An ultrasonic tomography spiral down-sampling data recovery method based on deep learning

Abstract

The application discloses a kind of based on deep learning's ultrasonic tomography spiral subsampling data recovery method, comprising: constitute complete data set, the complete data body obtained is carried out spiral sparse subsampling, establishes and trains neural network model, using spiral subsampling strategy acquisition radio frequency data body, subsampling data are input into the neural network model trained, obtain recovery data body and using the recovery data body is reconstructed by traditional ultrasonic delay superposition image reconstruction algorithm high-quality image etc. Step, it has solved the problem that the result error is caused by the fact that the scanning time is relatively long and the motion and breathing artifact is easily introduced in the current ultrasonic tomography imaging due to containing a large number of elements, and the data transmission, storage and processing cost is improved due to the large amount of original data.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a method for recovering ultrasound tomographic spiral downsampling data based on deep learning. Background Technology

[0002] Ultrasound tomography (UTC) systems offer standardized operation and advantages such as high repeatability and the ability to provide higher resolution reflective tomographic images, thus holding broad clinical application prospects. Reflection modes are common in ultrasound imaging. UTC systems typically employ a step-by-step scanning approach to acquire information from different tissue layers, and then generate two-dimensional images of the tissue through subsequent reconstruction processes, visually displaying lesion information. Compared to traditional ultrasound imaging, the standardized operating procedures of UTC systems reduce reliance on the experience of sonographers, effectively promoting the standardized application of ultrasound technology. However, UTC systems contain a large number of array elements, and obtaining a complete dataset under traditional data acquisition modes presents the following problems: 1) Long scan time: Limited by the propagation speed of ultrasound waves in human tissue (average 1540 m / s), the scan time is proportional to the number of emission events, resulting in a single-layer scan taking several seconds. 2) Huge raw data volume: The size of a complete dataset can reach 10 GB, which places higher demands on the system's data transmission, storage, and processing capabilities, and also increases the cost of the UTC system. 3) Generally, ultrasound tomography systems require sufficient transmit and receive data pairs to reconstruct high-quality images. However, the longer the scan time, the larger the data volume, which can easily introduce motion and breathing artifacts.

[0003] To address these issues, imaging speed can be accelerated by reducing the amount of data acquired. However, using downsampled sparse data for reconstruction introduces noise and sidelobe artifacts, reducing the resolution and contrast of the final image. One approach is to utilize deep learning techniques to restore the image reconstructed from sparse data; however, such images often exhibit over-smoothing, erasing detailed textures.

[0004] To address the aforementioned problems, the inventors, combining the characteristics of data transmission and acquisition in ultrasound computed tomography systems, proposed a spiral data downsampling strategy and a deep learning-based data recovery method to reconstruct a complete dataset from downsampled data. This scheme employs an improved data acquisition mode, namely sparse data acquisition. Compared to traditional complete data acquisition, sparse data acquisition selects only a subset of array elements for data acquisition, thus significantly reducing scan time and data volume. However, sparse data acquisition can lead to noise and sidelobe artifacts in the image. To solve this problem, this invention employs deep learning data recovery technology. Specifically, sparse data is input into a deep neural network for data recovery by training the network. The deep neural network can learn and infer information from the missing data, thereby recovering the data and reconstructing a more accurate image. In this way, even using sparse data, images with good resolution and contrast can be obtained, and detailed textures are preserved. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based method for recovering ultrasonic tomographic spiral downsampling data, in order to solve the problems mentioned in the background art, such as the long scanning time in ultrasonic tomographic imaging due to the large number of array elements, which easily introduces motion and breathing artifacts and thus leads to result errors, and the large amount of original data, which increases the cost of data transmission, storage and processing.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for recovering ultrasonic tomographic spiral downsampling data based on deep learning, comprising the following steps:

[0007] Step S1: Obtain a sufficient number of complete radio frequency data volumes containing complete information for reconstructing high-quality images, thus forming a complete dataset:

[0008] Step S2: Perform spiral sparse downsampling on the acquired complete data volume;

[0009] Step S3: Using spiral sparse downsampled data as input and the corresponding complete data volume as target, train a deep neural network model through supervised learning; through training with a sufficient number of sets of data, the network model can learn the mapping relationship between downsampled data and complete data;

[0010] Step S4: Acquire radio frequency data using a spiral downsampling strategy;

[0011] Step S5: Input the downsampled data into the trained neural network model to obtain the recovered data volume;

[0012] Step S6: Using the recovered data volume, a high-quality image is reconstructed using a traditional ultrasonic time-lapse image reconstruction algorithm.

