without undue experimentation.

[0054] 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 preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. Differentiated CT image enhancement method based on noise level, It is characterized in that The following steps are involved: S1. Construct training dataset: obtain CT images in dicom format with various noise levels; divide the dataset into high noise level sub-dataset and low noise level sub-dataset according to the clinician's suggestion; S2, data preprocessing: normalize the CT image and adjust the data range to between 0 and 1; S3. Build and train the denoising network: Select SCUNet as the denoising network; use a variety of traditional unsupervised denoising methods to manually adjust the parameters of the CT images in the high-noise level sub-dataset for denoising. After the doctor confirms that it meets the clinical requirements, it is used as the gold standard image to obtain a paired training data set; train the network and obtain the final denoising model after convergence; S4. Build and train a super-resolution network: Select the LIFF network as the super-resolution network; use the pre-trained model of the SwinIR image super-resolution network to perform super-resolution on the CT images in the low-noise level sub-dataset, and use them as gold standard images after doctors confirm that they meet clinical requirements; Train the network and obtain the final super-resolution model after convergence; S5. Construct and train a noise level classification network: Design a multi-layer perceptron as a classification network; read the attributes representing CT dose in the DICOM in the training data set, match the attribute values ​​with the corresponding noise levels determined by the doctor, and make a classification network training set; Train the network and obtain the final noise level classification model after convergence; S6, CT image enhancement: given a CT image in dicom format, normalize it using the data preprocessing method described in S1; Read the attribute representing the CT dose in DICOM, and use the noise level classification network described in S5 to determine its noise level; if the noise level is high, use the denoising model for reasoning, otherwise use the super-resolution model for reasoning to obtain the output CT image; S7, data post-processing: For the output CT image, first adjust its size to be consistent with the input; then crop its data range to between 0 and 1, and denormalize it to the data range of the input image; then use the MeanShift algorithm to adjust the local CT value so that the difference between it and the CT value of the corresponding area on the input image is within the clinically acceptable range, and obtain the enhanced CT image; save the final result in dicom format, the calculation method of the MeanShift algorithm is: output=MeanShift(input-output')+output', where MeanShift is a local mean convolution kernel with a kernel size of 51 and a step size of 1. Input, output', and output represent the input image, the output CT image after the denormalization step, and the enhanced CT image, respectively.

2. The method for differential CT image enhancement based on noise level according to claim 1, Features: When constructing the training data set, the data set contains a sufficient number of CT images scanned by machines from various manufacturers, with various resolutions and doses. Doctors will divide all CT images into high-noise level sub-datasets and low-noise level sub-datasets from a diagnostic perspective based on their clinical film reading experience.

3. The method for differential CT image enhancement based on noise level as claimed in claim 1, Features: During data preprocessing, a fixed minimum-maximum normalization operation is used for all CT images in the dataset, that is, img=(img’-v min ) / (v max -v min ) Among them, img', img represent the images before and after normalization, v min， v max They are 0 and 4095 respectively, representing the maximum data range of CT images in di com format.

4. The method for differential CT image enhancement based on noise level according to claim 1, Features: After each CT image is optimized using traditional methods and confirmed by doctors to meet clinical requirements, it is used as the corresponding gold standard. A deep learning network SCUNet with strong generalization ability and short processing time is trained and set as a denoising network.

5. The noise level-based differential CT image enhancement method according to claim 1, Features: A super-resolution network is used to improve the quality of CT images with low noise levels, and the SwinIR natural image super-resolution network is used to train the network using a pre-trained model to perform super-resolution on CT images in the low-noise level sub-dataset as the gold standard.

6. The method for differential CT image enhancement based on noise level according to claim 1, Features: When training the noise level classification network, the attributes representing the CT dose in the DICOM are read, including but not limited to the exposure type, exposure time, tube voltage, and tube current. These attributes are matched with the noise level classification results determined by the doctor and used as the training set to train the MLP network to obtain the noise level classification model.

7. The method for differential CT image enhancement based on noise level according to claim 1, Features: The CT value reflects the degree of X-ray absorption by different tissues. Different tissues and lesions, including but not limited to fat, muscle, blood, bone and tumor, have different CT values. The MeanShit algorithm is used to reduce the local CT value differences between the input and output images through convolution filtering to meet the clinically accepted difference range standards.

Images