Abstract
Automated detection of informative Regions of Interest (ROIs) is crucial for accurate medical image classification, as relevant signals are often localized within specific anatomical structures. We propose a framework based on multi-atlas label fusion that learns voxel-wise discriminativeness scores to identify these regions automatically. By applying iterative smoothing, the system focuses on consistent anatomical landmarks across a class. Evaluation on a dataset of 519 chest CT slices shows that ROI-based classification significantly outperforms global approaches, with accuracy improvements observed for both handcrafted features and CNN-based representations.
Introduction
In medical imaging, the diagnostic "signal" identifying a specific body part or disease is often confined to small, localized anatomical structures. Standard machine learning approaches frequently utilize global features from the entire image, which introduces significant noise from irrelevant tissues. While manual ROI labeling can guide models, it is time-consuming and requires expert knowledge. This paper addresses the challenge of automatic ROI detection: identifying specific regions most useful for distinguishing classes without explicit spatial supervision.
Methodology
The proposed framework leverages Multi-atlas Label Fusion to detect ROIs through several key steps:
- Voxel-wise Discriminativeness: For each voxel in a training image, a score is calculated representing its class-specific relevance. This is achieved by "warping" training atlases onto the target image via deformable registration to identify matching local patches.
- Iterative Smoothing: To ensure robustness against noise, a coordinate descent technique iteratively updates scores by comparing images within the same class. This forces the model to converge on consistent anatomical landmarks.
- ROI Extraction: Discriminativeness maps are thresholded to create binary masks for feature extraction.
- Classification: The detected ROIs are used either to weight voting in multi-atlas classification or to extract patches for learning-based classifiers (e.g., SVM with HoG, LBP, or CNN features).
The study utilized 75 axial chest CT scans, providing 519 representative 2D slices categorized into nine semantic classes representing cardiac anatomy, such as the aortic arch and pulmonary trunk.
Results and Analysis
The inclusion of automatically detected ROIs provided a substantial performance boost across all tested classification methods.
Key findings include:
- Performance Gains: Multi-atlas classification accuracy increased from 84.8% (global) to 92.1% (ROI-based). Learning-based approaches using HoG features saw an even larger leap from 64.7% to 81.9%.
- Feature Agnosticism: The method consistently improves results regardless of the feature descriptor used, demonstrating that providing a clean, ROI-focused patch is superior to using noisy global images.
- Precision: Most classification errors occurred between adjacent anatomical levels (within 22mm), indicating high accuracy in capturing the superior-inferior progression of the heart.
Conclusion
This multi-atlas approach effectively leverages the spatial consistency of human anatomy to automate ROI detection. By treating ROI identification as a label fusion problem and employing iterative smoothing, the model generalizes well across patients. Future work includes extending the framework to 3D volumes to capture more context and optimizing the computational overhead of deformable registration for larger datasets.
Summary updated