Classification of Structural Images via High-Dimensional Image Warping, Robust Feature Extraction, and SVM

Yong Fan, Dinggang Shen, Christos DavatzikosMICCAI 2005 · DOI

Introduction

In medical imaging, identifying structural differences between healthy and diseased brains is a fundamental challenge. Traditional approaches typically fall into two categories: Region of Interest (ROI) Volumetry, which requires prior knowledge to manually measure specific structures, and Voxel-Based Morphometry (VBM), which analyzes every voxel but is highly sensitive to noise and anatomical variations.

This paper addresses the "curse of dimensionality"—the discrepancy between millions of voxels and small clinical sample sizes—by introducing a framework that automatically groups voxels into meaningful, adaptive regions. By combining high-dimensional image warping with adaptive watershed segmentation and Support Vector Machine-Recursive Feature Elimination (SVM-RFE), the authors achieved a 91.8% accuracy in distinguishing schizophrenia patients from healthy controls.

Methodology

The proposed pipeline consists of several sophisticated stages designed to extract robust features from high-dimensional data:

  1. Morphological Profiling: Images are segmented into Gray Matter (GM), White Matter (WM), and Cerebrospinal Fluid (CSF). A "mass-preserving" warping technique (Jacobian modulation) maps these to a common template, ensuring local volume information is preserved even after spatial normalization.
  2. Adaptive Region Generation: Instead of static ROIs, the system generates a "score map" for every voxel based on discriminative power (Pearson correlation with bagging) and spatial consistency. A watershed segmentation algorithm then clusters these voxels into distinct, disease-relevant regions.
  3. Feature Selection and Classification: SVM-RFE is employed to rank these regions and identify the optimal subset for classification. A nonlinear SVM with a Gaussian radial basis function (RBF) kernel performs the final diagnosis.
Figure 1: The multi-stage segmentation and adaptive region generation process, moving from raw tissue segmentation to clustered disease-relevant regions.

Experimental Results

The framework was evaluated on a dataset of 61 MR T1 brain images (23 schizophrenia patients and 38 normal controls) using leave-one-out cross-validation (LOOCV).

Performance and Stability

The system reached a peak average classification accuracy of 91.8% using an optimal set of 37 automatically generated features. A key finding was the stability of the SVM-RFE method; performance remained consistent across various feature counts and kernel sizes, suggesting the model captured genuine biological signals rather than noise.

Visualization of Pathology

By projecting SVM results back onto a template brain, the authors visualized the specific regions most representative of schizophrenia. These regions, including parts of the ventricles and cortex, align with known pathological markers of the disease.

Figure 6: Visualization of the discriminative regions identified by the model, highlighting the anatomical areas most predictive of schizophrenia.

Discussion and Conclusion

The study demonstrates that disease-affected regions are best understood as clusters of contiguous voxels rather than isolated points. The adaptive watershed approach allows the algorithm to "learn" the shape of pathology without human bias.

While the results are promising, the authors note limitations regarding sample specificity (female-only cohort) and size (N=61). Future work aims to incorporate higher-order shape characteristics, such as cortical folding patterns, and to validate the framework on larger, more diverse multi-site datasets. Currently, this framework serves as a powerful research tool for identifying group differences in neuroanatomical studies.

Summary updated

BibTeX
@inproceedings{fan2005classification,
  author = {Yong Fan and Dinggang Shen and Christos Davatzikos},
  title = {{Classification of Structural Images via High-Dimensional Image Warping, Robust Feature Extraction, and SVM}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2005},
  doi = {10.1007/11566465\_1},
}
Figure