Classifier Selection Strategies for Label Fusion Using Large Atlas Databases

Paul Aljabar, Rolf A. Heckemann, Alexander Hammers, Joseph V. Hajnal, Daniel RueckertMICCAI 2007 · DOI

Introduction

Multi-atlas label fusion has emerged as a robust technique for automated brain MRI segmentation. By propagating labels from multiple expert-annotated "atlases" to a target image and fusing them, researchers can achieve high accuracy. However, as the availability of large-scale neuroimaging databases grows, this approach faces significant challenges. Registering hundreds of atlases to a single target is computationally prohibitive, and including irrelevant atlases—such as matching a pediatric atlas to a geriatric target—can degrade segmentation quality. This study investigates strategies for selecting an optimal subset of "expert" classifiers to improve both efficiency and accuracy.

Methodology

The core objective was to identify the most effective ranking strategies for atlas selection. The researchers evaluated several approaches:

  1. Image Similarity Metrics: Four metrics were tested to quantify the anatomical resemblance between the target and the atlases: Sum of Squared Differences (SSD), Cross-Correlation (CC), Mutual Information (MI), and Normalised Mutual Information (NMI).
  2. Demographic Matching: Atlases were ranked based on the chronological age proximity of the subjects.
  3. Segmentation Similarity (The "Oracle"): As a theoretical benchmark, atlases were ranked by the actual overlap (Dice coefficient) between their propagated labels and the target's manual labels.

To maintain computational efficiency, all images were registered to a common MNI Brainweb reference space. Once the top-ranked atlases were selected, their labels were fused using a simple majority voting scheme.

Experimental Setup

The study utilized a diverse dataset of 275 T1-weighted MR images, spanning an age range from 4 to 83 years. Expert manual delineations of 35 brain structures were provided by the Centre for Morphometric Analysis (MGH). The evaluation focused on three representative structures: the hippocampus, lateral ventricles, and thalamus.

Key Results

The analysis revealed that segmentation accuracy does not scale linearly with the number of atlases. Instead, there is a "sweet spot" typically between 10 and 20 atlases. Beyond this point, adding more classifiers introduces "noise" from less relevant matches, leading to a gradual decline in performance.

Dice coefficient performance for Hippocampus and Lateral Ventricle across different selection methods.

Among the image similarity metrics, Normalised Mutual Information (NMI) consistently outperformed others. As shown in Figure 1, NMI-based selection (pink diamonds) closely tracks the performance of the "Oracle" (blue triangles) and significantly exceeds the median random selection (box plots). For the thalamus, NMI-based selection was superior to random selection in every single case tested. Conversely, SSD proved to be an unreliable metric, often performing worse than random selection.

Key Insights: Age as a Proxy for Anatomy

One of the most compelling findings was the efficacy of demographic matching. In older subjects, selecting atlases based on age was often more effective than using image similarity metrics.

Comparison of z-scores for demographic and image similarity selection strategies across different subject ages.

As illustrated in Figure 4, for subjects aged 74 and 83, the demographic selection strategy (green bars) yielded significantly higher z-scores than image similarity (yellow bars). This suggests that chronological age serves as a powerful indicator of age-related anatomical changes, such as ventricular enlargement and cortical atrophy, which might be harder for standard intensity-based metrics to capture in a single reference space.

Conclusion and Future Work

This research demonstrates that intelligent atlas selection can significantly enhance multi-atlas label fusion. By using NMI or age-based ranking to select a small subset of 10–20 atlases, researchers can achieve higher segmentation accuracy with a fraction of the computational cost.

Future work aims to refine these strategies by performing the final label fusion in the target subject's native space to minimize registration errors. Additionally, the authors suggest that combining multiple demographic factors—such as sex and clinical diagnosis—with image similarity metrics could further optimize the selection of "expert" subsets for diverse patient populations.

Summary updated

BibTeX
@inproceedings{aljabar2007classifier,
  author = {Paul Aljabar and Rolf A. Heckemann and Alexander Hammers and Joseph V. Hajnal and Daniel Rueckert},
  title = {{Classifier Selection Strategies for Label Fusion Using Large Atlas Databases}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2007},
  doi = {10.1007/978-3-540-75757-3\_64},
}
Figure