CN118096920B - Abdominal organ imaging method and device based on deep learning and coil sensitivity - Google Patents

Abstract

The present invention discloses an abdominal organ imaging method and device based on deep learning and coil sensitivity, which belongs to the field of magnetic resonance imaging, and the steps include: collecting high-resolution diffusion-weighted data; reconstructing images by traditional high-resolution diffusion reconstruction methods; selecting high-quality images; generating high-resolution images contaminated by artifacts corresponding to the selected images; inputting the generated paired images into a deep learning network for generalization and training; inputting the abdominal organ multiple-excitation diffusion-weighted images based on traditional high-resolution diffusion reconstruction into a network model constructed by S5, and obtaining high-resolution diffusion-weighted images of abdominal organs with high signal-to-noise ratio, high resolution and few residual artifacts; and an imaging device is also disclosed. The present invention adopts the above-mentioned abdominal organ imaging method and device based on deep learning and coil sensitivity, improves the quality of the reconstructed high-resolution diffusion-weighted images, and provides more detailed anatomical information for clinical diagnosis.

Description

Abdominal organ imaging method and device based on deep learning and coil sensitivity

Technical Field

The present invention relates to the technical field of magnetic resonance imaging, and in particular to an abdominal organ imaging method and device based on deep learning and coil sensitivity.

Background Art

The conventional diffusion-weighted image acquisition method in clinical practice is achieved through single-shot echo planar imaging (SSH-EPI), but its image spatial resolution is low, resulting in severe geometric distortion and low image clarity after reconstruction. Therefore, the current improvement strategy is to achieve it through multiple excitations (along the phase encoding direction or the frequency encoding direction), such as the multi-shot echo planar imaging diffusion weighted imaging (MSH-EPIDWI) technology, which completes the acquisition with higher spatial resolution, thereby reconstructing less geometric deformation and high-resolution images, providing more detailed anatomical information for clinical diagnosis.

However, for multiple excitation technology, there will be linear and nonlinear phase changes between different excitations, and the diffusion gradient will amplify this phase difference and thus cause aliasing artifacts in the image reconstruction of multiple excitations. Compared with other parts such as the head and limbs, the abdominal organs will inevitably have this phase difference aggravated during imaging due to the presence of movements such as breathing and heartbeat, thereby introducing more motion artifacts, making high-resolution diffusion imaging of abdominal organs extremely challenging.

At present, traditional high-resolution diffusion reconstruction methods, such as MUSE and a series of variant algorithms: POCS-MUSE, self-feeding MUSE and DL-MUSE are all reconstruction methods for high-resolution diffusion imaging of parts with very small movements such as the brain. However, in applications of abdominal organs with violent and complex movements, the effect is not ideal and there are still very serious image artifacts.

Summary of the invention

The purpose of the present invention is to provide an abdominal organ imaging method and device based on deep learning and coil sensitivity. By performing deep learning network training and model construction on paired high-quality images with fewer artifacts and corresponding images with artifacts, the quality of reconstructed high-resolution diffusion-weighted images is improved, thereby providing more detailed anatomical information for clinical diagnosis.

To achieve the above object, the present invention provides an abdominal organ imaging method based on deep learning and coil sensitivity, the steps comprising:

S1, collecting high-resolution diffusion-weighted data of abdominal organs with multiple excitations;

S2, reconstructing the collected data using a traditional high-resolution diffusion reconstruction method;

S3, select images with fewer artifacts and higher quality after reconstruction;

S4, using the image selected in step S3 and the coil sensitivity map corresponding to the selected image, performing motion artifact simulation of abdominal organs and aliasing artifact simulation of the image reconstruction process itself, and generating a corresponding high-resolution diffusion image contaminated by artifacts;

S5, inputting the paired images generated in step S3 and step S4 into the deep learning network for generalization and training, and constructing a network model for high-resolution diffuse image artifact correction;

S6. Inputting the multi-excitation diffusion-weighted image of the abdominal organs based on traditional high-resolution diffusion reconstruction into the network model constructed in step S5, to obtain a high-resolution diffusion-weighted image of the abdominal organs with high signal-to-noise ratio, high resolution and few residual artifacts.

