A method for generating prostate ultra-high b-value DWI images based on an encoder-decoder structure

Abstract

The present invention discloses a method for generating ultra-high b-value DWI images of the prostate based on an encoder-decoder structure. The method comprises the following steps: first, acquiring and preprocessing images; second, building a deep learning network consisting of three encoder-decoder groups, with the encoder-decoder structure based on a U-shaped design, wherein the encoder-decoder networks on both sides serve as feature extraction layers, and the encoder-decoder network in the middle layer serves as a feature fusion layer; third, selecting a network loss function and optimizing network model parameters; fourth, training and optimizing the network model; and fifth, inputting a low-b-value diffusion-weighted prostate image to be processed into the optimized network model to generate ultra-high b-value DWI images. The algorithm of the present invention has a clear structure and operates end-to-end, efficiently generating high-quality ultra-high b-value DWI images.

Description

Technical Field

[0001] The invention relates to a method for generating a prostate ultra-high b-value DWI image based on an encoder-decoder structure, and belongs to the field of image processing. Background Art

[0002] Diffusion-weighted imaging (DWI) noninvasively reflects the diffusion of water molecules within living tissue, providing qualitative and quantitative analysis of prostate disease, thereby improving diagnostic accuracy. It has become a key diagnostic technique for the detection and characterization of prostate disease. Standard b-value DWI images (800–1000 s / mm²) often exhibit unsuppressed isointense or slightly hyperintense signals, potentially obscuring prostate cancer foci by normal prostate tissue or benign lesions. This phenomenon is known as the "T2 shine-through effect." In contrast, high b-value DWI (>=1000 s / mm²) can detect subtle changes in diffusion, suppressing the signal in non-lesion areas and enhancing the contrast between cancerous and non-cancerous tissue. Ultra-high b-values (≥2000 s / mm²) provide stronger diffusion weighting and suppression of benign prostate tissue, thereby increasing tumor-background contrast. Currently, scholars generally agree that ultra-high b-value DWI images can effectively improve the diagnosis of prostate cancer. However, the limitations of ultra-high b-value DWI include: MRI scanner gradient performance requirements, extended acquisition time, sensitivity artifacts and low signal-to-noise ratio (SNR).

[0003] In daily practice, to meet clinical and scientific research requirements, it is generally necessary to acquire images at multiple b-values, which places high demands on the performance of the MRI instrument and results in long scan times. Furthermore, conventional diffusion-weighted imaging uses an EPI (Echo Planar Imaging) readout method, which results in high noise levels and high accumulation of SAR (Specific Absorption Ratio) values when the powerful gradient field is rapidly switched. Consequently, the clinical adoption of ultra-high b-value DWI has been hindered.

[0004] Taking into account the information correlation between DWI images with different b-values, this application proposes a deep learning network with three sets of encoder-decoder structures to generate ultra-high b-value DWI images of the prostate. The algorithm has a clear structure and runs end-to-end, which can efficiently generate high-quality ultra-high b-value DWI images. Summary of the Invention

[0005] In order to solve the problem of promoting ultra-high b-value DWI images, the present invention provides a method for generating ultra-high b-value DWI images of the prostate based on an encoder-decoder structure, which efficiently generates high-quality ultra-high b-value DWI images.

[0006] The technical solution adopted by the present invention is: a method for generating a prostate ultra-high b-value DWI image based on an encoder-decoder structure, which comprises the following steps:

[0007] Step 1: Acquire and preprocess the diffusion-weighted image of the prostate at a lower b value

[0008] Step 2: Build a deep learning network. This deep learning network consists of three sets of encoder-decoders, and the encoder-decoder structure is designed based on a U-shaped structure. The encoder-decoder networks on both sides serve as feature extraction layers, used to extract low-b-value diffusion-weighted images and capture the difference information of images with different b-values. The encoder-decoder network in the middle layer serves as a feature fusion layer, used to fuse the extracted image features to generate ultra-high b-value images.

[0009] Step 3: Select the loss function of the encoder-decoder network and optimize the encoder-decoder network model parameters;

[0010] Step 4: Train and optimize the deep learning network model;

[0011] Step 5: Input the prostate diffusion-weighted image with a lower b value to be processed into the optimized deep learning network model to generate an ultra-high b value DWI image.

