Segmentation of MR images via discriminative dictionary learning and sparse coding: Application to hippocampus labeling

Tong Tong, Robin Wolz, Pierrick Coupé, Joseph V. Hajnal, Daniel RueckertNeuroImage 2013 · DOI

TL;DR

This paper introduces a novel method for automatically segmenting brain MRI images, specifically focusing on the hippocampus. By leveraging discriminative dictionary learning and sparse coding, the authors achieve accurate and efficient segmentation, outperforming existing methods in terms of both speed and accuracy.

What Problem Does This Paper Address?

Accurate segmentation of anatomical structures in brain MRI is crucial for quantitative analysis, especially in understanding how brain anatomy changes with aging or disease. Manual segmentation is time-consuming and prone to variability. While automated methods exist, they often struggle with the complexity of subcortical structures and can be computationally expensive. Existing atlas-based methods, for instance, rely on image registration and comparison, which can lead to errors and significant computational burden. This paper aims to address these challenges by proposing a more efficient and accurate automated segmentation approach.

Methods

The core of the proposed method lies in discriminative dictionary learning (DDLS) and sparse coding. Instead of relying on traditional patch similarity, the authors learn dictionaries and classifiers simultaneously from a set of brain atlases. This allows for the reconstruction and segmentation of unseen target images.

Key aspects of their approach include:

  • Patch Reconstruction: The method is based on reconstructing image patches rather than direct similarity comparisons.
  • Simultaneous Learning: Dictionaries and classifiers are learned together, capturing both reconstructive and discriminative information.
  • Voxel-based Dictionaries: A different dictionary is learned for each target voxel, allowing for fine-grained adaptation.
  • Sparse Representation Classification (SRC): This initial approach uses a template patch library as a predefined dictionary.
  • Discriminative Dictionary Learning for Segmentation (DDLS): This refined approach learns a smaller, task-specific dictionary and a linear classifier.
  • Fixed Discriminative Dictionary Learning for Segmentation (F-DDLS): To further improve efficiency, this strategy learns fixed dictionaries offline, enabling rapid online segmentation.

Datasets

The proposed methods were evaluated on two datasets:

  • ADNI (Alzheimer's Disease Neuroimaging Initiative) database: This dataset consists of 202 subjects, including cognitively normal individuals, those with Mild Cognitive Impairment (MCI), and early Alzheimer's Disease (AD) patients.
  • ICBM (International Consortium for Brain Mapping) database: This dataset comprises 80 healthy subjects.

The paper does not introduce new benchmarks but utilizes existing datasets with reference segmentations.

Key Insights and Ideas

  • Dictionary Learning for Segmentation: Applying discriminative dictionary learning to subcortical brain segmentation is a novel approach, moving beyond traditional intensity-based or patch-similarity methods.
  • Reconstruction Error as a Key Metric: Minimizing patch reconstruction error, rather than solely relying on image similarity, proves to be a powerful strategy for accurate segmentation.
  • Offline Dictionary Training for Speed: The F-DDLS strategy, by learning dictionaries offline, significantly accelerates the segmentation process, making it practical for clinical applications.
  • Generalization Across Datasets: The F-DDLS method demonstrates the ability to generalize from one type of data to morphologically different datasets, highlighting its robustness.

Key Results

  • High Accuracy: The method achieved competitive results, with the highest median Dice coefficients of 0.879 on the ADNI dataset and 0.890 on the ICBM dataset.
  • Computational Efficiency: The F-DDLS strategy allows for segmentation of an unseen subject in less than 1 minute after offline dictionary training. This is a significant speedup compared to other methods like patch-based (10 minutes) and SRC (40 minutes).
  • Robustness: The method shows robustness to intensity normalization issues, outperforming patch-based methods on the ADNI dataset.
  • Comparison to State-of-the-Art: The proposed method achieved results comparable to or better than existing state-of-the-art methods.

Limitations and Future Work

  • Computational Cost of Training: While F-DDLS offers fast online segmentation, the initial training of dictionaries can be computationally expensive, especially for larger dictionary sizes or when learning for every voxel.
  • Generalizability to Different Modalities/Protocols: The learned dictionaries are trained on specific imaging modalities and segmentation protocols. If these change, the dictionaries would need to be retrained.
  • Future Work:
    • Segmentation of Multiple Structures: Extending the approach to segment multiple brain structures by incorporating more label information.
  • Hippocampal Atrophy Measurement and Patient Classification: Applying the method to these clinical applications.
  • More Complex Classifiers: Investigating more sophisticated classifiers beyond linear ones to potentially improve accuracy.
  • Non-registration: Exploring the use of non-rigid registration instead of affine registration for potentially improved results.
  • Intensity Models and Learning-Based Methods: Incorporating these to correct systematic errors.

Summary updated

BibTeX
@article{tong2013segmentation,
  author = {Tong Tong and Robin Wolz and Pierrick Coupé and Joseph V. Hajnal and Daniel Rueckert},
  title = {{Segmentation of MR images via discriminative dictionary learning and sparse coding: Application to hippocampus labeling}},
  journal = {NeuroImage},
  year = {2013},
  doi = {10.1016/j.neuroimage.2013.02.069},
}
Figure