Preferably, the step S3 specifically includes:

S31, segmenting the reconstructed image;

S32, calculating image evaluation indicators based on the pre-trained segmentation model;

S33, performing image screening according to a certain threshold of the image evaluation index.

Preferably, the image screening is based on a combination of one or more similar image evaluation indicators such as artifact signal ratio GSR, signal-to-noise ratio SNR, and contrast-to-noise ratio CNR as an image quality screening method.

Preferably, problems existing in the abdominal organ multiple excitation diffusion weighted image based on traditional high-resolution diffusion reconstruction in step S6 include but are not limited to severe artifacts, low signal-to-noise ratio, signal loss due to motion, and image blur.

An abdominal organ imaging device based on deep learning and coil sensitivity, comprising an image preparation module, a motion simulation module and an image correction module;

An image preparation module, used for exciting, collecting and reconstructing magnetic resonance signals of the target object for multiple excitation diffusion data;

Motion simulation module, used to simulate various motion conditions of the target under test;

The image correction module is used to correct image artifacts of the high-resolution diffusion-weighted image using a correction algorithm or a network model according to the information provided by the image preparation module and the motion simulation module.

Preferably, the image preparation module includes a high-resolution diffusion data reconstruction unit and an image screening unit;

A high-resolution diffusion data reconstruction unit, used for reconstructing a high-resolution diffusion-weighted image;

The image screening unit is used to screen high-quality high-resolution diffusion-weighted images.

Preferably, the motion simulation module includes an object motion simulation unit and a reconstruction motion simulation unit;

An object motion simulation unit, used to directly simulate the influence of the motion of the measured object on the collected signal by using simulated motion;

The reconstruction motion simulation unit is used to use simulated motion to simulate the influence of the motion of the abdominal organ under test on the high-resolution diffusion reconstruction method itself.

Preferably, the image correction module includes a module for calculating and correcting the amplitude, artifacts and phase of the image itself.

Therefore, the present invention adopts the above-mentioned structure based on deep learning and coil sensitivity of the abdominal organ imaging method and device, which has the following beneficial effects:

(1) Using deep learning to replace the approximate solution method based on mathematical models in traditional high-resolution diffusion reconstruction methods can more accurately and effectively eliminate artifacts in the high-resolution diffusion-weighted image reconstruction of more challenging abdominal organs, and the reconstruction results are of higher quality and faster speed;

(2) The image quality and acceleration factor of the traditional high-resolution diffusion imaging reconstruction method are restricted by the g-factor. Therefore, in the application of multiple diffusion-weighted sequences, the number of excitations is limited, which limits the improvement of resolution. The deep learning-based method proposed in the present invention is not restricted by this model calculation and can be effective for applications with higher excitation times and higher resolution.

(3) A mature and robust network model will greatly reduce the reconstruction time of abdominal organs and improve their reconstruction accuracy.

The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is an overall flow chart and a specific network structure diagram of liver imaging according to an embodiment of the present invention;

FIG2 is a schematic diagram showing the comparison of the results of different reconstruction methods at multiple b values (0, 600 and 800 s/mm ² ) of an angiomyolipoma patient with dyslipidemia according to an embodiment of the present invention;

FIG3 is a comparison of representative reconstruction results of different reconstruction methods at multiple b values (0, 600 and 800 s/mm ² ) of two healthy subjects in an embodiment of the present invention;

FIG4 is a schematic diagram of the structure of an abdominal organ imaging device based on deep learning and coil sensitivity according to an embodiment of the present invention.

DETAILED DESCRIPTION

Example

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

1-3, a method for abdominal organ imaging based on deep learning and coil sensitivity includes the steps of:

S1. High-resolution diffusion-weighted data of abdominal organs were collected from 50 healthy subjects with multiple excitations, and single and four-shot data were collected for each subject.