[0013] As a preferred technical solution, the size of the complete radio frequency data volume is (N / 4)*N*M, where N is the number of array elements of the ring transducer and M is the number of sampling points for each data acquisition.

[0014] As a preferred technical solution, spiral downsampling means that for each array element's transmission event, the system only collects the echo data received by a specific array element. That is, for the nth array element's transmission, only the echo data received by the nth, n+m, n+2m... array elements are collected, where n is a positive integer, in order to reduce the amount of data and ensure that the data from each receiving channel can participate in data reconstruction.

[0015] As a preferred technical solution, m=4.

[0016] As a preferred technical solution, in step S3, the neural network updates the network model parameters by calculating the error between the recovered data and the complete data output by the network and backpropagating the error.

[0017] As a preferred technical solution, in step S4, a spiral downsampling strategy is used to acquire an RF data volume of size (N / 4)*(N / 4)*M, and then the downsampled data is input into a trained neural network model to obtain a recovered data volume of size (N / 4)*N*M.

[0018] Compared with the prior art, the beneficial effects of the present invention are:

[0019] 1) The spiral downsampling strategy reduces the amount of data acquired by the ultrasound tomography system, lowers the requirements for data transmission, storage and processing capabilities, and also reduces the cost of the system.

[0020] 2) The sparse data acquisition mode reduces scanning time, thereby speeding up imaging and improving imaging efficiency.

[0021] 3) A novel deep learning-based data restoration method is proposed, which reconstructs high-quality images using a small amount of data. Even when using sparse data for reconstruction, images with good resolution and contrast can be obtained. Compared with conventional image restoration methods, it can effectively avoid the problem of overly smooth images. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the complete data, uniform downsampling data, and spiral downsampling data of the present invention;

[0023] Figure 2 This is a schematic diagram of spiral downsampling data recovery according to the present invention;

[0024] Figure 3 This is a flowchart of the network model training process of the present invention;

[0025] Figure 4 This is a schematic diagram illustrating the application of the network model of the present invention;

[0026] Figure 5 This is the result of recovering the received signal from a single channel during a single transmission, as shown in the example.

[0027] Figure 6 This is a diagram showing the signal recovery results of all channels received in a single transmission, as illustrated in the example.

[0028] Figure 7 This is a comparison image of the complete data volume reconstruction image, the spiral downsampling data volume reconstruction image, and the network output data volume reconstruction image in the embodiment.

[0029] Figure 8 This is a flowchart of a deep learning-based ultrasonic tomographic spiral downsampling data recovery method according to the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Please see Figure 8 This invention provides a technical solution: a deep learning-based method for recovering ultrasonic tomographic spiral downsampling data, comprising the following steps: acquiring a sufficient number of complete radio frequency data volumes containing complete information for reconstructing high-quality images, forming a complete dataset; performing spiral sparse downsampling on the acquired complete data volumes; using the spiral sparse downsampling data as input and the corresponding complete data volumes as targets, training a deep neural network model through supervised learning; through training with a sufficient number of data sets, the network model can learn the mapping relationship between downsampling data and complete data; acquiring radio frequency data volumes using a spiral downsampling strategy; inputting the downsampling data into the trained neural network model to obtain the recovered data volume; and using the recovered data volume to reconstruct a high-quality image using a traditional ultrasonic time-lapse image reconstruction algorithm.

[0032] The specific plan is as follows:

[0033] 1. Obtain the complete dataset: See [link / reference] Figure 1A sufficient number of complete radiofrequency data volumes were acquired using an ultrasonic tomography system. These data volumes were (N / 4)*N*M in size, where N is the number of elements in the ring transducer and M is the number of sampling points in each acquisition. The complete dataset contains comprehensive information and can be used to reconstruct high-quality images.

[0034] 2. Spiral sparse downsampling of data: See [link / reference] Figure 2 The acquired complete data volume is then subjected to spiral sparse downsampling. Spiral downsampling is a strategy where, for each array element's transmission event, the system only collects the echo data received by that specific array element. Specifically, for the transmission of array element 1, only the echo data received by array elements 1, 5, 9, ... are collected; for the transmission of array element 2, only the echo data received by array elements 2, 6, 10, ... are collected, and so on. This downsampling strategy reduces the amount of data while ensuring that data from each receiving channel can participate in data reconstruction.

