Spatially Localized Atlas Network Tiles Enables 3D Whole Brain Segmentation from Limited Data (1806.00546v2)

Abstract: Whole brain segmentation on a structural magnetic resonance imaging (MRI) is essential in non-invasive investigation for neuroanatomy. Historically, multi-atlas segmentation (MAS) has been regarded as the de facto standard method for whole brain segmentation. Recently, deep neural network approaches have been applied to whole brain segmentation by learning random patches or 2D slices. Yet, few previous efforts have been made on detailed whole brain segmentation using 3D networks due to the following challenges: (1) fitting entire whole brain volume into 3D networks is restricted by the current GPU memory, and (2) the large number of targeting labels (e.g., > 100 labels) with limited number of training 3D volumes (e.g., < 50 scans). In this paper, we propose the spatially localized atlas network tiles (SLANT) method to distribute multiple independent 3D fully convolutional networks to cover overlapped sub-spaces in a standard atlas space. This strategy simplifies the whole brain learning task to localized sub-tasks, which was enabled by combing canonical registration and label fusion techniques with deep learning. To address the second challenge, auxiliary labels on 5111 initially unlabeled scans were created by MAS for pre-training. From empirical validation, the state-of-the-art MAS method achieved mean Dice value of 0.76, 0.71, and 0.68, while the proposed method achieved 0.78, 0.73, and 0.71 on three validation cohorts. Moreover, the computational time reduced from > 30 hours using MAS to ~15 minutes using the proposed method. The source code is available online https://github.com/MASILab/SLANTbrainSeg

• The paper introduces the SLANT method which divides 3D brain segmentation into spatially localized network tiles to overcome GPU memory and data limitations.
• It leverages canonical registration and auxiliary labels from 5111 scans to enhance training, achieving improved mean Dice coefficients across cohorts.
• The approach reduces segmentation time from 30 hours to 15 minutes, offering a scalable, efficient alternative to traditional multi-atlas segmentation.

Spatially Localized Atlas Network Tiles for 3D Whole Brain Segmentation

The paper "Spatially Localized Atlas Network Tiles Enables 3D Whole Brain Segmentation from Limited Data" tackles the challenges associated with whole brain segmentation using structural magnetic resonance imaging (MRI). Specifically, the research introduces the Spatially Localized Atlas Network Tiles (SLANT) method as an innovative approach to overcome the inherent limitations of 3D network applications for detailed brain segmentation due to constraints in GPU memory and limited sample size for training.

Historically, multi-atlas segmentation (MAS) has been the standard for segmenting more than 100 anatomical regions in the brain. However, it is computationally intensive, often requiring upwards of 30 hours for segmentation tasks. Furthermore, recent attempts to apply deep convolutional neural networks (DCNNs) have largely relied on patch-based methods, which, while useful, still have significant limitations in terms of spatial consistency and resource demands.

The proposed SLANT method employs multiple independent 3D fully convolutional networks—referred to as "network tiles"—that spatially cover overlapping sub-spaces within a standard atlas space. This design simplifies the task into more manageable sub-tasks, thus facilitating training. Canonical registration and label fusion techniques form the methodological backbone for addressing the crucial issue of limited training data. Auxiliary labels on 5111 initially unlabeled scans were generated using MAS, significantly enhancing the pre-training process.

Empirical results indicate that the SLANT approach provides a clear performance advantage. The proposed method demonstrated improved mean Dice coefficients across three validation cohorts when compared to the state-of-the-art MAS. Specifically, SLANT achieved mean Dice values of 0.78, 0.73, and 0.71 across different cohorts, showing a noticeable improvement from the MAS method, which achieved 0.76, 0.71, and 0.68, respectively. Furthermore, SLANT reduced the computational time to approximately 15 minutes while maintaining the segmentation accuracy, showcasing a substantial computational efficiency improvement over traditional MAS techniques.

The implications of this research are multifaceted. Practically, the significant reduction in computational time without sacrificing accuracy positions SLANT as a viable tool for efficient, large-scale brain segmentation tasks. Theoretically, the incorporation of network tiling can inspire subsequent studies on optimizing neural networks for handling high-dimensional data spaces with limited GPU resources. Additionally, the method's success highlights the potency of leveraging initially unlabeled data to enhance deep learning models, potentially informing strategies in other domains dealing with analogous challenges.

Looking ahead, this paper opens up potential avenues for dynamic network designs that can adapt to hardware limitations and data availability while continuing to improve segmentation accuracy. Further developments may explore integrating SLANT with more advanced 3D network architectures, which could yield even greater segmentation fidelity and additional reductions in computational demands. Overall, the SLANT method provides a robust, scalable framework for the efficient and accurate segmentation of brain MRIs, potentially advancing both research and clinical applications in neuroimaging.