S2. The collected data is reconstructed using the traditional high-resolution diffusion reconstruction method. After reconstruction, the image has many aliasing or motion artifacts and cannot be used for clinical diagnosis.

S3, select the image with few artifacts and high quality after reconstruction as the preparation image for step S4. Specifically, for each subject's liver diffusion-weighted image data, take the image data containing 3 b values (taking 0, 600 and 800s/ ^mm2 as examples) and covering 20 layers of the whole liver scan along the head angle direction as an example, for the reconstructed diffusion-weighted image data, select its b=0s/ ^mm2 non-diffusion-weighted image as an example to perform image segmentation to obtain the liver mask, as shown in Figure 1, using the pre-trained nnUNet model as an example for graphic segmentation; at the same time, for this batch of healthy subjects, all liver diffusion-weighted data with different b value weightings and different layers are mixed, and the pre-trained nnUNet model is used to calculate the GSR evaluation index of all high-resolution diffusion-weighted images, and the 20% threshold of the GSR evaluation index of the liver area is used as an example for screening.

S4, using the image obtained in step S3 and the coil sensitivity map corresponding to the selected image, perform abdominal organ motion simulation and aliasing artifact simulation in the image reconstruction process to generate a corresponding high-resolution image contaminated by artifacts, specifically:

For the simulation of abdominal organ motion, as shown in the liver motion simulation part of FIG1, various rigid and non-rigid motion simulations based on geometric space (including but not limited to translation, rotation, deformation, acceleration, etc.) are used in consideration of the non-uniformity of abdominal organ motion (respiration and heartbeat, etc.) during data acquisition. At the same time, in order to ensure the high quality of the final image reconstruction, the motion of the coil sensitivity map (CSM) corresponding to the abdominal image is added in consideration of the influence of the magnetic field caused by such motion on the non-uniformity of the reconstructed image. Finally, a high-resolution diffusion-weighted image contaminated by artifacts that fully simulates the influence of motion is generated. In this embodiment, three types of motion simulation are used: the displacement of the image along the x or y axis is less than 3 pixels, the rotation of the image along the x or y axis is less than 3 degrees, and the deformation simulation with elastic Gaussian characteristics. In view of the influence of abdominal breathing and cardiac motion on field non-uniformity during image acquisition, three types of motion simulation are also used for the coil sensitivity map CSM: the displacement of the CSM along the x or y axis is less than 3 pixels, the rotation of the CSM along the x or y axis is less than 3 degrees, and the elastic Gaussian deformation simulation of the CSM. Taking into account the inconsistency of the actual motion between different excitations during data acquisition, for the reconstruction of the data images of the four excitations, two of the three motions were randomly selected for each excitation image and CSM motion to ensure a more realistic presentation of the physiological dynamic characteristics.

S5. Input the paired images (high quality and motion contaminated) of 50 healthy subjects generated in steps S3 and S4 into the deep learning network for generalization and training, and construct a network model for high-resolution diffusion image artifact correction.

Specifically, the overall structure of the artifact correction network can be implemented by Pytorch using a 4-level encoder-decoder supervision network, and the training can be performed on an NVIDIA Tesla P100×4 GPU (each processor has 16GB of memory) as an example, or it can be implemented using other programming or training platforms familiar to those skilled in the art. The weights of the convolutional layer are initialized using a Gaussian random distribution as an example, and the AdamW optimizer is used to remove the artifact network as an example, or other deep learning optimizers familiar to those skilled in the art can be used. The training process includes using L1 loss for 150,000 iterations as an example, starting from an initial learning rate of 3× ^10-4 as an example, and gradually reducing it to ^10-6 through cosine annealing as an example. All weights can be initialized using the Kaiming method or other initialization methods familiar to those skilled in the art, and updated under the supervision of L1 loss between high-quality images and motion simulation images. The entire training time is about 46 hours, and the model is selected based on the loss metric.

S6. Input the multiple-excitation diffusion-weighted images of abdominal organs based on traditional high-resolution diffusion reconstruction, which have serious artifacts, low signal-to-noise ratio, signal loss and image blur due to movement, etc., into the network model trained in step S5, and perform training and generalization performance optimization to obtain high-resolution diffusion-weighted images of abdominal organs with high signal-to-noise ratio, high resolution and few residual artifacts.