[0012] Furthermore, the preprocessing in step 1 includes preliminary screening, and then unified operations are performed through Matlab, including cutting out the target area, unifying the size to 224×224 resolution, and normalization operations.

[0013] Furthermore, in step 2, a network framework is built using Pytorch, wherein the backbone network is a Transformer and Resnet structure.

[0014] Furthermore, in step 2, the extracted features are adaptively weighted fused using three strategies: element summation, element product, and maximum value selection. The generated ultra-high b-value DWI is compared with the label to correct the reconstruction loss.

[0015] Furthermore, in the step three, the loss functions L1 loss and MSE loss are used to perform reconstruction loss between the generated image and the reference image and feed it back to the network. At the same time, Focal Frequency Loss is introduced to optimize the reconstruction of frequency domain information and adaptively focus the model on difficult-to-process frequency components to improve the quality of the generated image. At the same time, combined with Total Variation Loss, some regularization terms are added to the optimization problem model to maintain the smoothness of the image.

[0016] Furthermore, in step 4, the network training method is changed to alternating training of the generator and the discriminator, and PatchGAN is introduced as the discriminator to continuously correct the ultra-high b-value DWI images generated by the generator composed of three sets of encoders and decoders.

[0017] The present invention has the following advantages: the network structure designed by this method is novel, easy to train, and exhibits excellent reconstruction performance. This method generates ultra-high b-value DWI images of the prostate through a deep learning network with three encoder-decoder structures. The algorithm has a clear structure and operates end-to-end, efficiently generating high-quality ultra-high b-value DWI images. Furthermore, the model demonstrates excellent overall image quality and image information reading capabilities, improving imaging efficiency.

[0018] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a structural framework containing three groups of encoders and decoders.

[0020] Figure 2 Schematic diagram of the structure of two modules combining optimization regularization and convolutional neural network.

[0021] Figure 3 It is the MFFB feature fusion module.

[0022] Figure 4 For the integration strategy.

[0023] Figure 5 A comparison chart of different models.

[0024] Figure 6 Comparison of the performance of different models. DETAILED DESCRIPTION

[0025] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0026] The present invention provides a method for generating a prostate ultra-high b-value DWI image based on an encoder-decoder structure, which comprises the following steps:

[0027] Step 1: Clinical imaging data were collected by hospital radiographers to obtain clinical and imaging data of patients undergoing multi-parameter prostate MRI examinations. The collected data were screened, and the inclusion criteria were: (1) elevated prostate-specific antigen level or clinical suspicion of prostate cancer with preliminary clinical diagnosis, such as elevated prostate-specific antigen level; (2) clear PI-RADS V2 score.

[0028] Step 2: After the initial screening, we used Matlab to perform unified operations on the image data, including cropping the target area, unifying the size to 224×224 resolution, and normalizing.

[0029] Step 3: We build a network framework through Pytorch, such as Figure 2 , according to three sets of encoder-decoder structures. For the backbone network, we combined the Swin-Transformer structure and residual links to construct the STRB structure, which enables the network to effectively capture the differences in DWI images with different b-values. We use a U-shaped structure for the encoder-decoder networks on both sides to capture the difference information of images with different b-values. The middle backbone network serves as the feature fusion layer. According to the currently popular fusion strategy, we use three strategies, element summation, element product, and maximum value selection, to perform adaptive weighted fusion on the extracted features. The generated ultra-high b-value DWI will be compared with the label to correct the reconstruction loss.

[0030] Step 4: For the network's loss function, the reconstruction loss between the generated image and the reference image is fed back to the network. This loss is primarily L1 loss and MSE loss. Furthermore, to further improve the quality of the generated images, we introduce FocalFrequency Loss to optimize the reconstruction of frequency domain information. This adaptively focuses the model on difficult-to-process frequency components, thereby improving the quality of the generated images. Due to the low signal-to-noise ratio of ultra-high b-value DWI images, we propose combining it with TotalVariation Loss and adding regularization terms to the optimization model to maintain image smoothness.

