Deep Feature Engineering in Colposcopy Image Recognition: A Comparative Study.

Area of Science:

• Medical imaging analysis
• Machine learning in healthcare
• Computational pathology

Background:

• Feature fusion enhances diagnostic accuracy in medical applications.
• Cervical cancer classification benefits from integrated data sources.
• Deep learning models extract complex features from medical data.

Purpose of the Study:

• To compare six feature fusion techniques for optimal cervical cancer classification.
• To evaluate the performance of different fusion methods using deep learning features.
• To identify the most effective technique for improving diagnostic accuracy.

Main Methods:

• Generated ten feature datasets using transfer learning from ten popular deep learning models.
• Applied six feature fusion techniques: CCA, DCA, LASSO, ICA, PCA, and concatenation.
• Classified four cervical cancer types (Negative, HISL, LSIL, SCC) using a support vector machine.

Main Results:

• Canonical correlation analysis (CCA) yielded the highest accuracy at 99.7%.
• Independent component analysis (ICA) and least absolute shrinkage and selection operator (LASSO) achieved 98.3% accuracy.
• Principal component analysis (PCA) showed the lowest accuracy at 90%.

Conclusions:

• CCA is the most effective feature fusion technique for cervical cancer classification among those tested.
• The study demonstrates the potential of feature fusion for enhancing medical diagnostic performance.
• The developed approach is applicable to other medical classification problems requiring feature reduction and accuracy improvement.