Specifically, the high-resolution diffusion-weighted image data of the first batch of 10 patients and 10 healthy subjects were used as an example to test the multi-channel data of liver diffusion-weighted data. The traditional high-resolution liver diffusion image was used for reconstruction, and the image would have many artifacts and could not be used for clinical diagnosis. The data of this new batch of livers and the remaining 80% of the data after high-quality image screening of the first batch of 50 healthy subjects were used as parameter inputs of the deep learning network model for generalization and training, and finally high-resolution, high-precision, and artifact-free liver high-resolution diffusion-weighted images that can be used for clinical diagnosis were obtained.

The high-resolution diffusion-weighted image data with four excitations were used to perform traditional high-resolution reconstruction on the data corresponding to the second batch of 10 healthy and 10 patient subjects and 80% of the healthy subjects after the first screening, and the high-resolution diffusion-weighted image of the liver was obtained through the trained network model.

The high-resolution diffusion-weighted image and low-resolution image of the liver were subjected to qualitative and quantitative comparative analysis. Through observation, it was found that the image quality obtained by the method provided by the present invention was significantly improved (with the increase of b value, the occurrence of artifacts and image blurring caused by breathing and cardiac movement, especially the signal loss in the left lobe of the liver, etc., was significantly reduced and eliminated).

Figures 2 and 3 show the comparison results on a patient and two healthy subjects, respectively. Figure 2 is a schematic diagram showing the comparison of the results of different reconstruction methods at multiple b values (0, 600 and 800 s/mm2) based on the traditional high-resolution diffusion reconstruction method (C-E), the clinically commonly used low-resolution reconstruction method (G-I) and the high-resolution diffusion reconstruction method (K-M) using the motion artifact correction of the present invention for a patient with lipid dyslipidemia. The reference images are T2-weighted (A), T1-weighted (B) and DCE images (F is the arterial phase, J is the portal phase, and N is the delayed phase) of routine clinical scans. The solid red arrow points to a smaller lesion at the edge of the liver, the hollow red arrow points to an obvious motion artifact, and the hollow blue arrow points to a place where the signal is missing and the signal intensity is wrong due to the presence of motion artifacts. In Figure 3, the high-resolution diffusion reconstruction method using the motion artifact correction network (the first row of A and the first row of B) and the low-resolution diffusion reconstruction method commonly used in clinical practice are shown. The hollow red arrows point to the areas where the traditional high-resolution method causes signal loss and signal intensity errors due to motion artifacts in the abdomen, especially in the left lobe of the liver near the heart. By comparing the reconstruction results, it can be seen that the method provided by the present invention provides excellent image quality for the diagnosis of clinical liver diseases, and can accurately depict small cysts, malignant and benign tumors, even in severe iron overload and poor liver contrast. Its enhanced diagnostic capability is also extended to small-sized lesions or regions of interest located in the left lobe of the liver, near blood vessels, or affected by heart and blood vessel pulsation.

Referring to FIG. 4 , an abdominal organ imaging device based on deep learning and coil sensitivity includes:

The image preparation module is used to excite, collect and reconstruct the magnetic resonance signals of the target object for multiple excitation diffusion data, and prepare high-quality high-resolution diffusion data at the same time. The image preparation module includes a high-resolution diffusion data reconstruction unit, which is used to reconstruct traditional high-resolution liver organ diffusion weighted images; an image screening unit, which uses a screening dimension based on the artifact signal ratio GSR to screen high-quality high-resolution diffusion images with a threshold of 20% as an example.

The motion simulation module is used to simulate various motion conditions of the target to be measured, so as to simulate the influence of the motion of the measured object during the magnetic resonance signal acquisition process on the actual image reconstruction. The motion simulation module includes an object motion simulation unit, which is used to directly simulate the influence of the liver's breathing, heartbeat and other movements on the data acquisition signal by using the image itself, including but not limited to translation, rotation, deformation or acceleration and other motion simulation forms; a reconstruction motion simulation unit, which is used to simulate the influence of the above rigid or non-rigid body motion on the reconstruction process itself by using the coil sensitivity CSM, including but not limited to translation, rotation, deformation or acceleration and other motion simulation forms.

