3D Whole Brain Segmentation using Spatially Localized Atlas Network Tiles (1903.12152v1)

Abstract: Detailed whole brain segmentation is an essential quantitative technique, which provides a non-invasive way of measuring brain regions from a structural magnetic resonance imaging (MRI). Recently, deep convolution neural network (CNN) has been applied to whole brain segmentation. However, restricted by current GPU memory, 2D based methods, downsampling based 3D CNN methods, and patch-based high-resolution 3D CNN methods have been the de facto standard solutions. 3D patch-based high resolution methods typically yield superior performance among CNN approaches on detailed whole brain segmentation (>100 labels), however, whose performance are still commonly inferior compared with multi-atlas segmentation methods (MAS) due to the following challenges: (1) a single network is typically used to learn both spatial and contextual information for the patches, (2) limited manually traced whole brain volumes are available (typically less than 50) for training a network. In this work, we propose the spatially localized atlas network tiles (SLANT) method to distribute multiple independent 3D fully convolutional networks (FCN) for high-resolution whole brain segmentation. To address the first challenge, multiple spatially distributed networks were used in the SLANT method, in which each network learned contextual information for a fixed spatial location. To address the second challenge, auxiliary labels on 5111 initially unlabeled scans were created by multi-atlas segmentation for training. Since the method integrated multiple traditional medical image processing methods with deep learning, we developed a containerized pipeline to deploy the end-to-end solution. From the results, the proposed method achieved superior performance compared with multi-atlas segmentation methods, while reducing the computational time from >30 hours to 15 minutes (https://github.com/MASILab/SLANTbrainSeg).

• The paper presents SLANT, a method that combines multi-atlas segmentation with independent 3D fully convolutional networks to achieve high-resolution brain segmentation on MRI scans.
• The approach leverages 5111 auxiliary labels to supplement limited manual annotations, greatly enhancing the training capacity of the network.
• SLANT reduces segmentation time from over 30 hours to approximately 15 minutes while delivering superior accuracy measured by Dice similarity coefficients.

Overview of 3D Whole Brain Segmentation using Spatially Localized Atlas Network Tiles

The paper describes a novel approach for high-resolution 3D whole brain segmentation using a method denoted as Spatially Localized Atlas Network Tiles (SLANT). This technique addresses some of the primary limitations impeding existing convolutional neural network (CNN) methods in medical image segmentation, particularly the constraints imposed by the graphics processing unit (GPU) memory and the sparse availability of manually traced brain scans for network training. The SLANT methodology integrates classical medical image processing paradigms with deep learning methods to offer enhanced performance in brain segmentation tasks.

Key Contributions

1. Integration of Multi-Atlas Segmentation with CNNs: Traditionally, multi-atlas segmentation methods set the standard for the detailed delineation of over 100 brain labels in MRI scans, owing to their high accuracy and reproducibility. The SLANT approach bridges this powerful technique with deep learning by using multiple independent 3D fully convolutional networks (FCNs). Each network is specialized to learn the spatial context for a designated location within the brain, effectively distributing data complexity and addressing the GPU memory limitations commonly faced by high-resolution imaging.
2. Use of Auxiliary Labels for Network Training: Data scarcity impedes effective training for deep learning segmentation models. SLANT circumvents this by synthesizing auxiliary labels from 5111 initially unlabeled brain scans via multi-atlas segmentation. These auxiliary labels supplement the scant 45 manually traced training volumes, significantly enhancing the training dataset without requiring extensive manual annotation.
3. Efficient Computation: The proposed SLANT method enhances computation efficiency. It substantially reduces the time required for segmentation from over 30 hours, typical of conventional multi-atlas approaches, to approximately 15 minutes. This performance gain is attributed to the efficient deployment of independent network tiles and the integration of optimized image processing techniques within a streamlined pipeline.
4. Freely Accessible Tool: The implementation of SLANT is containerized using Docker, facilitating its adoption and use in varied environments. Such accessibility ensures reproducibility and permits straightforward segregation on new MRI datasets.

Experimental Results

The experiments demonstrate that SLANT outperforms traditional methods and other CNN-based approaches in multiple test datasets in terms of both Dice similarity coefficients (DSC) and surface distance metrics. For instance, SLANT-27, after being fine-tuned with the auxiliary labeled dataset of 5111 samples and the manually labeled training set, achieved superior segmentation accuracy relative to other benchmark methods.

Implications and Future Prospects

The successful fusion of multi-atlas segmentation accuracy with the computational efficiencies of deep learning techniques in the SLANT method indicates a promising trajectory for real-time, high-resolution medical image analysis. The ability to operate within the confines of available GPU memory while leveraging expansive auxiliary datasets positions SLANT as a viable solution for large-scale clinical deployment.

There are further implications for the method's adaptability to other medical imaging modalities and increased resolution scenarios, as GPU capabilities continue to expand. Additionally, the prospect of integrating advanced registration techniques driven by deep learning could further streamline the workflow and reduce preprocessing time, enhancing the method's applicability in time-sensitive diagnostic environments.

Conclusion

In summary, the SLANT approach brings forth an innovative and effective solution to some fundamental challenges in whole brain segmentation. It is a testament to the potential of integrating classical image processing approaches with cutting-edge deep learning frameworks to produce efficient and highly accurate segmentation outcomes. The paper provides substantial groundwork for advancing how segmentation tasks are executed in medical imaging, foreshadowing enhanced capabilities in diagnostic workflows and research methodologies in neuroimaging domains.