An Unbiased Penalty for Sparse Classification with Application to Neuroimaging Data

Li Zhang, Dana Cobzas, Alan H. Wilman, Linglong KongMICCAI 2017 · DOI

TL;DR

This paper introduces a novel penalty function, SCADTV, for sparse classification in neuroimaging data. By combining the Smoothly Clipped Absolute Deviation (SCAD) penalty for unbiased variable selection with a Total Variation (TV) penalty for spatial continuity, they achieve more stable and accurate identification of brain regions associated with diseases like multiple sclerosis compared to existing methods.

What Problem Does This Paper Address?

Neuroimaging studies, particularly those investigating diseases like multiple sclerosis, often aim to identify specific brain regions that differ between groups of individuals. Traditional voxel-based analysis (VBA) methods can be limited by their inability to account for correlations between voxels and regions, leading to issues with multiple comparisons and difficulty in identifying complex, distributed pathologies. While sparse classification methods (like LASSO) have shown promise by simultaneously selecting features and classifying data, they can produce unstable and anatomically uninterpretable results due to inherent biases in their penalties. This paper addresses the need for a method that offers both accurate variable selection and anatomically meaningful, stable region identification in high-dimensional neuroimaging data.

Methods

The core contribution of this paper is the development of a new penalty function called SCADTV. This penalty is designed to overcome the limitations of existing methods by combining two key components:

  • SCAD Penalty: This is a non-concave penalty function that aims to provide "unbiased" variable selection. Unlike the LASSO (l1) penalty, which tends to shrink all coefficients, SCAD is designed to shrink small coefficients to zero while leaving larger, more important coefficients largely unchanged. This leads to more stable and accurate identification of relevant features.
  • Total Variation (TV) Penalty: This penalty is applied to the gradients of the coefficients across the image dimensions. It encourages neighboring voxels to have similar coefficient values, thereby enforcing spatial continuity and ensuring that the identified regions are anatomically compact and meaningful, rather than isolated voxels.

The authors formulate a sparse classification model using logistic regression and incorporate the SCAD and SCADTV penalties. They propose three variations: SCAD+SCADTV, SCAD+TV, and l1+SCADTV, and compare them against the classical l1+TV approach. Optimization of these non-convex problems is handled using the Alternating Direction Method of Multipliers (ADMM) algorithm.

Datasets

The authors evaluate their proposed methods using two types of datasets:

  • Synthetic Data: They generate synthetic neuroimaging data with known ground truth for the significant regions. This allows for a direct quantitative comparison of variable selection accuracy. The synthetic data is designed to mimic real neuroimaging data characteristics, including spatial coherence and noise.
  • Real MRI Data: They use data from an in-house multiple sclerosis (MS) study. This dataset consists of quantitative traverse relaxation rate (R2*) images from subjects with MS and age- and gender-matched controls. The focus is on deep gray matter structures. The data is pre-processed and aligned to an unbiased template.

Key Insights and Ideas

  • Unbiasedness Matters for Stability: The SCAD penalty's ability to provide more unbiased estimates of coefficients is crucial for achieving stable and reliable identification of important brain regions, especially in high-dimensional data.
  • Spatial Continuity is Key for Interpretability: Simply selecting sparse features isn't enough; enforcing spatial continuity through penalties like TV is essential for generating anatomically meaningful and interpretable results in neuroimaging.
  • Combining SCAD and TV is Powerful: The SCADTV penalty effectively balances the need for accurate variable selection with the requirement for anatomically coherent regions, leading to superior performance over methods that rely solely on l1 or TV penalties.
  • SCADTV Outperforms LASSO-based Methods: In both simulated and real data experiments, the SCAD-based approaches, particularly SCAD+SCADTV, consistently demonstrate better performance in terms of variable selection accuracy and region stability.

Key Results

  • Synthetic Data: The SCAD+SCADTV penalty achieved the highest dice score (a measure of overlap between detected and true regions) and the lowest Sum of Absolute Error (SAE) for recovered coefficients, indicating superior variable selection accuracy compared to l1+TV and other SCAD variants. Visualizations confirmed that SCAD+SCADTV produced the cleanest and closest-to-ground-truth variable selection results.
  • Real MRI Data: While classification accuracy was comparable across methods, the SCAD-based approaches (SCAD+SCADTV and SCAD+TV) showed significantly better stability of detected regions, as measured by between-folds dice scores. SCAD+SCADTV achieved the highest Dice Folds score (0.79), indicating more consistent region detection across different data splits. Visualizations showed that SCAD+SCADTV and SCAD+TV recovered compact regions in similar brain locations, with SCAD+SCADTV producing the most robust results.

Limitations and Future Work

  • Computational Complexity: While ADMM was used for optimization, the paper notes that SCAD is not convex, which can still present computational challenges.
  • Theoretical Guarantees: The authors mention that deriving theoretical results on coefficient bounds and accuracy of variable selection for SCAD-based models is a part of their future work.
  • Generalization to Other Tasks: They suggest that similar penalizers could be extended to other tasks like regression or data representation (e.g., PCA, CCA).
  • Parameter Tuning: The tuning parameters (λ, γ) were chosen using the generalized information criterion (GIC). Further exploration of optimal parameter selection strategies could be beneficial.

Summary updated

BibTeX
@inproceedings{zhang2017unbiased,
  author = {Li Zhang and Dana Cobzas and Alan H. Wilman and Linglong Kong},
  title = {{An Unbiased Penalty for Sparse Classification with Application to Neuroimaging Data}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2017},
  doi = {10.1007/978-3-319-66179-7\_7},
}
Figure