3D fully convolutional networks for subcortical segmentation in MRI: A large-scale study (1612.03925v2)

Abstract: This study investigates a 3D and fully convolutional neural network (CNN) for subcortical brain structure segmentation in MRI. 3D CNN architectures have been generally avoided due to their computational and memory requirements during inference. We address the problem via small kernels, allowing deeper architectures. We further model both local and global context by embedding intermediate-layer outputs in the final prediction, which encourages consistency between features extracted at different scales and embeds fine-grained information directly in the segmentation process. Our model is efficiently trained end-to-end on a graphics processing unit (GPU), in a single stage, exploiting the dense inference capabilities of fully CNNs. We performed comprehensive experiments over two publicly available datasets. First, we demonstrate a state-of-the-art performance on the ISBR dataset. Then, we report a {\em large-scale} multi-site evaluation over 1112 unregistered subject datasets acquired from 17 different sites (ABIDE dataset), with ages ranging from 7 to 64 years, showing that our method is robust to various acquisition protocols, demographics and clinical factors. Our method yielded segmentations that are highly consistent with a standard atlas-based approach, while running in a fraction of the time needed by atlas-based methods and avoiding registration/normalization steps. This makes it convenient for massive multi-site neuroanatomical imaging studies. To the best of our knowledge, our work is the first to study subcortical structure segmentation on such large-scale and heterogeneous data.

• The paper introduces a novel 3D FCNN architecture for subcortical segmentation in MRI, achieving state-of-the-art performance on benchmark datasets.
• The methodology leverages small convolutional kernels and multiscale integration to efficiently manage computational and memory challenges.
• Experimental validation on ISBR and ABIDE datasets demonstrates its robustness and potential to reduce segmentation time in large-scale neuroimaging studies.

3D Fully Convolutional Networks for Subcortical Segmentation in MRI: A Comprehensive Study

The paper investigates the efficacy of a three-dimensional fully convolutional neural network (3D FCNN) in segmenting subcortical brain structures from MRI scans. Given the importance of accurate subcortical segmentation in diagnosing and studying neurological disorders such as schizophrenia and autism, the authors present a model that addresses the computational and memory challenges typically associated with 3D convolutions.

Methodology

The authors propose a 3D FCNN architecture to tackle these segmentation tasks. This approach leverages small convolutional kernels, which facilitates deeper network architectures without incurring prohibitive memory costs. The architecture further incorporates both local and global context by using intermediate-layer outputs directly in the final prediction, which enhances feature consistency across different scales and embeds precise spatial details into the segmentation process. The entire model is optimized for execution on GPUs, permitting efficient end-to-end training.

Experimental Validation

The model's performance was validated on two datasets: the ISBR dataset, to establish baseline capabilities, and the large-scale ABIDE dataset, comprising 1112 subject datasets from 17 different acquisition sites. Notably, the model achieves state-of-the-art accuracy on ISBR data and demonstrates robustness across diverse acquisition protocols in the ABIDE dataset, reflecting its capacity to generalize well.

Results

In terms of quantitative results, the model achieved state-of-the-art segmentation performance with high agreement to an atlas-based approach, delivering segmentation quality that aligns closely with manual analyses. For the ABIDE dataset, the method proved consistent across varying demographics and clinical characteristics, handling unregistered data effectively and bypassing the need for registration or normalization steps.

Implications and Future Work

Practically, this approach significantly reduces the time required for brain structure segmentation, crucial for large-scale neuroanatomical studies. Theoretically, the successful integration of small kernels and multiscale features into the segmentation process could be applied to other medical image analysis tasks, potentially extending to other anatomical structures or imaging modalities.

Future research might delve into scaling the method for even larger datasets, further optimizing the architecture for specific neurological conditions, and exploring the potential of transfer learning to improve segmentations across different domains or populations. The integration of more advanced regularization techniques, such as those based on generative models, could further enhance segmentation performance.

In conclusion, this paper provides compelling evidence supporting the efficacy and efficiency of 3D FCNNs in subcortical brain MRI segmentation, offering a promising direction for future research in the domain of automatic medical imaging segmentation.