The image correction module is used to correct image artifacts of high-resolution diffusion-weighted images using correction algorithms or network models based on the information provided by the image preparation module and the motion simulation module to improve the quality of the image. The image correction module includes a module for calculating and correcting the amplitude, artifacts, and phase of the image itself.

In the description of the present invention, it is necessary to understand that the orientations or positional relationships indicated by terms such as "x-axis", "y-axis", "translation", "rotation", and "deformation" are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the methods, devices, or elements referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solution 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 they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims (7)

1. An abdominal organ imaging method based on deep learning and coil sensitivity, characterized in that the steps include: S1, collecting high-resolution diffusion-weighted data of abdominal organs with multiple excitations; S2, reconstructing the collected data using a traditional high-resolution diffusion reconstruction method; S3, select images with fewer artifacts and higher quality after reconstruction; S4, using the image selected in step S3 and the coil sensitivity map corresponding to the selected image, performing motion artifact simulation of abdominal organs and aliasing artifact simulation of the image reconstruction process itself, and generating a corresponding high-resolution diffusion image contaminated by artifacts; S5, inputting the paired images generated in step S3 and step S4 into the deep learning network for generalization and training, and constructing a network model for high-resolution diffuse image artifact correction; S6. Inputting the multi-excitation diffusion-weighted image of the abdominal organs based on traditional high-resolution diffusion reconstruction into the network model constructed in step S5, to obtain a high-resolution diffusion-weighted image of the abdominal organs with high signal-to-noise ratio, high resolution and few residual artifacts. 2. The abdominal organ imaging method based on deep learning and coil sensitivity according to claim 1, characterized in that step S3 specifically comprises: S31, segmenting the reconstructed image; S32, calculating image evaluation indicators based on the pre-trained segmentation model; S33, performing image screening according to a certain threshold of the image evaluation index. 3. The abdominal organ imaging method based on deep learning and coil sensitivity according to claim 2 is characterized in that: the image screening is based on one or more combinations of artifact signal ratio GSR, signal-to-noise ratio SNR, and contrast-to-noise ratio CNR as an image quality screening method. 4. An abdominal organ imaging device based on deep learning and coil sensitivity, applying the abdominal organ imaging method based on deep learning and coil sensitivity as described in any one of claims 1 to 3, characterized in that it includes an image preparation module, a motion simulation module and an image correction module; An image preparation module, used for exciting, collecting and reconstructing magnetic resonance signals of the target object for multiple excitation diffusion data; Motion simulation module, used to simulate various motion conditions of the target under test; The image correction module is used to correct image artifacts of the high-resolution diffusion-weighted image using a correction algorithm or a network model according to the information provided by the image preparation module and the motion simulation module. 5. The abdominal organ imaging device based on deep learning and coil sensitivity according to claim 4, characterized in that: the image preparation module includes a high-resolution diffusion data reconstruction unit and an image screening unit; A high-resolution diffusion data reconstruction unit, used for reconstructing a high-resolution diffusion-weighted image; The image screening unit is used to screen high-quality high-resolution diffusion-weighted images. 6. The abdominal organ imaging device based on deep learning and coil sensitivity according to claim 4, characterized in that: the motion simulation module includes an object motion simulation unit and a reconstruction motion simulation unit; An object motion simulation unit, used to directly simulate the influence of the motion of the measured object on the collected signal by using simulated motion; The reconstruction motion simulation unit is used to use simulated motion to simulate the influence of the motion of the abdominal organ under test on the high-resolution diffusion reconstruction method itself. 7. The abdominal organ imaging device based on deep learning and coil sensitivity according to claim 4 is characterized in that: the image correction module includes a module for calculating and correcting the amplitude, artifacts, and phase of the image itself.