CN-122023594-A - Method and system for generating synthetic SWI images from T1 weighted MRI scans for characterizing brain diseases

Abstract

A system for generating a simulated SWI image of a human brain based on SNC MR images of the human brain includes an input image module for storing the SNC MR images, a preprocessing module for receiving the SNC MR images, the preprocessing module for preparing and generating the SNC MR images into a standard format for an AI model to extract and classify features of the SNC MR images, a simulated SWI generation model module for receiving the SNC MR images in a cradle format and generating a simulated SWI image corresponding to each SNC MR image, a deep learning platform for operating an AI model, the AI model being in a connected state, a training module for receiving training data and transmitting the training data to the deep learning platform for adjusting adjustable parameters of the AI model to optimize generation of the simulated SWI image, a test module for communicating with the training module and the deep learning platform for receiving test data, the test module for verifying the simulated SWI image using pre-trained performance criteria, and an output storage module for receiving and storing the synthesized SWI image.

Inventors

• CHEN WEIEN
• Abdul-Mojid Olabisi Ilyas
• HUANG JIANPAN
• Jamal Philmat Banqi

Assignees

• 香港心脑血管健康工程研究中心有限公司

Dates

Publication Date

20260512

Application Date

20250512

Priority Date

20241106

Claims (20)

1. 1. A system for generating a simulated SWI image of a human brain based on SNC MR images of the human brain, comprising: An input image module for storing SNC MR images; A preprocessing module for receiving the SNC MR images, wherein the preprocessing module is for preparing and generating the SNC MR images into a standard format for AI models to extract and classify features of the SNC MR images; a simulated SWI generation model module for receiving the SNC MR images in the standard format and generating a simulated SWI image corresponding to each of the SNC MR images; A deep learning platform for operating the AI model, wherein the AI model is in a connected state; A training module for receiving training data and transmitting the training data to the deep learning platform so that adjustable parameters of the AI model may be adjusted for optimization to generate the simulated SWI image; a test module for communicating with the training module and the deep learning platform to receive test data, wherein the test module is for validating the simulated SWI image with pre-trained performance criteria; And the output storage module is used for receiving and storing the synthesized SWI image.
2. 2. A method for generating a simulated SWI image using an acquired SNC MR image (T1 weighted MR image) by a system having at least a processor and a memory for executing instructions of an artificial intelligence engine configured as a UNet model stored within the memory of the system, wherein the UNet model comprises: an encoder having a plurality of layer blocks, each layer block of the encoder comprising one or more convolutional layers, each of the convolutional layers being associated with an active layer and a downsampling layer; A decoder having a plurality of layer blocks, each layer block of the decoder comprising an upsampling layer and one or more convolutional layers, each of the convolutional layers being associated with an active layer; A skip connection for associating one of the layer blocks of the encoder with one of the layer blocks of the decoder at a corresponding multi-scale resolution level; Wherein the encoder is configured to extract features from the T1 weighted MR image for the decoder to combine the output of the encoder and the extracted image features at a multi-scale resolution level over the skip connection to generate the simulated SWI image.
3. 3. The method of claim 2, wherein the encoder and decoder are configured to perform a cross-sequence conversion from T1 weighted images to SWI images consisting of 19 convolutional layers.
4. 4. A method according to claim 3, wherein the encoder is adapted to receive an image comprising three-dimensional and one or more color channels, wherein one or more layer blocks of the encoder comprise a repeated realization of two 3 x3 convolutional layers, wherein 2 voxels span five layer blocks, and wherein the layer block of the encoder immediately preceding the decoder comprises a single convolutional layer.
5. 5. The method of claim 4, wherein the activation layer is configured to perform a linear correction function via one or more correction linear units (relus).
6. 6. The method of claim 5, wherein the downsampling comprises a 2 x 2 max pooling operation of step size 2 voxels, wherein each convolutional layer is configured to process input data with a plurality of convolutional filters.
7. 7. The method of claim 6, wherein the max pooling operation after activation layer reduces the spatial size of image feature map by a factor of 2 and the number of convolution filters doubles, increasing from 16 in the first block to 1024 in the last block, so as to allow the UNet model to learn the hierarchical relationship over a substantial receptive field of the SNC MR image.
8. 8. The method of claim 7, wherein the upsampling layer of the decoder is used to perform nearest neighbor interpolation to increase the image size through each layer block by a factor of 2 through each layer within the decoder.
9. 9. The method of claim 8, wherein the one or more convolutional layers with the decoder use random initialization and unequal kernel sizes.
10. 10. The method of claim 9, wherein the skip connection is used to copy and concatenate features generated from one of the layer blocks of the encoder to one of the layer blocks of the decoder at a corresponding multi-scale resolution level such that both high-level and low-level features of the encoder are used as additional inputs in the decoder to provide an efficient and stable image representation.
11. 11. The method of claim 10, wherein the output layer comprises a single output convolutional layer immediately following an output active layer, wherein the single output convolutional layer is a1 x 1 convolutional layer with a step size of 1, and wherein the output active layer is configured to perform a tanh operation.
12. 12. The method of claim 11, wherein the system further comprises a diagnostic model for classifying abnormalities in the simulated SWI images to characterize brain disease.
13. 13. A method for generating a simulated SWI image of a human brain based on SNC MR images of the human brain without injecting a contrast agent into the body, comprising: Collecting SNC MR images; Inputting the SNC MR images into a training module; Collecting SWI images corresponding to each subject in the SNC MR images; storing the SWI image into the training module; inputting the SNC MR images and corresponding SWI images into an AI model; Training the AI model to generate the simulated SWI image based on the SNC MR image input into the AI model and the corresponding SWI image output as a target; The simulated SWI image is tested against the corresponding SWI image previously input into the AI model, and the AI model input is optimized.
14. 14. The method of claim 13, wherein the SNC MR image is a T1 weighted image.
15. 15. The method of claim 14, wherein the SNC MR images are acquired from any type of MRI scanner including GE, siemens, and Philips.
16. 16. The method of claim 15, wherein the SNC MR image is acquired from an MRI scanner that includes any field strengths of 1.5T, 3T, and 7T.
17. 17. The method of claim 16, wherein training the AI model to generate the simulated SWI image includes applying a deep learning technique.
18. 18. The method of claim 17, wherein testing the simulated SWI image against the corresponding SWI image previously input into the AI model and optimizing the AI model input includes applying the deep learning technique.
19. 19. The method as recited in claim 18, further comprising: acquiring SNC MR images of the person on an MRI scanner; registering the SNC MR images to a standard non-contrast image template; transmitting the registered SNC MR images to a storage module; inputting the SNC MR images into a trained AI model; a simulated SWI image corresponding to the SNC MR input image is generated and the image is viewed using existing software.
20. 20. The method of claim 19, wherein the AI model is trained using an Adam random optimization algorithm with a learning rate of 0.002 that is used to minimize a Mean Square Error (MSE) loss function in a stepwise manner and is updated stepwise at each training step until the AI model reaches convergence.

