AtlasNet: Multi-atlas Non-linear Deep Networks for Medical Image Segmentation

Maria Vakalopoulou, Guillaume Chassagnon, Norbert Bus, Rafael Marini, Evangelia I. Zacharaki, Marie-Pierre Revel, Nikos ParagiosMICCAI 2018 · DOI

TL;DR

This paper introduces AtlasNet, a novel multi-network deep learning architecture for medical image segmentation. By leveraging multiple "atlases" (reference images) and non-linear registration, AtlasNet effectively addresses the challenges of limited data and anatomical variability in medical imaging, achieving human-level accuracy on interstitial lung disease segmentation.

What Problem Does This Paper Address?

Medical image segmentation, the process of identifying and outlining specific organs or tissues, is crucial for diagnosis and treatment planning. However, deep learning approaches in this domain face significant hurdles:

  • Limited Data: Acquiring large, annotated medical datasets is expensive and time-consuming. This is exacerbated in 3D imaging.
  • Anatomical Variability: Human anatomy varies significantly between individuals, and even within the same individual, tissues can exhibit diverse textures and patterns, making it difficult for models to generalize.
  • High Dimensionality: Medical images, especially 3D scans, are inherently high-dimensional, increasing computational demands.

Existing methods often struggle to overcome these limitations, leading to suboptimal segmentation accuracy and robustness.

Methods

AtlasNet tackles these challenges with a novel multi-network architecture that combines deep learning with multi-atlas based registration:

  • Multi-Atlas Registration: The core idea is to map input medical images to multiple pre-defined "atlases" (reference anatomies). This mapping is achieved using a non-linear deformable transformer that employs a robust, multi-metric, multi-modal graph-based registration framework. This process aims to reduce anatomical variability by bringing different anatomies into a common reference space.
  • Multiple Segmentation Networks: For each mapped image (corresponding to an atlas), a separate deep neural network (specifically, a SegNet architecture, a type of encoder-decoder FCN) is trained. This creates a redundant ensemble of networks.
  • De-Transformer for Projection: After segmentation by each individual network, a backward de-transformer is used to project the segmentation results back into the original image space.
  • Consensus for Final Segmentation: The final segmentation is obtained by combining the predictions from all the individual networks, often through a majority voting mechanism. This consensus approach enhances robustness and accuracy.

The novelty lies in the integration of non-linear deformable registration with a multi-network deep learning framework, allowing for data augmentation through anatomical mapping and leveraging domain knowledge from multiple atlases.

Datasets

The primary dataset used for evaluation is a collection of volumetric CT images for interstitial lung disease (ILD) segmentation. This dataset consists of:

  • 17 patients for training, with 6000 slices in total (512x512 dimension), annotated by a medical expert.
  • 29 additional patients for testing, with annotations on selected CT slices (n=20) by three different observers.

The data is noted to be multi-vendor (GE & Siemens). The paper does not introduce a new benchmark dataset but evaluates on a challenging real-world medical imaging task.

Key Insights and Ideas

  • Leveraging Diversity for Robustness: By mapping images to multiple atlases, the model effectively creates a form of implicit data augmentation, exposing the segmentation networks to a wider range of anatomical variations. This diversity, when fused, leads to more robust and generalizable segmentations.
  • Registration as a Preprocessing Step for Deep Learning: The paper demonstrates that sophisticated non-linear registration, when integrated into a deep learning pipeline, can significantly improve performance by normalizing anatomical differences, a long-standing challenge in medical image analysis.
  • Ensemble Learning through Atlases: The multi-network approach, where each network learns on a registered version of the data, acts as an ensemble method. The consensus from multiple networks trained on slightly different "views" of the anatomy leads to superior results compared to single networks.
  • Reduced Data Requirements: The proposed architecture achieves competitive or superior performance with significantly less training data (e.g., 30% of samples) compared to state-of-the-art methods, highlighting the efficiency gained from the registration and multi-atlas strategy.

Key Results

  • Human-Level Accuracy: AtlasNet achieved results comparable to human accuracy on the challenging task of interstitial lung disease segmentation.
  • Improved Performance: The method demonstrated superior performance compared to conventional networks and standard data augmentation techniques, achieving higher Dice coefficients and better accuracy across various metrics (sensitivity, precision, Hausdorff distance, average contour distance).
  • Efficiency with Fewer Samples: AtlasNet achieved comparable or better results using only 30% of the training data compared to state-of-the-art methods. For instance, with 30% of samples, AtlasNet achieved a mean Dice coefficient of 0.613, while a previous method achieved 0.434.
  • Robustness to Number of Atlases: The Dice coefficient initially increased with the number of atlases and tended to stabilize after approximately 5 atlases.
  • Computational Efficiency: While involving multiple networks, the registration step is efficient (3-5 seconds per subject using a graph-based GPU algorithm), and the overall testing time for a volumetric CT on a single GPU is around 3-4 minutes.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that inter-subject image registration is inherently ill-posed, and regularization is applied to ensure smooth solutions and maintain topological relationships.
  • While the method operates on 2D slices, it achieves smooth segmentation across all three axes due to the fusion of atlas predictions. However, explicitly accounting for 3D connections could be a future improvement.
  • Future Work:
    • Investigating techniques for soft parameter sharing of hidden layers and information transfer between networks and atlases.
  • Extending the framework to multi-organ segmentation, which would involve handling multiple classes and potentially more complex loss functions.

Summary updated

BibTeX
@inproceedings{vakalopoulou2018atlasnet,
  author = {Maria Vakalopoulou and Guillaume Chassagnon and Norbert Bus and Rafael Marini and Evangelia I. Zacharaki and Marie-Pierre Revel and Nikos Paragios},
  title = {{AtlasNet: Multi-atlas Non-linear Deep Networks for Medical Image Segmentation}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2018},
  doi = {10.1007/978-3-030-00937-3\_75},
}
Figure