Feature Selection for the Automated Detection of Metaphase Chromosomes: Performance Comparison Using a Receiver Operating Characteristic Method

Yuchen Qiu, Jie Song, Xianglan Lu, Yuhua Li, Bin Zheng, Shibo Li, Hong LiuAnalytical 2014 · DOI

TL;DR

This paper tackles the challenge of efficiently and accurately identifying analyzable metaphase chromosomes from high-throughput microscopy scans. By systematically evaluating nine different image features using Receiver Operating Characteristic (ROC) analysis, the authors identify optimal feature sets for both real-time (on-line) and post-processing (off-line) computer-aided detection (CAD) schemes, paving the way for more robust and efficient automated chromosome analysis.

What Problem Does This Paper Address?

Chromosome analysis, particularly karyotyping, is crucial for diagnosing genetic disorders and cancers. Traditionally, this involves a laborious two-step process: initial low-magnification scanning to locate potential cells, followed by manual high-magnification acquisition by clinicians. This manual step is time-consuming, prone to interobserver variability due to subjective cell selection, and can lead to missed diagnoses if cells with suboptimal morphology are chosen. While automated scanning systems have emerged to alleviate these issues, they require sophisticated computer-aided detection (CAD) schemes to automatically identify and select clinically relevant metaphase chromosome images from the vast amount of data generated by high-throughput scanners. The core problem this paper addresses is how to select the most effective image features for these CAD schemes to ensure both speed and accuracy in identifying analyzable metaphase chromosomes.

Methods

The authors employed a systematic approach to evaluate and compare nine different image features for their ability to distinguish between clinically analyzable and unanalyzable metaphase chromosome images. These features are primarily based on morphology, including:

  • Number of labeled regions: Counting distinct chromosome-like structures.
  • Average region pixel intensity: The average brightness of detected regions.
  • Standard deviation (STD) of region pixel intensity: Variability in brightness within detected regions.
  • Average region area: The average size of detected regions.
  • STD of region area: Variability in size of detected regions.
  • Average region circularity: How closely detected regions resemble circles.
  • STD of region circularity: Variability in circularity of detected regions.
  • Average region distance: The average distance of detected regions from the image's global center of mass.
  • STD of region distance: Variability in the distances of detected regions from the global center of mass.

The performance of each feature was assessed using Receiver Operating Characteristic (ROC) curve analysis. This method is preferred over simpler accuracy metrics because it evaluates the trade-off between true positive fraction (sensitivity) and false positive fraction (1-specificity) across a range of thresholds, providing a more comprehensive understanding of a feature's discriminatory power. The area under the ROC curve (AUC) was used as the primary metric for comparison. For the off-line CAD scheme, a two-level classifier combining selected features was also proposed and evaluated.

Datasets

The study utilized a dataset of 200 bone marrow cell images acquired using a high-throughput scanning microscopy prototype. These images were captured at a 100x objective lens with a pixel size of 7 µm. The dataset comprised 67 "clinically meaningful" (analyzable) metaphase chromosome images and 133 "clinically meaningless" (unanalyzable) images. The unanalyzable images included those with interphase cells only or with overlapping metaphase chromosomes. This dataset appears to be newly acquired for the purpose of this study, serving as a benchmark for evaluating features under the specific conditions of their high-throughput system.

Key Insights and Ideas

  • Feature selection is application-dependent: The study emphasizes that optimal features for chromosome detection are not universal and depend on the specific requirements of the CAD scheme (e.g., real-time vs. post-processing).
  • "Number of labeled regions" is a strong contender for on-line detection: This feature alone demonstrated high discriminatory power and is computationally efficient, making it suitable for real-time processing during high-speed scanning.
  • Morphology features are sufficient and efficient: The paper argues that morphology-based features strike a good balance between effectiveness and computational cost, avoiding the complexity of features designed for detailed band pattern analysis.
  • Combining features enhances performance: For the off-line CAD scheme, a combination of features significantly outperformed any single feature, highlighting the power of ensemble methods.
  • Correlated features can be redundant: The study noted that "STD of the region distance" and "STD of the region circularity" are correlated, suggesting that only one might be needed in a combined feature set.

Key Results

  • On-line CAD: The "number of labeled regions" was identified as the most suitable feature for the on-line CAD scheme, achieving an AUC of 0.896 ± 0.023.
  • Off-line CAD: For the off-line CAD scheme, a combination of features was recommended. Specifically, combining "number of labeled regions," "average region area," "average region pixel value," and either "STD of region distance" or "STD of region circularity" was suggested.
  • Combined Classifier Performance: A two-level classifier utilizing "average region area," "average region pixel intensity," "STD of the region distance," and "number of labeled regions" achieved a superior AUC of 0.924 ± 0.026.
  • Individual Feature Performance: While "number of labeled regions" performed best, other features like "average region area" (AUC 0.666 ± 0.037) and "average region pixel value" (AUC 0.592 ± 0.039) showed moderate discriminatory ability, but were significantly lower than the top feature. Features like "STD of region circularity" and "STD of region distance" had AUCs close to 0.5, indicating poor performance as individual discriminators.

Limitations and Future Work

  • Classifier Performance Not Fully Assessed: The authors acknowledge that while feature selection was the focus, the performance of different machine learning classifiers (e.g., decision trees, SVMs) using these features was not extensively discussed. They suggest that uniform assessment of various classifiers is needed for optimal off-line CAD scheme development.
  • Comparison with Conventional Scanners: The paper notes that a direct comparison of the overall performance between their high-throughput system and conventional scanners was not the primary focus, although previous work suggested the superiority of their system. A more comprehensive study in this area is proposed.
  • Application-Specific Feature Selection: The authors reiterate that the identified optimal features are specific to their application and system. They suggest that the proposed feature evaluation methodology, however, is generalizable and can be applied to optimize CAD schemes in other contexts.

Summary updated

Abstract

This study demonstrates an effective method of feature selection and comparison to facilitate the optimization of the CAD schemes for high throughput scanning microscope in the future.
BibTeX
@inproceedings{qiu2014feature,
  author = {Yuchen Qiu and Jie Song and Xianglan Lu and Yuhua Li and Bin Zheng and Shibo Li and Hong Liu},
  title = {{Feature Selection for the Automated Detection of Metaphase Chromosomes: Performance Comparison Using a Receiver Operating Characteristic Method}},
  booktitle = {Analytical},
  year = {2014},
  doi = {10.1155/2014/565392},
}
Figure