[0035] 3. Training the model: See [link / reference] Figure 3 Using downsampled data as input and the corresponding complete data volume as the target, a deep neural network model is trained through supervised learning. The neural network updates its parameters by calculating the error between the recovered data and the complete data in the network output and backpropagating this error. Through training on a sufficient number of sets of data, the network model can learn the mapping relationship between the downsampled data and the complete data.

[0036] 4. Application of spiral descent sampling data recovery: See [link / reference] Figure 4 In practical applications, a spiral downsampling strategy is used to acquire radio frequency data volumes of size (N / 4)*(N / 4)*M. The downsampled data is then input into a trained neural network model to obtain recovered data volumes of size (N / 4)*N*M. These recovered data volumes carry the same information as the complete data volumes, and high-quality images can be reconstructed using traditional ultrasonic time-delay overlay image reconstruction algorithms.

[0037] Please refer to the results of the verification test. Figures 5-7 , Figure 5 This is the result of recovering the signal received from a single transmission and a single channel. Figure 6 This is a diagram showing the signal recovery results for all channels received in a single transmission. Figure 7From left to right, the images are: reconstructed image of the complete data volume, reconstructed image of the spiral downsampled data volume (SSIM: 0.71, PSNR: 17.76), and reconstructed image of the network output data volume (SSIM: 0.79, PSNR: 24.28). SSIM refers to structural similarity, and PSNR refers to peak signal-to-noise ratio. These results verify that this technical solution achieves the goal of maintaining image quality while reducing data volume. The spiral downsampling strategy fully utilizes data redundancy within the data volume, while the deep learning network model can recover complete data from downsampled data by learning the mapping relationships between data. The combination of the spiral downsampling strategy and the deep learning-based data recovery method enables the ultrasound tomography system to reconstruct high-quality images while reducing data acquisition time and data volume. Compared to traditional uniform downsampling methods, the spiral downsampling strategy better preserves data integrity and reduces information loss. Furthermore, the data recovery process using the deep learning network effectively restores details and textures in the downsampled data, improving image contrast and resolution.

[0038] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for recovering ultrasonic tomographic spiral downsampling data based on deep learning, characterized in that, Includes the following steps: Step S1: Obtain a sufficient number of complete radio frequency data volumes containing complete information for reconstructing high-quality images, thus forming a complete dataset: Step S2: Perform spiral sparse downsampling on the acquired complete data volume; Step S3: Using spiral sparse downsampled data as input and the corresponding complete data volume as target, train a deep neural network model through supervised learning; through training with a sufficient number of sets of data, the network model can learn the mapping relationship between downsampled data and complete data; Step S4: Acquire radio frequency data using a spiral downsampling strategy; Step S5: Input the downsampled data into the trained neural network model to obtain the recovered data volume; Step S6: Using the recovered data volume, a high-quality image is reconstructed using a traditional ultrasonic time-lapse image reconstruction algorithm; For each array element's transmission event, the spiral downsampling system only collects the echo data received by a specific array element. That is, for the nth array element's transmission, only the echo data received by the nth, n+m, n+2m, ... array elements are collected, where n is a positive integer, in order to reduce the amount of data and ensure that the data from each receiving channel can participate in data reconstruction. In step S4, a spiral downsampling strategy is used to acquire an RF data volume of size (N / 4)*(N / 4)*M. Then, the downsampled data is input into the trained neural network model to obtain a recovered data volume of size (N / 4)*N*M.

2. The method for recovering ultrasonic tomographic spiral downsampling data based on deep learning according to claim 1, characterized in that, The size of the complete radio frequency data body is (N / 4)*N*M, where N is the number of array elements of the ring transducer and M is the number of sampling points for each data acquisition.

3. The method for recovering ultrasonic tomographic spiral downsampling data based on deep learning according to claim 1, characterized in that, The value of m is 4.

4. The method for recovering ultrasonic tomographic spiral downsampling data based on deep learning according to claim 1, characterized in that, In step S3, the neural network updates the network model parameters by calculating the error between the recovered data and the complete data output by the network and backpropagating the error.

Images