[0031] Step 5: Integrating the training concept of generative adversarial networks, we changed the network training method to alternating training of the generator and discriminator, introduced PatchGAN as the discriminator, and continuously corrected the ultra-high b-value DWI images generated by the generator composed of three sets of encoders and decoders.

[0032] Step 6: Input the prostate diffusion-weighted image with a lower b value to be processed into the optimized network model to generate an ultra-high b value DWI image.

[0033] In summary, the present invention obtains the image information of the lower b-value DWI through two sets of symmetrical U-shaped encoders and decoders, and outputs different levels of feature information of the lower b-value DWI image. The third set of encoders and decoders is used to receive the output feature information and perform image fusion to generate an ultra-high b-value DWI image. The third set of encoders and decoders consists of a feature fusion module. Figure 3 Provides a diagram of the MFFB module. Figure 5Following currently popular fusion strategies, we adaptively weighted-fuse the extracted features using three strategies: element-wise summation, element-wise product, and maximum value selection. The resulting ultra-high b-value DWI is then compared with the labels to correct the reconstruction loss. For the backbone network, we constructed a STRB (Swin-Transformer and Residual Block) architecture combining the Swin-Transformer structure and residual links, enabling the network to effectively capture the differences between DWI images with different b-values.

[0034] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the scope of protection of the present invention in any form, and all technical solutions obtained by equivalent replacement and other methods fall within the scope of protection of the present invention.

[0035] The parts not involved in the present invention are the same as the existing technology or can be implemented by using the existing technology.

Claims

1. A method for generating ultra-high b-value DWI images of the prostate based on an encoder-decoder network structure, characterized in that The steps include: Step 1: Acquire and preprocess prostate diffusion-weighted images with lower b-values; Step 2: Build a deep learning network model. The deep learning network model consists of three sets of encoder-decoder networks, and the encoder-decoder network is designed based on a U-shaped structure. The encoder-decoder networks on both sides serve as feature extraction layers, used to extract low b-value diffusion-weighted images and capture the difference information of images with different b-values. The encoder-decoder network in the middle layer serves as a feature fusion layer, used to fuse the extracted image features to generate ultra-high b-value images. Step 3: Select the loss function of the encoder-decoder network and optimize the encoder-decoder network model parameters; Step 4: Train and optimize the deep learning network model; Step 5: Input the prostate diffusion-weighted image with a lower b value to be processed into the optimized deep learning network model to generate an ultra-high b value DWI image.

2. The method for generating a prostate ultra-high b-value DWI image based on an encoder-decoder network structure according to claim 1, characterized in that: The preprocessing in step 1 includes preliminary screening, and then unified operation is performed through Matlab to intercept the target area and unify the size to 224 224 resolution and normalization operations.

3. The method for generating a prostate ultra-high b-value DWI image based on an encoder-decoder network structure according to claim 1, characterized in that: In step 2, a network framework is built using Pytorch, where the backbone network is a Transformer and Resnet structure.

4. The method for generating a prostate ultra-high b-value DWI image based on an encoder-decoder network structure according to claim 1, characterized in that: In the second step, the extracted features are adaptively weighted fused using three strategies: element summation, element product, and maximum value selection. The generated ultra-high b-value DWI is compared with the label to correct the reconstruction loss.

5. The method for generating a prostate ultra-high b-value DWI image based on an encoder-decoder network structure according to claim 1, characterized in that: In the step three, the loss functions L1 loss and MSE loss are used to perform reconstruction loss between the generated image and the reference image and feed it back to the encoder-decoder network. At the same time, the focus frequency loss is introduced to optimize the reconstruction of frequency domain information and adaptively focus the model on difficult-to-process frequency components to improve the quality of the generated image. At the same time, combined with the total variation loss, some regularization terms are added to the optimization problem model to maintain the smoothness of the image.

6. The method for generating a prostate ultra-high b-value DWI image based on an encoder-decoder network structure according to claim 1, characterized in that: In step 4, the training method of the deep learning network model is changed to alternating training of the generator and the discriminator, and PatchGAN is introduced as the discriminator to continuously correct the ultra-high b-value DWI images generated by the generator composed of three sets of encoder-decoder networks.

Images