Shulou(Shulou.com)06/01 Report--

How to deeply learn the brain imaging tool FastSurfer, I believe that many inexperienced people do not know what to do, so this paper summarizes the causes of the problem and solutions, through this article I hope you can solve this problem.

It's like doing a jigsaw puzzle is easy for you, but building a complex Lego may be more difficult, not to mention irregular brain imaging. For many years, the graduate school of biological imaging has been committed to improving the cost and efficiency of brain imaging testing, which is very important to solve brain medical problems.

Previous researchers have generally used FreeSurfer for brain imaging analysis. FastSurfer provides a complete FreeSurfer alternative for volume analysis and surface-based thickness analysis, including:

FastSurferCNN- is an advanced deep learning architecture that divides the whole brain into 95 categories in less than a minute, mimicking FreeSurfer's anatomical segmentation and cortical fragmentation (DKTatlas).

Recon-surf-- 's FreeSurfer-based total surface reconstruction workflow allows cortical surface reconstruction, cortical label mapping, and traditional point-by-point and ROI thickness analysis to be completed in about 60 minutes.

Rebuild the brain? You can't just know FreeSurfer. Maybe you should try FastSurfer!

The traditional neural image analysis is faced with the problems of large amount of computation and time-consuming, so it is difficult to popularize. FreeSurfer is a widely used software in current medical imaging, but its convenient performance still needs to be optimized.

FastSurfer may be able to solve this problem. German researchers have proposed a fast and accurate neuroimaging method based on deep learning for automatic processing of structured human brain MRI scans. This method provides a complete FreeSurfer alternative for volume analysis (less than 1 minute) and surface-based thickness analysis (only in about 1 hour run time).

In terms of speed, it takes about 7 hours for a complete FreeSurfer to run on CPU (4 hours in parallel), depending on image quality, disease severity, and so on. FastSurfer achieved volume segmentation (subcutaneous and cortical areas) in just 1 minute (on GPU, 14 minutes on CPU), surface treatments including cortical ROI thickness measurements, including spherical registration, and potential subsequent group analysis of surface maps on CPU within 3.7hours (parallel 1.6hours).

In the test of OASIS1 retest data set, the performance of FastSurfer is relatively good, and its ICC (intra-group correlation coefficient) is mostly between 0.8-1, and high.

In particular, sensitivity to population differences in dementia is an important part of our assessment. The sensitivity difference between FreeSurfer and FastSurfer was determined by evaluating their ability to separate diagnostic groups in OASIS1 (AD and CN).

In the flow of FastSurfer analysis, the difference of cortical thickness is more obvious, that is, the sensitivity is more intense.

In the test, FastSurfer is not only several orders of magnitude faster than the traditional method, but also improves the reliability and sensitivity. Therefore, it is a reliable tool for large-scale analysis tasks in the future.

FastSurferCNN: deep learning architecture allows machines to divide the whole brain into 95 categories

The first thing the deep learning framework needs to do is to provide accurate 3D whole brain segmentation. The researchers used a novel spectral method (using the Laplace eigenfunction to quickly draw the cortex) to perform cortical surface reconstruction and fast spherical mapping.

The figure above shows the FastSurfer network architecture. The network consists of four competitive dense blocks (CDB) in the encoder and decoder parts and is separated by a bottleneck layer. Except for the first encoder block, each block consists of three sequences of parameter rectifier linear unit (PReLU), convolution (Conv) and batch normalization (BN). In the first block, replace PReLU with BN to standardize the original input.

For FastSurferCNN, the first thing this deep learning framework needs to do is to provide accurate 3D whole brain segmentation. The researchers used a novel spectral method (using the Laplace eigenfunction to quickly draw the cortex) to perform cortical surface reconstruction and fast spherical mapping.

The data set was trained using the information of 140 representative subjects from ABIDE-II,ADNI,LA5C and OASIS, and 20 subjects from MIRIAD were used for verification. The training set is balanced in terms of gender, age, diagnosis, and covers a variety of other parameters (i.e. scanner, field strength and acquisition parameters).

Sufficient anatomical structure and collection diversity in the training image can improve the robustness and versatility of the network, and finally improve the segmentation accuracy in most invisible scans without fine-tuning the model weight.

Recon-surf: rebuilding the brain

After analyzing the information of the brain using the deep learning framework, it needs to be reconstructed. First, the moving cube algorithm is used to reconstruct the surface. Then, a new, fast spectrum is mapped to the sphere. Here, we will use the two steps of the original FreeSurfer module (mri_tessellate and mris_sphere), which are the same as FastSurfer in other aspects. What needs to be quantified are the number of topological surface defects and the average quality of the resulting surface triangular mesh.

When using moving cubes to construct surfaces, the average number of defects in the workflow of the FreeSurfer module (27.2 defects per hemisphere) decreased by 12% (24.0 defects), while in the proposed workflow (FastSurfer: moving cubes + spectral spherical projection) decreased by 15.3% (23.1 defects). The average surface treatment time of FastSurfer was significantly reduced by 15 minutes per hemisphere.

After reading the above, have you mastered how to learn the brain imaging tool FastSurfer in depth? If you want to learn more skills or want to know more about it, you are welcome to follow the industry information channel, thank you for reading!

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.