Heterogeneous ensemble with information theoretic diversity measure for human epithelial cell image classification

Area of Science:

• Computer Vision
• Biomedical Image Analysis
• Machine Learning

Background:

• Accurate classification of human epithelial (HEp-2) cells is crucial for disease diagnosis.
• Indirect immunofluorescence (IIF) imaging generates complex patterns requiring sophisticated analysis.
• Existing classification methods may struggle with the diversity and complexity of HEp-2 cell images.

Purpose of the Study:

• To develop a robust and high-performance HEp-2 cell image classification system.
• To investigate the efficacy of ensemble methods combining diverse feature representations.
• To optimize member selection and fusion strategies for improved classification accuracy.

Main Methods:

• Proposed a heterogeneous ensemble of Convolutional Neural Network (CNN)-based (ResNet, DenseNet, Inception) and traditional classifiers.
• Employed class-specific and texture features for traditional members.
• Utilized an information-theoretic measure for selecting diverse and discriminating members.
• Investigated various fusion techniques (voting, product, Bayes, Dempster-Shafer) for combining member outputs.

Main Results:

• Achieved state-of-the-art performance on the ICPR-2014 HEp-2 cell image dataset with 99.80% accuracy.
• Demonstrated comparable performance on a new large-scale dataset (63K images) with 86.03% accuracy, even with fewer training samples.
• Validated the effectiveness of the ensemble approach with diverse feature representations.

Conclusions:

• Heterogeneous ensembles combining diverse feature types significantly improve HEp-2 cell image classification.
• The proposed member selection and fusion strategy effectively leverages the strengths of individual classifiers.
• The methodology shows strong generalization capabilities across different datasets and sample sizes.