Description

Method and system for generating synthetic SWI images from T1 weighted MRI scans for characterizing brain diseases Technical Field The present disclosure relates to a method and system for generating synthetic SWI images obtained from T1-weighted MRI scans for characterizing brain diseases, in particular parkinson-related diseases. Background Magnetic resonance imaging (Magnetic Resonance Imaging, MRI) is a non-invasive technique that exploits the relaxation properties of water protons in a magnetic field to enable visualization of internal body structures. MRI is considered safe, reproducible and free of known hazards, for use in keeping with well-defined technical constraints. MRI images can be generated by different contrast modes to reflect proton density, T1 and T2 relaxation times, changes in tissue magnetic sensitivity, diffusion, temperature, motion fields, biomechanical properties, tissue perfusion, current, oxygen levels, and spectra of key biochemicals. Among the many imaging sequences, T1 weighted imaging (T1-w) is widely used clinically because of its ability to reveal brain morphology changes, but this imaging modality fails to identify certain specific pathological changes that require finer imaging techniques to be displayed. Susceptibility weighted imaging (susceptability WEIGHTED IMAGING, SWI) uses the amplitude information and the filtered phase information, alone or in combination, to generate new endogenous contrast enhancement modes. SWI is highly sensitive to deoxygenated blood, ferrioxacin, ferritin and calcium. This makes SWI of great value in diagnosing a variety of neurological disorders including aging-related disorders, multiple sclerosis, stroke, cerebral reperfusion, traumatic brain injury, cerebral vascular malformations, intracranial arterial stenosis and smoothies, cerebral microhemorrhages, central nervous system primary vasculitis, fungal aneurysms, brain tumors, and neurodegenerative disorders such as alzheimer's disease and parkinson's disease. SWI plays a key role in the detection of parkinson's disease, and early recognition can be achieved by visualizing "dovetails" (Swallow Tail Sign, STS). This sign is due to the presence of Nigrosome-1 in the lateral part of the substantia nigra dorsal part, nigrosome-1 being the largest cluster in dopaminergic neurons. However, studies have shown that the acquisition process of SWI takes longer and that magnetically sensitive artifacts may lead to significant signal loss, thereby reducing diagnostic accuracy. Although technological advances have somewhat simplified the acquisition procedure of SWI, artifact problems still exist, easily resulting in massive signal cancellation and loss of anatomical detail. In addition, SWI operations are complex, often extending diagnostic time, and introducing artifacts that further affect accuracy. Thus, its usability may be limited compared to conventional T1-weighted (T1-w) MRI scans. In recent years, machine learning technology has made remarkable progress, particularly the advent of convolutional neural networks (Convolutional Neural Network, CNN), and has shown great potential in various medical image analysis tasks, including lesion detection, brain tumor segmentation, medical image super-resolution reconstruction, intra-and inter-modality synthesis of medical images, and automatic extraction methods of blood vessels, etc. For example, CNN-based automated diagnostic systems are used to classify Parkinson's Disease (PD) and healthy controls (Healthy Control, HC). The Parkinson's disease progression marker initiative (Parkinson's Progression MARKERS INITIATIVE, PPMI) was used as baseline T2-weighted MRI data for PD and HC. Midbrain slice was selected from 500T 2-weighted MRI scans and aligned by image registration techniques. The performance of the proposed method is assessed by indexes such as Accuracy (Accuracy), sensitivity (Sensitivity), specificity (SPECIFICITY), area Under Curve (AUC), etc. However, diagnosis at the early stages of parkinson's disease remains a challenge. The disease affects mainly the largest cluster of dopaminergic neurons located on the lateral part of the substantia nigra dorsal part-Nigrosome-1. In 3D SWI sequences, this region exhibits a high signal structure relative to the overall low signal substantia nigra, forming a typical "dovetailed" (STS). Thus, researchers have explored various image-to-image conversion methods to achieve the synthesis of scanned images, such as PD, T2-FLAIR, and MRA, which can be generated based on single-modality image input or multi-modality image input. Some studies have employed 2D or 3D generation of the challenge Network (GENERATIVE ADVERSARIAL networks, GANs) with high computational cost, as well as studies have employed U-Net deep learning architectures of different variants. Some of these methods utilize multiple input modalities to enhance image composition. But limited by the limited verification data from truly ill patient